refactor : llama-model.cpp (#16252)

* Sqashed: llama-model.cpp refactoring

* Fix formatting of attn / ffn / ffn_moe calls

* Fix import regression / unify spacing in models.h

* totally DID NOT miss those!

* Add missing qwen3vl(moe) models

* Add missing new .cpp files to build

* Remove extra semicolons

* Editor checker

* Update src/models/models.h

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

.github/workflows/docker.yml

@@ -40,7 +40,7 @@ jobs:
           # https://github.com/ggml-org/llama.cpp/issues/11888
           #- { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: false }
           - { tag: "cpu",    dockerfile: ".devops/cpu.Dockerfile",    platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
-          - { tag: "cuda",   dockerfile: ".devops/cuda.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
+          - { tag: "cuda",   dockerfile: ".devops/cuda.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
           - { tag: "musa",   dockerfile: ".devops/musa.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
           - { tag: "intel",  dockerfile: ".devops/intel.Dockerfile",  platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
           - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }

common/arg.cpp

@@ -3203,7 +3203,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
     add_opt(common_arg(
         {"--parse-special"},
-        string_format("prase special tokens (chat, tool, etc) (default: %s)", params.parse_special ? "true" : "false"),
+        string_format("parse special tokens (chat, tool, etc) (default: %s)", params.parse_special ? "true" : "false"),
         [](common_params & params) {
             params.parse_special = true;
         }

convert_hf_to_gguf.py

@@ -1054,6 +1054,9 @@ class TextModel(ModelBase):
         if chkhsh == "53e325976a6e142379c19b09afcae354f2f496f147afa8f9e189a33fe4e3024e":
             # ref: https://huggingface.co/ibm-granite/granite-docling-258M
             res = "granite-docling"
+        if chkhsh == "f4f37b6c8eb9ea29b3eac6bb8c8487c5ab7885f8d8022e67edc1c68ce8403e95":
+            # ref: https://huggingface.co/MiniMaxAI/MiniMax-M2
+            res = "minimax-m2"
 
         if res is None:
             logger.warning("\n")
@@ -1528,7 +1531,7 @@ class MmprojModel(ModelBase):
             self.gguf_writer.add_vision_embedding_length(self.find_vparam(["hidden_size"]))
             self.gguf_writer.add_vision_feed_forward_length(self.find_vparam(["intermediate_size"]))
             self.gguf_writer.add_vision_block_count(self.find_vparam(self.n_block_keys))
-            self.gguf_writer.add_vision_head_count(self.find_vparam(["num_attention_heads"]))
+            self.gguf_writer.add_vision_head_count(self.find_vparam(["num_attention_heads", "num_heads"]))
 
             # preprocessor config
             image_mean = _MISTRAL_COMMON_DATASET_MEAN if self.is_mistral_format else self.preprocessor_config["image_mean"]
@@ -3852,7 +3855,43 @@ class Qwen2MoeModel(TextModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # process the experts separately
         name = name.replace("language_model.", "") # InternVL
-        if name.startswith("mlp") or name.startswith("vision_model") or name.startswith("model.vision_tower") or name.startswith("model.multi_modal_projector"):
+
+        # handle aggregated expert tensors
+        # GGUF stores dimensions reversed from PyTorch, so:
+        # PyTorch (A,B,C) -> GGUF writes [C,B,A] -> GGML reads ne={C,B,A}
+        # Input shapes from HF: (n_expert, n_ff_exp, n_embd) or (n_expert, n_embd, n_ff_exp)
+        # Expected GGML ne: {n_embd, n_ff_exp, n_expert} for gate/up, {n_ff_exp, n_embd, n_expert} for down
+        if name.endswith("mlp.experts.down_proj") or name.endswith("mlp.experts.down_proj.weight"):
+            mapped = f"{name}.weight" if not name.endswith(".weight") else name
+            # Input: (n_expert=128, n_ff_exp=768, n_embd=2048)
+            # Want GGML ne: {n_ff_exp, n_embd, n_expert} = {768, 2048, 128}
+            # Need PyTorch: (128, 2048, 768) [reversed of GGML]
+            # So: permute(0, 2, 1): (128, 768, 2048) -> (128, 2048, 768)
+            permuted = data_torch.permute(0, 2, 1).contiguous()
+            return [(self.map_tensor_name(mapped), permuted)]
+
+        if name.endswith("mlp.experts.gate_up_proj") or name.endswith("mlp.experts.gate_up_proj.weight"):
+            if data_torch.ndim < 3 or data_torch.shape[-1] % 2 != 0:
+                raise ValueError(f"Unexpected gate_up_proj shape for {name}: {tuple(data_torch.shape)}")
+            split_dim = data_torch.shape[-1] // 2
+            gate = data_torch[..., :split_dim].contiguous()
+            up = data_torch[..., split_dim:].contiguous()
+            # Input gate/up: (n_expert=128, n_embd=2048, n_ff_exp=768)
+            # Want GGML ne: {n_embd, n_ff_exp, n_expert} = {2048, 768, 128}
+            # Need PyTorch: (128, 768, 2048) [reversed of GGML]
+            # So: permute(0, 2, 1): (128, 2048, 768) -> (128, 768, 2048)
+            base_name = name.removesuffix(".weight")
+            base = base_name.rsplit('.', 1)[0]
+            mapped_gate = f"{base}.gate_proj.weight"
+            mapped_up = f"{base}.up_proj.weight"
+            perm_gate = gate.permute(0, 2, 1).contiguous()
+            perm_up = up.permute(0, 2, 1).contiguous()
+            return [
+                (self.map_tensor_name(mapped_gate), perm_gate),
+                (self.map_tensor_name(mapped_up), perm_up),
+            ]
+
+        if name.startswith("mlp") or name.startswith("vision_model") or name.startswith("model.vision_tower") or name.startswith("model.multi_modal_projector") or name.startswith("model.visual"):
             # skip visual tensors
             return []
         if name.find("experts") != -1:
@@ -4004,6 +4043,187 @@ class Qwen3MoeModel(Qwen2MoeModel):
         super().set_vocab()
 
 
+@ModelBase.register("Qwen3VLForConditionalGeneration", "Qwen3VLMoeForConditionalGeneration")
+class Qwen3VLVisionModel(MmprojModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None
+        # Compute image_size if not present
+        if "image_size" not in self.hparams_vision:
+            # For Qwen3VL/Qwen3VLMoe, compute from num_position_embeddings
+            num_pos = self.hparams_vision.get("num_position_embeddings", 2304)
+            patch_size = self.hparams_vision.get("patch_size", 16)
+            # num_position_embeddings = (image_size / patch_size) ** 2
+            # So image_size = sqrt(num_position_embeddings) * patch_size
+            image_size = int(num_pos**0.5 * patch_size)
+            self.hparams_vision["image_size"] = image_size
+
+        # Rename config values for compatibility
+        self.hparams_vision["num_attention_heads"] = self.hparams_vision.get("num_heads")
+        self.hparams_vision["num_hidden_layers"] = self.hparams_vision.get("depth")
+
+        self.is_deepstack_layers = [False] * int(self.hparams_vision["num_hidden_layers"] or 0)
+        for idx in self.hparams_vision.get("deepstack_visual_indexes", []):
+            self.is_deepstack_layers[idx] = True
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.QWEN3VL)
+        self.gguf_writer.add_vision_use_gelu(True)
+
+        if self.hparams_vision is not None:
+            merge_size = self.hparams_vision.get("spatial_merge_size")
+            if merge_size is not None:
+                self.gguf_writer.add_vision_spatial_merge_size(int(merge_size))
+
+        # Use text config's rms_norm_eps for vision attention layernorm eps
+        rms_norm_eps = self.global_config.get("text_config", {}).get("rms_norm_eps", 1e-6)
+        self.gguf_writer.add_vision_attention_layernorm_eps(rms_norm_eps)
+
+        if self.is_deepstack_layers:
+            self.gguf_writer.add_vision_is_deepstack_layers(self.is_deepstack_layers)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        assert self.hparams_vision is not None
+        # Skip text model tensors - they go in the text model file
+        if name.startswith("model.language_model.") or name.startswith("lm_head."):
+            return []
+
+        if name.startswith("model.visual."):
+            name = name.replace("model.visual.", "visual.", 1)
+
+        if name.startswith("visual.deepstack_merger_list."):
+            prefix, rest = name.split(".", maxsplit=3)[2:]
+            # prefix is the layer index, convert to absolute clip layer index!
+            idx = self.hparams_vision.get("deepstack_visual_indexes", [])[int(prefix)]
+            target = rest
+
+            tensor_type: gguf.MODEL_TENSOR
+            if target.startswith("norm."):
+                tensor_type = gguf.MODEL_TENSOR.V_DS_NORM
+                suffix = target.split(".", 1)[1]
+            elif target.startswith("linear_fc1."):
+                tensor_type = gguf.MODEL_TENSOR.V_DS_FC1
+                suffix = target.split(".", 1)[1]
+            elif target.startswith("linear_fc2."):
+                tensor_type = gguf.MODEL_TENSOR.V_DS_FC2
+                suffix = target.split(".", 1)[1]
+            else:
+                raise ValueError(f"Unexpected deepstack tensor: {name}")
+
+            new_name = self.format_tensor_name(tensor_type, idx, suffix=f".{suffix}")
+            return [(new_name, data_torch)]
+
+        if name.startswith("visual.merger."):
+            suffix = name.split(".", 2)[2]
+            if suffix.startswith("linear_fc"):
+                fc_idx_str, tail = suffix.split(".", 1)
+                fc_num = int(fc_idx_str.replace("linear_fc", ""))
+                # Qwen3VL has linear_fc1 and linear_fc2
+                # Map to indices 0 and 2 (matching Qwen2VL which uses indices 0 and 2)
+                if fc_num == 1:
+                    fc_idx = 0
+                elif fc_num == 2:
+                    fc_idx = 2
+                else:
+                    raise ValueError(f"unexpected fc index {fc_num} in {name}")
+                new_name = self.format_tensor_name(gguf.MODEL_TENSOR.V_MMPROJ, fc_idx, suffix=f".{tail}")
+            elif suffix.startswith("norm."):
+                new_name = self.format_tensor_name(gguf.MODEL_TENSOR.V_POST_NORM, suffix=f".{suffix.split('.', 1)[1]}")
+            else:
+                raise ValueError(f"Unexpected merger tensor: {name}")
+            return [(new_name, data_torch)]
+
+        if name == "visual.patch_embed.proj.weight":
+            # split Conv3D into Conv2Ds along temporal dimension
+            c1, c2, kt, _, _ = data_torch.shape
+            del c1, c2
+            if kt != 2:
+                raise ValueError("Current implementation only supports temporal_patch_size of 2")
+            return [
+                (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight", data_torch[:, :, 0, ...]),
+                (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight.1", data_torch[:, :, 1, ...]),
+            ]
+
+        if name == "visual.patch_embed.proj.bias":
+            # Include the bias - it's used by the C++ code
+            return [(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".bias", data_torch)]
+
+        if name.startswith("visual."):
+            return [(self.map_tensor_name(name), data_torch)]
+
+        # Fall back to parent class for other tensors
+        return super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("Qwen3VLForConditionalGeneration")
+class Qwen3VLTextModel(Qwen3Model):
+    model_arch = gguf.MODEL_ARCH.QWEN3VL
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # Handle MRoPE (Multi-axis Rotary Position Embedding) for Qwen3-VL
+        text_config = self.hparams.get("text_config", {})
+        # rope_scaling is deprecated in V5, use rope_parameters instead
+        rope_scaling = text_config.get("rope_scaling") or text_config.get("rope_parameters") or {}
+
+        if rope_scaling.get("mrope_section"):
+            # mrope_section contains [time, height, width] dimensions
+            mrope_section = rope_scaling["mrope_section"]
+            # Pad to 4 dimensions [time, height, width, extra]
+            while len(mrope_section) < 4:
+                mrope_section.append(0)
+            self.gguf_writer.add_rope_dimension_sections(mrope_section[:4])
+
+            logger.info(f"MRoPE sections: {mrope_section[:4]}")
+
+        vision_config = self.hparams.get("vision_config", {})
+        deepstack_layer_num = len(vision_config.get("deepstack_visual_indexes", []))
+        self.gguf_writer.add_num_deepstack_layers(deepstack_layer_num)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Skip vision tensors - they go in the mmproj file
+        if name.startswith("model.visual."):
+            return []
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("Qwen3VLMoeForConditionalGeneration")
+class Qwen3VLMoeTextModel(Qwen3MoeModel):
+    model_arch = gguf.MODEL_ARCH.QWEN3VLMOE
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # Handle MRoPE (Multi-axis Rotary Position Embedding) for Qwen3-VL
+        text_config = self.hparams.get("text_config", {})
+        # rope_scaling is deprecated in V5, use rope_parameters instead
+        rope_scaling = text_config.get("rope_scaling") or text_config.get("rope_parameters") or {}
+
+        if rope_scaling.get("mrope_section"):
+            # mrope_section contains [time, height, width] dimensions
+            mrope_section = rope_scaling["mrope_section"]
+            # Pad to 4 dimensions [time, height, width, extra]
+            while len(mrope_section) < 4:
+                mrope_section.append(0)
+            self.gguf_writer.add_rope_dimension_sections(mrope_section[:4])
+
+            logger.info(f"MRoPE sections: {mrope_section[:4]}")
+
+        vision_config = self.hparams.get("vision_config", {})
+        deepstack_layer_num = len(vision_config.get("deepstack_visual_indexes", []))
+        self.gguf_writer.add_num_deepstack_layers(deepstack_layer_num)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Skip vision tensors - they go in the mmproj file
+        if name.startswith("model.visual."):
+            return []
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("GPT2LMHeadModel")
 class GPT2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.GPT2
@@ -6909,6 +7129,64 @@ class DeepseekV2Model(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("MiniMaxM2ForCausalLM")
+class MiniMaxM2Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.MINIMAXM2
+    _experts_cache: dict[int, dict[str, Tensor]] = {}
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.hparams["num_experts"] = self.hparams["num_local_experts"]
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if self.hparams["scoring_func"] == "sigmoid":
+            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+        elif self.hparams["scoring_func"] == "softmax":
+            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX)
+        else:
+            raise ValueError(f"Unsupported scoring_func value: {self.hparams['scoring_func']}")
+
+        self.gguf_writer.add_expert_feed_forward_length(self.find_hparam(["intermediate_size"]))
+        self.gguf_writer.add_rope_dimension_count(self.find_hparam(["rotary_dim"]))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        if name.endswith("e_score_correction_bias"):
+            name = name.replace("e_score_correction_bias", "e_score_correction.bias")
+
+        # merge expert weights
+        if 'experts' in name:
+            n_experts = self.hparams["num_experts"]
+            assert bid is not None
+
+            expert_cache = self._experts_cache.setdefault(bid, {})
+            expert_cache[name] = data_torch
+            expert_weights = ["w1", "w2", "w3"]
+
+            # not enough expert weights to merge
+            if len(expert_cache) < n_experts * len(expert_weights):
+                return []
+
+            tensors: list[tuple[str, Tensor]] = []
+            for w_name in expert_weights:
+                datas: list[Tensor] = []
+
+                for xid in range(n_experts):
+                    ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{w_name}.weight"
+                    datas.append(expert_cache[ename])
+                    del expert_cache[ename]
+
+                data_torch = torch.stack(datas, dim=0)
+                merged_name = f"model.layers.{bid}.block_sparse_moe.experts.{w_name}.weight"
+                new_name = self.map_tensor_name(merged_name)
+                tensors.append((new_name, data_torch))
+
+            del self._experts_cache[bid]
+            return tensors
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("Dots1ForCausalLM")
 class Dots1Model(Qwen2MoeModel):
     model_arch = gguf.MODEL_ARCH.DOTS1
@@ -9493,6 +9771,37 @@ class KimiVLModel(MmprojModel):
 
         return [] # skip other tensors
 
+
+@ModelBase.register("CogVLMForCausalLM")
+class CogVLMVisionModel(MmprojModel):
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams.get("layer_norm_eps", 1e-6))
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.COGVLM)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        if not name.startswith("model.vision."):
+            return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
+@ModelBase.register("CogVLMForCausalLM")
+class CogVLMModel(LlamaModel):
+    model_arch = gguf.MODEL_ARCH.COGVLM
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        # block vision tensors
+        if name.startswith("model.vision."):
+            return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
 ###### CONVERSION LOGIC ######
 
 

convert_hf_to_gguf_update.py

@@ -141,6 +141,7 @@ models = [
     {"name": "mellum",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/JetBrains/Mellum-4b-base", },
     {"name": "bailingmoe2",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-mini-base-2.0", },
     {"name": "granite-docling",  "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ibm-granite/granite-docling-258M", },
+    {"name": "minimax-m2",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/MiniMaxAI/MiniMax-M2", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions
@@ -435,7 +436,7 @@ for model in models:
             tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False)
         else:
             tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
-    except OSError as e:
+    except (OSError, TypeError) as e:
         logger.error(f"Failed to load tokenizer for model {name}. Error: {e}")
         continue  # Skip this model and continue with the next one in the loop
 

ggml/include/ggml.h

@@ -242,6 +242,7 @@
 #define GGML_ROPE_TYPE_NEOX   2
 #define GGML_ROPE_TYPE_MROPE  8
 #define GGML_ROPE_TYPE_VISION 24
+#define GGML_ROPE_TYPE_IMROPE 40 // binary: 101000
 
 #define GGML_MROPE_SECTIONS   4
 

ggml/src/ggml-cpu/ggml-cpu.c

@@ -1613,13 +1613,8 @@ static void ggml_compute_forward_mul_mat_id(
             chunk_size = 64;
         }
 
-#if defined(__aarch64__)
-        // disable for ARM
-        const bool disable_chunking = true;
-#else
         // disable for NUMA
         const bool disable_chunking = ggml_is_numa();
-#endif // defined(__aarch64__)
 
         int64_t nchunk0 = (nr0 + chunk_size - 1) / chunk_size;
         int64_t nchunk1 = (nr1 + chunk_size - 1) / chunk_size;

ggml/src/ggml-cpu/ops.cpp

@@ -5474,7 +5474,7 @@ static void ggml_rope_cache_init(
 }
 
 static void ggml_mrope_cache_init(
-     float theta_base_t, float theta_base_h, float theta_base_w, float theta_base_e, int sections[4], bool indep_sects,
+     float theta_base_t, float theta_base_h, float theta_base_w, float theta_base_e, int sections[4], bool is_imrope, bool indep_sects,
      float freq_scale, const float * freq_factors, float corr_dims[2], int64_t ne0, float ext_factor, float mscale,
      float * cache, float sin_sign, float theta_scale) {
     // ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py
@@ -5509,14 +5509,26 @@ static void ggml_mrope_cache_init(
         }
 
         float theta = theta_t;
-        if (sector >= sections[0] && sector < sec_w) {
-            theta = theta_h;
-        }
-        else if (sector >= sec_w && sector < sec_w + sections[2]) {
-            theta = theta_w;
-        }
-        else if (sector >= sec_w + sections[2]) {
-            theta = theta_e;
+        if (is_imrope) { // qwen3vl apply interleaved mrope
+            if (sector % 3 == 1 && sector < 3 * sections[1]) {
+                theta = theta_h;
+            } else if (sector % 3 == 2 && sector < 3 * sections[2]) {
+                theta = theta_w;
+            } else if (sector % 3 == 0 && sector < 3 * sections[0]) {
+                theta = theta_t;
+            } else {
+                theta = theta_e;
+            }
+        } else {
+            if (sector >= sections[0] && sector < sec_w) {
+                theta = theta_h;
+            }
+            else if (sector >= sec_w && sector < sec_w + sections[2]) {
+                theta = theta_w;
+            }
+            else if (sector >= sec_w + sections[2]) {
+                theta = theta_e;
+            }
         }
 
         rope_yarn(
@@ -5589,6 +5601,7 @@ static void ggml_compute_forward_rope_f32(
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
     const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;  // ggml_rope_multi, multimodal rotary position embedding
+    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE; // qwen3vl apply interleaved mrope
     const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
 
     if (is_mrope) {
@@ -5627,7 +5640,7 @@ static void ggml_compute_forward_rope_f32(
                 const int64_t p_w = pos[i2 + ne2 * 2];
                 const int64_t p_e = pos[i2 + ne2 * 3];
                 ggml_mrope_cache_init(
-                    p_t, p_h, p_w, p_e, sections, is_vision,
+                    p_t, p_h, p_w, p_e, sections, is_imrope, is_vision,
                     freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
             }
 
@@ -5775,6 +5788,7 @@ static void ggml_compute_forward_rope_f16(
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
     const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE;
     const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
 
     if (is_mrope) {
@@ -5813,7 +5827,7 @@ static void ggml_compute_forward_rope_f16(
                 const int64_t p_w = pos[i2 + ne2 * 2];
                 const int64_t p_e = pos[i2 + ne2 * 3];
                 ggml_mrope_cache_init(
-                    p_t, p_h, p_w, p_e, sections, is_vision,
+                    p_t, p_h, p_w, p_e, sections, is_imrope, is_vision,
                     freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
             }
 
@@ -7909,10 +7923,10 @@ void ggml_compute_forward_argsort(
 
 // ggml_compute_forward_flash_attn_ext
 
-static void ggml_compute_forward_flash_attn_ext_f16(
+static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
         const ggml_compute_params * params,
-        ggml_tensor * dst) {
-
+        ggml_tensor * dst,
+        int ir0, int ir1) {
     const ggml_tensor * q     = dst->src[0];
     const ggml_tensor * k     = dst->src[1];
     const ggml_tensor * v     = dst->src[2];
@@ -7928,9 +7942,6 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
     GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
 
-    const int ith = params->ith;
-    const int nth = params->nth;
-
     const int64_t DK = nek0;
     const int64_t DV = nev0;
     const int64_t N  = neq1;
@@ -7964,16 +7975,6 @@ static void ggml_compute_forward_flash_attn_ext_f16(
 
     // parallelize by q rows using ggml_vec_dot_f32
 
-    // total rows in q
-    const int nr = neq1*neq2*neq3;
-
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
-
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
-
     float scale         = 1.0f;
     float max_bias      = 0.0f;
     float logit_softcap = 0.0f;
@@ -8000,6 +8001,8 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     GGML_ASSERT((                            q_to_vec_dot) && "fattn: unsupported K-type");
     GGML_ASSERT((v->type == GGML_TYPE_F32 || v_to_float  ) && "fattn: unsupported V-type");
 
+    int ith = params->ith;
+
     // loop over n_batch and n_head
     for (int ir = ir0; ir < ir1; ++ir) {
         // q indices
@@ -8147,6 +8150,91 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     }
 }
 
+static void ggml_compute_forward_flash_attn_ext_f16(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * q     = dst->src[0];
+    const ggml_tensor * k     = dst->src[1];
+    const ggml_tensor * v     = dst->src[2];
+
+    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nek, k,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbk, k,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nev, v,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbv, v,   nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
+
+    const int64_t DK = nek0;
+    const int64_t DV = nev0;
+    const int64_t N  = neq1;
+
+    GGML_ASSERT(ne0 == DV);
+    GGML_ASSERT(ne2 == N);
+
+    // input tensor rows must be contiguous
+    GGML_ASSERT(nbq0 == ggml_type_size(q->type));
+    GGML_ASSERT(nbk0 == ggml_type_size(k->type));
+    GGML_ASSERT(nbv0 == ggml_type_size(v->type));
+
+    GGML_ASSERT(neq0 == DK);
+    GGML_ASSERT(nek0 == DK);
+    GGML_ASSERT(nev0 == DV);
+
+    GGML_ASSERT(neq1 == N);
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    // parallelize by q rows using ggml_vec_dot_f32
+
+    // total rows in q
+    const int64_t nr = neq1*neq2*neq3;
+
+    // rows per thread
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    // disable for NUMA
+    const bool disable_chunking = ggml_is_numa();
+
+    // 4x chunks per thread
+    int nth_scaled = nth * 4;
+    int64_t chunk_size = (nr + nth_scaled - 1) / nth_scaled;
+    int64_t nchunk     = (nr + chunk_size - 1) / chunk_size;
+
+    if (nth == 1 || nchunk < nth || disable_chunking) {
+        nchunk = nth;
+    }
+
+    if (ith == 0) {
+        // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
+        ggml_threadpool_chunk_set(params->threadpool, nth);
+    }
+
+    ggml_barrier(params->threadpool);
+
+    // The number of elements in each chunk
+    const int64_t dr = (nr + nchunk - 1) / nchunk;
+
+    // The first chunk comes from our thread_id, the rest will get auto-assigned.
+    int current_chunk = ith;
+
+    while (current_chunk < nchunk) {
+        const int64_t ir0 = dr * current_chunk;
+        const int64_t ir1 = MIN(ir0 + dr, nr);
+
+        ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1);
+
+        current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
+    }
+}
+
 void ggml_compute_forward_flash_attn_ext(
         const ggml_compute_params * params,
         ggml_tensor * dst) {

ggml/src/ggml-cpu/repack.cpp

@@ -1600,6 +1600,32 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
         return false;
     }
 
+    void forward_mul_mat_one_chunk(ggml_compute_params * params, ggml_tensor * op, int64_t src0_start, int64_t src0_end) {
+        const ggml_tensor * src0 = op->src[0];
+        const ggml_tensor * src1 = op->src[1];
+        ggml_tensor *       dst  = op;
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        const void * src1_wdata      = params->wdata;
+        const size_t src1_col_stride = ggml_row_size(PARAM_TYPE, ne10);
+
+        // If there are more than three rows in src1, use gemm; otherwise, use gemv.
+        if (ne11 > 3) {
+            gemm<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
+                    (float *) ((char *) dst->data) + src0_start, ne01,
+                    (const char *) src0->data + src0_start * nb01,
+                    (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start);
+        }
+        for (int iter = ne11 - ne11 % 4; iter < ne11; iter++) {
+            gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
+                    (float *) ((char *) dst->data + (iter * nb1)) + src0_start, ne01,
+                    (const char *) src0->data + src0_start * nb01,
+                    (const char *) src1_wdata + (src1_col_stride * iter), 1,
+                    src0_end - src0_start);
+        }
+    }
+
     void forward_mul_mat(ggml_compute_params * params, ggml_tensor * op) {
         const ggml_tensor * src0 = op->src[0];
         const ggml_tensor * src1 = op->src[1];
@@ -1643,31 +1669,41 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
             from_float((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), ne10);
         }
 
+        // disable for NUMA
+        const bool disable_chunking = ggml_is_numa();
+
+        // 4x chunks per thread
+        int64_t nr = ggml_nrows(op->src[0]);
+        int nth_scaled = nth * 4;
+        int64_t chunk_size = (nr + nth_scaled - 1) / nth_scaled;
+        int64_t nchunk     = (nr + chunk_size - 1) / chunk_size;
+
+        if (nth == 1 || nchunk < nth || disable_chunking) {
+            nchunk = nth;
+        }
+
+        if (ith == 0) {
+            // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
+            ggml_threadpool_chunk_set(params->threadpool, nth);
+        }
+
         ggml_barrier(params->threadpool);
 
-        const void * src1_wdata      = params->wdata;
-        const size_t src1_col_stride = ggml_row_size(PARAM_TYPE, ne10);
-        int64_t      src0_start      = (ith * ne01) / nth;
-        int64_t      src0_end        = ((ith + 1) * ne01) / nth;
-        src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start;
-        src0_end   = (src0_end   % NB_COLS) ? src0_end   + NB_COLS - (src0_end   % NB_COLS) : src0_end;
-        if (src0_start >= src0_end) {
-            return;
-        }
+        // The first chunk comes from our thread_id, the rest will get auto-assigned.
+        int current_chunk = ith;
 
-        // If there are more than three rows in src1, use gemm; otherwise, use gemv.
-        if (ne11 > 3) {
-            gemm<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
-                    (float *) ((char *) dst->data) + src0_start, ne01,
-                    (const char *) src0->data + src0_start * nb01,
-                    (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start);
-        }
-        for (int iter = ne11 - ne11 % 4; iter < ne11; iter++) {
-            gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
-                    (float *) ((char *) dst->data + (iter * nb1)) + src0_start, ne01,
-                    (const char *) src0->data + src0_start * nb01,
-                    (const char *) src1_wdata + (src1_col_stride * iter), 1,
-                    src0_end - src0_start);
+        while (current_chunk < nchunk) {
+            int64_t src0_start = (current_chunk * ne01) / nchunk;
+            int64_t src0_end   = ((current_chunk + 1) * ne01) / nchunk;
+            src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start;
+            src0_end   = (src0_end   % NB_COLS) ? src0_end   + NB_COLS - (src0_end   % NB_COLS) : src0_end;
+            if (src0_start >= src0_end) {
+                break;
+            }
+
+            forward_mul_mat_one_chunk(params, dst, src0_start, src0_end);
+
+            current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
         }
     }
 

ggml/src/ggml-cuda/argsort.cu

@@ -87,7 +87,7 @@ template<ggml_sort_order order>
 static __global__ void k_argsort_f32_i32(const float * x, int * dst, const int ncols, int ncols_pad) {
     // bitonic sort
     int col = threadIdx.x;
-    int row = blockIdx.y;
+    int row = blockIdx.x;
 
     if (col >= ncols_pad) {
         return;
@@ -151,7 +151,7 @@ static void argsort_f32_i32_cuda_bitonic(const float *   x,
     const int ncols_pad = next_power_of_2(ncols);
 
     const dim3 block_dims(ncols_pad, 1, 1);
-    const dim3 block_nums(1, nrows, 1);
+    const dim3 block_nums(nrows, 1, 1);
     const size_t shared_mem = ncols_pad * sizeof(int);
 
     // FIXME: this limit could be raised by ~2-4x on Ampere or newer

ggml/src/ggml-cuda/common.cuh

@@ -224,6 +224,11 @@ static const char * cu_get_error_str(CUresult err) {
 #define AMD_MFMA_AVAILABLE
 #endif // defined(GGML_USE_HIP) && defined(CDNA) && !defined(GGML_HIP_NO_MMQ_MFMA)
 
+// The Volta instructions are in principle available on Turing or newer but they are effectively unusable:
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#define VOLTA_MMA_AVAILABLE
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+
 #if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 #define TURING_MMA_AVAILABLE
 #endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
@@ -278,7 +283,10 @@ static bool amd_mfma_available(const int cc) {
 #endif //!defined(GGML_HIP_NO_MMQ_MFMA)
 }
 
-// Volta technically had FP16 tensor cores but they work very differently compared to Turing and later.
+static bool volta_mma_available(const int cc) {
+    return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_VOLTA;
+}
+
 static bool turing_mma_available(const int cc) {
     return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING;
 }

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -27,6 +27,7 @@
 #include "ggml-cuda/mmq.cuh"
 #include "ggml-cuda/mmvf.cuh"
 #include "ggml-cuda/mmvq.cuh"
+#include "ggml-cuda/moe-expert-reduce.cuh"
 #include "ggml-cuda/norm.cuh"
 #include "ggml-cuda/opt-step-adamw.cuh"
 #include "ggml-cuda/opt-step-sgd.cuh"
@@ -3169,6 +3170,31 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                         continue;
                     }
 
+                    if (node->op == GGML_OP_MUL) {
+                        int current_node = i + 1;
+                        int num_views    = 0;
+                        int num_adds     = 0;
+                        while (current_node < cgraph->n_nodes && cgraph->nodes[current_node]->op == GGML_OP_VIEW) {
+                            num_views++;
+                            current_node++;
+                        }
+
+                        while (current_node < cgraph->n_nodes && cgraph->nodes[current_node]->op == GGML_OP_ADD &&
+                                num_adds < num_views - 1) {
+                            num_adds++;
+                            current_node++;
+                        }
+
+                        if (num_adds == num_views - 1 && num_views > 0) {
+                            ggml_tensor * dst_node = cgraph->nodes[current_node - 1];
+                            if (ggml_cuda_should_use_moe_expert_reduce(cgraph, i, current_node)) {
+                                ggml_cuda_op_moe_expert_reduce(*cuda_ctx, node->src[0], node->src[1], dst_node);
+                                i += num_views + num_adds;
+                                continue;
+                            }
+                        }
+                    }
+
                     if (node->op == GGML_OP_ADD) {
                         int n_fuse = 0;
                         ggml_op ops[8];

ggml/src/ggml-cuda/mma.cuh

@@ -18,6 +18,10 @@
 
 #include "common.cuh"
 
+// On Volta each warp is doing 4 8x8 mma operations in parallel.
+// The basic memory layout for a 32x8 output tile is to stack 4 input tiles in I direction and to mirror the B tile.
+// However, the i indices in this file are by default permuted to simplify the index calculations.
+// #define GGML_CUDA_MMA_NO_VOLTA_PERM
 
 #if CUDART_VERSION >= 11080
 
@@ -73,6 +77,15 @@ namespace ggml_cuda_mma {
         static constexpr int ne = I * J / 64;
         T x[ne] = {0};
 
+        static constexpr __device__ bool supported() {
+            if (I == 64 && J ==  2) return true;
+            if (I == 16 && J ==  8) return true;
+            if (I == 32 && J ==  4) return true;
+            if (I == 16 && J == 16) return true;
+            if (I == 32 && J == 32) return true;
+            return false;
+        }
+
         static __device__ __forceinline__ int get_i(const int l) {
             if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
                 return threadIdx.x % 16;
@@ -85,7 +98,8 @@ namespace ggml_cuda_mma {
             } else if constexpr (I == 32 && J == 32) {
                 return 4 * (threadIdx.x / 32) + 8 * (l / 4) + (l % 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 
@@ -101,22 +115,67 @@ namespace ggml_cuda_mma {
             } else if constexpr (I == 32 && J == 32) {
                 return threadIdx.x % 32;
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+#elif __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        static constexpr int ne = I * J / 32;
+        T x[ne] = {0};
+
+        static constexpr __device__ bool supported() {
+            if (I == 32 && J ==  8) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 32 && J == 8) {
+#ifdef GGML_CUDA_MMA_NO_VOLTA_PERM
+                return (((threadIdx.x % 16) / 4) * 8) | ((threadIdx.x / 16) * 4) | (l & 2) | (threadIdx.x % 2);
+#else
+                return (l & 2) | (threadIdx.x & ~2);
+#endif // GGML_CUDA_MMA_NO_VOLTA_PERM
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr (I == 32 && J == 8) {
+                return (threadIdx.x & 2) | (l & (4 + 1));
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 #else
         static constexpr int ne = I * J / 32;
         T x[ne] = {0};
 
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  4) return true;
+            if (I ==  8 && J ==  8) return true;
+            if (I == 16 && J ==  8) return true;
+            if (I == 16 && J == 16) return true;
+            if (I == 32 && J ==  8) return true;
+            return false;
+        }
+
         static __device__ __forceinline__ int get_i(const int l) {
-            if constexpr (I == 8 && (J == 4 || J == 8)) {
+            if constexpr (I == 8 && J == 4) {
+                return threadIdx.x / 4;
+            } else if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 8) {
-                return (l / 2) * 8 + threadIdx.x / 4;
+                return ((l / 2) * 8) | (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 16) {
-                return ((l / 2) % 2) * 8 + threadIdx.x / 4;
+                return (((l / 2) % 2) * 8) | (threadIdx.x / 4);
+            } else if constexpr (I == 32 && J == 8) {
+                return tile<16, 8, T>::get_i(l); // Memory layout simply repeated with same pattern in i direction.
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 
@@ -124,13 +183,16 @@ namespace ggml_cuda_mma {
             if constexpr (I == 8 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 8 && J == 8) {
-                return 4 * l + threadIdx.x % 4;
+                return (l * 4) | (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 8) {
-                return 2 * (threadIdx.x % 4) + l % 2;
+                return ((threadIdx.x % 4) * 2) | (l % 2);
             } else if constexpr (I == 16 && J == 16) {
-                return 8 * (l / 4) + 2 * (threadIdx.x % 4) + l % 2;
+                return ((l / 4) * 8) | ((threadIdx.x % 4) * 2) | (l % 2);
+            } else if constexpr (I == 32 && J == 8) {
+                return tile<16, 8, T>::get_j(l); // Memory layout simply repeated with same pattern in i direction.
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 #endif // defined(GGML_USE_HIP)
@@ -140,32 +202,83 @@ namespace ggml_cuda_mma {
     struct tile<I_, J_, half2> {
         static constexpr int I  = I_;
         static constexpr int J  = J_;
+
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        static constexpr int ne = I == 8 && J == 8 ? I * J / (WARP_SIZE/4) : I * J / WARP_SIZE;
+        half2 x[ne] = {{0.0f, 0.0f}};
+
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  8) return true;
+            if (I == 32 && J ==  8) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 8 && J == 8) {
+                return ((threadIdx.x / 16) * 4) | (threadIdx.x % 4);
+            } else if constexpr (I == 32 && J == 8) {
+#ifdef GGML_CUDA_MMA_NO_VOLTA_PERM
+                return (((threadIdx.x % 16) / 4) * 8) | ((threadIdx.x / 16) * 4) | (threadIdx.x % 4);
+#else
+                return threadIdx.x;
+#endif // GGML_CUDA_MMA_NO_VOLTA_PERM
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr ((I == 8 || I == 32) && J == 8) {
+                return l;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+#else
         static constexpr int ne = I * J / WARP_SIZE;
         half2 x[ne] = {{0.0f, 0.0f}};
 
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  4) return true;
+            if (I ==  8 && J ==  8) return true;
+            if (I == 16 && J ==  8) return true;
+            if (I == 16 && J == 16) return true;
+            if (I == 32 && J ==  8) return true;
+            return false;
+        }
+
         static __device__ __forceinline__ int get_i(const int l) {
             if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 4) {
-                return l * 8 + threadIdx.x / 4;
+                return (l * 8) | (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 8) {
-                return (l % 2) * 8 + threadIdx.x / 4;
+                return ((l % 2) * 8) | (threadIdx.x / 4);
+            } else if constexpr (I == 32 && J == 8) {
+                return ((l / 4) * 16) | ((l % 2) * 8) | (threadIdx.x / 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 8 && J == 8) {
-                return l * 4 + threadIdx.x % 4;
+                return (l * 4) | (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 16 && J == 8) {
-                return (l / 2) * 4 + threadIdx.x % 4;
+                return ((l / 2) * 4) | (threadIdx.x % 4);
+            } else if constexpr (I == 32 && J == 8) {
+                return ((l & 2) * 2) | (threadIdx.x % 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
     };
 
     template <int I_, int J_>
@@ -175,27 +288,36 @@ namespace ggml_cuda_mma {
         static constexpr int ne = I * J / WARP_SIZE;
         nv_bfloat162 x[ne] = {{0.0f, 0.0f}};
 
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  8) return true;
+            if (I == 16 && J ==  4) return true;
+            if (I == 16 && J ==  8) return true;
+            return false;
+        }
+
         static __device__ __forceinline__ int get_i(const int l) {
             if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 4) {
-                return l * 8 + threadIdx.x / 4;
+                return (l * 8) | (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 8) {
-                return (l % 2) * 8 + threadIdx.x / 4;
+                return ((l % 2) * 8) | (threadIdx.x / 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 8 && J == 8) {
-                return l * 4 + threadIdx.x % 4;
+                return (l * 4) | (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 16 && J == 8) {
-                return (l / 2) * 4 + threadIdx.x % 4;
+                return ((l / 2) * 4) | (threadIdx.x % 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
     };
@@ -263,8 +385,12 @@ namespace ggml_cuda_mma {
             : "=r"(xi[0]), "=r"(xi[1])
             : "l"(xs));
 #else
-        load_generic(xs0, stride);
-        GGML_UNUSED(t);
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        GGML_UNUSED_VARS(t, xs0, stride);
+        NO_DEVICE_CODE;
+#else
+        load_generic(t, xs0, stride);
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
 #endif // TURING_MMA_AVAILABLE
     }
 
@@ -277,11 +403,35 @@ namespace ggml_cuda_mma {
         asm volatile("ldmatrix.sync.aligned.m8n8.x4.b16 {%0, %1, %2, %3}, [%4];"
             : "=r"(xi[0]), "=r"(xi[1]), "=r"(xi[2]), "=r"(xi[3])
             : "l"(xs));
+#else
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        GGML_UNUSED_VARS(t, xs0, stride);
+        NO_DEVICE_CODE;
 #else
         load_generic(t, xs0, stride);
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
 #endif // TURING_MMA_AVAILABLE
     }
 
+    template <typename T>
+    static __device__ __forceinline__ void load_ldmatrix(
+            tile<32, 8, T> & t, const T * __restrict__ xs0, const int stride) {
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if 1
+        // TODO: more generic handling
+        static_assert(sizeof(T) == 4, "bad type size");
+        ggml_cuda_memcpy_1<4*sizeof(T)>(t.x + 0, xs0 + t.get_i(0)*stride + 0);
+        ggml_cuda_memcpy_1<4*sizeof(T)>(t.x + 4, xs0 + t.get_i(4)*stride + 4);
+#else
+        load_generic(t, xs0, stride);
+#endif // 1
+#else
+        tile<16, 8, T> * t16 = (tile<16, 8, T> *) &t;
+        load_ldmatrix(t16[0], xs0 +  0*stride, stride);
+        load_ldmatrix(t16[1], xs0 + 16*stride, stride);
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+    }
+
     template <typename T>
     static __device__ __forceinline__ void load_ldmatrix_trans(
             tile<16, 8, T> & t, const T * __restrict__ xs0, const int stride) {
@@ -546,4 +696,43 @@ namespace ggml_cuda_mma {
         NO_DEVICE_CODE;
 #endif // AMD_MFMA_AVAILABLE
     }
+
+    template <typename T1, typename T2, int J, int K>
+    static __device__ __forceinline__ void mma(
+            tile<32, J, T1> & D, const tile<32, K, T2> & A, const tile<J, K, T2> & B) {
+        tile<16, J, T1> * D16 = (tile<16, J, T1> *) &D;
+        tile<16, K, T2> * A16 = (tile<16, K, T2> *) &A;
+        mma(D16[0], A16[0], B);
+        mma(D16[1], A16[1], B);
+    }
+
+    static __device__ __forceinline__ void mma(
+            tile<32, 8, float> & D, const tile<32, 8, half2> & A, const tile<8, 8, half2> & B) {
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        const int * Axi = (const int *) A.x;
+        const int * Bxi = (const int *) B.x;
+        int       * Dxi = (int       *) D.x;
+        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
+            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[0]), "r"(Axi[1]), "r"(Bxi[0]), "r"(Bxi[1]));
+        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
+            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[2]), "r"(Bxi[3]));
+        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
+            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[4]), "r"(Axi[5]), "r"(Bxi[4]), "r"(Bxi[5]));
+        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
+            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[6]), "r"(Axi[7]), "r"(Bxi[6]), "r"(Bxi[7]));
+#else
+        tile<16, 8, float> * D16 = (tile<16, 8, float> *) &D;
+        tile<16, 8, half2> * A16 = (tile<16, 8, half2> *) &A;
+        mma(D16[0], A16[0], B);
+        mma(D16[1], A16[1], B);
+#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+    }
 }

ggml/src/ggml-cuda/mmf.cu

@@ -148,7 +148,7 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
         case GGML_TYPE_F32:
             return ampere_mma_available(cc);
         case GGML_TYPE_F16:
-            return turing_mma_available(cc);
+            return volta_mma_available(cc) || turing_mma_available(cc);
         case GGML_TYPE_BF16:
             return ampere_mma_available(cc);
         default:

ggml/src/ggml-cuda/mmf.cuh

@@ -28,9 +28,19 @@ static __global__ void mul_mat_f(
         const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
         const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
 #if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-    typedef tile<16, 8, T>     tile_A;
-    typedef tile< 8, 8, T>     tile_B;
-    typedef tile<16, 8, float> tile_C;
+    constexpr bool I_16_supported = tile<16, 8, T>::supported() && tile<16, 8, float>::supported();
+    constexpr bool I_32_supported = tile<32, 8, T>::supported() && tile<32, 8, float>::supported();
+
+    if (!I_16_supported && !I_32_supported) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    constexpr int I_preferred = I_16_supported ? 16 : 32; // For Turing MMA both work but 16 is ~1% faster.
+
+    typedef tile<I_preferred, 8, T>     tile_A;
+    typedef tile<8,           8, T>     tile_B;
+    typedef tile<I_preferred, 8, float> tile_C;
 
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     constexpr int tile_k_padded = warp_size + 4;
@@ -232,7 +242,6 @@ static __global__ void mul_mat_f(
 #endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 }
 
-
 //This kernel is for larger batch sizes of mul_mat_id
 template <typename T, int rows_per_block, int cols_per_block, int nwarps>
 __launch_bounds__(ggml_cuda_get_physical_warp_size()*nwarps, 1)
@@ -245,9 +254,19 @@ static __global__ void mul_mat_f_ids(
         const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
         const uint3 sis1_fd, const uint3 nch_fd) {
 #if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-    typedef tile<16, 8, T>     tile_A;
-    typedef tile< 8, 8, T>     tile_B;
-    typedef tile<16, 8, float> tile_C;
+    constexpr bool I_16_supported = tile<16, 8, T>::supported() && tile<16, 8, float>::supported();
+    constexpr bool I_32_supported = tile<32, 8, T>::supported() && tile<32, 8, float>::supported();
+
+    if (!I_16_supported && !I_32_supported) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    constexpr int I_preferred = I_16_supported ? 16 : 32; // For Turing MMA both work butr 16 is ~1% faster.
+
+    typedef tile<I_preferred, 8, T>     tile_A;
+    typedef tile<8,           8, T>     tile_B;
+    typedef tile<I_preferred, 8, float> tile_C;
 
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     constexpr int tile_k_padded = warp_size + 4;
@@ -533,7 +552,8 @@ void mul_mat_f_cuda(
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
         cudaStream_t stream, const mmf_ids_data * ids_data) {
-    typedef tile<16, 8, T>     tile_A;
+    typedef tile<16, 8, T>     tile_A_16;
+    typedef tile<32, 8, T>     tile_A_32;
     typedef tile< 8, 8, T>     tile_B;
 
     GGML_ASSERT(ncols_x      % 2 == 0);
@@ -544,7 +564,8 @@ void mul_mat_f_cuda(
     const int64_t channel_ratio = nchannels_dst / nchannels_x;
     const int64_t sample_ratio  = nsamples_dst  / nsamples_x;
 
-    const int device = ggml_cuda_get_device();
+    const int device    = ggml_cuda_get_device();
+    const int cc        = ggml_cuda_info().devices[device].cc;
     const int warp_size = ggml_cuda_info().devices[device].warp_size;
 
     int64_t nwarps_best     = 1;
@@ -559,7 +580,7 @@ void mul_mat_f_cuda(
     }
 
     constexpr int rows_per_block = MMF_ROWS_PER_BLOCK;
-    const int nbytes_shared_iter = nwarps_best * tile_A::I * (warp_size + 4) * 4;
+    const int nbytes_shared_iter = nwarps_best * (volta_mma_available(cc) ? tile_A_32::I : tile_A_16::I) * (warp_size + 4) * 4;
     const int nbytes_shared_combine = GGML_PAD(cols_per_block, tile_B::I) * (nwarps_best*rows_per_block + 4) * 4;
     const int nbytes_shared = std::max(nbytes_shared_iter, nbytes_shared_combine);
     const int nbytes_slotmap = ids ? GGML_PAD(cols_per_block, 16) * sizeof(int) : 0;

ggml/src/ggml-cuda/mmvq.cu

@@ -190,12 +190,28 @@ static __global__ void mul_mat_vec_q(
 
     const uint32_t channel_bias = ids ? channel_x : channel_dst;
 
+    float x_biases[ncols_dst][rows_per_cuda_block]    = { { 0.0f } };
+    float gate_biases[ncols_dst][rows_per_cuda_block] = { { 0.0f } };
     if constexpr (has_fusion) {
         if (use_bias) {
             x_bias = x_bias + sample_dst*stride_sample_dst + channel_bias*stride_channel_dst + row0;
+            // 1. Hide latency by prefetching bias and gate here
+            // 2. load only on threads that won't die after partial sum calculation
+            if (threadIdx.x < rows_per_cuda_block && threadIdx.y == 0 &&
+                (rows_per_cuda_block == 1 || uint32_t(row0 + threadIdx.x) < stride_col_dst)) {
+                for (int j = 0; j < ncols_dst; ++j) {
+                    x_biases[j][threadIdx.x] = x_bias[j * stride_col_dst + threadIdx.x];
+                }
+            }
         }
         if (use_gate_bias) {
             gate_bias = gate_bias + sample_dst*stride_sample_dst + channel_bias*stride_channel_dst + row0;
+            if (threadIdx.x < rows_per_cuda_block && threadIdx.y == 0 &&
+                (rows_per_cuda_block == 1 || uint32_t(row0 + threadIdx.x) < stride_col_dst)) {
+                for (int j = 0; j < ncols_dst; ++j) {
+                    gate_biases[j][threadIdx.x] = gate_bias[j * stride_col_dst + threadIdx.x];
+                }
+            }
         }
     }
 
@@ -283,12 +299,12 @@ static __global__ void mul_mat_vec_q(
             float result = tmp[j][threadIdx.x];
             if constexpr (has_fusion) {
                 if (use_bias) {
-                    result += x_bias[j*stride_col_dst + threadIdx.x];
+                    result += x_biases[j][threadIdx.x];
                 }
                 if (use_gate) {
                     float gate_value = tmp_gate[j][threadIdx.x];
                     if (use_gate_bias) {
-                        gate_value += gate_bias[j*stride_col_dst + threadIdx.x];
+                        gate_value += gate_biases[j][threadIdx.x];
                     }
                     switch (active_glu) {
                         case GGML_GLU_OP_SWIGLU:

ggml/src/ggml-cuda/moe-expert-reduce.cu

@@ -0,0 +1,168 @@
+#include "moe-expert-reduce.cuh"
+
+// This kernel is a fusion of the expert weight reduce, common in MoE models
+
+template <int n_expert_used_template>
+__global__ void moe_expert_reduce_cuda(const float * __restrict__ experts,
+                                       const float * __restrict__ weights,
+                                       float * __restrict__ dst,
+                                       const int n_expert_used,
+                                       const int n_cols) {
+    const int row = blockIdx.x;
+    const int col = blockIdx.y * blockDim.x + threadIdx.x;
+    if (col >= n_cols) {
+        return;
+    }
+
+    experts += row * n_cols * n_expert_used;
+    weights += row * n_expert_used;
+    dst += row * n_cols;
+
+    float acc = 0.f;
+    if constexpr (n_expert_used_template == 0) {
+        for (int expert = 0; expert < n_expert_used; ++expert) {
+            ggml_cuda_mad(acc, experts[col], weights[expert]);
+            experts += n_cols;
+        }
+        dst[col] = acc;
+    } else {
+#pragma unroll
+        for (int i = 0; i < n_expert_used_template; ++i) {
+            ggml_cuda_mad(acc, experts[col], weights[i]);
+            experts += n_cols;
+        }
+        dst[col] = acc;
+    }
+}
+
+static void launch_moe_expert_reduce(ggml_backend_cuda_context & ctx,
+                                     const float *               experts,
+                                     const float *               weights,
+                                     float *                     dst,
+                                     const int                   n_expert_used,
+                                     const int                   n_cols,
+                                     const int                   n_rows) {
+    const int block_size = 32;
+
+    const int n_blocks_x = n_rows;
+    const int n_blocks_y = (n_cols + block_size - 1) / block_size;
+
+    dim3 block_dims(block_size);
+    dim3 grid_dims(n_blocks_x, n_blocks_y);
+
+    cudaStream_t stream = ctx.stream();
+    switch (n_expert_used) {
+        case 1:
+            moe_expert_reduce_cuda<1>
+                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
+            break;
+        case 2:
+            moe_expert_reduce_cuda<2>
+                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
+            break;
+        case 4:
+            moe_expert_reduce_cuda<4>
+                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
+            break;
+        case 6:
+            moe_expert_reduce_cuda<6>
+                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
+            break;
+        case 8:
+            moe_expert_reduce_cuda<8>
+                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
+            break;
+        case 16:
+            moe_expert_reduce_cuda<16>
+                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
+            break;
+        case 32:
+            moe_expert_reduce_cuda<32>
+                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
+            break;
+        case 64:
+            moe_expert_reduce_cuda<64>
+                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
+            break;
+        case 128:
+            moe_expert_reduce_cuda<128>
+                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
+            break;
+        default:
+            moe_expert_reduce_cuda<0>
+                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
+            break;
+    }
+}
+
+bool ggml_cuda_should_use_moe_expert_reduce(const ggml_cgraph * cgraph, int start_index, int end_index) {
+    const ggml_tensor * mul = cgraph->nodes[start_index];
+
+    if (mul->op != GGML_OP_MUL || !ggml_is_contiguous(mul->src[0]) || !ggml_is_contiguous(mul->src[1])) {
+        return false;
+    }
+
+    int    current_node   = start_index + 1;
+    size_t current_offset = 0;
+
+    std::vector<const ggml_tensor *> view_nodes;
+    //check if all are views of the expert in increasing order
+    while (current_node < end_index && cgraph->nodes[current_node]->op == GGML_OP_VIEW) {
+        const ggml_tensor * node = cgraph->nodes[current_node];
+        if (node->view_src != mul) {
+            return false;
+        }
+        if (node->view_offs < current_offset) {
+            return false;
+        }
+        current_offset = node->view_offs;
+        current_node++;
+        view_nodes.push_back(node);
+    }
+
+    //check if all the adds are in increasing order
+    const ggml_tensor * prev_add_src = view_nodes.empty() ? nullptr : view_nodes[0];
+    int                 num_adds     = 0;
+    int                 num_views    = view_nodes.size();
+    while (current_node < end_index && cgraph->nodes[current_node]->op == GGML_OP_ADD) {
+        const ggml_tensor * add_node = cgraph->nodes[current_node];
+
+        bool is_first_op_ok  = num_views > num_adds ? add_node->src[0] == prev_add_src : false;
+        bool is_second_op_ok = num_views > num_adds ? add_node->src[1] == view_nodes[num_adds + 1] : false;
+
+        if (!is_first_op_ok || !is_second_op_ok) {
+            return false;
+        }
+        prev_add_src = add_node;
+
+        num_adds++;
+        current_node++;
+    }
+
+    if (num_views != num_adds + 1) {
+        return false;
+    }
+
+    return true;
+}
+
+void ggml_cuda_op_moe_expert_reduce(ggml_backend_cuda_context & ctx,
+                                    const ggml_tensor *         experts,
+                                    const ggml_tensor *         weights,
+                                    ggml_tensor *               dst) {
+    const int n_rows        = experts->ne[2];
+    const int n_expert_used = experts->ne[1];
+    const int n_cols        = experts->ne[0];
+
+    GGML_ASSERT(experts->type == GGML_TYPE_F32);
+    GGML_ASSERT(weights->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(experts));
+    GGML_ASSERT(ggml_is_contiguous(weights));
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    const float * experts_d = (const float *) experts->data;
+    const float * weights_d = (const float *) weights->data;
+    float *       dst_d     = (float *) dst->data;
+
+    launch_moe_expert_reduce(ctx, experts_d, weights_d, dst_d, n_expert_used, n_cols, n_rows);
+}

ggml/src/ggml-cuda/moe-expert-reduce.cuh

@@ -0,0 +1,11 @@
+#include "common.cuh"
+#include "ggml.h"
+
+#include <initializer_list>
+
+void ggml_cuda_op_moe_expert_reduce(ggml_backend_cuda_context & ctx,
+                                    const ggml_tensor *         experts,
+                                    const ggml_tensor *         weights,
+                                    ggml_tensor *               dst);
+
+bool ggml_cuda_should_use_moe_expert_reduce(const ggml_cgraph * cgraph, int start_index, int end_index);

ggml/src/ggml-cuda/rope.cu

@@ -125,7 +125,7 @@ template<bool forward, bool has_ff, typename T>
 static __global__ void rope_multi(
         const T * x, T * dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2,
         const int n_dims, const int32_t * pos, const float freq_scale, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float theta_scale, const float * freq_factors, const mrope_sections sections) {
+        const rope_corr_dims corr_dims, const float theta_scale, const float * freq_factors, const mrope_sections sections, const bool is_imrope) {
     const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
 
     if (i0 >= ne0) {
@@ -152,17 +152,29 @@ static __global__ void rope_multi(
     const int sector = (i0 / 2) % sect_dims;
 
     float theta_base = 0.0;
-    if (sector < sections.v[0]) {
-        theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
-    }
-    else if (sector >= sections.v[0] && sector < sec_w) {
-        theta_base = pos[channel_x + ne2 * 1]*powf(theta_scale, i0/2.0f);
-    }
-    else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
-        theta_base = pos[channel_x + ne2 * 2]*powf(theta_scale, i0/2.0f);
-    }
-    else if (sector >= sec_w + sections.v[2]) {
-        theta_base = pos[channel_x + ne2 * 3]*powf(theta_scale, i0/2.0f);
+    if (is_imrope) {
+        if (sector % 3 == 1 && sector < 3 * sections.v[1]) { // h
+            theta_base = pos[channel_x + ne2 * 1]*powf(theta_scale, i0/2.0f);
+        } else if (sector % 3 == 2 && sector < 3 * sections.v[2]) { // w
+            theta_base = pos[channel_x + ne2 * 2]*powf(theta_scale, i0/2.0f);
+        } else if (sector % 3 == 0 && sector < 3 * sections.v[0]) { // t
+            theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
+        } else {
+            theta_base = pos[channel_x + ne2 * 3]*powf(theta_scale, i0/2.0f);
+        }
+    } else {
+        if (sector < sections.v[0]) {
+            theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
+        }
+        else if (sector >= sections.v[0] && sector < sec_w) {
+            theta_base = pos[channel_x + ne2 * 1]*powf(theta_scale, i0/2.0f);
+        }
+        else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
+            theta_base = pos[channel_x + ne2 * 2]*powf(theta_scale, i0/2.0f);
+        }
+        else if (sector >= sec_w + sections.v[2]) {
+            theta_base = pos[channel_x + ne2 * 3]*powf(theta_scale, i0/2.0f);
+        }
     }
 
     const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
@@ -276,7 +288,7 @@ template<bool forward, typename T>
 static void rope_multi_cuda(
         const T * x, T * dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2, const int n_dims, const int nr,
         const int32_t * pos, const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float * freq_factors, const mrope_sections sections, cudaStream_t stream) {
+        const rope_corr_dims corr_dims, const float * freq_factors, const mrope_sections sections, const bool is_imrope, cudaStream_t stream) {
     GGML_ASSERT(ne0 % 2 == 0);
     const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
     const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
@@ -287,11 +299,11 @@ static void rope_multi_cuda(
     if (freq_factors == nullptr) {
         rope_multi<forward, false, T><<<block_nums, block_dims, 0, stream>>>(
             x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, sections);
+            attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope);
     } else {
         rope_multi<forward, true, T><<<block_nums, block_dims, 0, stream>>>(
             x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, sections);
+            attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope);
     }
 }
 
@@ -369,6 +381,7 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
     const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE;
     const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
 
     if (is_mrope) {
@@ -406,11 +419,11 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
         if (src0->type == GGML_TYPE_F32) {
             rope_multi_cuda<forward>(
                 (const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02, n_dims, nr, pos, freq_scale,
-                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, stream);
+                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, is_imrope, stream);
         } else if (src0->type == GGML_TYPE_F16) {
             rope_multi_cuda<forward>(
                 (const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02, n_dims, nr, pos, freq_scale,
-                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, stream);
+                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, is_imrope, stream);
         } else {
             GGML_ABORT("fatal error");
         }

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -676,6 +676,15 @@ static void repack_q4_0_q4x4x2(ggml_tensor * t, const void * data, size_t size)
     size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_Q4_0x4x2));  // extra elements for the pad
     size_t row_size_rp = row_size * 2;  // extra space for tmp pad (if any)
 
+    // Ensure we don't try to read more data than is available in the source buffer 'data'
+    // or write more than the tensor can hold.
+    const size_t total_tensor_size = (size_t)nrows * row_size;
+    const size_t n_bytes_to_copy = size < total_tensor_size ? size : total_tensor_size;
+
+    // Calculate how many full rows and how many remaining bytes we need to process.
+    const int64_t n_full_rows = n_bytes_to_copy / row_size;
+    const size_t  n_rem_bytes = n_bytes_to_copy % row_size;
+
     void * buf_pd = ggml_aligned_malloc(row_size_pd);
     GGML_ASSERT(buf_pd != NULL);
 
@@ -687,7 +696,8 @@ static void repack_q4_0_q4x4x2(ggml_tensor * t, const void * data, size_t size)
 
     init_row_q4x4x2((block_q4_0 *) buf_pd, t->ne[0]);  // init padded buffer to make sure the tail is all zeros
 
-    for (int64_t i = 0; i < nrows; i++) {
+    // 1. Process all the full rows
+    for (int64_t i = 0; i < n_full_rows; i++) {
         const uint8_t * src = (const uint8_t *) data + (i * row_size);
         uint8_t *       dst = (uint8_t *) t->data + (i * row_size);
 
@@ -696,6 +706,25 @@ static void repack_q4_0_q4x4x2(ggml_tensor * t, const void * data, size_t size)
         memcpy(dst, buf_rp, row_size);
     }
 
+    // 2. Process the final, potentially partial, row
+    if (n_rem_bytes > 0) {
+        const int64_t i = n_full_rows;
+        const uint8_t * src = (const uint8_t *) data + (i * row_size);
+        uint8_t *       dst = (uint8_t *) t->data + (i * row_size);
+
+        // re-init the row because we are potentially copying a partial row
+        init_row_q4x4x2((block_q4_0 *) buf_pd, t->ne[0]);
+
+        // Copy only the remaining bytes from the source.
+        memcpy(buf_pd, src, n_rem_bytes);
+
+        // Repack the entire buffer
+        repack_row_q4x4x2((uint8_t *) buf_rp, (const block_q4_0 *) buf_pd, t->ne[0]);
+
+        // Write only the corresponding remaining bytes to the destination tensor.
+        memcpy(dst, buf_rp, n_rem_bytes);
+    }
+
     ggml_aligned_free(buf_pd, row_size_pd);
     ggml_aligned_free(buf_rp, row_size_rp);
 }
@@ -708,6 +737,14 @@ static void repack_q4x4x2_q4_0(void * data, const ggml_tensor * t, size_t size)
     size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_Q4_0x4x2));  // extra elements for the pad
     size_t row_size_rp = row_size * 2;  // extra space for tmp pad (if any)
 
+    // Ensure we don't try to copy more data than the tensor actually contains.
+    const size_t total_tensor_size = (size_t)nrows * row_size;
+    const size_t n_bytes_to_copy = size < total_tensor_size ? size : total_tensor_size;
+
+    // Calculate how many full rows and how many remaining bytes we need to process.
+    const int64_t n_full_rows = n_bytes_to_copy / row_size;
+    const size_t  n_rem_bytes = n_bytes_to_copy % row_size;
+
     void * buf_pd = ggml_aligned_malloc(row_size_pd);
     GGML_ASSERT(buf_pd != NULL);
 
@@ -719,7 +756,8 @@ static void repack_q4x4x2_q4_0(void * data, const ggml_tensor * t, size_t size)
 
     memset(buf_pd, 0, row_size_pd);  // clear-out padded buffer to make sure the tail is all zeros
 
-    for (int64_t i = 0; i < nrows; i++) {
+    // 1. Process all the full rows
+    for (int64_t i = 0; i < n_full_rows; i++) {
         const uint8_t * src = (const uint8_t *) t->data + (i * row_size);
         uint8_t *       dst = (uint8_t *) data + (i * row_size);
 
@@ -728,6 +766,20 @@ static void repack_q4x4x2_q4_0(void * data, const ggml_tensor * t, size_t size)
         memcpy(dst, buf_rp, row_size);
     }
 
+    // 2. Process the final, potentially partial, row
+    if (n_rem_bytes > 0) {
+        const int64_t i = n_full_rows;
+        const uint8_t * src = (const uint8_t *) t->data + (i * row_size);
+        uint8_t *       dst = (uint8_t *) data + (i * row_size);
+
+        // We still need to read and unpack the entire source row because quantization is block-based.
+        memcpy(buf_pd, src, row_size);
+        unpack_row_q4x4x2((block_q4_0 *) buf_rp, (const uint8_t *) buf_pd, t->ne[0]);
+
+        // But we only copy the remaining number of bytes to the destination.
+        memcpy(dst, buf_rp, n_rem_bytes);
+    }
+
     ggml_aligned_free(buf_pd, row_size_pd);
     ggml_aligned_free(buf_rp, row_size_rp);
 }
@@ -950,6 +1002,15 @@ static void repack_q8_0_q8x4x2(ggml_tensor * t, const void * data, size_t size)
     size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_Q8_0x4x2));  // extra elements for the pad
     size_t row_size_rp = row_size * 2;  // extra space for tmp pad (if any)
 
+    // Ensure we don't try to read more data than is available in the source buffer 'data'
+    // or write more than the tensor can hold.
+    const size_t total_tensor_size = (size_t)nrows * row_size;
+    const size_t n_bytes_to_copy = size < total_tensor_size ? size : total_tensor_size;
+
+    // Calculate how many full rows and how many remaining bytes we need to process.
+    const int64_t n_full_rows = n_bytes_to_copy / row_size;
+    const size_t  n_rem_bytes = n_bytes_to_copy % row_size;
+
     void * buf_pd = ggml_aligned_malloc(row_size_pd);
     GGML_ASSERT(buf_pd != NULL);
 
@@ -961,7 +1022,8 @@ static void repack_q8_0_q8x4x2(ggml_tensor * t, const void * data, size_t size)
 
     init_row_q8x4x2((block_q8_0 *) buf_pd, t->ne[0]);  // init padded buffer to make sure the tail is all zeros
 
-    for (int64_t i = 0; i < nrows; i++) {
+    // 1. Process all the full rows
+    for (int64_t i = 0; i < n_full_rows; i++) {
         const uint8_t * src = (const uint8_t *) data + (i * row_size);
         uint8_t *       dst = (uint8_t *) t->data + (i * row_size);
 
@@ -970,6 +1032,25 @@ static void repack_q8_0_q8x4x2(ggml_tensor * t, const void * data, size_t size)
         memcpy(dst, buf_rp, row_size);
     }
 
+    // 2. Process the final, potentially partial, row
+    if (n_rem_bytes > 0) {
+        const int64_t i = n_full_rows;
+        const uint8_t * src = (const uint8_t *) data + (i * row_size);
+        uint8_t *       dst = (uint8_t *) t->data + (i * row_size);
+
+        // re-init the row because we are potentially copying a partial row
+        init_row_q8x4x2((block_q8_0 *) buf_pd, t->ne[0]);
+
+        // Copy only the remaining bytes from the source.
+        memcpy(buf_pd, src, n_rem_bytes);
+
+        // Repack the entire buffer
+        repack_row_q8x4x2((uint8_t *) buf_rp, (const block_q8_0 *) buf_pd, t->ne[0]);
+
+        // Write only the corresponding remaining bytes to the destination tensor.
+        memcpy(dst, buf_rp, n_rem_bytes);
+    }
+
     ggml_aligned_free(buf_pd, row_size_pd);
     ggml_aligned_free(buf_rp, row_size_rp);
 }
@@ -982,6 +1063,14 @@ static void repack_q8x4x2_q8_0(void * data, const ggml_tensor * t, size_t size)
     size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_Q8_0x4x2));  // extra elements for the pad
     size_t row_size_rp = row_size * 2;  // extra space for tmp pad (if any)
 
+    // Ensure we don't try to copy more data than the tensor actually contains.
+    const size_t total_tensor_size = (size_t)nrows * row_size;
+    const size_t n_bytes_to_copy = size < total_tensor_size ? size : total_tensor_size;
+
+    // Calculate how many full rows and how many remaining bytes we need to process.
+    const int64_t n_full_rows = n_bytes_to_copy / row_size;
+    const size_t  n_rem_bytes = n_bytes_to_copy % row_size;
+
     void * buf_pd = ggml_aligned_malloc(row_size_pd);
     GGML_ASSERT(buf_pd != NULL);
 
@@ -993,7 +1082,8 @@ static void repack_q8x4x2_q8_0(void * data, const ggml_tensor * t, size_t size)
 
     memset(buf_pd, 0, row_size_pd);  // clear-out padded buffer to make sure the tail is all zeros
 
-    for (int64_t i = 0; i < nrows; i++) {
+    // 1. Process all the full rows
+    for (int64_t i = 0; i < n_full_rows; i++) {
         const uint8_t * src = (const uint8_t *) t->data + (i * row_size);
         uint8_t *       dst = (uint8_t *) data + (i * row_size);
 
@@ -1002,6 +1092,20 @@ static void repack_q8x4x2_q8_0(void * data, const ggml_tensor * t, size_t size)
         memcpy(dst, buf_rp, row_size);
     }
 
+    // 2. Process the final, potentially partial, row
+    if (n_rem_bytes > 0) {
+        const int64_t i = n_full_rows;
+        const uint8_t * src = (const uint8_t *) t->data + (i * row_size);
+        uint8_t *       dst = (uint8_t *) data + (i * row_size);
+
+        // We still need to read and unpack the entire source row because quantization is block-based.
+        memcpy(buf_pd, src, row_size);
+        unpack_row_q8x4x2((block_q8_0 *) buf_rp, (const uint8_t *) buf_pd, t->ne[0]);
+
+        // But we only copy the remaining number of bytes to the destination.
+        memcpy(dst, buf_rp, n_rem_bytes);
+    }
+
     ggml_aligned_free(buf_pd, row_size_pd);
     ggml_aligned_free(buf_rp, row_size_rp);
 }
@@ -1249,6 +1353,15 @@ static void repack_mxfp4_mxfp4x4x2(ggml_tensor * t, const void * data, size_t si
     size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_MXFP4x4x2));  // extra elements for the pad
     size_t row_size_rp = row_size * 2;  // extra space for tmp pad (if any)
 
+    // Ensure we don't try to read more data than is available in the source buffer 'data'
+    // or write more than the tensor can hold.
+    const size_t total_tensor_size = (size_t)nrows * row_size;
+    const size_t n_bytes_to_copy = size < total_tensor_size ? size : total_tensor_size;
+
+    // Calculate how many full rows and how many remaining bytes we need to process.
+    const int64_t n_full_rows = n_bytes_to_copy / row_size;
+    const size_t  n_rem_bytes = n_bytes_to_copy % row_size;
+
     void * buf_pd = ggml_aligned_malloc(row_size_pd);
     GGML_ASSERT(buf_pd != NULL);
 
@@ -1260,7 +1373,8 @@ static void repack_mxfp4_mxfp4x4x2(ggml_tensor * t, const void * data, size_t si
 
     init_row_mxfp4x4x2((block_mxfp4 *) buf_pd, t->ne[0]);  // init padded buffer to make sure the tail is all zeros
 
-    for (int64_t i = 0; i < nrows; i++) {
+    // 1. Process all the full rows
+    for (int64_t i = 0; i < n_full_rows; i++) {
         const uint8_t * src = (const uint8_t *) data + (i * row_size);
         uint8_t *       dst = (uint8_t *) t->data + (i * row_size);
 
@@ -1269,6 +1383,25 @@ static void repack_mxfp4_mxfp4x4x2(ggml_tensor * t, const void * data, size_t si
         memcpy(dst, buf_rp, row_size);
     }
 
+    // 2. Process the final, potentially partial, row
+    if (n_rem_bytes > 0) {
+        const int64_t i = n_full_rows;
+        const uint8_t * src = (const uint8_t *) data + (i * row_size);
+        uint8_t *       dst = (uint8_t *) t->data + (i * row_size);
+
+        // re-init the row because we are potentially copying a partial row
+        init_row_mxfp4x4x2((block_mxfp4 *) buf_pd, t->ne[0]);
+
+        // Copy only the remaining bytes from the source.
+        memcpy(buf_pd, src, n_rem_bytes);
+
+        // Repack the entire buffer (partial data + zero padding).
+        repack_row_mxfp4x4x2((uint8_t *) buf_rp, (const block_mxfp4 *) buf_pd, t->ne[0]);
+
+        // Write only the corresponding remaining bytes to the destination tensor.
+        memcpy(dst, buf_rp, n_rem_bytes);
+    }
+
     ggml_aligned_free(buf_pd, row_size_pd);
     ggml_aligned_free(buf_rp, row_size_rp);
 }
@@ -1281,6 +1414,14 @@ static void repack_mxfp4x4x2_mxfp4(void * data, const ggml_tensor * t, size_t si
     size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_MXFP4x4x2));  // extra elements for the pad
     size_t row_size_rp = row_size * 2;  // extra space for tmp pad (if any)
 
+    // Ensure we don't try to copy more data than the tensor actually contains.
+    const size_t total_tensor_size = (size_t)nrows * row_size;
+    const size_t n_bytes_to_copy = size < total_tensor_size ? size : total_tensor_size;
+
+    // Calculate how many full rows and how many remaining bytes we need to process.
+    const int64_t n_full_rows = n_bytes_to_copy / row_size;
+    const size_t  n_rem_bytes = n_bytes_to_copy % row_size;
+
     void * buf_pd = ggml_aligned_malloc(row_size_pd);
     GGML_ASSERT(buf_pd != NULL);
 
@@ -1292,7 +1433,8 @@ static void repack_mxfp4x4x2_mxfp4(void * data, const ggml_tensor * t, size_t si
 
     memset(buf_pd, 0, row_size_pd);  // clear-out padded buffer to make sure the tail is all zeros
 
-    for (int64_t i = 0; i < nrows; i++) {
+    // 1. Process all the full rows
+    for (int64_t i = 0; i < n_full_rows; i++) {
         const uint8_t * src = (const uint8_t *) t->data + (i * row_size);
         uint8_t *       dst = (uint8_t *) data + (i * row_size);
 
@@ -1301,6 +1443,20 @@ static void repack_mxfp4x4x2_mxfp4(void * data, const ggml_tensor * t, size_t si
         memcpy(dst, buf_rp, row_size);
     }
 
+    // 2. Process the final, potentially partial, row
+    if (n_rem_bytes > 0) {
+        const int64_t i = n_full_rows;
+        const uint8_t * src = (const uint8_t *) t->data + (i * row_size);
+        uint8_t *       dst = (uint8_t *) data + (i * row_size);
+
+        // We still need to read and unpack the entire source row because the format is block-based.
+        memcpy(buf_pd, src, row_size);
+        unpack_row_mxfp4x4x2((block_mxfp4 *) buf_rp, (const uint8_t *) buf_pd, t->ne[0]);
+
+        // But we only copy the remaining number of bytes to the destination to respect the size limit.
+        memcpy(dst, buf_rp, n_rem_bytes);
+    }
+
     ggml_aligned_free(buf_pd, row_size_pd);
     ggml_aligned_free(buf_rp, row_size_rp);
 }
@@ -1319,19 +1475,19 @@ static void ggml_backend_hexagon_buffer_set_tensor(ggml_backend_buffer_t buffer,
     switch (tensor->type) {
         case GGML_TYPE_Q4_0:
             GGML_ASSERT(offset == 0);
-            GGML_ASSERT(size == ggml_nbytes(tensor));
+            GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
             repack_q4_0_q4x4x2(tensor, data, size);
             break;
 
         case GGML_TYPE_Q8_0:
             GGML_ASSERT(offset == 0);
-            GGML_ASSERT(size == ggml_nbytes(tensor));
+            GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
             repack_q8_0_q8x4x2(tensor, data, size);
             break;
 
         case GGML_TYPE_MXFP4:
             GGML_ASSERT(offset == 0);
-            GGML_ASSERT(size == ggml_nbytes(tensor));
+            GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
             repack_mxfp4_mxfp4x4x2(tensor, data, size);
             break;
 
@@ -1355,19 +1511,19 @@ static void ggml_backend_hexagon_buffer_get_tensor(ggml_backend_buffer_t buffer,
     switch (tensor->type) {
         case GGML_TYPE_Q4_0:
             GGML_ASSERT(offset == 0);
-            GGML_ASSERT(size == ggml_nbytes(tensor));
+            GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
             repack_q4x4x2_q4_0(data, tensor, size);
             break;
 
         case GGML_TYPE_Q8_0:
             GGML_ASSERT(offset == 0);
-            GGML_ASSERT(size == ggml_nbytes(tensor));
+            GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
             repack_q8x4x2_q8_0(data, tensor, size);
             break;
 
         case GGML_TYPE_MXFP4:
             GGML_ASSERT(offset == 0);
-            GGML_ASSERT(size == ggml_nbytes(tensor));
+            GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
             repack_mxfp4x4x2_mxfp4(data, tensor, size);
             break;
 

ggml/src/ggml-metal/ggml-metal-device.cpp

@@ -677,7 +677,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm_id_map0(ggml_metal_
     char name[256];
 
     snprintf(base, 256, "kernel_mul_mm_id_map0_ne20_%d", ne20);
-    snprintf(name, 256, "%s", base);
+    snprintf(name, 256, "%s_ne02=%d", base, ne02);
 
     ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
     if (res) {
@@ -1332,11 +1332,12 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_rope(ggml_metal_library_t
 
     const bool is_neox   = mode & GGML_ROPE_TYPE_NEOX;
     const bool is_mrope  = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE;
     const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
 
     if (is_neox) {
         snprintf(base, 256, "kernel_rope_neox_%s", ggml_type_name(op->src[0]->type));
-    } else if (is_mrope && !is_vision) {
+    } else if ((is_mrope || is_imrope) && !is_vision) {
         GGML_ASSERT(op->src[1]->ne[0]*4 >= op->src[0]->ne[2]); // need at least 4 pos per token
         snprintf(base, 256, "kernel_rope_multi_%s", ggml_type_name(op->src[0]->type));
     } else if (is_vision) {
@@ -1346,14 +1347,20 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_rope(ggml_metal_library_t
         snprintf(base, 256, "kernel_rope_norm_%s", ggml_type_name(op->src[0]->type));
     }
 
-    snprintf(name, 256, "%s", base);
+    snprintf(name, 256, "%s_imrope=%d", base, is_imrope ? 1 : 0);
 
     ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
     if (res) {
         return res;
     }
 
-    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+    ggml_metal_cv_set_bool(cv, is_imrope, FC_ROPE + 0);
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+    ggml_metal_cv_free(cv);
 
     return res;
 }

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -76,6 +76,7 @@
 #define FC_FLASH_ATTN_EXT_VEC_REDUCE   500
 #define FC_MUL_MV                      600
 #define FC_MUL_MM                      700
+#define FC_ROPE                        800
 
 // op-specific constants
 #define OP_FLASH_ATTN_EXT_NQPTG 8

ggml/src/ggml-metal/ggml-metal.metal

@@ -3709,6 +3709,8 @@ template [[host_name("kernel_mul_mv_bf16_f32_short")]]  kernel mul_mv_t_t_short_
 template [[host_name("kernel_mul_mv_bf16_bf16_short")]] kernel mul_mv_t_t_short_t kernel_mul_mv_t_t_short<bfloat, bfloat>;
 #endif
 
+constant bool FC_rope_is_imrope [[function_constant(FC_ROPE + 0)]];
+
 static float rope_yarn_ramp(const float low, const float high, const int i0) {
     const float y = (i0 / 2 - low) / max(0.001f, high - low);
     return 1.0f - min(1.0f, max(0.0f, y));
@@ -3889,14 +3891,26 @@ kernel void kernel_rope_multi(
             const int sector    = ic % sect_dims;
 
             float theta_base;
-            if (sector < args.sect_0) {
-                theta_base = (float) pos[i2];
-            } else if (sector < sec_w01) {
-                theta_base = (float) pos[i2 + args.ne02];
-            } else if (sector < sec_w012) {
-                theta_base = (float) pos[i2 + args.ne02 * 2];
+            if (FC_rope_is_imrope) {
+                if (sector % 3 == 1 && sector < 3 * args.sect_1) { // h
+                    theta_base = (float) pos[i2 + args.ne02 * 1];
+                } else if (sector % 3 == 2 && sector < 3 * args.sect_2) { // w
+                    theta_base = (float) pos[i2 + args.ne02 * 2];
+                } else if (sector % 3 == 0 && sector < 3 * args.sect_0) { // t
+                    theta_base = (float) pos[i2 + args.ne02 * 0];
+                } else { // e
+                    theta_base = (float) pos[i2 + args.ne02 * 3];
+                }
             } else {
-                theta_base = (float) pos[i2 + args.ne02 * 3];
+                if (sector < args.sect_0) {
+                    theta_base = (float) pos[i2];
+                } else if (sector < sec_w01) {
+                    theta_base = (float) pos[i2 + args.ne02 * 1];
+                } else if (sector < sec_w012) {
+                    theta_base = (float) pos[i2 + args.ne02 * 2];
+                } else {
+                    theta_base = (float) pos[i2 + args.ne02 * 3];
+                }
             }
             // end of mrope
 

ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl

@@ -79,8 +79,8 @@ kernel void kernel_mul_mm_f16_f32_l4_lm(
 
     for (int block = 0; block < ne00; block += BK) {
         for (int l = 0; l < BM; l += loadstride_a) {
-            if (loadc_a + l < ne01) {
-            const int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
+            if (ir*BM + loadc_a + l < ne01) {
+                const int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
                 buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = src0[idx].s0;
                 buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = src0[idx].s1;
                 buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = src0[idx].s2;
@@ -94,7 +94,7 @@ kernel void kernel_mul_mm_f16_f32_l4_lm(
         }
 
         for (int l = 0; l < BN; l += loadstride_b) {
-            if (loadc_b + l < ne11) {
+            if (ic*BN + loadc_b + l < ne11) {
                 const int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
                 buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
                 buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;

ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl

@@ -79,7 +79,7 @@ kernel void kernel_mul_mm_f32_f32_l4_lm(
 
     for (int block = 0; block < ne00; block += BK) {
         for (int l = 0; l < BM; l += loadstride_a) {
-            if (loadc_a + l < ne01) {
+            if (ir*BM + loadc_a + l < ne01) {
                 const int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
                 buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = src0[idx].s0;
                 buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = src0[idx].s1;
@@ -94,7 +94,7 @@ kernel void kernel_mul_mm_f32_f32_l4_lm(
         }
 
         for (int l = 0; l < BN; l += loadstride_b) {
-            if (loadc_b + l < ne11) {
+            if (ic*BN + loadc_b + l < ne11) {
                 const int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
                 buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
                 buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;

ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_l4_lm.cl

@@ -78,7 +78,7 @@ kernel void kernel_mul_mm_q8_0_f32_l4_lm(
 
     for (int block = 0; block < ne00; block += BK) {
         for (int l = 0; l < BM; l += loadstride_a) {
-            if (loadc_a + l < ne01) {
+            if (ir*BM + loadc_a + l < ne01) {
                 int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
                 int ib  = idx / 8;
                 int iqs = idx % 8;
@@ -101,7 +101,7 @@ kernel void kernel_mul_mm_q8_0_f32_l4_lm(
         }
 
         for (int l = 0; l < BN; l += loadstride_b) {
-            if (loadc_b + l < ne11) {
+            if (ic*BN + loadc_b + l < ne11) {
                 int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
                 buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
                 buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;

ggml/src/ggml-sycl/rope.cpp

@@ -119,7 +119,7 @@ static void rope_multi(const T * x, T * dst, const int ne0, const int ne1, const
                         const size_t s2, const int n_dims, const int32_t * pos, const float freq_scale,
                         const float ext_factor, const float attn_factor, const rope_corr_dims corr_dims,
                         const float theta_scale, const float * freq_factors, const mrope_sections sections,
-                        const sycl::nd_item<3> & item_ct1) {
+                        const bool is_imrope, const sycl::nd_item<3> & item_ct1) {
     // get index pos
     const int i0 = 2 * (item_ct1.get_group(1) * item_ct1.get_local_range(1) + item_ct1.get_local_id(1));
     if (i0 >= ne0) {
@@ -143,17 +143,29 @@ static void rope_multi(const T * x, T * dst, const int ne0, const int ne1, const
 
 
     float theta_base = 0.0;
-    if (sector < sections.v[0]) {
-        theta_base = pos[channel_x]*sycl::pow(theta_scale, i0/2.0f);
-    }
-    else if (sector >= sections.v[0] && sector < sec_w) {
-        theta_base = pos[channel_x + ne2 * 1]*sycl::pow(theta_scale, i0/2.0f);
-    }
-    else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
-        theta_base = pos[channel_x + ne2 * 2]*sycl::pow(theta_scale, i0/2.0f);
-    }
-    else if (sector >= sec_w + sections.v[2]) {
-        theta_base = pos[channel_x + ne2 * 3]*sycl::pow(theta_scale, i0/2.0f);
+    if (is_imrope) {
+        if (sector % 3 == 1 && sector < 3 * sections.v[1]) {
+            theta_base = pos[channel_x + ne2 * 1]*sycl::pow(theta_scale, i0/2.0f);
+        } else if (sector % 3 == 2 && sector < 3 * sections.v[2]) {
+            theta_base = pos[channel_x + ne2 * 2]*sycl::pow(theta_scale, i0/2.0f);
+        } else if (sector % 3 == 0 && sector < 3 * sections.v[0]) {
+            theta_base = pos[channel_x]*sycl::pow(theta_scale, i0/2.0f);
+        } else {
+            theta_base = pos[channel_x + ne2 * 3]*sycl::pow(theta_scale, i0/2.0f);
+        }
+    } else {
+        if (sector < sections.v[0]) {
+            theta_base = pos[channel_x]*sycl::pow(theta_scale, i0/2.0f);
+        }
+        else if (sector >= sections.v[0] && sector < sec_w) {
+            theta_base = pos[channel_x + ne2 * 1]*sycl::pow(theta_scale, i0/2.0f);
+        }
+        else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
+            theta_base = pos[channel_x + ne2 * 2]*sycl::pow(theta_scale, i0/2.0f);
+        }
+        else if (sector >= sec_w + sections.v[2]) {
+            theta_base = pos[channel_x + ne2 * 3]*sycl::pow(theta_scale, i0/2.0f);
+        }
     }
 
     const float freq_factor = has_ff ? freq_factors[i0 / 2] : 1.0f;
@@ -281,7 +293,7 @@ static void rope_multi_sycl(const T * x, T * dst, const int ne0, const int ne1,
                              const size_t s2, const int n_dims, const int nr, const int32_t * pos,
                              const float freq_scale, const float freq_base, const float ext_factor,
                              const float attn_factor, const rope_corr_dims corr_dims, const float * freq_factors,
-                             const mrope_sections sections, queue_ptr stream) {
+                             const mrope_sections sections, const bool is_imrope, queue_ptr stream) {
     GGML_ASSERT(ne0 % 2 == 0);
     const sycl::range<3>    block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1);
     const int               n_blocks_y = ceil_div(ne0, (2 * SYCL_ROPE_BLOCK_SIZE));
@@ -297,12 +309,12 @@ static void rope_multi_sycl(const T * x, T * dst, const int ne0, const int ne1,
     if (freq_factors == nullptr) {
         stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
             rope_multi<T, false>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
-                                  corr_dims, theta_scale, freq_factors, sections, item_ct1);
+                                  corr_dims, theta_scale, freq_factors, sections, is_imrope, item_ct1);
         });
     } else {
         stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
             rope_multi<T, true>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
-                                 corr_dims, theta_scale, freq_factors, sections, item_ct1);
+                                 corr_dims, theta_scale, freq_factors, sections, is_imrope, item_ct1);
         });
     }
 }
@@ -381,6 +393,7 @@ inline void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst)
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
     const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE;
     const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
 
     if (is_mrope) {
@@ -422,11 +435,11 @@ inline void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst)
         if (dst->src[0]->type == GGML_TYPE_F16) {
             rope_multi_sycl((const sycl::half *)dst->src[0]->data, (sycl::half *)dst->data, ne00, ne01, ne02, s01,
                 s02, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
-                freq_factors, sections, main_stream);
+                freq_factors, sections, is_imrope, main_stream);
         } else if (dst->src[0]->type == GGML_TYPE_F32) {
             rope_multi_sycl((const float *) dst->src[0]->data, (float *) dst->data, ne00, ne01, ne02, s01, s02, n_dims,
                              nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections,
-                             main_stream);
+                             is_imrope, main_stream);
         } else {
             GGML_ABORT("Fatal error: Tensor type unsupported!");
         }

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -145,8 +145,13 @@ static void ggml_vk_destroy_pipeline(vk::Device& device, vk_pipeline& pipeline);
 struct vk_matmul_pipeline_struct {
     vk_pipeline l, m, s;
     vk_pipeline a_l, a_m, a_s;
+    // Returns true when all unaligned pipelines are null.
+    // We only check for unaligned variants since one of the unaligned pipelines must exist
+    // while aligned pipelines are optional
+    bool is_empty() const {
+        return l == nullptr && m == nullptr && s == nullptr;
+    }
 };
-
 typedef std::shared_ptr<vk_matmul_pipeline_struct> vk_matmul_pipeline;
 
 struct vk_matmul_pipeline2 {
@@ -456,6 +461,11 @@ static topk_moe_mode ggml_vk_num_additional_ops_to_topk_moe_mode(uint32_t num) {
     return mode;
 }
 
+static constexpr std::initializer_list<std::array<int, 3>> rope_view_set_rows_edges {
+    { 1, 0, 0 }, // view->src[0]     == rope
+    { 2, 0, 1 }, // set_rows->src[0] == view
+};
+
 struct vk_device_struct {
     std::recursive_mutex mutex;
 
@@ -638,8 +648,8 @@ struct vk_device_struct {
     vk_pipeline pipeline_soft_max_f32, pipeline_soft_max_f32_f16;
     vk_pipeline pipeline_soft_max_f32_wg512, pipeline_soft_max_f32_f16_wg512;
     vk_pipeline pipeline_soft_max_back_f32;
-    vk_pipeline pipeline_rope_norm_f32, pipeline_rope_norm_f16;
-    vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16;
+    vk_pipeline pipeline_rope_norm_f32, pipeline_rope_norm_f16, pipeline_rope_norm_f32_f16;
+    vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16, pipeline_rope_neox_f32_f16;
     vk_pipeline pipeline_rope_multi_f32, pipeline_rope_multi_f16;
     vk_pipeline pipeline_rope_vision_f32, pipeline_rope_vision_f16;
     vk_pipeline pipeline_argsort_f32[num_argsort_pipelines];
@@ -1051,7 +1061,9 @@ struct vk_op_rope_push_constants {
     uint32_t s1;
     uint32_t s2;
     int32_t sections[4];
+    uint32_t is_imrope;
     uint32_t is_back;
+    uint32_t set_rows_stride;
 };
 
 struct vk_op_soft_max_push_constants {
@@ -1076,6 +1088,7 @@ struct vk_op_soft_max_push_constants {
 
 struct vk_op_argsort_push_constants {
     uint32_t ncols;
+    uint32_t nrows;
     int32_t order;
 };
 
@@ -1562,6 +1575,10 @@ struct ggml_backend_vk_context {
     // number of additional consecutive nodes that are being fused with the
     // node currently being processed
     int num_additional_fused_ops {};
+    // Bitmask of which fused ops need to write an intermediate value to memory.
+    // Bit 'i' means nodes[start_of_fusion + i] writes to memory.
+    // If there's no fusion, bit 0 is still set.
+    int fused_ops_write_mask {};
 };
 
 static void * const vk_ptr_base = (void *)(uintptr_t) 0x1000;  // NOLINT
@@ -3695,21 +3712,27 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16_wg512, "soft_max_f32_f16_wg512", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 4, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);
     ggml_vk_create_pipeline(device, device->pipeline_soft_max_back_f32, "soft_max_back_f32", soft_max_back_f32_len, soft_max_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1, true);
 
-    ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32, "rope_norm_f32", rope_norm_f32_len, rope_norm_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f32, "rope_neox_f32", rope_neox_f32_len, rope_neox_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_rope_multi_f32, "rope_multi_f32", rope_multi_f32_len, rope_multi_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f32, "rope_vision_f32", rope_vision_f32_len, rope_vision_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32, "rope_norm_f32", rope_norm_f32_len, rope_norm_f32_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f32, "rope_neox_f32", rope_neox_f32_len, rope_neox_f32_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_rope_multi_f32, "rope_multi_f32", rope_multi_f32_len, rope_multi_f32_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f32, "rope_vision_f32", rope_vision_f32_len, rope_vision_f32_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
 
     if (device->float_controls_rte_fp16) {
-        ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_rte_len, rope_norm_f16_rte_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_rte_len, rope_neox_f16_rte_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_rope_multi_f16, "rope_multi_f16", rope_multi_f16_rte_len, rope_multi_f16_rte_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f16, "rope_vision_f16", rope_vision_f16_rte_len, rope_vision_f16_rte_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_rte_len, rope_norm_f16_rte_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_rte_len, rope_neox_f16_rte_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_rope_multi_f16, "rope_multi_f16", rope_multi_f16_rte_len, rope_multi_f16_rte_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f16, "rope_vision_f16", rope_vision_f16_rte_len, rope_vision_f16_rte_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+
+        ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32_f16, "rope_norm_f32_f16", rope_norm_f32_f16_rte_len, rope_norm_f32_f16_rte_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f32_f16, "rope_neox_f32_f16", rope_neox_f32_f16_rte_len, rope_neox_f32_f16_rte_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
     } else {
-        ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_len, rope_norm_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_len, rope_neox_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_rope_multi_f16, "rope_multi_f16", rope_multi_f16_len, rope_multi_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f16, "rope_vision_f16", rope_vision_f16_len, rope_vision_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_len, rope_norm_f16_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_len, rope_neox_f16_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_rope_multi_f16, "rope_multi_f16", rope_multi_f16_len, rope_multi_f16_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f16, "rope_vision_f16", rope_vision_f16_len, rope_vision_f16_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+
+        ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32_f16, "rope_norm_f32_f16", rope_norm_f32_f16_len, rope_norm_f32_f16_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f32_f16, "rope_neox_f32_f16", rope_neox_f32_f16_len, rope_neox_f32_f16_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
     }
 
     for (uint32_t i = 0; i < num_argsort_pipelines; ++i) {
@@ -5061,7 +5084,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
     if (src1_type == GGML_TYPE_Q8_1) {
         vk_matmul_pipeline pipelines = ctx->device->pipeline_dequant_mul_mat_mat_q8_1[src0_type].f32acc;
 
-        if (pipelines->s == nullptr && pipelines->m == nullptr && pipelines->l == nullptr) {
+        if (pipelines->is_empty()) {
             return nullptr;
         }
 
@@ -5210,7 +5233,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
     if (src1_type == GGML_TYPE_Q8_1) {
         vk_matmul_pipeline pipelines = ctx->device->pipeline_dequant_mul_mat_mat_id_q8_1[src0_type].f32acc;
 
-        if (pipelines->s == nullptr && pipelines->m == nullptr && pipelines->l == nullptr) {
+        if (pipelines->is_empty()) {
             return nullptr;
         }
 
@@ -5245,16 +5268,17 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
             return nullptr;
     }
 
+    vk_matmul_pipeline2& mmp = ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type];
     // XXX TODO 'prec' is not actually allowed in mul_mat_id.
     bool prefer_fp16acc = ctx->device->fp16 /*&& prec == GGML_PREC_DEFAULT*/;
-    bool support_fp16acc = ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f16acc != nullptr;
-    bool support_fp32acc = ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f32acc != nullptr;
+    bool support_fp16acc = !mmp.f16acc->is_empty();
+    bool support_fp32acc = !mmp.f32acc->is_empty();
 
     if (support_fp16acc && (prefer_fp16acc || !support_fp32acc)) {
-        return ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f16acc;
+        return mmp.f16acc;
     } else {
         GGML_ASSERT(support_fp32acc);
-        return ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f32acc;
+        return mmp.f32acc;
     }
 }
 
@@ -8168,7 +8192,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
     case GGML_OP_ROPE:
     case GGML_OP_ROPE_BACK:
         {
-            const int mode = ((const int32_t *) dst->op_params)[2];
+            const ggml_tensor *rope = ctx->num_additional_fused_ops == 2 ? dst->src[0]->src[0] : dst;
+            const int mode = ((const int32_t *) rope->op_params)[2];
             const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
             const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
             const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
@@ -8177,6 +8202,9 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
                 if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
                     return ctx->device->pipeline_rope_neox_f32;
                 }
+                if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
+                    return ctx->device->pipeline_rope_neox_f32_f16;
+                }
                 if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
                     return ctx->device->pipeline_rope_neox_f16;
                 }
@@ -8198,6 +8226,9 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
                 if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
                     return ctx->device->pipeline_rope_norm_f32;
                 }
+                if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
+                    return ctx->device->pipeline_rope_norm_f32_f16;
+                }
                 if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
                     return ctx->device->pipeline_rope_norm_f16;
                 }
@@ -8407,20 +8438,22 @@ static uint32_t get_misalign_bytes(ggml_backend_vk_context * ctx, const ggml_ten
     return ((vk_tensor_offset(t) + t->view_offs) & (ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1));;
 }
 
-template <typename T> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, T &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+template <typename T> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, T &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
     GGML_UNUSED(p);
     GGML_UNUSED(src0);
     GGML_UNUSED(src1);
     GGML_UNUSED(src2);
+    GGML_UNUSED(src3);
     GGML_UNUSED(dst);
     static_assert(!std::is_const<T>::value, "unexpected type");
     GGML_ASSERT(!src0 || get_misalign_bytes(ctx, src0) == 0);
     GGML_ASSERT(!src1 || get_misalign_bytes(ctx, src1) == 0);
     GGML_ASSERT(!src2 || get_misalign_bytes(ctx, src2) == 0);
+    GGML_ASSERT(!src3 || get_misalign_bytes(ctx, src3) == 0);
     GGML_ASSERT(!dst  || get_misalign_bytes(ctx, dst) == 0);
 }
 
-template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_unary_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_unary_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
     const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
     const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
 
@@ -8428,9 +8461,10 @@ template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk
 
     GGML_UNUSED(src1);
     GGML_UNUSED(src2);
+    GGML_UNUSED(src3);
 }
 
-template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_sum_rows_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_sum_rows_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
     const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
     const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
 
@@ -8438,9 +8472,10 @@ template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk
 
     GGML_UNUSED(src1);
     GGML_UNUSED(src2);
+    GGML_UNUSED(src3);
 }
 
-template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_pad_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_pad_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
     const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
     const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
 
@@ -8448,9 +8483,10 @@ template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk
 
     GGML_UNUSED(src1);
     GGML_UNUSED(src2);
+    GGML_UNUSED(src3);
 }
 
-template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_im2col_3d_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_im2col_3d_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
     const uint32_t a_offset = get_misalign_bytes(ctx, src1) / ggml_type_size(src1->type);
     const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
 
@@ -8458,9 +8494,10 @@ template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk
 
     GGML_UNUSED(src0);
     GGML_UNUSED(src2);
+    GGML_UNUSED(src3);
 }
 
-template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_binary_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_binary_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
     const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
     const uint32_t b_offset = get_misalign_bytes(ctx, src1) / ggml_type_size(src1->type);
     const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
@@ -8470,9 +8507,10 @@ template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk
     p.misalign_offsets = (a_offset << 16) | (b_offset << 8) | d_offset;
 
     GGML_UNUSED(src2);
+    GGML_UNUSED(src3);
 }
 
-template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_upscale_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_upscale_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
     const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
     const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
 
@@ -8481,10 +8519,11 @@ template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk
 
     GGML_UNUSED(src1);
     GGML_UNUSED(src2);
+    GGML_UNUSED(src3);
 }
 
 template<typename PC>
-static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op, PC&& pc, bool dryrun = false) {
+static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst, ggml_op op, PC&& pc, bool dryrun = false) {
     VK_LOG_DEBUG("ggml_vk_op_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     if (src1 != nullptr) {
         std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
@@ -8492,6 +8531,9 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     if (src2 != nullptr) {
         std::cerr << "), (" << src2 << ", name=" << src2->name << ", type=" << src2->type << ", ne0=" << src2->ne[0] << ", ne1=" << src2->ne[1] << ", ne2=" << src2->ne[2] << ", ne3=" << src2->ne[3] << ", nb0=" << src2->nb[0] << ", nb1=" << src2->nb[1] << ", nb2=" << src2->nb[2] << ", nb3=" << src2->nb[3];
     }
+    if (src3 != nullptr) {
+        std::cerr << "), (" << src3 << ", name=" << src3->name << ", type=" << src3->type << ", ne0=" << src3->ne[0] << ", ne1=" << src3->ne[1] << ", ne2=" << src3->ne[2] << ", ne3=" << src3->ne[3] << ", nb0=" << src3->nb[0] << ", nb1=" << src3->nb[1] << ", nb2=" << src3->nb[2] << ", nb3=" << src3->nb[3];
+    }
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
     std::cerr << "), " << ggml_op_name(op) << ", " << (dryrun ? "dryrun" : "") << ")");
     GGML_ASSERT(op == GGML_OP_GET_ROWS || op == GGML_OP_CPY || (!ggml_is_quantized(src0->type) && (src1 == nullptr || !ggml_is_quantized(src1->type))));  // NOLINT
@@ -8518,6 +8560,13 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     const uint64_t ne23 = use_src2 ? src2->ne[3] : 0;
     const uint64_t ne2 = ne20 * ne21;
 
+    const bool use_src3 = src3 != nullptr;
+    const uint64_t ne30 = use_src3 ? src3->ne[0] : 0;
+    const uint64_t ne31 = use_src3 ? src3->ne[1] : 0;
+    const uint64_t ne32 = use_src3 ? src3->ne[2] : 0;
+    const uint64_t ne33 = use_src3 ? src3->ne[3] : 0;
+    const uint64_t ne3 = ne30 * ne31;
+
     const uint64_t ned0 = dst->ne[0];
     const uint64_t ned1 = dst->ne[1];
     const uint64_t ned2 = dst->ne[2];
@@ -8548,6 +8597,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
     ggml_backend_vk_buffer_context * src1_buf_ctx = use_src1 ? (ggml_backend_vk_buffer_context *)src1->buffer->context : nullptr;
     ggml_backend_vk_buffer_context * src2_buf_ctx = use_src2 ? (ggml_backend_vk_buffer_context *)src2->buffer->context : nullptr;
+    ggml_backend_vk_buffer_context * src3_buf_ctx = use_src3 ? (ggml_backend_vk_buffer_context *)src3->buffer->context : nullptr;
 
     vk_buffer d_X = nullptr;
     size_t x_buf_offset = 0;
@@ -8555,10 +8605,13 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     size_t y_buf_offset = 0;
     vk_buffer d_Z = nullptr;
     size_t z_buf_offset = 0;
+    vk_buffer d_W = nullptr;
+    size_t w_buf_offset = 0;
 
     bool src0_uma = false;
     bool src1_uma = false;
     bool src2_uma = false;
+    bool src3_uma = false;
 
     if (ctx->device->uma) {
         ggml_vk_host_get(ctx->device, src0->data, d_X, x_buf_offset);
@@ -8571,6 +8624,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
             ggml_vk_host_get(ctx->device, src2->data, d_Z, z_buf_offset);
             src2_uma = d_Z != nullptr;
         }
+        if (use_src3) {
+            ggml_vk_host_get(ctx->device, src3->data, d_W, w_buf_offset);
+            src3_uma = d_W != nullptr;
+        }
     }
 
     vk_buffer d_D = dst_buf_ctx->dev_buffer;
@@ -8592,11 +8649,17 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         z_buf_offset = vk_tensor_offset(src2) + src2->view_offs;
         GGML_ASSERT(d_Z != nullptr);
     }
+    if (use_src3 && !src3_uma) {
+        d_W = src3_buf_ctx->dev_buffer;
+        w_buf_offset = vk_tensor_offset(src3) + src3->view_offs;
+        GGML_ASSERT(d_W != nullptr);
+    }
     // Compute misalignment offset for descriptors and store it in in push constants, then align the descriptor offsets.
-    init_pushconst_tensor_offsets(ctx, pc, src0, src1, src2, dst);
+    init_pushconst_tensor_offsets(ctx, pc, src0, src1, src2, src3, dst);
     x_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
     y_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
     z_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
+    w_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
     d_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
 
     std::array<uint32_t, 3> elements;
@@ -8653,6 +8716,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         break;
     case GGML_OP_ARGSORT:
         elements = { (uint32_t)ne00, (uint32_t)ggml_nrows(src0), 1 };
+        elements[1] = std::min(elements[1], ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
         break;
     case GGML_OP_IM2COL:
         {
@@ -8797,12 +8861,13 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         break;
     }
 
-    uint64_t x_sz, y_sz, z_sz, d_sz;
+    uint64_t x_sz, y_sz, z_sz, w_sz, d_sz;
 
     if (op_supports_incontiguous) {
         x_sz = ggml_nbytes(src0) + get_misalign_bytes(ctx, src0);
         y_sz = use_src1 ? ggml_nbytes(src1) + get_misalign_bytes(ctx, src1) : 0;
         z_sz = use_src2 ? ggml_nbytes(src2) + get_misalign_bytes(ctx, src2) : 0;
+        w_sz = use_src3 ? ggml_nbytes(src3) + get_misalign_bytes(ctx, src3) : 0;
         d_sz = ggml_nbytes(dst) + get_misalign_bytes(ctx, dst);
 
         if (x_buf_offset + x_sz >= d_X->size) {
@@ -8814,6 +8879,9 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         if (use_src2 && z_buf_offset + z_sz >= d_Z->size) {
             z_sz = ggml_vk_get_max_buffer_range(ctx, d_Z, z_buf_offset);
         }
+        if (use_src3 && w_buf_offset + w_sz >= d_W->size) {
+            w_sz = ggml_vk_get_max_buffer_range(ctx, d_W, w_buf_offset);
+        }
         if (d_buf_offset + d_sz >= d_D->size) {
             d_sz = ggml_vk_get_max_buffer_range(ctx, d_D, d_buf_offset);
         }
@@ -8821,6 +8889,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         x_sz = ggml_type_size(src0->type)/ggml_blck_size(src0->type) * ne0 * ne02 * ne03;
         y_sz = use_src1 ? ggml_type_size(src1->type) * ne1 * ne12 * ne13 : 0;
         z_sz = use_src2 ? ggml_type_size(src2->type) * ne2 * ne22 * ne23 : 0;
+        w_sz = use_src3 ? ggml_type_size(src3->type) * ne3 * ne32 * ne33 : 0;
         d_sz = ggml_type_size(dst->type) * ned * ned2 * ned3;
     }
 
@@ -8862,14 +8931,19 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (op == GGML_OP_ROPE || op == GGML_OP_ROPE_BACK) {
         // Empty src2 is possible in rope, but the shader needs a buffer
-        vk_subbuffer subbuf_z;
+        vk_subbuffer subbuf_z, subbuf_w;
         if (use_src2) {
             subbuf_z = { d_Z, z_buf_offset, z_sz };
         } else {
             subbuf_z = { d_X, 0, x_sz };
         }
+        if (use_src3) {
+            subbuf_w = { d_W, w_buf_offset, w_sz };
+        } else {
+            subbuf_w = { d_X, 0, x_sz };
+        }
 
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz }, subbuf_w }, pc, elements);
     } else if (op == GGML_OP_IM2COL || op == GGML_OP_IM2COL_3D) {
         if (ctx->device->shader_int64 && ctx->device->buffer_device_address) {
             // buffer device address path doesn't use dst buffer
@@ -8885,6 +8959,8 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     } else if (op == GGML_OP_OPT_STEP_SGD) {
         // OPT_STEP_SGD works on src0, it does not need dst
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz } }, pc, elements);
+    } else if (use_src3) {
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz }, vk_subbuffer{ d_W, w_buf_offset, w_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (use_src2) {
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (use_src1) {
@@ -8899,7 +8975,7 @@ static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context& subctx,
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
-    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_GET_ROWS, {
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_GET_ROWS, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
@@ -8919,7 +8995,7 @@ static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const
     // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
     int offset = dst->op_params[3] / 4; // offset in bytes
 
-    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_ACC, {
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_ACC, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)nb1, (uint32_t)nb2, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
@@ -9044,7 +9120,7 @@ static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context& subctx, const
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
-    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_ADD, {
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_ADD, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
@@ -9059,7 +9135,7 @@ static void ggml_vk_sub(ggml_backend_vk_context * ctx, vk_context& subctx, const
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
-    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SUB, {
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_SUB, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
@@ -9074,7 +9150,7 @@ static void ggml_vk_mul(ggml_backend_vk_context * ctx, vk_context& subctx, const
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
-    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_MUL, {
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_MUL, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
@@ -9089,7 +9165,7 @@ static void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context& subctx, const
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
-    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_DIV, {
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_DIV, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
@@ -9104,7 +9180,7 @@ static void ggml_vk_add_id(ggml_backend_vk_context * ctx, vk_context& subctx, co
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t src2_type_size = ggml_type_size(src2->type);
 
-    ggml_vk_op_f32<vk_op_add_id_push_constants>(ctx, subctx, src0, src1, src2, dst, GGML_OP_ADD_ID, {
+    ggml_vk_op_f32<vk_op_add_id_push_constants>(ctx, subctx, src0, src1, src2, nullptr, dst, GGML_OP_ADD_ID, {
         (uint32_t)dst->ne[0],
         (uint32_t)dst->ne[1],
         (uint32_t)src0->nb[1] / src0_type_size,
@@ -9337,7 +9413,7 @@ static void ggml_vk_ssm_conv(ggml_backend_vk_context * ctx, vk_context& subctx,
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
 
-    ggml_vk_op_f32<vk_op_ssm_conv_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SSM_CONV, {
+    ggml_vk_op_f32<vk_op_ssm_conv_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_SSM_CONV, {
         (uint32_t)src0->nb[1], (uint32_t)src0->nb[2],
         (uint32_t)src1->nb[1],
         (uint32_t)dst->nb[0], (uint32_t)dst->nb[1], (uint32_t)dst->nb[2],
@@ -9455,7 +9531,7 @@ static void ggml_vk_opt_step_adamw(ggml_backend_vk_context * ctx, vk_context& su
 static void ggml_vk_opt_step_sgd(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) {
     const size_t n = ggml_nelements(dst->src[0]);
 
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, src2, dst, GGML_OP_OPT_STEP_SGD, { (uint32_t)n, 0, 0.0f, 0.0f }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, src2, nullptr, dst, GGML_OP_OPT_STEP_SGD, { (uint32_t)n, 0, 0.0f, 0.0f }, dryrun);
 }
 
 static void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -9465,7 +9541,7 @@ static void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, co
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
-    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONCAT, {
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONCAT, {
         (uint32_t)ggml_nelements(dst),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
@@ -9493,7 +9569,7 @@ static void ggml_vk_upscale(ggml_backend_vk_context * ctx, vk_context& subctx, c
         pixel_offset = 0.0f;
     }
 
-    ggml_vk_op_f32<vk_op_upscale_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UPSCALE, {
+    ggml_vk_op_f32<vk_op_upscale_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_UPSCALE, {
         (uint32_t)ggml_nelements(dst), 0, 0,
         (uint32_t)ne00, (uint32_t)ne01,
         (uint32_t)nb00 / src0_type_size, (uint32_t)nb01 / src0_type_size, (uint32_t)nb02 / src0_type_size, (uint32_t)nb03 / src0_type_size,
@@ -9507,23 +9583,23 @@ static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context& subctx, con
     p.param1 = ggml_get_op_params_f32(dst, 0);
     p.param2 = ggml_get_op_params_f32(dst, 1);
 
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SCALE, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SCALE, std::move(p), dryrun);
 }
 
 static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SQR, vk_op_unary_push_constants_init(src0, dst), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SQR, vk_op_unary_push_constants_init(src0, dst), dryrun);
 }
 
 static void ggml_vk_sqrt(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SQRT, vk_op_unary_push_constants_init(src0, dst), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SQRT, vk_op_unary_push_constants_init(src0, dst), dryrun);
 }
 
 static void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SIN, vk_op_unary_push_constants_init(src0, dst), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SIN, vk_op_unary_push_constants_init(src0, dst), dryrun);
 }
 
 static void ggml_vk_cos(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_COS, vk_op_unary_push_constants_init(src0, dst), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_COS, vk_op_unary_push_constants_init(src0, dst), dryrun);
 }
 
 static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
@@ -9531,12 +9607,12 @@ static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context& subctx, con
     p.param1 = ggml_get_op_params_f32(dst, 0);
     p.param2 = ggml_get_op_params_f32(dst, 1);
 
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CLAMP, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_CLAMP, std::move(p), dryrun);
 }
 
 static void ggml_vk_pad(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     vk_op_pad_push_constants p = vk_op_pad_push_constants_init(src0, dst);
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_PAD, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_PAD, std::move(p), dryrun);
 }
 
 static void ggml_vk_roll(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
@@ -9551,17 +9627,17 @@ static void ggml_vk_roll(ggml_backend_vk_context * ctx, vk_context& subctx, cons
     memcpy(&p.param1, &s01_packed, sizeof(float));
     memcpy(&p.param2, &s23_packed, sizeof(float));
 
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ROLL, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_ROLL, std::move(p), dryrun);
 }
 
 static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst, ggml_nelements(dst));
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_REPEAT, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_REPEAT, std::move(p), dryrun);
 }
 
 static void ggml_vk_repeat_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst, ggml_nelements(dst));
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_REPEAT_BACK, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_REPEAT_BACK, std::move(p), dryrun);
 }
 
 static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
@@ -9577,7 +9653,7 @@ static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const
     }
 
     vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst, ne);
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CPY, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_CPY, std::move(p), dryrun);
 }
 
 static void ggml_vk_set_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -9592,7 +9668,7 @@ static void ggml_vk_set_rows(ggml_backend_vk_context * ctx, vk_context& subctx,
         return;
     }
 
-    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SET_ROWS, {
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_SET_ROWS, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
@@ -9603,13 +9679,13 @@ static void ggml_vk_set_rows(ggml_backend_vk_context * ctx, vk_context& subctx,
 }
 
 static void ggml_vk_silu_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SILU_BACK, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_SILU_BACK, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
 }
 
 static void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     float * op_params = (float *)dst->op_params;
 
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
 }
 
 static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
@@ -9620,7 +9696,7 @@ static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx
     const float eps = float_op_params[1];
     const uint32_t group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
 
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_GROUP_NORM, { group_size, 0, eps, 0.0f }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_GROUP_NORM, { group_size, 0, eps, 0.0f }, dryrun);
 }
 
 static uint32_t ggml_vk_rms_num_partials(ggml_backend_vk_context * ctx, const ggml_tensor *node) {
@@ -9643,7 +9719,7 @@ static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx,
 
     uint32_t param3 = ctx->do_add_rms_partials ? ggml_vk_rms_num_partials(ctx, dst) : 0;
 
-    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_RMS_NORM, {
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_RMS_NORM, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
@@ -9660,16 +9736,16 @@ static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx,
 
 static void ggml_vk_rms_norm_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
     float * op_params = (float *)dst->op_params;
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_RMS_NORM_BACK, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_RMS_NORM_BACK, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
 }
 
 static void ggml_vk_l2_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     float * op_params = (float *)dst->op_params;
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_L2_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_L2_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
 }
 
 static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
 }
 
 static void ggml_vk_glu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -9692,7 +9768,7 @@ static void ggml_vk_glu(ggml_backend_vk_context * ctx, vk_context& subctx, const
 
     const uint32_t mode = split ? 2 : (swapped ? 1 : 0);
 
-    ggml_vk_op_f32<vk_op_glu_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_GLU,
+    ggml_vk_op_f32<vk_op_glu_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_GLU,
         {
             (uint32_t)ggml_nelements(dst),
             (uint32_t)src0->ne[0],
@@ -9705,7 +9781,7 @@ static void ggml_vk_glu(ggml_backend_vk_context * ctx, vk_context& subctx, const
 
 static void ggml_vk_diag_mask_inf(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     int32_t * op_params = (int32_t *)dst->op_params;
-    ggml_vk_op_f32<vk_op_diag_mask_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_DIAG_MASK_INF, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0] }, dryrun);
+    ggml_vk_op_f32<vk_op_diag_mask_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_DIAG_MASK_INF, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0] }, dryrun);
 }
 
 static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) {
@@ -9730,7 +9806,7 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
-    ggml_vk_op_f32<vk_op_soft_max_push_constants>(ctx, subctx, src0, src1, src2, dst, GGML_OP_SOFT_MAX, {
+    ggml_vk_op_f32<vk_op_soft_max_push_constants>(ctx, subctx, src0, src1, src2, nullptr, dst, GGML_OP_SOFT_MAX, {
         ncols,
         src1 != nullptr ? nrows_y : (uint32_t)0,
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],
@@ -9746,7 +9822,7 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
 
 static void ggml_vk_soft_max_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
     float * op_params = (float *)dst->op_params;
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SOFT_MAX_BACK, { (uint32_t)src0->ne[0], (uint32_t)ggml_nrows(src0), op_params[0], op_params[1] }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_SOFT_MAX_BACK, { (uint32_t)src0->ne[0], (uint32_t)ggml_nrows(src0), op_params[0], op_params[1] }, dryrun);
 }
 
 static void ggml_vk_topk_moe(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_cgraph * cgraph, int node_idx, bool dryrun = false) {
@@ -9837,7 +9913,12 @@ static void ggml_vk_topk_moe(ggml_backend_vk_context * ctx, vk_context& subctx,
         }, pc, elements);
 }
 
-static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool backprop, bool dryrun = false) {
+static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_cgraph * cgraph, int node_idx, bool backprop, bool dryrun = false) {
+    ggml_tensor * dst = cgraph->nodes[node_idx];
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const ggml_tensor * src2 = dst->src[2];
+    const ggml_tensor * src3 = nullptr;
     const int n_dims        = ((int32_t *) dst->op_params)[1];
     const int mode          = ((int32_t *) dst->op_params)[2];
     // const int n_ctx         = ((int32_t *) dst->op_params)[3];
@@ -9853,6 +9934,8 @@ static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, cons
         memcpy(sections, (int32_t *) dst->op_params + 11, sizeof(int)*4);
     }
 
+    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE;
+
     float corr_dims[2];
     ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
 
@@ -9861,11 +9944,20 @@ static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, cons
     uint32_t s1 = src0->nb[1] / ggml_type_size(src0->type);
     uint32_t s2 = src0->nb[2] / ggml_type_size(src0->type);
 
-    ggml_vk_op_f32<vk_op_rope_push_constants>(ctx, subctx, src0, src1, src2, dst, GGML_OP_ROPE, {
+    uint32_t set_rows_stride = 0;
+    // Fused rope + view + set_rows passes the set_rows destination stride in set_rows_stride
+    // and overrides the dst and sets src3=row_indices
+    if (ctx->num_additional_fused_ops > 0) {
+        set_rows_stride = cgraph->nodes[node_idx + 2]->nb[1] / ggml_type_size(cgraph->nodes[node_idx + 2]->type);
+        src3 = cgraph->nodes[node_idx + 2]->src[1];
+        dst = cgraph->nodes[node_idx + 2];
+    }
+
+    ggml_vk_op_f32<vk_op_rope_push_constants>(ctx, subctx, src0, src1, src2, src3, dst, GGML_OP_ROPE, {
         (uint32_t)src0->ne[0], (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1],
         freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale,
         src2 != nullptr, (uint32_t)src0->ne[2], s1, s2,
-        { sections[0], sections[1], sections[2], sections[3] }, backprop
+        { sections[0], sections[1], sections[2], sections[3] }, is_imrope, backprop, set_rows_stride,
     }, dryrun);
 }
 
@@ -9873,35 +9965,37 @@ static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context& subctx, c
     int32_t * op_params = (int32_t *)dst->op_params;
 
     uint32_t ncols = src0->ne[0];
+    uint32_t nrows = ggml_nrows(src0);
 
-    ggml_vk_op_f32<vk_op_argsort_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ARGSORT, {
+    ggml_vk_op_f32<vk_op_argsort_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_ARGSORT, {
         ncols,
+        nrows,
         op_params[0],
     }, dryrun);
 }
 
 static void ggml_vk_sum(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     vk_op_sum_rows_push_constants p = vk_op_sum_rows_push_constants_init(src0, dst, ggml_nelements(src0));
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SUM, p, dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SUM, p, dryrun);
 }
 
 static void ggml_vk_sum_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     vk_op_sum_rows_push_constants p = vk_op_sum_rows_push_constants_init(src0, dst, src0->ne[0]);
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SUM_ROWS, p, dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_SUM_ROWS, p, dryrun);
 }
 
 static void ggml_vk_mean(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     vk_op_sum_rows_push_constants p = vk_op_sum_rows_push_constants_init(src0, dst, src0->ne[0]);
     p.weight = 1.0f / (float)src0->ne[0];
-    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_MEAN, p, dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_MEAN, p, dryrun);
 }
 
 static void ggml_vk_argmax(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ARGMAX, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], 0.0f, 0.0f }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_ARGMAX, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], 0.0f, 0.0f }, dryrun);
 }
 
 static void ggml_vk_count_equal(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_COUNT_EQUAL, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_COUNT_EQUAL, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
 }
 
 static void ggml_vk_im2col(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -9934,7 +10028,7 @@ static void ggml_vk_im2col(ggml_backend_vk_context * ctx, vk_context& subctx, co
 
     const vk::DeviceAddress dst_addr = d_buf->bda_addr + vk_tensor_offset(dst) + dst->view_offs;
 
-    ggml_vk_op_f32<vk_op_im2col_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_IM2COL, {
+    ggml_vk_op_f32<vk_op_im2col_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_IM2COL, {
         dst_addr,
         batch_offset, offset_delta,
         IC, IW, IH, OW, OH, KW, KH,
@@ -10007,7 +10101,7 @@ static void ggml_vk_im2col_3d(ggml_backend_vk_context * ctx, vk_context& subctx,
     pc.OH_OW_IC_KD_KH_KW = OH*OW*IC*KD*KH*KW;
     pc.OW_IC_KD_KH_KW = OW*IC*KD*KH*KW;
 
-    ggml_vk_op_f32<vk_op_im2col_3d_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_IM2COL_3D, std::move(pc), dryrun);
+    ggml_vk_op_f32<vk_op_im2col_3d_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_IM2COL_3D, std::move(pc), dryrun);
 }
 
 static void ggml_vk_timestep_embedding(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
@@ -10015,7 +10109,7 @@ static void ggml_vk_timestep_embedding(ggml_backend_vk_context * ctx, vk_context
     const uint32_t max_period = dst->op_params[1];
     const uint32_t nb1 = dst->nb[1] / ggml_type_size(dst->type);
 
-    ggml_vk_op_f32<vk_op_timestep_embedding_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_TIMESTEP_EMBEDDING, {
+    ggml_vk_op_f32<vk_op_timestep_embedding_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_TIMESTEP_EMBEDDING, {
         nb1, dim, max_period,
     }, dryrun);
 }
@@ -10048,7 +10142,7 @@ static void ggml_vk_conv_transpose_1d(ggml_backend_vk_context * ctx, vk_context&
     p.nb1 = static_cast<uint32_t>(nb1 / nb0);
     p.s0 = static_cast<uint32_t>(s0);
 
-    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_TRANSPOSE_1D, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONV_TRANSPOSE_1D, std::move(p), dryrun);
 }
 
 static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
@@ -10071,7 +10165,7 @@ static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, c
 
     const uint32_t parallel_elements = N * OC * OH * OW;
 
-    ggml_vk_op_f32<vk_op_pool2d_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_POOL_2D, {
+    ggml_vk_op_f32<vk_op_pool2d_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_POOL_2D, {
         IW, IH, OW, OH, OC,
         parallel_elements,
         op,
@@ -10125,7 +10219,7 @@ static void ggml_vk_conv_2d(ggml_backend_vk_context * ctx, vk_context & subctx,
     GGML_ASSERT(ne03 == ne2);
     GGML_ASSERT(ne02 == ne12);
 
-    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_2D, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONV_2D, std::move(p), dryrun);
 }
 
 static void ggml_vk_conv_transpose_2d(ggml_backend_vk_context * ctx, vk_context & subctx, const ggml_tensor * src0,
@@ -10174,7 +10268,7 @@ static void ggml_vk_conv_transpose_2d(ggml_backend_vk_context * ctx, vk_context
     GGML_ASSERT(ne02 == ne2);
     GGML_ASSERT(ne03 == ne12);
 
-    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_TRANSPOSE_2D, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONV_TRANSPOSE_2D, std::move(p), dryrun);
 }
 
 static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -10198,12 +10292,12 @@ static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx
     GGML_ASSERT(src0->ne[3] == p.channels);
     GGML_ASSERT(src1->ne[3] == p.batches);
 
-    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_2D_DW, std::move(p), dryrun);
+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONV_2D_DW, std::move(p), dryrun);
 }
 
 static void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     const float * op_params = (const float *)dst->op_params;
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f }, dryrun);
 }
 
 #ifdef GGML_VULKAN_RUN_TESTS
@@ -11329,7 +11423,6 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_SOFT_MAX_BACK:
-        case GGML_OP_ROPE:
         case GGML_OP_ROPE_BACK:
         case GGML_OP_ARGSORT:
         case GGML_OP_SUM:
@@ -11403,9 +11496,12 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         // nodes require synchronization.
         for (int32_t i = 0; i < ctx->num_additional_fused_ops + 1 && !need_sync; ++i) {
             const ggml_tensor *cur_node = cgraph->nodes[node_idx + i];
-            if (overlaps_unsynced(cur_node, ctx->unsynced_nodes_read) || overlaps_unsynced(cur_node, ctx->unsynced_nodes_written)) {
-                need_sync = true;
-                break;
+            // If the node actually writes to memory, then check if it needs to sync
+            if (ctx->fused_ops_write_mask & (1 << i)) {
+                if (overlaps_unsynced(cur_node, ctx->unsynced_nodes_read) || overlaps_unsynced(cur_node, ctx->unsynced_nodes_written)) {
+                    need_sync = true;
+                    break;
+                }
             }
             for (uint32_t j = 0; j < GGML_MAX_SRC; ++j) {
                 if (!cur_node->src[j]) {
@@ -11432,7 +11528,9 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         for (int32_t i = 0; i < ctx->num_additional_fused_ops + 1; ++i) {
             const ggml_tensor *cur_node = cgraph->nodes[node_idx + i];
             // Multiple outputs could be written, e.g. in topk_moe. Add them all to the list.
-            ctx->unsynced_nodes_written.push_back(cur_node);
+            if (ctx->fused_ops_write_mask & (1 << i)) {
+                ctx->unsynced_nodes_written.push_back(cur_node);
+            }
             for (uint32_t j = 0; j < GGML_MAX_SRC; ++j) {
                 if (!cur_node->src[j]) {
                     continue;
@@ -11623,11 +11721,11 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
 
         break;
     case GGML_OP_ROPE:
-        ggml_vk_rope(ctx, compute_ctx, src0, src1, src2, node, false, dryrun);
+        ggml_vk_rope(ctx, compute_ctx, cgraph, node_idx, false, dryrun);
 
         break;
     case GGML_OP_ROPE_BACK:
-        ggml_vk_rope(ctx, compute_ctx, src0, src1, src2, node, true, dryrun);
+        ggml_vk_rope(ctx, compute_ctx, cgraph, node_idx, true, dryrun);
 
         break;
     case GGML_OP_ARGSORT:
@@ -12464,6 +12562,41 @@ static bool ggml_vk_can_fuse_topk_moe(ggml_backend_vk_context * ctx, const struc
     return true;
 }
 
+static bool ggml_vk_can_fuse_rope_set_rows(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph,
+                                           int node_idx) {
+    GGML_UNUSED(ctx);
+    const ggml_tensor *rope = cgraph->nodes[node_idx + 0];
+    const ggml_tensor *view = cgraph->nodes[node_idx + 1];
+    const ggml_tensor *set_rows = cgraph->nodes[node_idx + 2];
+
+    // ne3 not tested
+    if (rope->src[0]->ne[3] != 1) {
+        return false;
+    }
+
+    if (set_rows->type != GGML_TYPE_F32 && set_rows->type != GGML_TYPE_F16) {
+        return false;
+    }
+
+    if (set_rows->src[1]->type != GGML_TYPE_I64) {
+        return false;
+    }
+
+    // The view should flatten two dims of rope into one dim
+    if (!ggml_is_contiguous(view) ||
+        view->ne[0] != rope->ne[0] * rope->ne[1]) {
+        return false;
+    }
+
+    // Only norm/neox shaders have the fusion code
+    const int mode = ((const int32_t *) rope->op_params)[2];
+    if (mode != GGML_ROPE_TYPE_NORMAL && mode != GGML_ROPE_TYPE_NEOX) {
+        return false;
+    }
+
+    return true;
+}
+
 static uint32_t ggml_vk_fuse_multi_add(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph, int node_idx) {
 
     const ggml_tensor *first_node = cgraph->nodes[node_idx];
@@ -12539,6 +12672,10 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
                 ctx->num_additional_fused_ops = num_adds - 1;
             } else if (ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
                 ctx->num_additional_fused_ops = 1;
+            } else if (ggml_can_fuse_subgraph(cgraph, i, { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS }, { i + 2 }) &&
+                       ggml_check_edges(cgraph, i, rope_view_set_rows_edges) &&
+                       ggml_vk_can_fuse_rope_set_rows(ctx, cgraph, i)) {
+                ctx->num_additional_fused_ops = 2;
             } else if (ggml_can_fuse_subgraph(cgraph, i, topk_moe_early_softmax_norm, { i + 3, i + 9 }) &&
                        ggml_check_edges(cgraph, i, topk_moe_early_softmax_norm_edges) &&
                        ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, TOPK_MOE_EARLY_SOFTMAX_NORM)) {
@@ -12648,20 +12785,31 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
                 ctx->num_additional_fused_ops = num_adds - 1;
             } else if (ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
                 ctx->num_additional_fused_ops = 1;
+            } else if (ggml_can_fuse_subgraph(cgraph, i, { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS }, { i + 2 }) &&
+                       ggml_check_edges(cgraph, i, rope_view_set_rows_edges) &&
+                       ggml_vk_can_fuse_rope_set_rows(ctx, cgraph, i)) {
+                ctx->num_additional_fused_ops = 2;
             } else if (ggml_can_fuse_subgraph(cgraph, i, topk_moe_early_softmax_norm, { i + 3, i + 9 }) &&
                        ggml_check_edges(cgraph, i, topk_moe_early_softmax_norm_edges) &&
                        ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, TOPK_MOE_EARLY_SOFTMAX_NORM)) {
                 ctx->num_additional_fused_ops = topk_moe_early_softmax_norm.size() - 1;
+                // view of argsort writes to memory
+                ctx->fused_ops_write_mask |= 1 << 3;
             } else if (ggml_can_fuse_subgraph(cgraph, i, topk_moe_early_softmax, { i + 3, i + 4 }) &&
                        ggml_check_edges(cgraph, i, topk_moe_early_softmax_edges) &&
                        ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, TOPK_MOE_EARLY_SOFTMAX)) {
                 ctx->num_additional_fused_ops = topk_moe_early_softmax.size() - 1;
+                // view of argsort writes to memory
+                ctx->fused_ops_write_mask |= 1 << 3;
             } else if (ggml_can_fuse_subgraph(cgraph, i, topk_moe_late_softmax, { i + 1, i + 5 }) &&
                        ggml_check_edges(cgraph, i, topk_moe_late_softmax_edges) &&
                        ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, TOPK_MOE_LATE_SOFTMAX)) {
                 ctx->num_additional_fused_ops = topk_moe_late_softmax.size() - 1;
+                // view of argsort writes to memory
+                ctx->fused_ops_write_mask |= 1 << 1;
             }
         }
+        ctx->fused_ops_write_mask |= 1 << ctx->num_additional_fused_ops;
 
         // Signal the almost_ready fence when the graph is mostly complete (< 20% remaining)
         bool almost_ready = (cgraph->n_nodes - i) < cgraph->n_nodes / 5;
@@ -12707,6 +12855,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         }
         i += ctx->num_additional_fused_ops;
         ctx->num_additional_fused_ops = 0;
+        ctx->fused_ops_write_mask = 0;
     }
 
     if (vk_perf_logger_enabled) {
@@ -12863,6 +13012,32 @@ static void ggml_vk_graph_optimize(ggml_backend_t backend, struct ggml_cgraph *
             }
             if (ok) {
                 current_set.push_back(j);
+                // Look for ROPE + VIEW + SET_ROWS and make them consecutive
+                if (graph->nodes[j]->op == GGML_OP_ROPE) {
+                    int view_idx = -1;
+                    int set_rows_idx = -1;
+                    for (int k = j+1; k < std::min(j + 10, graph->n_nodes); ++k) {
+                        if (view_idx == -1 &&
+                            graph->nodes[k]->op == GGML_OP_VIEW &&
+                            graph->nodes[k]->src[0] == graph->nodes[j]) {
+                            view_idx = k;
+                            continue;
+                        }
+                        if (view_idx != -1 &&
+                            set_rows_idx == -1 &&
+                            graph->nodes[k]->op == GGML_OP_SET_ROWS &&
+                            graph->nodes[k]->src[0] == graph->nodes[view_idx]) {
+                            set_rows_idx = k;
+                            break;
+                        }
+                    }
+                    if (set_rows_idx != -1) {
+                        current_set.push_back(view_idx);
+                        current_set.push_back(set_rows_idx);
+                        used[view_idx] = true;
+                        used[set_rows_idx] = true;
+                    }
+                }
             }
         }
         // Second pass grabs view nodes.

ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp

@@ -14,6 +14,7 @@ layout (binding = 1)          buffer D {int data_d[];};
 
 layout (push_constant) uniform parameter {
     uint ncols;
+    uint nrows;
     uint order;
 } p;
 
@@ -26,10 +27,9 @@ void swap(uint idx0, uint idx1) {
     dst_row[idx1] = tmp;
 }
 
-void argsort(bool needs_bounds_check) {
+void argsort(bool needs_bounds_check, const uint row) {
     // bitonic sort
     const int col = int(gl_LocalInvocationID.x);
-    const uint row = gl_WorkGroupID.y;
 
     const uint row_offset = row * p.ncols;
 
@@ -72,8 +72,16 @@ void argsort(bool needs_bounds_check) {
 
 void main() {
     if (p.ncols == BLOCK_SIZE) {
-        argsort(false);
+        uint row = gl_WorkGroupID.y;
+        while (row < p.nrows) {
+            argsort(false, row);
+            row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+        }
     } else {
-        argsort(true);
+        uint row = gl_WorkGroupID.y;
+        while (row < p.nrows) {
+            argsort(true, row);
+            row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+        }
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp

@@ -82,9 +82,13 @@ layout (constant_id = 10) const uint WARP = 32;
 
 #include "mul_mmq_shmem_types.glsl"
 
+#ifdef MUL_MAT_ID
+#define BK_STEP 1
+#else
 #ifndef BK_STEP
 #define BK_STEP 4
 #endif
+#endif
 
 // Shared memory cache
 shared block_a_cache buf_a[BM * BK_STEP];

ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl

@@ -27,7 +27,7 @@ struct block_a_cache {
 #elif defined(DATA_A_Q8_0)
 #define QUANT_R_MMQ 1
 // AMD likes 4, Intel likes 1 and Nvidia likes 2
-#define BK_STEP 1
+// #define BK_STEP 1
 struct block_a_cache {
     int32_t qs[32/4];
     FLOAT_TYPE dm;

ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl

@@ -10,6 +10,7 @@ layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer Y {int data_pos[];};
 layout (binding = 2) readonly buffer Z {float data_ff[];};
 layout (binding = 3) writeonly buffer D {D_TYPE data_d[];};
+layout (binding = 4) readonly buffer I {uvec2 data_i[];}; // indices for set_rows
 
 layout (push_constant) uniform parameter {
     uint ncols;
@@ -26,7 +27,9 @@ layout (push_constant) uniform parameter {
     uint s1;
     uint s2;
     int sections[4];
+    uint is_imrope;
     uint is_back;
+    uint set_rows_stride;
 } p;
 
 float rope_yarn_ramp(const float low, const float high, const uint i0) {

ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp

@@ -32,17 +32,29 @@ void main() {
     const uint sector = (i0 / 2) % sect_dims;
 
     float theta_base = 0.0;
-    if (sector < p.sections[0]) {
-        theta_base = data_pos[channel_x]*pow(p.theta_scale, i0/2.0f);
-    }
-    else if (sector >= p.sections[0] && sector < sec_w) {
-        theta_base = data_pos[channel_x + ne2 * 1]*pow(p.theta_scale, i0/2.0f);
-    }
-    else if (sector >= sec_w && sector < sec_w + p.sections[2]) {
-        theta_base = data_pos[channel_x + ne2 * 2]*pow(p.theta_scale, i0/2.0f);
-    }
-    else if (sector >= sec_w + p.sections[2]) {
-        theta_base = data_pos[channel_x + ne2 * 3]*pow(p.theta_scale, i0/2.0f);
+    if (p.is_imrope != 0) {
+        if (sector % 3 == 1 && sector < 3 * p.sections[1]) {
+            theta_base = data_pos[channel_x + ne2 * 1]*pow(p.theta_scale, i0/2.0f);
+        } else if (sector % 3 == 2 && sector < 3 * p.sections[2]) {
+            theta_base = data_pos[channel_x + ne2 * 2]*pow(p.theta_scale, i0/2.0f);
+        } else if (sector % 3 == 0 && sector < 3 * p.sections[0]) {
+            theta_base = data_pos[channel_x]*pow(p.theta_scale, i0/2.0f);
+        } else {
+            theta_base = data_pos[channel_x + ne2 * 3]*pow(p.theta_scale, i0/2.0f);
+        }
+    } else {
+        if (sector < p.sections[0]) {
+            theta_base = data_pos[channel_x]*pow(p.theta_scale, i0/2.0f);
+        }
+        else if (sector >= p.sections[0] && sector < sec_w) {
+            theta_base = data_pos[channel_x + ne2 * 1]*pow(p.theta_scale, i0/2.0f);
+        }
+        else if (sector >= sec_w && sector < sec_w + p.sections[2]) {
+            theta_base = data_pos[channel_x + ne2 * 2]*pow(p.theta_scale, i0/2.0f);
+        }
+        else if (sector >= sec_w + p.sections[2]) {
+            theta_base = data_pos[channel_x + ne2 * 3]*pow(p.theta_scale, i0/2.0f);
+        }
     }
 
     const float freq_factor = p.has_ff != 0 ? data_ff[i0/2] : 1.0f;

ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp

@@ -16,12 +16,19 @@ void main() {
     const uint row_x     = row_dst % ne1;
     const uint channel_x = row_dst / ne1;
 
-    const uint idst = row_dst*ne0 + i0/2;
+    uint idst = row_dst*ne0 + i0/2;
     const uint ix   = channel_x*p.s2 + row_x*p.s1 + i0/2;
 
+    // Fusion optimization: ROPE + VIEW + SET_ROWS..
+    // The rope output is viewed as a 1D tensor and offset based on a row index in data_i.
+    if (p.set_rows_stride != 0) {
+        idst = row_x*ne0 + i0/2;
+        idst += data_i[channel_x].x * p.set_rows_stride;
+    }
+
     if (i0 >= p.n_dims) {
-        data_d[idst + i0/2 + 0] = data_a[ix + i0/2 + 0];
-        data_d[idst + i0/2 + 1] = data_a[ix + i0/2 + 1];
+        data_d[idst + i0/2 + 0] = D_TYPE(data_a[ix + i0/2 + 0]);
+        data_d[idst + i0/2 + 1] = D_TYPE(data_a[ix + i0/2 + 1]);
 
         return;
     }

ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp

@@ -16,12 +16,19 @@ void main() {
     const uint row_x     = row_dst % ne1;
     const uint channel_x = row_dst / ne1;
 
-    const uint idst = row_dst*ne0 + i0;
+    uint idst = row_dst*ne0 + i0;
     const uint ix   = channel_x*p.s2 + row_x*p.s1 + i0;
 
+    // Fusion optimization: ROPE + VIEW + SET_ROWS..
+    // The rope output is viewed as a 1D tensor and offset based on a row index in data_i.
+    if (p.set_rows_stride != 0) {
+        idst = row_x*ne0 + i0;
+        idst += data_i[channel_x].x * p.set_rows_stride;
+    }
+
     if (i0 >= p.n_dims) {
-        data_d[idst + 0] = data_a[ix + 0];
-        data_d[idst + 1] = data_a[ix + 1];
+        data_d[idst + 0] = D_TYPE(data_a[ix + 0]);
+        data_d[idst + 1] = D_TYPE(data_a[ix + 1]);
 
         return;
     }

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -317,7 +317,8 @@ void string_to_spv_func(std::string name, std::string in_path, std::string out_p
 
     // disable spirv-opt for coopmat shaders for https://github.com/ggerganov/llama.cpp/issues/10734
     // disable spirv-opt for bf16 shaders for https://github.com/ggml-org/llama.cpp/issues/15344
-    std::string opt_level = (coopmat || name.find("bf16") != std::string::npos) ? "" : "-O";
+    // disable spirv-opt for rope shaders for https://github.com/ggml-org/llama.cpp/issues/16860
+    std::string opt_level = (coopmat || name.find("bf16") != std::string::npos || name.find("rope") != std::string::npos) ? "" : "-O";
 
     #ifdef _WIN32
         std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, opt_level, "\"" + in_path + "\"", "-o", "\"" + out_path + "\""};
@@ -842,10 +843,14 @@ void process_shaders() {
     string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
     string_to_spv("rope_norm_f16_rte", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
+    string_to_spv("rope_norm_f32_f16", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
+    string_to_spv("rope_norm_f32_f16_rte", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
 
     string_to_spv("rope_neox_f32", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("rope_neox_f16", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
     string_to_spv("rope_neox_f16_rte", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
+    string_to_spv("rope_neox_f32_f16", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
+    string_to_spv("rope_neox_f32_f16_rte", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
 
     string_to_spv("rope_multi_f32", "rope_multi.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("rope_multi_f16", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});

ggml/src/ggml-webgpu/wgsl-shaders/rope.tmpl.wgsl

@@ -221,6 +221,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
 
     let is_neox = bool(params.mode & 2);
     let is_mrope = bool(params.mode & 8);
+    let is_imrope = params.mode == 40;
     let is_vision = params.mode == 24;
 
     var i = gid.x * 2; // start index for this thread
@@ -248,24 +249,36 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
         let sec_w = params.sections1 + params.sections0;
         let sec_e = params.sections2 + sec_w;
         let sector = (i0 / 2) % sect_dims;
-        if (sector >= params.sections0 && sector < sec_w) {
-            theta_base_mult = 1;
-            if (is_vision) {
-                theta_scale_pwr = sector - params.sections0;
-            }
-        } else if (sector >= sec_w && sector < sec_e) {
-            theta_base_mult = 2;
-            if (is_vision) {
-                theta_scale_pwr = sector - sec_w;
-            }
-        } else if (sector >= sec_e) {
-            if (is_vision) {
-                theta_scale_pwr = sector - sec_e;
-                theta_scale_pwr = (i0 / 2) % sec_e;
-            }
-            theta_base_mult = 3;
-        } else if (is_vision) {
-            theta_scale_pwr = sector;
+        if (is_imrope) {
+          if (sector % 3 == 1 && sector < 3 * params.sections1) {
+              theta_base_mult = 1;
+          } else if (sector % 3 == 2 && sector < 3 * params.sections2) {
+              theta_base_mult = 2;
+          } else if (sector % 3 == 0 && sector < 3 * params.sections0) {
+              theta_base_mult = 0;
+          } else {
+              theta_base_mult = 3;
+          }
+        } else {
+          if (sector >= params.sections0 && sector < sec_w) {
+              theta_base_mult = 1;
+              if (is_vision) {
+                  theta_scale_pwr = sector - params.sections0;
+              }
+          } else if (sector >= sec_w && sector < sec_e) {
+              theta_base_mult = 2;
+              if (is_vision) {
+                  theta_scale_pwr = sector - sec_w;
+              }
+          } else if (sector >= sec_e) {
+              if (is_vision) {
+                  theta_scale_pwr = sector - sec_e;
+                  theta_scale_pwr = (i0 / 2) % sec_e;
+              }
+              theta_base_mult = 3;
+          } else if (is_vision) {
+              theta_scale_pwr = sector;
+          }
         }
     }
     let theta_base = f32(src1[params.offset_src1 + i2 + params.ne2 * theta_base_mult]) * pow(params.theta_scale, f32(theta_scale_pwr));

gguf-py/gguf/constants.py

@@ -111,6 +111,7 @@ class Keys:
         EXPERTS_PER_GROUP                 = "{arch}.experts_per_group"
         MOE_EVERY_N_LAYERS                = "{arch}.moe_every_n_layers"
         NEXTN_PREDICT_LAYERS              = "{arch}.nextn_predict_layers"
+        NUM_DEEPSTACK_LAYERS              = "{arch}.n_deepstack_layers"
         POOLING_TYPE                      = "{arch}.pooling_type"
         LOGIT_SCALE                       = "{arch}.logit_scale"
         DECODER_START_TOKEN_ID            = "{arch}.decoder_start_token_id"
@@ -277,6 +278,7 @@ class Keys:
         USE_GELU            = "clip.use_gelu"
         USE_SILU            = "clip.use_silu"
         N_WA_PATTERN        = "clip.vision.n_wa_pattern" # used by qwen2.5vl
+        IS_DEEPSTACK_LAYERS = "clip.vision.is_deepstack_layers"
 
         class Attention:
             HEAD_COUNT      = "clip.vision.attention.head_count"
@@ -350,6 +352,8 @@ class MODEL_ARCH(IntEnum):
     QWEN2VL          = auto()
     QWEN3            = auto()
     QWEN3MOE         = auto()
+    QWEN3VL          = auto()
+    QWEN3VLMOE       = auto()
     PHI2             = auto()
     PHI3             = auto()
     PHIMOE           = auto()
@@ -420,6 +424,8 @@ class MODEL_ARCH(IntEnum):
     SEED_OSS         = auto()
     GROVEMOE         = auto()
     APERTUS          = auto()
+    COGVLM           = auto()
+    MINIMAXM2        = auto()
 
 
 class VISION_PROJECTOR_TYPE(IntEnum):
@@ -430,6 +436,8 @@ class VISION_PROJECTOR_TYPE(IntEnum):
     GLM_EDGE  = auto()
     MERGER    = auto()
     GEMMA3    = auto()
+    QWEN3VL   = auto()
+    COGVLM    = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -600,6 +608,11 @@ class MODEL_TENSOR(IntEnum):
     SHORTCONV_CONV       = auto()
     SHORTCONV_INPROJ     = auto()
     SHORTCONV_OUTPROJ    = auto()
+    VISEXP_ATTN_QKV      = auto()
+    VISEXP_ATTN_OUT      = auto()
+    VISEXP_GATE          = auto()
+    VISEXP_DOWN          = auto()
+    VISEXP_UP            = auto()
     # vision
     V_MMPROJ             = auto()
     V_MMPROJ_FC          = auto()
@@ -609,6 +622,7 @@ class MODEL_TENSOR(IntEnum):
     V_ENC_EMBD_PATCH     = auto()
     V_ENC_EMBD_POS       = auto()
     V_ENC_INPUT_NORM     = auto()
+    V_ENC_ATTN_QKV       = auto()
     V_ENC_ATTN_Q         = auto()
     V_ENC_ATTN_Q_NORM    = auto()
     V_ENC_ATTN_K         = auto()
@@ -640,6 +654,15 @@ class MODEL_TENSOR(IntEnum):
     V_RESMPL_QUERY       = auto() # minicpmv
     V_TOK_EMBD_IMG_BREAK = auto() # pixtral
     V_MM_PATCH_MERGER    = auto() # mistral small 3.1
+    V_DS_NORM            = auto() # qwen3vl
+    V_DS_FC1             = auto() # qwen3vl
+    V_DS_FC2             = auto() # qwen3vl
+    V_MM_POST_FC_NORM    = auto() # cogvlm
+    V_MM_UP              = auto() # cogvlm
+    V_MM_DOWN            = auto() # cogvlm
+    V_MM_GATE            = auto() # cogvlm
+    V_TOK_BOI            = auto() # cogvlm
+    V_TOK_EOI            = auto() # cogvlm
     # audio (mtmd)
     A_ENC_EMBD_POS       = auto()
     A_ENC_CONV1D         = auto()
@@ -695,6 +718,8 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.QWEN2VL:          "qwen2vl",
     MODEL_ARCH.QWEN3:            "qwen3",
     MODEL_ARCH.QWEN3MOE:         "qwen3moe",
+    MODEL_ARCH.QWEN3VL:          "qwen3vl",
+    MODEL_ARCH.QWEN3VLMOE:       "qwen3vlmoe",
     MODEL_ARCH.PHI2:             "phi2",
     MODEL_ARCH.PHI3:             "phi3",
     MODEL_ARCH.PHIMOE:           "phimoe",
@@ -766,6 +791,8 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.SEED_OSS:         "seed_oss",
     MODEL_ARCH.GROVEMOE:         "grovemoe",
     MODEL_ARCH.APERTUS:          "apertus",
+    MODEL_ARCH.MINIMAXM2:        "minimax-m2",
+    MODEL_ARCH.COGVLM:           "cogvlm",
 }
 
 VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -946,6 +973,11 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.SHORTCONV_CONV:            "blk.{bid}.shortconv.conv",
     MODEL_TENSOR.SHORTCONV_INPROJ:          "blk.{bid}.shortconv.in_proj",
     MODEL_TENSOR.SHORTCONV_OUTPROJ:         "blk.{bid}.shortconv.out_proj",
+    MODEL_TENSOR.VISEXP_ATTN_QKV:           "blk.{bid}.vis_attn_qkv",
+    MODEL_TENSOR.VISEXP_ATTN_OUT:           "blk.{bid}.vis_attn_output",
+    MODEL_TENSOR.VISEXP_GATE:               "blk.{bid}.vis_gate",
+    MODEL_TENSOR.VISEXP_DOWN:               "blk.{bid}.vis_down",
+    MODEL_TENSOR.VISEXP_UP:                 "blk.{bid}.vis_up",
     # vision
     MODEL_TENSOR.V_MMPROJ:                  "mm.{bid}",
     MODEL_TENSOR.V_MMPROJ_FC:               "mm.model.fc",
@@ -954,6 +986,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.V_ENC_EMBD_CLS:            "v.class_embd",
     MODEL_TENSOR.V_ENC_EMBD_PATCH:          "v.patch_embd",
     MODEL_TENSOR.V_ENC_EMBD_POS:            "v.position_embd",
+    MODEL_TENSOR.V_ENC_ATTN_QKV:            "v.blk.{bid}.attn_qkv",
     MODEL_TENSOR.V_ENC_ATTN_Q:              "v.blk.{bid}.attn_q",
     MODEL_TENSOR.V_ENC_ATTN_Q_NORM:         "v.blk.{bid}.attn_q_norm",
     MODEL_TENSOR.V_ENC_ATTN_K:              "v.blk.{bid}.attn_k",
@@ -986,6 +1019,15 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.V_RESMPL_QUERY:            "resampler.query",
     MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK:      "v.token_embd.img_break", # pixtral
     MODEL_TENSOR.V_MM_PATCH_MERGER:         "mm.patch_merger", # mistral small 3.1
+    MODEL_TENSOR.V_DS_NORM:                 "v.deepstack.{bid}.norm",
+    MODEL_TENSOR.V_DS_FC1:                  "v.deepstack.{bid}.fc1",
+    MODEL_TENSOR.V_DS_FC2:                  "v.deepstack.{bid}.fc2",
+    MODEL_TENSOR.V_MM_POST_FC_NORM:         "mm.post_fc_norm", # cogvlm
+    MODEL_TENSOR.V_MM_UP:                   "mm.up",
+    MODEL_TENSOR.V_MM_DOWN:                 "mm.down",
+    MODEL_TENSOR.V_MM_GATE:                 "mm.gate",
+    MODEL_TENSOR.V_TOK_BOI:                 "v.boi",
+    MODEL_TENSOR.V_TOK_EOI:                 "v.eoi",
     # audio (mtmd)
     MODEL_TENSOR.A_ENC_EMBD_POS:            "a.position_embd",
     MODEL_TENSOR.A_ENC_CONV1D:              "a.conv1d.{bid}",
@@ -1023,6 +1065,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.V_ENC_EMBD_PATCH,
         MODEL_TENSOR.V_ENC_EMBD_POS,
         MODEL_TENSOR.V_ENC_INPUT_NORM,
+        MODEL_TENSOR.V_ENC_ATTN_QKV,
         MODEL_TENSOR.V_ENC_ATTN_Q,
         MODEL_TENSOR.V_ENC_ATTN_Q_NORM,
         MODEL_TENSOR.V_ENC_ATTN_K,
@@ -1054,6 +1097,15 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.V_RESMPL_QUERY,
         MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK,
         MODEL_TENSOR.V_MM_PATCH_MERGER,
+        MODEL_TENSOR.V_DS_NORM,
+        MODEL_TENSOR.V_DS_FC1,
+        MODEL_TENSOR.V_DS_FC2,
+        MODEL_TENSOR.V_MM_POST_FC_NORM,
+        MODEL_TENSOR.V_MM_UP,
+        MODEL_TENSOR.V_MM_DOWN,
+        MODEL_TENSOR.V_MM_GATE,
+        MODEL_TENSOR.V_TOK_BOI,
+        MODEL_TENSOR.V_TOK_EOI,
         # audio
         MODEL_TENSOR.A_ENC_EMBD_POS,
         MODEL_TENSOR.A_ENC_CONV1D,
@@ -1495,6 +1547,40 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_EXP,
         MODEL_TENSOR.FFN_UP_EXP,
     ],
+    MODEL_ARCH.QWEN3VL: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.QWEN3VLMOE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+    ],
     MODEL_ARCH.PLAMO: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -2837,6 +2923,41 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_CHEXP,
         MODEL_TENSOR.FFN_UP_CHEXP,
     ],
+    MODEL_ARCH.MINIMAXM2: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_EXP_PROBS_B,
+    ],
+    MODEL_ARCH.COGVLM: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.VISEXP_ATTN_QKV,
+        MODEL_TENSOR.VISEXP_ATTN_OUT,
+        MODEL_TENSOR.VISEXP_GATE,
+        MODEL_TENSOR.VISEXP_UP,
+        MODEL_TENSOR.VISEXP_DOWN,
+    ],
     # TODO
 }
 
@@ -3055,6 +3176,7 @@ class VisionProjectorType:
     LLAMA4 = "llama4"
     QWEN2VL = "qwen2vl_merger"
     QWEN25VL = "qwen2.5vl_merger"
+    QWEN3VL = "qwen3vl_merger"
     ULTRAVOX = "ultravox"
     INTERNVL = "internvl"
     QWEN2A = "qwen2a" # audio
@@ -3063,6 +3185,7 @@ class VisionProjectorType:
     LFM2 = "lfm2"
     KIMIVL = "kimivl"
     LIGHTONOCR = "lightonocr"
+    COGVLM = "cogvlm"
 
 
 # Items here are (block size, type size)

gguf-py/gguf/gguf_writer.py

@@ -860,6 +860,9 @@ class GGUFWriter:
     def add_pooling_type(self, value: PoolingType) -> None:
         self.add_uint32(Keys.LLM.POOLING_TYPE.format(arch=self.arch), value.value)
 
+    def add_num_deepstack_layers(self, count: int) -> None:
+        self.add_uint32(Keys.LLM.NUM_DEEPSTACK_LAYERS.format(arch=self.arch), count)
+
     def add_rope_dimension_count(self, count: int) -> None:
         self.add_uint32(Keys.Rope.DIMENSION_COUNT.format(arch=self.arch), count)
 
@@ -1071,6 +1074,9 @@ class GGUFWriter:
     def add_vision_n_wa_pattern(self, value: int) -> None:
         self.add_uint32(Keys.ClipVision.N_WA_PATTERN, value)
 
+    def add_vision_is_deepstack_layers(self, layers: Sequence[bool]) -> None:
+        self.add_array(Keys.ClipVision.IS_DEEPSTACK_LAYERS, layers)
+
     # audio models
 
     def add_audio_projection_dim(self, value: int) -> None:

gguf-py/gguf/tensor_mapping.py

@@ -104,6 +104,7 @@ class TensorNameMap:
             "backbone.final_layer_norm",               # wavtokenizer
             "model.norm",                              # llama4
             "model.transformer.ln_f",                  # llada
+            "model.norm",                              # cogvlm
         ),
 
         # Rope frequencies
@@ -162,6 +163,7 @@ class TensorNameMap:
             "encoder.layer.{bid}.layer_norm_1",             # jina-v2-code
             "rwkv.blocks.{bid}.ln2",                        # rwkv6
             "model.layers.{bid}.ln2",                       # rwkv7
+            "model.layers.{bid}.post_attention_layernorm",  # cogvlm
         ),
 
         # Attention query-key-value
@@ -184,6 +186,7 @@ class TensorNameMap:
             "encoder.layers.{bid}.self_attention.query_key_value",                 # chatglm
             "transformer.layers.{bid}.attn.qkv_proj",                              # openelm
             "transformer_encoder.{bid}.qkv",                                       # neobert
+            "model.layers.{bid}.self_attn.language_expert_query_key_value",        # cogvlm
         ),
 
         # Attention query
@@ -279,6 +282,7 @@ class TensorNameMap:
             "model.transformer.blocks.{bid}.attn_out",                      # llada
             "layers.{bid}.self_attn.o_proj",                                # qwen3-embedding
             "backbone.layers.{bid}.mixer.o_proj",                           # nemotron-h
+            "model.layers.{bid}.self_attn.language_expert_dense",           # cogvlm
         ),
 
         # Attention output norm
@@ -377,6 +381,7 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.moe_statics.e_score_correction",        # ernie4.5-moe
             "model.layers.{bid}.mlp.gate.expert_bias",                      # bailingmoe2
             "model.layers.{bid}.feed_forward.expert_bias",                  # lfm2moe
+            "model.layers.{bid}.block_sparse_moe.e_score_correction",       # minimax-m2
         ),
 
         # Feed-forward up
@@ -418,6 +423,7 @@ class TensorNameMap:
             "model.transformer.blocks.{bid}.up_proj",                 # llada
             "layers.{bid}.mlp.up_proj",                               # qwen3-embedding
             "backbone.layers.{bid}.mixer.up_proj",                    # nemotron-h
+            "model.layers.{bid}.mlp.language_mlp.up_proj",            # cogvlm
         ),
 
         MODEL_TENSOR.FFN_UP_EXP: (
@@ -450,21 +456,22 @@ class TensorNameMap:
 
         # Feed-forward gate
         MODEL_TENSOR.FFN_GATE: (
-            "model.layers.{bid}.mlp.gate_proj",           # llama-hf refact olmo2
-            "layers.{bid}.mlp.gate_proj",                 # embeddinggemma
-            "layers.{bid}.feed_forward.w1",               # llama-pth
-            "transformer.h.{bid}.mlp.w2",                 # qwen
-            "transformer.h.{bid}.mlp.c_fc2",              # jais
-            "model.layers.layers.{bid}.mlp.gate_proj",    # plamo
-            "model.layers.{bid}.feed_forward.w1",         # internlm2
-            "encoder.layers.{bid}.mlp.fc12",              # nomic-bert
-            "encoder.layer.{bid}.mlp.gated_layers_w",     # jina-bert-v2 (split up/gate, no longer used)
-            "transformer.h.{bid}.mlp.linear_1",           # refact
-            "model.layers.{bid}.residual_mlp.w1",         # arctic
-            "transformer.h.{bid}.mlp.c_fc_0",             # exaone
-            "model.layers.{bid}.feed_forward.gate_proj",  # llama4 jamba granite-hybrid
-            "model.transformer.blocks.{bid}.ff_proj",     # llada
-            "layers.{bid}.mlp.gate_proj",                 # qwen3-embedding
+            "model.layers.{bid}.mlp.gate_proj",               # llama-hf refact olmo2
+            "layers.{bid}.mlp.gate_proj",                     # embeddinggemma
+            "layers.{bid}.feed_forward.w1",                   # llama-pth
+            "transformer.h.{bid}.mlp.w2",                     # qwen
+            "transformer.h.{bid}.mlp.c_fc2",                  # jais
+            "model.layers.layers.{bid}.mlp.gate_proj",        # plamo
+            "model.layers.{bid}.feed_forward.w1",             # internlm2
+            "encoder.layers.{bid}.mlp.fc12",                  # nomic-bert
+            "encoder.layer.{bid}.mlp.gated_layers_w",         # jina-bert-v2 (split up/gate, no longer used)
+            "transformer.h.{bid}.mlp.linear_1",               # refact
+            "model.layers.{bid}.residual_mlp.w1",             # arctic
+            "transformer.h.{bid}.mlp.c_fc_0",                 # exaone
+            "model.layers.{bid}.feed_forward.gate_proj",      # llama4 jamba granite-hybrid
+            "model.transformer.blocks.{bid}.ff_proj",         # llada
+            "layers.{bid}.mlp.gate_proj",                     # qwen3-embedding
+            "model.layers.{bid}.mlp.language_mlp.gate_proj",  # cogvlm
         ),
 
         MODEL_TENSOR.FFN_GATE_EXP: (
@@ -522,6 +529,7 @@ class TensorNameMap:
             "model.transformer.blocks.{bid}.ff_out",                  # llada
             "layers.{bid}.mlp.down_proj",                             # qwen3-embedding
             "backbone.layers.{bid}.mixer.down_proj",                  # nemotron-h
+            "model.layers.{bid}.mlp.language_mlp.down_proj",          # cogvlm
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (
@@ -1047,6 +1055,26 @@ class TensorNameMap:
             "encoder.block.{bid}.layer.1.DenseReluDense.wo", # t5
         ),
 
+        MODEL_TENSOR.VISEXP_UP: (
+            "model.layers.{bid}.mlp.vision_mlp.up_proj",  # cogvlm
+        ),
+
+        MODEL_TENSOR.VISEXP_GATE: (
+            "model.layers.{bid}.mlp.vision_mlp.gate_proj",  # cogvlm
+        ),
+
+        MODEL_TENSOR.VISEXP_DOWN: (
+            "model.layers.{bid}.mlp.vision_mlp.down_proj",  # cogvlm
+        ),
+
+        MODEL_TENSOR.VISEXP_ATTN_OUT: (
+            "model.layers.{bid}.self_attn.vision_expert_dense",  # cogvlm
+        ),
+
+        MODEL_TENSOR.VISEXP_ATTN_QKV: (
+            "model.layers.{bid}.self_attn.vision_expert_query_key_value",  # cogvlm
+        ),
+
         ############################################################################
         # TODO: these do not belong to block_mappings_cfg - move them to mappings_cfg
         MODEL_TENSOR.ENC_OUTPUT_NORM: (
@@ -1148,6 +1176,7 @@ class TensorNameMap:
 
         MODEL_TENSOR.V_MMPROJ_FC: (
             "model.connector.modality_projection.proj", # SmolVLM
+            "model.vision.linear_proj.linear_proj", # cogvlm
         ),
 
         MODEL_TENSOR.V_MMPROJ_MLP: (
@@ -1164,6 +1193,7 @@ class TensorNameMap:
             "vision_tower.vision_model.embeddings.class_embedding",
             "model.vision_tower.embeddings.cls_token", # Intern-S1
             "vision_model.class_embedding", # llama 4
+            "model.vision.patch_embedding.cls_embedding", # cogvlm
         ),
 
         MODEL_TENSOR.V_ENC_EMBD_PATCH: (
@@ -1176,6 +1206,7 @@ class TensorNameMap:
             "vision_model.patch_embedding.linear", # llama 4
             "visual.patch_embed.proj", # qwen2vl
             "vision_tower.patch_embed.proj", # kimi-vl
+            "model.vision.patch_embedding.proj", # cogvlm
         ),
 
         MODEL_TENSOR.V_ENC_EMBD_POS: (
@@ -1185,6 +1216,13 @@ class TensorNameMap:
             "model.vision_model.embeddings.position_embedding", # SmolVLM
             "vision_model.positional_embedding_vlm", # llama 4
             "vision_tower.patch_embed.pos_emb", # kimi-vl
+            "visual.pos_embed", # qwen3vl
+            "model.vision.patch_embedding.position_embedding", # cogvlm
+        ),
+
+        MODEL_TENSOR.V_ENC_ATTN_QKV: (
+            "visual.blocks.{bid}.attn.qkv", # qwen3vl
+            "model.vision.transformer.layers.{bid}.attention.query_key_value", # cogvlm
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_Q: (
@@ -1244,6 +1282,7 @@ class TensorNameMap:
             "vision_model.model.layers.{bid}.input_layernorm", # llama4
             "visual.blocks.{bid}.norm1", # qwen2vl
             "vision_tower.encoder.blocks.{bid}.norm0", # kimi-vl (norm0/norm1)
+            "model.vision.transformer.layers.{bid}.input_layernorm", # cogvlm
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_O: (
@@ -1257,6 +1296,7 @@ class TensorNameMap:
             "vision_encoder.transformer.layers.{bid}.attention.wo", # pixtral
             "visual.blocks.{bid}.attn.proj", # qwen2vl
             "vision_tower.encoder.blocks.{bid}.wo", # kimi-vl
+            "model.vision.transformer.layers.{bid}.attention.dense", # cogvlm
         ),
 
         MODEL_TENSOR.V_ENC_POST_ATTN_NORM: (
@@ -1270,6 +1310,7 @@ class TensorNameMap:
             "vision_encoder.transformer.layers.{bid}.ffn_norm", # pixtral
             "visual.blocks.{bid}.norm2", # qwen2vl
             "vision_tower.encoder.blocks.{bid}.norm1", # kimi-vl (norm0/norm1)
+            "model.vision.transformer.layers.{bid}.post_attention_layernorm", # cogvlm
         ),
 
         MODEL_TENSOR.V_ENC_FFN_UP: (
@@ -1282,7 +1323,9 @@ class TensorNameMap:
             "vision_model.model.layers.{bid}.mlp.fc1", # llama4
             "visual.blocks.{bid}.mlp.fc1", # qwen2vl
             "visual.blocks.{bid}.mlp.up_proj", # qwen2.5vl
+            "visual.blocks.{bid}.mlp.linear_fc1", # qwen3vl
             "vision_tower.encoder.blocks.{bid}.mlp.fc0", # kimi-vl (fc0/fc1)
+            "model.vision.transformer.layers.{bid}.mlp.fc1", # cogvlm
         ),
 
         MODEL_TENSOR.V_ENC_FFN_GATE: (
@@ -1301,7 +1344,9 @@ class TensorNameMap:
             "vision_model.model.layers.{bid}.mlp.fc2", # llama4
             "visual.blocks.{bid}.mlp.fc2", # qwen2vl
             "visual.blocks.{bid}.mlp.down_proj", # qwen2.5vl
+            "visual.blocks.{bid}.mlp.linear_fc2", # qwen3vl
             "vision_tower.encoder.blocks.{bid}.mlp.fc1", # kimi-vl (fc0/fc1)
+            "model.vision.transformer.layers.{bid}.mlp.fc2", # cogvlm
         ),
 
         MODEL_TENSOR.V_LAYER_SCALE_1: (
@@ -1338,6 +1383,7 @@ class TensorNameMap:
             "multi_modal_projector.layer_norm",
             "multi_modal_projector.pre_norm",
             "pre_mm_projector_norm",
+            "model.vision.linear_proj.norm1", # cogvlm
         ),
 
         MODEL_TENSOR.V_MM_SOFT_EMB_NORM: (
@@ -1397,6 +1443,42 @@ class TensorNameMap:
             "patch_merger.merging_layer", # mistral
         ),
 
+        MODEL_TENSOR.V_DS_NORM: (
+            "model.visual.deepstack_merger_list.{bid}.norm", # deepstack in qwen3vl
+        ),
+
+        MODEL_TENSOR.V_DS_FC1: (
+            "model.visual.deepstack_merger_list.{bid}.linear_fc1", # deepstack in qwen3vl
+        ),
+
+        MODEL_TENSOR.V_DS_FC2: (
+            "model.visual.deepstack_merger_list.{bid}.linear_fc2", # deepstack in qwen3vl
+        ),
+
+        MODEL_TENSOR.V_MM_POST_FC_NORM: (
+            "model.vision.linear_proj.norm1", # cogvlm
+        ),
+
+        MODEL_TENSOR.V_MM_UP: (
+            "model.vision.linear_proj.dense_h_to_4h", # cogvlm
+        ),
+
+        MODEL_TENSOR.V_MM_DOWN: (
+            "model.vision.linear_proj.dense_4h_to_h", # cogvlm
+        ),
+
+        MODEL_TENSOR.V_MM_GATE: (
+            "model.vision.linear_proj.gate_proj", # cogvlm
+        ),
+
+        MODEL_TENSOR.V_TOK_BOI: (
+            "model.vision.boi", # cogvlm
+        ),
+
+        MODEL_TENSOR.V_TOK_EOI: (
+            "model.vision.eoi", # cogvlm
+        ),
+
         # audio (mtmd)
 
         MODEL_TENSOR.A_ENC_EMBD_POS: (

include/llama.h

@@ -83,6 +83,7 @@ extern "C" {
         LLAMA_ROPE_TYPE_NORM   = 0,
         LLAMA_ROPE_TYPE_NEOX   = GGML_ROPE_TYPE_NEOX,
         LLAMA_ROPE_TYPE_MROPE  = GGML_ROPE_TYPE_MROPE,
+        LLAMA_ROPE_TYPE_IMROPE = GGML_ROPE_TYPE_IMROPE,
         LLAMA_ROPE_TYPE_VISION = GGML_ROPE_TYPE_VISION,
     };
 

requirements/requirements-convert_legacy_llama.txt

@@ -1,14 +1,7 @@
 numpy~=1.26.4
 sentencepiece~=0.2.0
 
-# Embedding Gemma is currently a preview release:
-# https://github.com/huggingface/transformers/releases/tag/v4.56.0-Embedding-Gemma-preview
-
-# The version is needed to be able to convert Embedding Gemma models to GGUF format:
-git+https://github.com/huggingface/transformers@v4.56.0-Embedding-Gemma-preview
-
-# Once Embedding Gemma is officially released, we can switch to:
-#transformers>=4.57.1,<5.0.0
+transformers>=4.57.1,<5.0.0
 
 gguf>=0.1.0
 protobuf>=4.21.0,<5.0.0

scripts/sync-ggml.last

@@ -1 +1 @@
-72632094336524a9c809e129e8b1c52154543a5a
+e02fb860ccbba8967905bceff23b677e88105280

src/CMakeLists.txt

@@ -35,6 +35,100 @@ add_library(llama
             unicode-data.cpp
             unicode.cpp
             unicode.h
+            models/apertus.cpp
+            models/arcee.cpp
+            models/arctic.cpp
+            models/arwkv7.cpp
+            models/baichuan.cpp
+            models/bailingmoe.cpp
+            models/bailingmoe2.cpp
+            models/bert.cpp
+            models/bitnet.cpp
+            models/bloom.cpp
+            models/chameleon.cpp
+            models/chatglm.cpp
+            models/codeshell.cpp
+            models/cogvlm.cpp
+            models/cohere2-iswa.cpp
+            models/command-r.cpp
+            models/dbrx.cpp
+            models/deci.cpp
+            models/deepseek.cpp
+            models/deepseek2.cpp
+            models/dots1.cpp
+            models/dream.cpp
+            models/ernie4-5-moe.cpp
+            models/ernie4-5.cpp
+            models/exaone.cpp
+            models/exaone4.cpp
+            models/falcon-h1.cpp
+            models/falcon.cpp
+            models/gemma-embedding.cpp
+            models/gemma.cpp
+            models/gemma2-iswa.cpp
+            models/gemma3-iswa.cpp
+            models/gemma3n-iswa.cpp
+            models/glm4-moe.cpp
+            models/glm4.cpp
+            models/gpt2.cpp
+            models/gptneox.cpp
+            models/granite-hybrid.cpp
+            models/granite.cpp
+            models/grok.cpp
+            models/grovemoe.cpp
+            models/hunyuan-dense.cpp
+            models/hunyuan-moe.cpp
+            models/internlm2.cpp
+            models/jais.cpp
+            models/jamba.cpp
+            models/lfm2.cpp
+            models/llada-moe.cpp
+            models/llada.cpp
+            models/llama-iswa.cpp
+            models/llama.cpp
+            models/mamba.cpp
+            models/minicpm3.cpp
+            models/minimax-m2.cpp
+            models/mpt.cpp
+            models/nemotron-h.cpp
+            models/nemotron.cpp
+            models/neo-bert.cpp
+            models/olmo.cpp
+            models/olmo2.cpp
+            models/olmoe.cpp
+            models/openai-moe-iswa.cpp
+            models/openelm.cpp
+            models/orion.cpp
+            models/phi2.cpp
+            models/phi3.cpp
+            models/plamo.cpp
+            models/plamo2.cpp
+            models/plm.cpp
+            models/qwen.cpp
+            models/qwen2.cpp
+            models/qwen2moe.cpp
+            models/qwen2vl.cpp
+            models/qwen3.cpp
+            models/qwen3vl.cpp
+            models/qwen3vl-moe.cpp
+            models/qwen3moe.cpp
+            models/refact.cpp
+            models/rwkv6-base.cpp
+            models/rwkv6.cpp
+            models/rwkv6qwen2.cpp
+            models/rwkv7-base.cpp
+            models/rwkv7.cpp
+            models/seed-oss.cpp
+            models/smallthinker.cpp
+            models/smollm3.cpp
+            models/stablelm.cpp
+            models/starcoder.cpp
+            models/starcoder2.cpp
+            models/t5-dec.cpp
+            models/t5-enc.cpp
+            models/wavtokenizer-dec.cpp
+            models/xverse.cpp
+            models/graph-context-mamba.cpp
             )
 
 target_include_directories(llama PRIVATE .)

src/llama-arch.cpp

@@ -32,6 +32,8 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_QWEN2VL,          "qwen2vl"          },
     { LLM_ARCH_QWEN3,            "qwen3"            },
     { LLM_ARCH_QWEN3MOE,         "qwen3moe"         },
+    { LLM_ARCH_QWEN3VL,          "qwen3vl"          },
+    { LLM_ARCH_QWEN3VLMOE,       "qwen3vlmoe"       },
     { LLM_ARCH_PHI2,             "phi2"             },
     { LLM_ARCH_PHI3,             "phi3"             },
     { LLM_ARCH_PHIMOE,           "phimoe"           },
@@ -103,6 +105,8 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_SEED_OSS,         "seed_oss"         },
     { LLM_ARCH_GROVEMOE,         "grovemoe"         },
     { LLM_ARCH_APERTUS,          "apertus"          },
+    { LLM_ARCH_MINIMAX_M2,       "minimax-m2"       },
+    { LLM_ARCH_COGVLM,           "cogvlm"           },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -145,6 +149,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_EXPERTS_PER_GROUP,                 "%s.experts_per_group"                 },
     { LLM_KV_MOE_EVERY_N_LAYERS,                "%s.moe_every_n_layers"                },
     { LLM_KV_NEXTN_PREDICT_LAYERS,              "%s.nextn_predict_layers"              },
+    { LLM_KV_NUM_DEEPSTACK_LAYERS,              "%s.n_deepstack_layers"                },
     { LLM_KV_POOLING_TYPE,                      "%s.pooling_type"                      },
     { LLM_KV_LOGIT_SCALE,                       "%s.logit_scale"                       },
     { LLM_KV_DECODER_START_TOKEN_ID,            "%s.decoder_start_token_id"            },
@@ -779,6 +784,45 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_QWEN3VL,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_QWEN3VLMOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_PHI2,
         {
@@ -2312,6 +2356,47 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_CHEXPS,      "blk.%d.ffn_up_chexps" },
         },
     },
+    {
+        LLM_ARCH_MINIMAX_M2,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B,    "blk.%d.exp_probs_b" },
+        },
+    },
+    {
+        LLM_ARCH_COGVLM,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_VISEXP_ATTN_QKV, "blk.%d.vis_attn_qkv" },
+            { LLM_TENSOR_VISEXP_ATTN_OUT, "blk.%d.vis_attn_output" },
+            { LLM_TENSOR_VISEXP_FFN_GATE, "blk.%d.vis_gate" },
+            { LLM_TENSOR_VISEXP_FFN_DOWN, "blk.%d.vis_down" },
+            { LLM_TENSOR_VISEXP_FFN_UP,   "blk.%d.vis_up" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -2488,6 +2573,11 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_SHORTCONV_CONV,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
     {LLM_TENSOR_SHORTCONV_INPROJ,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_SHORTCONV_OUTPROJ,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_VISEXP_ATTN_QKV,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_VISEXP_ATTN_OUT,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_VISEXP_FFN_GATE,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_VISEXP_FFN_DOWN,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_VISEXP_FFN_UP,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     // NextN/MTP tensors are currently ignored (reserved for future MTP support)
     // These tensors only exist in the last layer(s) and are treated as output tensors
     {LLM_TENSOR_NEXTN_EH_PROJ,              {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},

src/llama-arch.h

@@ -36,6 +36,8 @@ enum llm_arch {
     LLM_ARCH_QWEN2VL,
     LLM_ARCH_QWEN3,
     LLM_ARCH_QWEN3MOE,
+    LLM_ARCH_QWEN3VL,
+    LLM_ARCH_QWEN3VLMOE,
     LLM_ARCH_PHI2,
     LLM_ARCH_PHI3,
     LLM_ARCH_PHIMOE,
@@ -107,6 +109,8 @@ enum llm_arch {
     LLM_ARCH_SEED_OSS,
     LLM_ARCH_GROVEMOE,
     LLM_ARCH_APERTUS,
+    LLM_ARCH_MINIMAX_M2,
+    LLM_ARCH_COGVLM,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -149,6 +153,7 @@ enum llm_kv {
     LLM_KV_EXPERTS_PER_GROUP,
     LLM_KV_MOE_EVERY_N_LAYERS,
     LLM_KV_NEXTN_PREDICT_LAYERS,
+    LLM_KV_NUM_DEEPSTACK_LAYERS,
     LLM_KV_POOLING_TYPE,
     LLM_KV_LOGIT_SCALE,
     LLM_KV_DECODER_START_TOKEN_ID,
@@ -455,6 +460,11 @@ enum llm_tensor {
     LLM_TENSOR_SHORTCONV_CONV,
     LLM_TENSOR_SHORTCONV_INPROJ,
     LLM_TENSOR_SHORTCONV_OUTPROJ,
+    LLM_TENSOR_VISEXP_ATTN_QKV,
+    LLM_TENSOR_VISEXP_ATTN_OUT,
+    LLM_TENSOR_VISEXP_FFN_GATE,
+    LLM_TENSOR_VISEXP_FFN_DOWN,
+    LLM_TENSOR_VISEXP_FFN_UP,
     LLM_TENSOR_NEXTN_EH_PROJ,
     LLM_TENSOR_NEXTN_EMBED_TOKENS,
     LLM_TENSOR_NEXTN_ENORM,

src/llama-batch.cpp

@@ -261,15 +261,29 @@ bool llama_batch_allocr::init(
 
             const llama_pos p0 = memory ? memory->seq_pos_max(s) : -1;
 
-            if (p0 >= 0 && p0 >= seq_pos_min(s)) {
-                LLAMA_LOG_ERROR(
-                        "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n"
-                        " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n"
-                        " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n"
-                        " for M-RoPE, it is required that the position satisfies: X < Y\n",
-                        __func__, s, s, p0, s, seq_pos_min(s));
+            if (batch.token) {
+                if (p0 >= 0 && p0 >= seq_pos_min(s)) {
+                    LLAMA_LOG_ERROR(
+                            "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n"
+                            " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n"
+                            " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n"
+                            " for M-RoPE, it is required that the position satisfies: X < Y\n",
+                            __func__, s, s, p0, s, seq_pos_min(s));
 
-                return false;
+                    return false;
+                }
+            } else {
+                // embedding inputs can have overlapping positions
+                if (p0 >= 0 && p0 > seq_pos_min(s)) {
+                    LLAMA_LOG_ERROR(
+                            "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n"
+                            " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n"
+                            " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n"
+                            " for M-RoPE, it is required that the position satisfies: X <= Y\n",
+                            __func__, s, s, p0, s, seq_pos_min(s));
+
+                    return false;
+                }
             }
         }
     } else {

src/llama-graph.cpp

@@ -2035,7 +2035,7 @@ int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buck
 
     if (bidirectional) {
         relative_bucket += (relative_position > 0) * n_buckets;
-        relative_position = abs(relative_position);
+        relative_position = std::abs(relative_position);
     } else {
         relative_position = -std::min<int32_t>(relative_position, 0);
     }

src/llama-hparams.cpp

@@ -148,7 +148,7 @@ bool llama_hparams::is_recurrent(uint32_t il) const {
 }
 
 uint32_t llama_hparams::n_pos_per_embd() const {
-    return rope_type == LLAMA_ROPE_TYPE_MROPE ? 4 : 1;
+    return rope_type == LLAMA_ROPE_TYPE_MROPE || rope_type == LLAMA_ROPE_TYPE_IMROPE ? 4 : 1;
 }
 
 bool llama_hparams::is_swa(uint32_t il) const {

src/llama-hparams.h

@@ -183,6 +183,9 @@ struct llama_hparams {
     std::array<float, LLAMA_MAX_LAYERS> xielu_beta;
     std::array<float, LLAMA_MAX_LAYERS> xielu_eps;
 
+    // qwen3vl deepstack
+    uint32_t n_deepstack_layers = 0;
+
     // needed by encoder-decoder models (e.g. T5, FLAN-T5)
     // ref: https://github.com/ggerganov/llama.cpp/pull/8141
     llama_token dec_start_token_id = LLAMA_TOKEN_NULL;

src/llama-kv-cache.cpp

@@ -1375,7 +1375,7 @@ ggml_tensor * llama_kv_cache::build_rope_shift(
     const auto & yarn_beta_slow  = cparams.yarn_beta_slow;
 
     const auto & n_rot     = hparams.n_rot;
-    const auto & rope_type = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE
+    const auto & rope_type = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE || hparams.rope_type == LLAMA_ROPE_TYPE_IMROPE
                                 // @ngxson : this is a workaround
                                 // for M-RoPE, we want to rotate the whole vector when doing KV shift
                                 // a normal RoPE should work, we just need to use the correct ordering

src/llama-model.h

@@ -114,6 +114,7 @@ enum llm_type {
     LLM_TYPE_30B_A3B,
     LLM_TYPE_100B_A6B,
     LLM_TYPE_106B_A12B, // GLM-4.5-Air
+    LLM_TYPE_230B_A10B, // Minimax M2
     LLM_TYPE_235B_A22B,
     LLM_TYPE_300B_A47B, // Ernie MoE big
     LLM_TYPE_355B_A32B, // GLM-4.5
@@ -384,6 +385,13 @@ struct llama_layer {
     // openai-moe
     struct ggml_tensor * attn_sinks = nullptr;
 
+    // cogvlm
+    struct ggml_tensor * visexp_attn_wqkv = nullptr;
+    struct ggml_tensor * visexp_attn_wo   = nullptr;
+    struct ggml_tensor * visexp_ffn_gate  = nullptr;
+    struct ggml_tensor * visexp_ffn_down  = nullptr;
+    struct ggml_tensor * visexp_ffn_up    = nullptr;
+
     // xIELU activation parameters for Apertus
     struct ggml_tensor * ffn_act_alpha_n = nullptr;
     struct ggml_tensor * ffn_act_alpha_p = nullptr;

src/llama-quant.cpp

@@ -653,7 +653,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
                 gguf_set_val_f32(ctx_out.get(), o.key, o.val_f64);
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_INT) {
                 // Setting type to UINT32. See https://github.com/ggml-org/llama.cpp/pull/14182 for context
-                gguf_set_val_u32(ctx_out.get(), o.key, (uint32_t)abs(o.val_i64));
+                gguf_set_val_u32(ctx_out.get(), o.key, (uint32_t)std::abs(o.val_i64));
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_BOOL) {
                 gguf_set_val_bool(ctx_out.get(), o.key, o.val_bool);
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_STR) {

src/llama-vocab.cpp

@@ -401,6 +401,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                 };
                 break;
             case LLAMA_VOCAB_PRE_TYPE_GPT4O:
+            case LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2:
                 regex_exprs = {
                     // original regex from tokenizer.json
                     // "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
@@ -1992,6 +1993,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "grok-2") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_GROK_2;
                 clean_spaces = false;
+            } else if (
+                tokenizer_pre == "minimax-m2") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2;
+                clean_spaces = false;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }

src/llama-vocab.h

@@ -49,6 +49,7 @@ enum llama_vocab_pre_type {
     LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE   = 38,
     LLAMA_VOCAB_PRE_TYPE_GROK_2          = 39,
     LLAMA_VOCAB_PRE_TYPE_GRANITE_DOCLING = 40,
+    LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2      = 41,
 };
 
 struct LLM_KV;

src/models/apertus.cpp

@@ -0,0 +1,125 @@
+#include "models.h"
+
+
+
+llm_build_apertus::llm_build_apertus(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    ggml_tensor * inp_pos  = build_inp_pos();
+    auto *        inp_attn = build_attn_inp_kv();
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur_pos", il);
+            cb(Kcur, "Kcur_pos", il);
+            cb(Vcur, "Vcur_pos", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network with xIELU activation
+        {
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            // Up projection
+            ggml_tensor * up = build_lora_mm(model.layers[il].ffn_up, cur);
+            cb(up, "ffn_up", il);
+
+            float alpha_n_val = hparams.xielu_alpha_n[il];
+            float alpha_p_val = hparams.xielu_alpha_p[il];
+            float beta_val    = hparams.xielu_beta[il];
+            float eps_val     = hparams.xielu_eps[il];
+
+            // Apply xIELU activation
+            ggml_tensor * activated = ggml_xielu(ctx0, up, alpha_n_val, alpha_p_val, beta_val, eps_val);
+            cb(activated, "ffn_xielu", il);
+
+            // Down projection
+            cur = build_lora_mm(model.layers[il].ffn_down, activated);
+            cb(cur, "ffn_down", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/arcee.cpp

@@ -0,0 +1,135 @@
+#include "models.h"
+
+
+llm_build_arcee::llm_build_arcee(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        // ARCEE uses relu^2 instead of silu
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                NULL,                      NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/arctic.cpp

@@ -0,0 +1,138 @@
+#include "models.h"
+
+
+llm_build_arctic::llm_build_arctic(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        ggml_tensor * ffn_out = ggml_add(ctx0, cur, ffn_inp);
+        cb(ffn_out, "ffn_out", il);
+
+        // MoE
+        cur = build_norm(inpSA,
+                model.layers[il].ffn_norm_exps, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm_exps", il);
+
+        cur = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, true,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+        cb(cur, "ffn_moe_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_out);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/arwkv7.cpp

@@ -0,0 +1,86 @@
+#include "models.h"
+
+
+llm_build_arwkv7::llm_build_arwkv7(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv7_base(model, params) {
+    GGML_ASSERT(n_embd == hparams.n_embd_r());
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+    ggml_tensor * v_first = nullptr;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    auto * rs_inp = build_rs_inp();
+
+    const auto n_embd = hparams.n_embd;
+    const auto n_seq_tokens = ubatch.n_seq_tokens;
+    const auto n_seqs = ubatch.n_seqs;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const llama_layer * layer = &model.layers[il];
+        inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
+
+        ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il);
+
+        ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, il);
+        cb(att_norm, "attn_norm", il);
+
+        ggml_tensor * x_prev = ggml_concat(
+                ctx0,
+                token_shift,
+                ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0),
+                1
+                );
+
+        cur = build_rwkv7_time_mix(rs_inp, att_norm, x_prev, v_first, ubatch, il);
+
+        token_shift = ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm));
+        ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il));
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur     = ggml_reshape_2d(ctx0, cur,     n_embd, n_tokens);
+        ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur     = ggml_get_rows(ctx0, cur,     inp_out_ids);
+            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
+        }
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+    cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/baichuan.cpp

@@ -0,0 +1,122 @@
+#include "models.h"
+
+
+llm_build_baichuan::llm_build_baichuan(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = model.type == LLM_TYPE_7B ? build_inp_pos() : nullptr;
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            switch (model.type) {
+                case LLM_TYPE_7B:
+                    Qcur = ggml_rope_ext(
+                            ctx0, Qcur, inp_pos, nullptr,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow
+                            );
+                    Kcur = ggml_rope_ext(
+                            ctx0, Kcur, inp_pos, nullptr,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow
+                            );
+                    break;
+                case LLM_TYPE_13B:
+                    break;
+                default:
+                    GGML_ABORT("fatal error");
+            }
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/bailingmoe.cpp

@@ -0,0 +1,144 @@
+#include "models.h"
+
+
+llm_build_bailingmoe::llm_build_bailingmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_rot, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_rot, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        ggml_tensor * moe_out =
+            build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, hparams.expert_weights_norm,
+                    false, hparams.expert_weights_scale,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+        cb(moe_out, "ffn_moe_out", il);
+
+        // FFN shared expert
+        {
+            ggml_tensor * ffn_shexp = build_ffn(cur,
+                    model.layers[il].ffn_up_shexp,   NULL, NULL,
+                    model.layers[il].ffn_gate_shexp, NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(ffn_shexp, "ffn_shexp", il);
+
+            cur = ggml_add(ctx0, moe_out, ffn_shexp);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/bailingmoe2.cpp

@@ -0,0 +1,135 @@
+#include "models.h"
+
+
+
+llm_build_bailingmoe2::llm_build_bailingmoe2(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
+    for (int il = 0; il < n_transformer_layers; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float),
+                                              cur->nb[1], 0 * sizeof(float) * (n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                              cur->nb[1], 1 * sizeof(float) * (n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                              cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_transformer_layers - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpSA);
+        cb(sa_out, "sa_out", il);
+
+        // MoE branch
+        cur = build_norm(sa_out, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                model.layers[il].ffn_exp_probs_b,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, hparams.expert_weights_norm,
+                true, hparams.expert_weights_scale,
+                (llama_expert_gating_func_type) hparams.expert_gating_func,
+                il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            }
+        }
+
+        cur = ggml_add(ctx0, cur, sa_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/bert.cpp

@@ -0,0 +1,176 @@
+#include "models.h"
+
+
+
+llm_build_bert::llm_build_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+    ggml_tensor * inp_pos = nullptr;
+
+    if (model.arch != LLM_ARCH_JINA_BERT_V2) {
+        inp_pos = build_inp_pos();
+    }
+
+    // construct input embeddings (token, type, position)
+    inpL = build_inp_embd(model.tok_embd);
+
+    // token types are hardcoded to zero ("Sentence A")
+    if (model.type_embd) {
+        ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0);
+        inpL                    = ggml_add(ctx0, inpL, type_row0);
+    }
+    if (model.arch == LLM_ARCH_BERT) {
+        inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL);
+    }
+    cb(inpL, "inp_embd", -1);
+
+    // embed layer norm
+    inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1);
+    cb(inpL, "inp_norm", -1);
+
+    auto * inp_attn = build_attn_inp_no_cache();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * cur = inpL;
+
+        {
+            ggml_tensor * Qcur;
+            ggml_tensor * Kcur;
+            ggml_tensor * Vcur;
+
+            // self-attention
+            if (model.layers[il].wqkv) {
+                cur = build_lora_mm(model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                if (model.layers[il].bqkv) {
+                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+                }
+
+                Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), cur->nb[1],
+                                    0 * sizeof(float) * (n_embd));
+                Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                    cur->nb[1], 1 * sizeof(float) * (n_embd));
+                Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                    cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
+            } else {
+                Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, cur), model.layers[il].bq);
+                Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, cur), model.layers[il].bk);
+                Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, cur), model.layers[il].bv);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+            }
+
+            if (model.layers[il].attn_q_norm) {
+                Qcur = ggml_reshape_2d(ctx0, Qcur, n_embd_head * n_head, n_tokens);
+
+                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, model.layers[il].attn_q_norm_b, LLM_NORM, il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            }
+
+            if (model.layers[il].attn_k_norm) {
+                Kcur = ggml_reshape_2d(ctx0, Kcur, n_embd_head * n_head_kv, n_tokens);
+
+                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, model.layers[il].attn_k_norm_b, LLM_NORM, il);
+
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            }
+
+            // RoPE
+            if (model.arch == LLM_ARCH_NOMIC_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE ||
+                model.arch == LLM_ARCH_JINA_BERT_V3) {
+                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                     ext_factor, attn_factor, beta_fast, beta_slow);
+
+                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                     ext_factor, attn_factor, beta_fast, beta_slow);
+            }
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+            cb(cur, "kqv_out", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // re-add the layer input
+        cur = ggml_add(ctx0, cur, inpL);
+
+        // attention layer norm
+        cur = build_norm(cur, model.layers[il].attn_out_norm, model.layers[il].attn_out_norm_b, LLM_NORM, il);
+
+        if (model.layers[il].attn_norm_2 != nullptr) {
+            cur = ggml_add(ctx0, cur, inpL);  // re-add the layer input
+            cur = build_norm(cur, model.layers[il].attn_norm_2, model.layers[il].attn_norm_2_b, LLM_NORM, il);
+        }
+
+        ggml_tensor * ffn_inp = cur;
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        if (hparams.moe_every_n_layers > 0 && il % hparams.moe_every_n_layers == 1) {
+            // MoE branch
+            cur = build_moe_ffn(cur, model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps, nullptr,
+                                model.layers[il].ffn_down_exps, nullptr, hparams.n_expert, hparams.n_expert_used,
+                                LLM_FFN_GELU, false, false, 0.0f, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
+            cb(cur, "ffn_moe_out", il);
+        } else if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE ||
+                   model.arch == LLM_ARCH_JINA_BERT_V3) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+                    NULL, NULL, NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        } else if (model.arch == LLM_ARCH_JINA_BERT_V2) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, NULL,
+                    model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_GEGLU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        // attentions bypass the intermediate layer
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        // output layer norm
+        cur = build_norm(cur, model.layers[il].layer_out_norm, model.layers[il].layer_out_norm_b, LLM_NORM, il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cb(cur, "result_embd", -1);
+    res->t_embd = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/bitnet.cpp

@@ -0,0 +1,160 @@
+#include "models.h"
+
+
+llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            if (model.layers[il].wq_scale) {
+                Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_scale);
+            }
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+
+            // B1.K
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            if (model.layers[il].wk_scale) {
+                Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_scale);
+            }
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+
+            // B1.V
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            if (model.layers[il].wv_scale) {
+                Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_scale);
+            }
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    NULL, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+
+            cur = build_norm(cur,
+                    model.layers[il].attn_sub_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_sub_norm", il);
+
+            cur = build_lora_mm(model.layers[il].wo, cur);
+            if (model.layers[il].wo_scale) {
+                cur = ggml_mul(ctx0, cur, model.layers[il].wo_scale);
+            }
+            if (model.layers[il].bo) {
+                cur = ggml_add(ctx0, cur, model.layers[il].bo);
+            }
+            cb(cur, "attn_out", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward forward
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, model.layers[il].ffn_up_scale,
+                model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_scale,
+                NULL,                      NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_sub_out", il);
+
+        cur = build_norm(cur,
+                model.layers[il].ffn_sub_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_sub_norm", il);
+
+        cur = build_lora_mm(model.layers[il].ffn_down, cur);
+        if (model.layers[il].ffn_down_scale) {
+            cur = ggml_mul(ctx0, cur, model.layers[il].ffn_down_scale);
+        }
+        cb(cur, "ffn_down", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    // FIXME: do not use model.tok_embd directly, duplicate as model.output
+    cur = build_lora_mm(model.tok_embd, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/bloom.cpp

@@ -0,0 +1,101 @@
+#include "models.h"
+
+llm_build_bloom::llm_build_bloom(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    inpL = build_norm(inpL,
+            model.tok_norm,
+            model.tok_norm_b,
+            LLM_NORM, -1);
+    cb(inpL, "inp_norm", -1);
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+            cb(cur, "bqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // Add the input
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    model.layers[il].ffn_norm_b,
+                    LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    NULL,                      NULL,                        NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = build_norm(inpL,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/chameleon.cpp

@@ -0,0 +1,178 @@
+#include "models.h"
+
+#include <float.h>
+
+llm_build_chameleon::llm_build_chameleon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        if (hparams.swin_norm) {
+            cur = inpL;
+        } else {
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+        }
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            if (model.layers[il].attn_q_norm) {
+                Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens,
+                        ggml_element_size(Qcur) * n_embd_head,
+                        ggml_element_size(Qcur) * n_embd_head * n_head,
+                        0);
+                cb(Qcur, "Qcur", il);
+
+                Qcur = build_norm(Qcur,
+                        model.layers[il].attn_q_norm,
+                        model.layers[il].attn_q_norm_b,
+                        LLM_NORM, il);
+                cb(Qcur, "Qcur", il);
+            }
+
+            if (model.layers[il].attn_k_norm) {
+                Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens,
+                        ggml_element_size(Kcur) * n_embd_head,
+                        ggml_element_size(Kcur) * n_embd_head * n_head_kv,
+                        0);
+                cb(Kcur, "Kcur", il);
+
+                Kcur = build_norm(Kcur,
+                        model.layers[il].attn_k_norm,
+                        model.layers[il].attn_k_norm_b,
+                        LLM_NORM, il);
+                cb(Kcur, "Kcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, nullptr,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        if (hparams.swin_norm) {
+            cur = build_norm(cur,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        if (!hparams.swin_norm) {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+        }
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        if (hparams.swin_norm) {
+            cur = build_norm(cur,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output_with_img_logits", -1);
+
+    // TODO: this suppresses the output of image tokens, which is required to enable text-only outputs.
+    // Needs to be removed once image outputs are supported.
+    int img_token_end_idx = 8196;
+    int img_token_start_idx = 4;
+    int num_img_tokens = img_token_end_idx - img_token_start_idx;
+    // creates 1d tensor of size num_img_tokens and values -FLT_MAX,
+    // which ensures that text token values are always at least larger than image token values
+    ggml_tensor * img_logits = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, num_img_tokens);
+    img_logits = ggml_clamp(ctx0, img_logits, -FLT_MAX, -FLT_MAX);
+    cb(img_logits, "img_logits", -1);
+
+    cur = ggml_set_1d(ctx0, cur, img_logits, ggml_element_size(cur) * img_token_start_idx);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/chatglm.cpp

@@ -0,0 +1,132 @@
+#include "models.h"
+
+
+llm_build_chatglm::llm_build_chatglm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = nullptr;
+            ggml_tensor * Kcur = nullptr;
+            ggml_tensor * Vcur = nullptr;
+
+            if (model.layers[il].wqkv == nullptr) {
+                Qcur = build_lora_mm(model.layers[il].wq, cur);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                }
+                Kcur = build_lora_mm(model.layers[il].wk, cur);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                }
+                Vcur = build_lora_mm(model.layers[il].wv, cur);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                }
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+            } else {
+                cur = build_lora_mm(model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+                if (model.layers[il].bqkv) {
+                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+                }
+                Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+                Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+                Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+            }
+
+            //printf("freq_base: %f freq_scale: %f ext_factor: %f attn_factor: %f\n", freq_base, freq_scale, ext_factor, attn_factor);
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        // Add the input
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    NULL,                      NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+
+        }
+
+        inpL = ggml_add(ctx0, cur, ffn_inp);
+        cb(inpL, "l_out", il);
+    }
+
+    cur = build_norm(inpL,
+            model.output_norm,
+            NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/codeshell.cpp

@@ -0,0 +1,111 @@
+#include "models.h"
+
+llm_build_codeshell::llm_build_codeshell(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+            cb(cur, "bqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // add the input
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    model.layers[il].ffn_norm_b,
+                    LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    NULL,                      NULL,                        NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = build_norm(inpL,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/cogvlm.cpp

@@ -0,0 +1,100 @@
+#include "models.h"
+
+llm_build_cogvlm::llm_build_cogvlm(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    float         kq_scale    = 1.0f / sqrtf(float(n_embd_head));
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor *inpL, *cur;
+    inpL = build_inp_embd(model.tok_embd);
+
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    // check ubatch to see if we have input tokens (text)
+    // or an input embedding vector (image)
+    bool is_text;
+    if (ubatch.token) {
+        is_text = true;
+    } else {
+        is_text = false;
+    }
+
+    for (int il = 0; il < n_layer; ++il) {
+        // get either the text or image weight tensors
+        ggml_tensor *wqkv, *wo;
+        ggml_tensor *ffn_gate, *ffn_down, *ffn_up;
+
+        if (is_text) {
+            wqkv     = model.layers[il].wqkv;
+            wo       = model.layers[il].wo;
+            ffn_gate = model.layers[il].ffn_gate;
+            ffn_down = model.layers[il].ffn_down;
+            ffn_up   = model.layers[il].ffn_up;
+        } else {
+            wqkv     = model.layers[il].visexp_attn_wqkv;
+            wo       = model.layers[il].visexp_attn_wo;
+            ffn_gate = model.layers[il].visexp_ffn_gate;
+            ffn_down = model.layers[il].visexp_ffn_down;
+            ffn_up   = model.layers[il].visexp_ffn_up;
+        }
+
+        ggml_tensor * inpSA = inpL;
+        cur                 = build_norm(inpSA, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+
+        // build self attention
+        {
+            ggml_tensor * qkv = build_lora_mm(wqkv, cur);
+
+            // split qkv into Q, K, V along the first dimension
+            ggml_tensor * Qcur =
+                ggml_view_3d(ctx0, qkv, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), qkv->nb[1], 0);
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                              qkv->nb[1], n_embd * ggml_element_size(qkv));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, qkv, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                              qkv->nb[1], 2 * n_embd * ggml_element_size(qkv));
+
+            Qcur = ggml_rope(ctx0, Qcur, inp_pos, n_embd_head, rope_type);
+            Kcur = ggml_rope(ctx0, Kcur, inp_pos, n_embd_head, rope_type);
+
+            cur = build_attn(inp_attn,
+                wo, nullptr,
+                Qcur, Kcur, Vcur,
+                nullptr, nullptr, nullptr,
+                kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                ffn_up, NULL, NULL,
+                ffn_gate, NULL, NULL,
+                ffn_down, NULL, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+    ggml_build_forward_expand(gf, cur);
+}

src/models/cohere2-iswa.cpp

@@ -0,0 +1,131 @@
+#include "models.h"
+
+llm_build_cohere2_iswa::llm_build_cohere2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    const float f_logit_scale = hparams.f_logit_scale;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv_iswa();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const bool is_swa = hparams.is_swa(il);
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+        ggml_tensor * ffn_inp = cur;
+
+        // self-attention
+        {
+            // rope freq factors for 128k context
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            if (is_swa) {
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+            }
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur     = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpL    = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
+        }
+
+        ggml_tensor * attn_out = cur;
+
+        // feed-forward network
+        {
+            cur = build_ffn(ffn_inp,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        // add together residual + FFN + self-attention
+        cur = ggml_add(ctx0, cur, inpL);
+        cur = ggml_add(ctx0, cur, attn_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    if (f_logit_scale) {
+        cur = ggml_scale(ctx0, cur, f_logit_scale);
+    }
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/command-r.cpp

@@ -0,0 +1,122 @@
+#include "models.h"
+
+
+
+llm_build_command_r::llm_build_command_r(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    const float f_logit_scale = hparams.f_logit_scale;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        ggml_tensor * ffn_inp = cur;
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            if (model.layers[il].attn_q_norm) {
+                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM, il);
+                cb(Qcur, "Qcur", il);
+            }
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            if (model.layers[il].attn_k_norm) {
+                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM, il);
+                cb(Kcur, "Kcur", il);
+            }
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur     = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpL    = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
+        }
+        ggml_tensor * attn_out = cur;
+
+        // feed-forward network
+        {
+            cur = build_ffn(ffn_inp,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        // add together residual + FFN + self-attention
+        cur = ggml_add(ctx0, cur, inpL);
+        cur = ggml_add(ctx0, cur, attn_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    if (f_logit_scale) {
+        cur = ggml_scale(ctx0, cur, f_logit_scale);
+    }
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/dbrx.cpp

@@ -0,0 +1,123 @@
+#include "models.h"
+
+
+llm_build_dbrx::llm_build_dbrx(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = nullptr;
+            ggml_tensor * Kcur = nullptr;
+            ggml_tensor * Vcur = nullptr;
+
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
+            cb(cur, "wqkv_clamped", il);
+
+            Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        // MoE branch
+        cur = build_norm(ffn_inp,
+                model.layers[il].attn_out_norm, NULL,
+                LLM_NORM, il);
+        cb(cur, "attn_out_norm", il);
+
+        cur = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, true,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+        cb(cur, "ffn_moe_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/deci.cpp

@@ -0,0 +1,135 @@
+#include "models.h"
+
+
+
+llm_build_deci::llm_build_deci(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA     = inpL;
+        const int64_t n_head_kv = hparams.n_head_kv(il);
+        const int64_t n_head    = hparams.n_head(il);
+        const int64_t n_ff      = hparams.n_ff(il);
+
+        if (n_head == 0) {
+            // attention-free layer of Llama-3_1-Nemotron-51B
+            cur = inpL;
+        } else {
+            // norm
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+        }
+        if (n_head > 0 && n_head_kv == 0) {
+            // "linear attention" of Llama-3_1-Nemotron-51B
+            cur = build_lora_mm(model.layers[il].wo, cur);
+            cb(cur, "wo", il);
+        } else if (n_head > 0) {
+            // self-attention
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        // FFN-free layer of Llama-3_1-Nemotron-Ultra-253B
+        if (n_ff == 0) {
+            continue;
+        }
+        // modified to support attention-free layer of Llama-3_1-Nemotron-51B
+        ggml_tensor * ffn_inp = cur;
+        if (n_head > 0) {
+            ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+        }
+        // feed-forward network
+        if (model.layers[il].ffn_gate_inp == nullptr) {
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/deepseek.cpp

@@ -0,0 +1,144 @@
+#include "models.h"
+
+
+
+llm_build_deepseek::llm_build_deepseek(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if ((uint32_t) il < hparams.n_layer_dense_lead) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, false,
+                false, hparams.expert_weights_scale,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // FFN shared expert
+            {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            }
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/deepseek2.cpp

@@ -0,0 +1,236 @@
+#include "models.h"
+
+
+
+llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    bool is_lite = (hparams.n_layer == 27);
+
+    const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
+
+    // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
+    const int64_t n_embd_head_k = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
+    const int64_t n_embd_head_v = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
+
+    const int64_t n_embd_head_qk_rope = hparams.n_rot;
+    const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
+
+    const uint32_t kv_lora_rank = hparams.n_lora_kv;
+
+    // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
+    // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
+    const float mscale      = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
+    const float kq_scale    = 1.0f * mscale * mscale / sqrtf(float(n_embd_head_k));
+    const float attn_factor = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    // {n_embd, n_tokens}
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            ggml_tensor * q = NULL;
+            if (!is_lite) {
+                q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
+                cb(q, "q", il);
+
+                q = build_norm(q, model.layers[il].attn_q_a_norm, nullptr, LLM_NORM_RMS, il);
+                cb(q, "q", il);
+
+                q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
+                cb(q, "q", il);
+            } else {
+                q = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+                cb(q, "q", il);
+            }
+            // split into {n_embd_head_qk_nope, n_head, n_tokens}
+            ggml_tensor * q_nope =
+                ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k),
+                             ggml_row_size(q->type, n_embd_head_k) * n_head, 0);
+            cb(q_nope, "q_nope", il);
+
+            // and {n_embd_head_qk_rope, n_head, n_tokens}
+            ggml_tensor * q_pe = ggml_view_3d(
+                ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k),
+                ggml_row_size(q->type, n_embd_head_k) * n_head, ggml_row_size(q->type, n_embd_head_qk_nope));
+            cb(q_pe, "q_pe", il);
+
+            ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
+            cb(kv_cmpr_pe, "kv_cmpr_pe", il);
+
+            // split into {kv_lora_rank, n_tokens}
+            ggml_tensor * kv_cmpr =
+                ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens,
+                             ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);
+            cb(kv_cmpr, "kv_cmpr", il);
+
+            // and {n_embd_head_qk_rope, 1, n_tokens}
+            ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens,
+                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
+            cb(k_pe, "k_pe", il);
+
+            q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+            cb(q_pe, "q_pe", il);
+
+            k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+            cb(k_pe, "k_pe", il);
+
+            kv_cmpr = build_norm(kv_cmpr, model.layers[il].attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
+            cb(kv_cmpr, "kv_cmpr", il);
+
+            if (is_mla) {
+                // {n_embd_head_qk_nope, n_tokens, n_head}
+                q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
+                cb(q_nope, "q_nope_perm", il);
+
+                // {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head}
+                ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);
+                cb(q_nope_absorbed, "q_nope_absorbed", il);
+
+                // {kv_lora_rank, n_head, n_tokens}
+                q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
+                cb(q_nope_absorbed, "q_nope_absorbed_perm", il);
+
+                // {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens}
+                // note: rope must go first for in-place context shifting in build_rope_shift()
+                ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope_absorbed, 0);
+                cb(Qcur, "Qcur", il);
+
+                kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
+                cb(kv_cmpr, "kv_cmpr_reshape", il);
+
+                // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}
+                ggml_tensor * Kcur = ggml_concat(ctx0, k_pe, kv_cmpr, 0);
+                cb(Kcur, "Kcur", il);
+
+                // {kv_lora_rank, 1, n_tokens}
+                ggml_tensor * Vcur = kv_cmpr;
+                cb(Vcur, "Vcur", il);
+
+                // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, NULL,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il);
+            } else {
+                ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_cmpr);
+                cb(kv, "kv", il);
+
+                // split into {n_embd_head_qk_nope, n_head, n_tokens}
+                ggml_tensor * k_nope =
+                    ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
+                                 ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),
+                                 ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head, 0);
+                cb(k_nope, "k_nope_view", il);
+
+                // and {n_embd_head_v, n_head, n_tokens}
+                ggml_tensor * Vcur = ggml_view_3d(ctx0, kv, n_embd_head_v, n_head, n_tokens,
+                                                  ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),
+                                                  ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head,
+                                                  ggml_row_size(kv->type, n_embd_head_qk_nope));
+                cb(Vcur, "Vcur_view", il);
+
+                Vcur = ggml_cont(ctx0, Vcur);
+                cb(Vcur, "Vcur_cont", il);
+
+                // note: rope must go first for in-place context shifting in build_rope_shift()
+                ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope, 0);
+                cb(Qcur, "Qcur", il);
+
+                ggml_tensor * Kcur = ggml_concat(ctx0, ggml_repeat(ctx0, k_pe, q_pe), k_nope, 0);
+                cb(Kcur, "Kcur", il);
+
+                // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
+                cur = build_attn(inp_attn,
+                            model.layers[il].wo, NULL,
+                            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            }
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if ((uint32_t) il < hparams.n_layer_dense_lead) {
+            cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                model.layers[il].ffn_exp_probs_b,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, hparams.expert_weights_norm,
+                true, hparams.expert_weights_scale,
+                (llama_expert_gating_func_type) hparams.expert_gating_func,
+                il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // FFN shared expert
+            {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            }
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = ggml_mul_mat(ctx0, model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/dots1.cpp

@@ -0,0 +1,134 @@
+#include "models.h"
+
+
+
+llm_build_dots1::llm_build_dots1(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // MoE branch
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if ((uint32_t) il < hparams.n_layer_dense_lead) {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                model.layers[il].ffn_exp_probs_b,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, hparams.expert_weights_norm,
+                true, hparams.expert_weights_scale,
+                (llama_expert_gating_func_type) hparams.expert_gating_func,
+                il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                cb(cur, "ffn_out", il);
+            }
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/dream.cpp

@@ -0,0 +1,105 @@
+#include "models.h"
+
+
+
+llm_build_dream::llm_build_dream(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    //copied from qwen2
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_no_cache();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            Qcur               = ggml_add(ctx0, Qcur, model.layers[il].bq);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            Kcur               = ggml_add(ctx0, Kcur, model.layers[il].bk);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            Vcur               = ggml_add(ctx0, Vcur, model.layers[il].bv);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+            model.layers[il].ffn_up, NULL, NULL,
+            model.layers[il].ffn_gate, NULL, NULL,
+            model.layers[il].ffn_down, NULL, NULL,
+            NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/ernie4-5-moe.cpp

@@ -0,0 +1,150 @@
+#include "models.h"
+
+
+
+llm_build_ernie4_5_moe::llm_build_ernie4_5_moe(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    GGML_ASSERT(hparams.n_moe_layer_step > 0 && "Ernie 4.5 MoE requires n_moe_layer_step > 0");
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+        // norm
+        {
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+        }
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+            cb(cur, "attn_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        bool is_moe_layer =
+            static_cast<uint32_t>(il) >= hparams.n_layer_dense_lead && (il + 1) % hparams.n_moe_layer_step == 0;
+
+        if (!is_moe_layer) {
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                                        model.layers[il].ffn_gate_inp,
+                                        model.layers[il].ffn_up_exps,
+                                        model.layers[il].ffn_gate_exps,
+                                        model.layers[il].ffn_down_exps,
+                                        model.layers[il].ffn_exp_probs_b,
+                                        n_expert, n_expert_used,
+                                        LLM_FFN_SILU, true,
+                                        false, 0.0,
+                                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                                        il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // Shared expert (if present)
+            if (hparams.n_ff_shexp > 0) {
+                ggml_tensor * ffn_shexp =
+                    build_ffn(cur,
+                        model.layers[il].ffn_up_shexp, NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(ffn_shexp, "ffn_shexp", il);
+
+                cur = ggml_add(ctx0, moe_out, ffn_shexp);
+            } else {
+                cur = moe_out;
+            }
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/ernie4-5.cpp

@@ -0,0 +1,111 @@
+#include "models.h"
+
+
+
+llm_build_ernie4_5::llm_build_ernie4_5(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        {
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+        }
+        // self-attention
+        {
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1) {
+            // skip computing output for unused tokens
+            ggml_tensor * inp_out_ids = build_inp_out_ids();
+            cur                       = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA                     = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        {
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/exaone.cpp

@@ -0,0 +1,114 @@
+#include "models.h"
+
+
+
+llm_build_exaone::llm_build_exaone(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/exaone4.cpp

@@ -0,0 +1,123 @@
+#include "models.h"
+
+
+template <bool iswa>
+llm_build_exaone4<iswa>::llm_build_exaone4(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_k;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    using inp_attn_type      = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
+    inp_attn_type * inp_attn = nullptr;
+
+    if constexpr (iswa) {
+        inp_attn = build_attn_inp_kv_iswa();
+    } else {
+        inp_attn = build_attn_inp_kv();
+    }
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // use RoPE for SWA layers or non-SWA models
+        const bool use_rope = hparams.is_swa(il) || hparams.swa_type == LLAMA_SWA_TYPE_NONE;
+
+        cur = inpL;
+
+        // self-attention
+        {
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+            cb(Kcur, "Kcur_normed", il);
+
+            if (use_rope) {
+                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,
+                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+
+                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,
+                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+            }
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+            cb(cur, "attn_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_ffn(ffn_inp,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL, NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, -1);
+        cb(cur, "ffn_post_norm", -1);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+// Explicit template instantiations
+template struct llm_build_exaone4<false>;
+template struct llm_build_exaone4<true>;

src/models/falcon-h1.cpp

@@ -0,0 +1,113 @@
+#include "models.h"
+
+
+
+llm_build_falcon_h1::llm_build_falcon_h1(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context_mamba(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // Build the inputs in the recurrent & kv cache
+    auto * inp = build_inp_mem_hybrid();
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+        cb(Qcur, "Qcur", il);
+
+        ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+        cb(Kcur, "Kcur", il);
+
+        ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+        cb(Vcur, "Vcur", il);
+
+        Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+        Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+
+        Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+        Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, hparams.rope_type, n_ctx_orig, freq_base, freq_scale,
+                             ext_factor, attn_factor, beta_fast, beta_slow);
+
+        Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, hparams.rope_type, n_ctx_orig, freq_base, freq_scale,
+                             ext_factor, attn_factor, beta_fast, beta_slow);
+
+        cb(Qcur, "Qcur-post-rope", il);
+        cb(Kcur, "Kcur-post-rope", il);
+        cb(Vcur, "Vcur-post-rope", il);
+
+        ggml_tensor * attn_out = build_attn(inp->get_attn(),
+                                    model.layers[il].wo, NULL,
+                                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+        cb(attn_out, "attn_out", il);
+
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        // Mamba2 layer
+        cb(cur, "ssm_in", il);
+
+        ggml_tensor * ssm_out = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il);
+        cb(ssm_out, "ssm_out", il);
+
+        // // Aggregation
+        cur   = ggml_add(ctx0, attn_out, ssm_out);
+        inpSA = ggml_add(ctx0, cur, inpSA);
+        cb(cur, "layer_out", il);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = inpSA;
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, inpSA);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/falcon.cpp

@@ -0,0 +1,120 @@
+#include "models.h"
+
+
+llm_build_falcon::llm_build_falcon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * attn_norm;
+
+        attn_norm = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(attn_norm, "attn_norm", il);
+
+        // self-attention
+        {
+            if (model.layers[il].attn_norm_2) {
+                // Falcon-40B
+                cur = build_norm(inpL,
+                        model.layers[il].attn_norm_2,
+                        model.layers[il].attn_norm_2_b,
+                        LLM_NORM, il);
+                cb(cur, "attn_norm_2", il);
+            } else {
+                cur = attn_norm;
+            }
+
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            // using mode = 2 for neox mode
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur       = ggml_get_rows(ctx0,       cur, inp_out_ids);
+            inpL      = ggml_get_rows(ctx0,      inpL, inp_out_ids);
+            attn_norm = ggml_get_rows(ctx0, attn_norm, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = cur;
+
+        // feed forward
+        {
+            cur = build_ffn(attn_norm, // !! use the attn norm, not the result
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    NULL,                      NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cur = ggml_add(ctx0, cur, inpL);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    // norm
+    cur = build_norm(cur,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/gemma-embedding.cpp

@@ -0,0 +1,120 @@
+#include "models.h"
+
+
+
+llm_build_gemma_embedding::llm_build_gemma_embedding(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_k;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
+    if (ubatch.token) {
+        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+        cb(inpL, "inp_scaled", -1);
+    }
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_no_cache();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        const float freq_base_l  = model.get_rope_freq_base(cparams, il);
+        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315
+            Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
+
+            cur =
+                build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+        cb(sa_out, "sa_out", il);
+
+        cur = build_norm(sa_out, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        {
+            cur = build_ffn(cur,
+                model.layers[il].ffn_up, NULL, NULL,
+                model.layers[il].ffn_gate, NULL, NULL,
+                model.layers[il].ffn_down, NULL, NULL,
+                NULL, LLM_FFN_GELU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, -1);
+        cb(cur, "ffn_post_norm", -1);
+
+        cur = ggml_add(ctx0, cur, sa_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/gemma.cpp

@@ -0,0 +1,112 @@
+#include "models.h"
+
+
+llm_build_gemma::llm_build_gemma(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+    cb(inpL, "inp_scaled", -1);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head)));
+            cb(Qcur, "Qcur_scaled", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+        cb(sa_out, "sa_out", il);
+
+        cur = build_norm(sa_out,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        {
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, sa_out);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/gemma2-iswa.cpp

@@ -0,0 +1,125 @@
+#include "models.h"
+
+llm_build_gemma2_iswa::llm_build_gemma2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_k;
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+        cb(inpL, "inp_scaled", -1);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv_iswa();
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+        for (int il = 0; il < n_layer; ++il) {
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
+
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, NULL,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+            }
+            if (il == n_layer - 1 && inp_out_ids) {
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            }
+            cur = build_norm(cur,
+                    model.layers[il].attn_post_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_post_norm", il);
+
+            ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+            cb(sa_out, "sa_out", il);
+
+            cur = build_norm(sa_out,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            // feed-forward network
+            {
+                cur = build_ffn(cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
+            }
+            cur = build_norm(cur,
+                    model.layers[il].ffn_post_norm, NULL,
+                    LLM_NORM_RMS, -1);
+            cb(cur, "ffn_post_norm", -1);
+
+            cur = ggml_add(ctx0, cur, sa_out);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        // final logit soft-capping
+        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
+        cur = ggml_tanh(ctx0, cur);
+        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+}

src/models/gemma3-iswa.cpp

@@ -0,0 +1,131 @@
+#include "models.h"
+
+llm_build_gemma3_iswa::llm_build_gemma3_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_k;
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
+        if (ubatch.token) {
+            inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+            cb(inpL, "inp_scaled", -1);
+        }
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        // TODO: is causal == true correct? might need some changes
+        auto * inp_attn = build_attn_inp_kv_iswa();
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+        for (int il = 0; il < n_layer; ++il) {
+            const float freq_base_l  = model.get_rope_freq_base (cparams, il);
+            const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
+            // norm
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+                cb(Qcur, "Qcur_normed", il);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+
+                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+                cb(Kcur, "Kcur_normed", il);
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315
+                Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
+
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, NULL,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+            }
+            if (il == n_layer - 1 && inp_out_ids) {
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            }
+            cur = build_norm(cur,
+                    model.layers[il].attn_post_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_post_norm", il);
+
+            ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+            cb(sa_out, "sa_out", il);
+
+            cur = build_norm(sa_out,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            // feed-forward network
+            {
+                cur = build_ffn(cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
+            }
+            cur = build_norm(cur,
+                    model.layers[il].ffn_post_norm, NULL,
+                    LLM_NORM_RMS, -1);
+            cb(cur, "ffn_post_norm", -1);
+
+            cur = ggml_add(ctx0, cur, sa_out);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+}

src/models/gemma3n-iswa.cpp

@@ -0,0 +1,377 @@
+#include "models.h"
+
+
+
+llm_build_gemma3n_iswa::llm_build_gemma3n_iswa(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params),
+    model(model),
+    n_embd_head(model.hparams.n_embd_head_k),
+    n_embd_altup(model.hparams.n_embd_altup),
+    n_altup(model.hparams.n_altup),
+    i_altup_act(model.hparams.i_altup_act) {
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
+    if (ubatch.token) {
+        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+        cb(inpL, "inp_scaled", -1);
+    }
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // TODO: is causal == true correct? might need some changes
+    auto * inp_attn = build_attn_inp_kv_iswa();
+
+    // inp_per_layer shape: [n_embd_altup, n_tokens, n_layer]
+    ggml_tensor * inp_per_layer = project_per_layer_inputs(inpL, get_per_layer_inputs());
+
+    // inpL now has only 1 altup, project it to the rest of the altups
+    // these "added" altups will be concat to the last dim of inpL
+    {
+        ggml_tensor * target_magnitude = calc_magnitude(inpL);
+        ggml_tensor * inp_repeated     = ggml_repeat_4d(ctx0, inpL, n_embd, n_tokens, n_altup - 1, 1);
+        ggml_tensor * altup_added =
+            ggml_mul_mat(ctx0, model.altup_proj, inp_repeated);  // shape: [n_embd, n_tokens, n_altup - 1]
+        ggml_tensor * new_magnitude = calc_magnitude(altup_added);
+        altup_added                 = ggml_div(ctx0, ggml_mul(ctx0, altup_added, target_magnitude), new_magnitude);
+        inpL                        = ggml_concat(ctx0, inpL, altup_added, 2);  // shape: [n_embd, n_tokens, n_altup]
+        cb(inpL, "inp_stacked", -1);
+    }
+    // inpL now has shape:          [n_embd,       n_tokens, n_altup]
+    // inp_per_layer now has shape: [n_embd_altup, n_tokens, n_layer]
+
+    for (int il = 0; il < n_layer; ++il) {
+        // this block is made to be closely resemble Gemma3p5DecoderLayer on python code
+        const float freq_base_l  = model.get_rope_freq_base(cparams, il);
+        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
+        ggml_tensor * cur         = inpL;                    // [n_embd, n_tokens, n_altup]
+        ggml_tensor * predictions = altup_predict(cur, il);  // [n_embd, n_tokens, n_altup]
+
+        // predicted value will go through self-attention and laurel
+        ggml_tensor * active_prediction = view_2d_slice(predictions, i_altup_act);  // [n_embd, n_tokens]
+        cur                             = active_prediction;
+        cb(cur, "active_prediction", il);
+
+        // norm
+        cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // laurel
+        ggml_tensor * laurel_out = laurel(cur, il);  // [n_embd, n_tokens]
+
+        // self-attention
+        if (hparams.has_kv(il)) {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            Vcur = ggml_rms_norm(ctx0, Vcur, hparams.f_norm_rms_eps);
+
+            cb(Qcur, "Qcur_normed", il);
+            cb(Kcur, "Kcur_normed", il);
+            cb(Vcur, "Vcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur_pos", il);
+            cb(Kcur, "Kcur_pos", il);
+
+            cur = build_attn(inp_attn, model.layers[il].wo,
+                    NULL, Qcur, Kcur, Vcur, nullptr, nullptr, nullptr,
+                    hparams.f_attention_scale, il);
+        } else {
+            // reuse KV cache of earlier layers
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+            cb(Qcur, "Qcur_pos", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, nullptr, nullptr, nullptr, nullptr, nullptr, hparams.f_attention_scale, il);
+        }
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_post_norm", il);
+
+        cur = ggml_add(ctx0, cur, active_prediction);  // [n_embd, n_tokens]
+        cb(cur, "attn_gated", il);
+
+        ggml_tensor * attn_laurel = ggml_scale(ctx0, ggml_add(ctx0, cur, laurel_out),
+                                               1.0f / sqrtf(2.0f));  // [n_embd, n_tokens]
+        cb(attn_laurel, "attn_laurel", il);
+
+        cur = build_norm(attn_laurel, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        {
+            ggml_tensor * up_proj   = build_lora_mm(model.layers[il].ffn_up, cur);
+            ggml_tensor * gate_proj = build_lora_mm(model.layers[il].ffn_gate, cur);
+
+            if (il < n_layer_sparsity) {
+                // apply activation sparsity
+                gate_proj = gaussian_topk(gate_proj);
+            }
+            gate_proj = ggml_gelu(ctx0, gate_proj);
+
+            cur = ggml_mul(ctx0, up_proj, gate_proj);
+            cur = build_lora_mm(model.layers[il].ffn_down, cur);
+            cb(cur, "ffn_out", il);
+        }
+        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, -1);
+        cb(cur, "ffn_post_norm", il);
+
+        ggml_tensor * attn_ffw_laurel_gated = ggml_add(ctx0, cur, attn_laurel);  // [n_embd, n_tokens]
+        cb(attn_ffw_laurel_gated, "attn_ffw_laurel_gated", il);
+
+        ggml_tensor * corrected = altup_correct(predictions, attn_ffw_laurel_gated, il);  // [n_embd, n_tokens, n_altup]
+
+        ggml_tensor * first_prediction;                                                   // [n_embd, n_tokens]
+        {
+            first_prediction = view_2d_slice(corrected, i_altup_act);                     // [n_embd, n_tokens]
+            first_prediction = ggml_mul(ctx0, first_prediction, model.layers[il].altup_correct_scale);
+            first_prediction = build_lora_mm(model.layers[il].per_layer_inp_gate, first_prediction);
+            first_prediction = ggml_gelu(ctx0, first_prediction);                 // [n_embd_altup, n_tokens]
+            cb(first_prediction, "first_prediction_gated", il);
+            ggml_tensor * inp_this_layer = view_2d_slice(inp_per_layer, il);      // [n_embd_altup, n_tokens]
+            first_prediction = ggml_mul(ctx0, first_prediction, inp_this_layer);  // [n_embd_altup, n_tokens]
+            cb(first_prediction, "first_prediction_scaled", il);
+
+            first_prediction = build_lora_mm(model.layers[il].per_layer_proj, first_prediction);  // [n_embd, n_tokens]
+            first_prediction =
+                build_norm(first_prediction, model.layers[il].per_layer_post_norm, NULL, LLM_NORM_RMS, il);
+            cb(first_prediction, "first_prediction_out", il);
+        }
+        // equivalent to python code: corrected_predictions[1:] += first_prediction
+        {
+            ggml_tensor * slice_first = view_2d_slice(corrected, 0);
+            ggml_tensor * slice_rest  = ggml_view_3d(
+                ctx0, corrected, n_embd, n_tokens, n_altup - 1, ggml_row_size(corrected->type, n_embd),
+                ggml_row_size(corrected->type, n_embd * n_tokens), n_embd * n_tokens * ggml_element_size(corrected));
+            ggml_tensor * tmp = ggml_add(ctx0, slice_rest, first_prediction);  // [n_embd, n_tokens, n_altup - 1]
+            corrected         = ggml_concat(ctx0, slice_first, tmp, 2);        // [n_embd, n_tokens, n_altup]
+        }
+        cur = corrected;                                                       // [n_embd, n_tokens, n_altup]
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;  // [n_embd, n_tokens, n_altup]
+
+    // cur now has multiple altup(s), we want to merge them back to 1 altup
+    {
+        ggml_tensor * target_magnitude = calc_magnitude(view_2d_slice(cur, i_altup_act));  // [n_embd, n_tokens]
+        // do a view to skip the first slice (active altup)
+        ggml_tensor * alt_slice =
+            ggml_view_3d(ctx0, cur, n_embd, n_tokens, n_altup - 1, ggml_row_size(cur->type, n_embd),
+                         ggml_row_size(cur->type, n_embd * n_tokens), n_embd * n_tokens * ggml_element_size(cur));
+        ggml_tensor * altup_unembd =
+            ggml_mul_mat(ctx0, model.altup_unembd_proj, alt_slice);  // shape: [n_embd, n_tokens, n_altup - 1]
+        ggml_tensor * new_magnitude = calc_magnitude(altup_unembd);
+        altup_unembd                = ggml_div(ctx0, ggml_mul(ctx0, altup_unembd, target_magnitude), new_magnitude);
+        cb(altup_unembd, "altup_unembd", -1);
+
+        // equivalent to torch.mean(hidden_states, dim=0)
+        cur = view_2d_slice(cur, 0);  // [n_embd, n_tokens]
+        for (int i = 0; i < n_altup - 1; ++i) {
+            cur = ggml_add(ctx0, cur, view_2d_slice(altup_unembd, i));
+        }
+        cur = ggml_scale(ctx0, cur, 1.0f / float(n_altup));  // [n_embd, n_tokens]
+        cb(cur, "unembd_merged", -1);
+    }
+    // cur now has shape: [n_embd, n_tokens]
+
+    // TODO: move this to right after the last KV layer
+    {
+        // skip computing output for unused tokens
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+        cur                       = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    {
+        // final logit soft-capping
+        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
+        cur = ggml_tanh(ctx0, cur);
+        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
+    }
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+ggml_tensor * llm_build_gemma3n_iswa::calc_magnitude(ggml_tensor * x) {
+    return ggml_sqrt(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, x)));
+}
+
+// get 2D slice view from a 3D tensor, the idx corresponds to the 3rd dim
+ggml_tensor * llm_build_gemma3n_iswa::view_2d_slice(ggml_tensor * x, int idx) {
+    GGML_ASSERT(idx < (int) x->ne[2]);
+    return ggml_view_2d(ctx0, x, x->ne[0], x->ne[1], ggml_row_size(x->type, x->ne[0]),
+                        idx * x->ne[0] * x->ne[1] * ggml_element_size(x));
+}
+
+// equivalent to get_per_layer_inputs() in python code
+// output shape: [n_embd_altup, n_layer, n_tokens]
+ggml_tensor * llm_build_gemma3n_iswa::get_per_layer_inputs() {
+    auto          inp = std::make_unique<llm_graph_input_embd>();
+    ggml_tensor * inp_per_layer;
+    if (ubatch.token) {
+        inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
+        ggml_set_input(inp->tokens);
+        res->t_tokens = inp->tokens;
+        inp_per_layer = ggml_get_rows(ctx0, model.tok_embd_per_layer, inp->tokens);
+        inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_altup, n_layer, n_tokens);
+        inp_per_layer = ggml_scale(ctx0, inp_per_layer, sqrtf((float) n_embd_altup));
+        cb(inp_per_layer, "inp_per_layer_selected", -1);
+    } else {
+        GGML_ABORT("TODO: support embd input");
+    }
+    res->add_input(std::move(inp));
+    return inp_per_layer;
+}
+
+// equivalent to project_per_layer_inputs() in python code
+// this calculates the per-layer inputs, so the final tensor shape will have n_layer as the last dim
+// output shape: [n_embd_altup, n_tokens, n_layer]
+ggml_tensor * llm_build_gemma3n_iswa::project_per_layer_inputs(ggml_tensor * inputs_embeds, ggml_tensor * inp_per_layer) {
+    const float per_layer_projection_scale = 1.0f / sqrtf((float) n_embd);
+    const float per_layer_input_scale      = 1.0f / sqrtf(2.0f);
+
+    ggml_tensor * per_layer_proj = ggml_mul_mat(ctx0, model.per_layer_model_proj, inputs_embeds);
+    per_layer_proj               = ggml_scale(ctx0, per_layer_proj, per_layer_projection_scale);
+    per_layer_proj               = ggml_reshape_3d(ctx0, per_layer_proj, n_embd_altup, n_layer, n_tokens);
+    per_layer_proj               = build_norm(per_layer_proj, model.per_layer_proj_norm, NULL, LLM_NORM_RMS,
+                                              -1);  // [n_embd_altup, n_layer, n_tokens]
+    cb(per_layer_proj, "per_layer_proj", -1);
+
+    inp_per_layer = ggml_add(ctx0, inp_per_layer, per_layer_proj);
+    inp_per_layer = ggml_scale(ctx0, inp_per_layer, per_layer_input_scale);
+    cb(inp_per_layer, "inp_per_layer", -1);
+
+    // permute to shape: [n_embd_altup, n_tokens, n_layer]
+    inp_per_layer = ggml_cont(ctx0, ggml_permute(ctx0, inp_per_layer, 0, 2, 1, 3));
+    return inp_per_layer;
+}
+
+// input cur shape: [n_altup, n_tokens]
+// output    shape: [n_altup, n_tokens]
+ggml_tensor * llm_build_gemma3n_iswa::laurel(ggml_tensor * cur, int il) {
+    ggml_tensor * tmp = cur;
+    tmp               = build_lora_mm(model.layers[il].laurel_l, tmp);
+    tmp               = build_lora_mm(model.layers[il].laurel_r, tmp);
+    tmp               = build_norm(tmp, model.layers[il].laurel_post_norm, NULL, LLM_NORM_RMS, il);
+    tmp               = ggml_add(ctx0, tmp, cur);
+    cb(tmp, "laurel_out", il);
+    return tmp;
+}
+
+// input x shape: [n_embd, n_tokens]
+// output  shape: [n_embd, n_tokens]
+ggml_tensor * llm_build_gemma3n_iswa::gaussian_topk(ggml_tensor * x) {
+    ggml_tensor * mean = ggml_mean(ctx0, x);
+    ggml_tensor * std  = ggml_sqrt(ctx0, ggml_scale(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, ggml_sub(ctx0, x, mean))),
+                                                    1.0f / (float) (x->ne[0] - 1)));
+    ggml_tensor * cutoff_x = ggml_add(ctx0, mean, ggml_scale(ctx0, std, f_sparsity_std_mul));
+    return ggml_relu(ctx0, ggml_sub(ctx0, x, cutoff_x));
+}
+
+//
+// altup functions
+//
+
+// equivalent to compute_router_modalities() in python code
+// input x shape: [n_embd,  n_tokens]
+// output  shape: [n_altup, n_tokens]
+ggml_tensor * llm_build_gemma3n_iswa::altup_compute_router_modalities(ggml_tensor * x, int il) {
+    ggml_tensor * router_inputs = build_norm(x, model.layers[il].altup_router_norm, NULL, LLM_NORM_RMS, il);
+
+    // router_input_scale
+    router_inputs = ggml_scale(ctx0, router_inputs, 1.0f / (float) n_embd);
+
+    ggml_tensor * output = ggml_mul_mat(ctx0, model.layers[il].altup_router, router_inputs);
+    return ggml_tanh(ctx0, output);  // [n_altup, n_tokens]
+}
+
+// input cur shape: [n_embd, n_tokens, n_altup]
+// output    shape: [n_embd, n_tokens, n_altup]
+ggml_tensor * llm_build_gemma3n_iswa::altup_predict(ggml_tensor * cur, int il) {
+    ggml_tensor * activated  = view_2d_slice(cur, i_altup_act);                 // [n_embd, n_tokens]
+    ggml_tensor * modalities = altup_compute_router_modalities(activated, il);  // [n_altup, n_tokens]
+    cb(modalities, "modalities", il);
+
+    ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_predict_coef, modalities);
+    cb(all_coefs, "all_coefs", il);
+    // first dim now having n_altup^2 elements, we reshape it to 2D (so we end up with 3D tensor)
+    all_coefs = ggml_reshape_3d(ctx0, all_coefs, n_altup, n_altup, n_tokens);
+
+    // permute to [n_altup, n_embd, n_tokens]
+    ggml_tensor * cur_permuted = ggml_cont(ctx0, ggml_permute(ctx0, cur, 1, 2, 0, 3));
+    ggml_tensor * predictions  = ggml_mul_mat(ctx0, cur_permuted, all_coefs);  // [n_altup, n_embd, n_tokens]
+
+    // final shape must be the same as cur: [n_embd, n_tokens, n_altup]
+    predictions = ggml_cont(ctx0, ggml_permute(ctx0, predictions, 0, 2, 1, 3));
+    predictions = ggml_add(ctx0, predictions, cur);
+    cb(predictions, "predictions", il);
+
+    return predictions;
+}
+
+// input predictions       shape: [n_embd, n_tokens, n_altup]
+// input activated         shape: [n_embd, n_tokens]
+// output                  shape: [n_embd, n_tokens, n_altup]
+ggml_tensor * llm_build_gemma3n_iswa::altup_correct(ggml_tensor * predictions, ggml_tensor * activated, int il) {
+    ggml_tensor * modalities = altup_compute_router_modalities(activated, il);  // [n_altup, n_tokens]
+    cb(modalities, "modalities", il);
+
+    ggml_tensor * active_prediction = view_2d_slice(predictions, i_altup_act);
+    ggml_tensor * innovation        = ggml_sub(ctx0, activated, active_prediction);  // [n_embd, n_tokens]
+    cb(innovation, "innovation", il);
+
+    ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_correct_coef, modalities);  // [n_altup, n_tokens]
+    all_coefs               = ggml_scale_bias(ctx0, all_coefs, 1.0f, 1.0f);                    // + 1.0
+    cb(all_coefs, "all_coefs", il);
+    all_coefs = ggml_transpose(ctx0, all_coefs);                                               // [n_tokens, n_altup]
+    all_coefs = ggml_cont_3d(ctx0, all_coefs, 1, n_tokens, n_altup);                           // [1, n_tokens, n_altup]
+
+    innovation              = ggml_repeat_4d(ctx0, innovation, n_embd, n_tokens, n_altup, 1);
+    ggml_tensor * corrected = ggml_mul(ctx0, innovation, all_coefs);   // [n_embd, n_tokens, n_altup]
+    corrected               = ggml_add(ctx0, corrected, predictions);  // [n_embd, n_tokens, n_altup]
+    cb(corrected, "corrected", il);
+
+    return corrected;
+}

src/models/glm4-moe.cpp

@@ -0,0 +1,153 @@
+#include "models.h"
+
+llm_build_glm4_moe::llm_build_glm4_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv();
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+        // Only process up to last layer (skip final NextN layer)
+        // Final layer tensors are loaded but not processed in forward pass
+        const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
+        for (int il = 0; il < n_transformer_layers; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // Pre-attention norm
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                }
+                cb(Qcur, "Qcur", il);
+
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                }
+                cb(Kcur, "Kcur", il);
+
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                }
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                // Apply Q/K norm if available (GLM-4.5 355B variant)
+                if (model.layers[il].attn_q_norm) {
+                    Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+                    cb(Qcur, "Qcur_normed", il);
+                }
+                if (model.layers[il].attn_k_norm) {
+                    Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+                    cb(Kcur, "Kcur_normed", il);
+                }
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, NULL,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+            }
+            if (il == n_transformer_layers - 1 && inp_out_ids) {
+                cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // Post-attention norm
+            cur = build_norm(ffn_inp, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "post_attn_norm", il);
+
+            // Check if this is a dense layer (n_layer_dense_lead=1, so layer 0 is dense)
+            if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {
+                // Dense FFN layer
+                cur = build_ffn(cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
+            } else {
+                // Process routed experts using existing MoE infrastructure
+                ggml_tensor * routed_out = build_moe_ffn(cur,
+                        model.layers[il].ffn_gate_inp,
+                        model.layers[il].ffn_up_exps,
+                        model.layers[il].ffn_gate_exps,
+                        model.layers[il].ffn_down_exps,
+                        model.layers[il].ffn_exp_probs_b,
+                        n_expert, n_expert_used,
+                        LLM_FFN_SILU, hparams.expert_weights_norm,
+                        true, hparams.expert_weights_scale,
+                        (llama_expert_gating_func_type) hparams.expert_gating_func,
+                        il);
+                cb(routed_out, "ffn_moe_out", il);
+
+                // Process shared expert on original input
+                ggml_tensor * shared_out = build_ffn(cur,
+                        model.layers[il].ffn_up_shexp,   NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(shared_out, "ffn_shexp_out", il);
+
+                // Final output: routed_output + shared_output
+                cur = ggml_add(ctx0, routed_out, shared_out);
+                cb(cur, "ffn_out", il);
+            }
+            cur = ggml_add(ctx0, cur, ffn_inp);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+        cur = inpL;
+        cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+}

src/models/glm4.cpp

@@ -0,0 +1,127 @@
+#include "models.h"
+
+
+
+llm_build_glm4::llm_build_glm4(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // Pre-attention norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            ggml_tensor * Qcur = nullptr;
+            ggml_tensor * Kcur = nullptr;
+            ggml_tensor * Vcur = nullptr;
+
+            if (model.layers[il].wqkv == nullptr) {
+                Qcur = build_lora_mm(model.layers[il].wq, cur);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                }
+                Kcur = build_lora_mm(model.layers[il].wk, cur);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                }
+                Vcur = build_lora_mm(model.layers[il].wv, cur);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                }
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+            } else {
+                cur = build_lora_mm(model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+                if (model.layers[il].bqkv) {
+                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+                }
+                Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), cur->nb[1],
+                                    0 * sizeof(float) * (n_embd));
+                Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                    cur->nb[1], 1 * sizeof(float) * (n_embd));
+                Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
+                                    cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
+            }
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        // Post-attention norm (new!)
+        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "post_attn_norm", il);
+
+        // Add the input (residual connection after post-attention norm)
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            // Pre-MLP norm
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            // MLP
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    NULL, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+
+            // Post-MLP norm
+            cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "post_mlp_norm", il);
+        }
+        // Add residual connection after post-MLP norm
+        inpL = ggml_add(ctx0, cur, ffn_inp);
+        cb(inpL, "l_out", il);
+    }
+    // Final norm
+    cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // Output projection
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/gpt2.cpp

@@ -0,0 +1,105 @@
+#include "models.h"
+
+llm_build_gpt2::llm_build_gpt2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * pos;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
+    cb(pos, "pos_embd", -1);
+
+    inpL = ggml_add(ctx0, inpL, pos);
+    cb(inpL, "inpL", -1);
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+            cb(cur, "bqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // add the input
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // FF
+        {
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    model.layers[il].ffn_norm_b,
+                    LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    NULL,                      NULL,                        NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = build_norm(inpL,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/gptneox.cpp

@@ -0,0 +1,144 @@
+#include "models.h"
+
+
+llm_build_gptneox::llm_build_gptneox(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            cur = build_lora_mm(model.layers[il].wqkv, cur);
+            cb(cur, "wqkv", il);
+
+            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+            cb(cur, "bqkv", il);
+
+            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // ffn
+        if (hparams.use_par_res) {
+            // attention and ffn are computed in parallel
+            // x = x + attn(ln1(x)) + ffn(ln2(x))
+
+            ggml_tensor * attn_out = cur;
+
+            cur = build_norm(inpL,
+                    model.layers[il].ffn_norm,
+                    model.layers[il].ffn_norm_b,
+                    LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    NULL,                      NULL,                        NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, inpL);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, attn_out);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        } else {
+            // attention and ffn are computed sequentially
+            // x = x + attn(ln1(x))
+            // x = x + ffn(ln2(x))
+
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+            cb(ffn_inp, "ffn_inp", il);
+
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm,
+                    model.layers[il].ffn_norm_b,
+                    LLM_NORM, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    NULL,                      NULL,                        NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+    }
+
+    cur = build_norm(inpL,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/granite-hybrid.cpp

@@ -0,0 +1,196 @@
+#include "models.h"
+
+
+llm_build_granite_hybrid::llm_build_granite_hybrid(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context_mamba(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    auto * inp = build_inp_mem_hybrid();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    // Positional embeddings populated if rope enabled
+    ggml_tensor * inp_pos = nullptr;
+    if (hparams.rope_finetuned) {
+        inp_pos = build_inp_pos();
+    }
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        if (hparams.is_recurrent(il)) {
+            // ssm layer //
+            cur = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il);
+        } else {
+            // attention layer //
+            cur = build_attention_layer(cur, inp_pos, inp->get_attn(), model, n_embd_head, il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        // ffn
+        cur = build_layer_ffn(cur, inpSA, model, il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    // For Granite architectures - scale logits
+    if (hparams.f_logit_scale) {
+        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
+    }
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+ggml_tensor * llm_build_granite_hybrid::build_attention_layer(ggml_tensor *             cur,
+                                                              ggml_tensor *             inp_pos,
+                                                              llm_graph_input_attn_kv * inp_attn,
+                                                              const llama_model &       model,
+                                                              const int64_t             n_embd_head,
+                                                              const int                 il) {
+    // compute Q and K and (optionally) RoPE them
+    ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+    cb(Qcur, "Qcur", il);
+    if (model.layers[il].bq) {
+        Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+        cb(Qcur, "Qcur", il);
+    }
+
+    ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+    cb(Kcur, "Kcur", il);
+    if (model.layers[il].bk) {
+        Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+        cb(Kcur, "Kcur", il);
+    }
+
+    ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+    cb(Vcur, "Vcur", il);
+    if (model.layers[il].bv) {
+        Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+        cb(Vcur, "Vcur", il);
+    }
+
+    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il), n_tokens);
+    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+
+    const bool use_rope = hparams.rope_finetuned;
+    if (use_rope) {
+        ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+        Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                             ext_factor, attn_factor, beta_fast, beta_slow);
+
+        Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                             ext_factor, attn_factor, beta_fast, beta_slow);
+    }
+
+    cb(Qcur, "Qcur", il);
+    cb(Kcur, "Kcur", il);
+    cb(Vcur, "Vcur", il);
+
+    const float kq_scale =
+        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+    cur = build_attn(inp_attn,
+            model.layers[il].wo, model.layers[il].bo,
+            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+    cb(cur, "attn_out", il);
+    return cur;
+}
+
+ggml_tensor * llm_build_granite_hybrid::build_layer_ffn(ggml_tensor *       cur,
+                                                        ggml_tensor *       inpSA,
+                                                        const llama_model & model,
+                                                        const int           il) {
+    // For Granite architectures - scale residual
+    if (hparams.f_residual_scale) {
+        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+    }
+    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+    cb(ffn_inp, "ffn_inp", il);
+
+    // feed-forward network (non-MoE)
+    if (model.layers[il].ffn_gate_inp == nullptr) {
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
+                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+    } else {
+        // MoE branch
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        ggml_tensor * moe_out =
+            build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, true,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+        cb(moe_out, "ffn_moe_out", il);
+
+        // For Granite MoE Shared
+        if (hparams.n_ff_shexp > 0) {
+            ggml_tensor * ffn_shexp =
+                build_ffn(cur,
+                    model.layers[il].ffn_up_shexp, NULL, NULL,
+                    model.layers[il].ffn_gate_shexp, NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(ffn_shexp, "ffn_shexp", il);
+
+            cur = ggml_add(ctx0, moe_out, ffn_shexp);
+            cb(cur, "ffn_out", il);
+        } else {
+            cur = moe_out;
+        }
+    }
+
+    // For Granite architectures - scale residual
+    if (hparams.f_residual_scale) {
+        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+    }
+    cur = ggml_add(ctx0, cur, ffn_inp);
+    cb(cur, "ffn_out", il);
+
+    cur = build_cvec(cur, il);
+    cb(cur, "l_out", il);
+
+    return cur;
+}

src/models/granite.cpp

@@ -0,0 +1,211 @@
+#include "models.h"
+
+
+llm_build_granite::llm_build_granite(
+    const llama_model & model,
+    const llm_graph_params & params)
+    : llm_graph_context(params) {
+
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - built only if rope enabled
+    ggml_tensor * inp_pos = nullptr;
+    if (hparams.rope_finetuned) {
+        inp_pos = build_inp_pos();
+    }
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        cur = build_attention_layer(
+            cur, inp_pos, inp_attn,
+            model, n_embd_head, il);
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        // ffn
+        cur = build_layer_ffn(cur, inpSA, model, il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    // For Granite architectures - scale logits
+    cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
+
+ggml_tensor * llm_build_granite::build_attention_layer(
+          ggml_tensor             * cur,
+          ggml_tensor             * inp_pos,
+          llm_graph_input_attn_kv * inp_attn,
+    const llama_model             & model,
+    const int64_t                 n_embd_head,
+    const int                     il) {
+
+    // compute Q and K and (optionally) RoPE them
+    ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+    cb(Qcur, "Qcur", il);
+    if (model.layers[il].bq) {
+        Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+        cb(Qcur, "Qcur", il);
+    }
+
+    ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+    cb(Kcur, "Kcur", il);
+    if (model.layers[il].bk) {
+        Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+        cb(Kcur, "Kcur", il);
+    }
+
+    ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+    cb(Vcur, "Vcur", il);
+    if (model.layers[il].bv) {
+        Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+        cb(Vcur, "Vcur", il);
+    }
+
+    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il),    n_tokens);
+    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
+
+    const bool use_rope = hparams.rope_finetuned;
+    if (use_rope) {
+        ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+        Qcur = ggml_rope_ext(
+                ctx0, Qcur, inp_pos, rope_factors,
+                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                ext_factor, attn_factor, beta_fast, beta_slow
+                );
+
+        Kcur = ggml_rope_ext(
+                ctx0, Kcur, inp_pos, rope_factors,
+                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                ext_factor, attn_factor, beta_fast, beta_slow
+                );
+    }
+
+    cb(Qcur, "Qcur", il);
+    cb(Kcur, "Kcur", il);
+    cb(Vcur, "Vcur", il);
+
+    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+    cur = build_attn(inp_attn,
+            model.layers[il].wo, model.layers[il].bo,
+            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+    return cur;
+}
+
+ggml_tensor * llm_build_granite::build_layer_ffn(
+          ggml_tensor       * cur,
+          ggml_tensor       * inpSA,
+    const llama_model       & model,
+    const int                 il) {
+
+    // For Granite architectures - scale residual
+    if (hparams.f_residual_scale) {
+        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+    }
+    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+    cb(ffn_inp, "ffn_inp", il);
+
+    // feed-forward network (non-MoE)
+    if (model.layers[il].ffn_gate_inp == nullptr) {
+
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+                cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
+
+    } else {
+        // MoE branch
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+                cb(cur, "ffn_norm", il);
+
+        ggml_tensor * moe_out = build_moe_ffn(cur,
+                model.layers[il].ffn_gate_inp,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, true,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il);
+        cb(moe_out, "ffn_moe_out", il);
+
+        // For Granite MoE Shared
+        if (hparams.n_ff_shexp > 0) {
+            ggml_tensor * ffn_shexp = build_ffn(cur,
+                model.layers[il].ffn_up_shexp,   NULL, NULL,
+                model.layers[il].ffn_gate_shexp, NULL, NULL,
+                model.layers[il].ffn_down_shexp, NULL, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(ffn_shexp, "ffn_shexp", il);
+
+            cur = ggml_add(ctx0, moe_out, ffn_shexp);
+            cb(cur, "ffn_out", il);
+        } else {
+            cur = moe_out;
+        }
+    }
+
+    // For Granite architectures - scale residual
+    if (hparams.f_residual_scale) {
+        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+    }
+    cur = ggml_add(ctx0, cur, ffn_inp);
+    cb(cur, "ffn_out", il);
+
+    cur = build_cvec(cur, il);
+    cb(cur, "l_out", il);
+
+    return cur;
+}

src/models/graph-context-mamba.cpp

@@ -0,0 +1,283 @@
+#include "models.h"
+
+llm_graph_context_mamba::llm_graph_context_mamba(const llm_graph_params & params) : llm_graph_context(params) {}
+
+ggml_tensor * llm_graph_context_mamba::build_mamba_layer(llm_graph_input_rs * inp,
+                                                         ggml_tensor *        cur,
+                                                         const llama_model &  model,
+                                                         const llama_ubatch & ubatch,
+                                                         int                  il) {
+    const auto * mctx_cur = inp->mctx;
+
+    const auto kv_head = mctx_cur->get_head();
+
+    const auto & layer = model.layers[il];
+
+    const int64_t d_conv         = hparams.ssm_d_conv;
+    const int64_t d_inner        = hparams.ssm_d_inner;
+    const int64_t d_state        = hparams.ssm_d_state;
+    const int64_t dt_rank        = hparams.ssm_dt_rank;
+    const int64_t n_head         = d_inner;
+    const int64_t head_dim       = 1;
+    const int64_t n_seqs         = ubatch.n_seqs;
+    // Some variants of Mamba arch (e.g. FalconMamba do apply layer norm on B and Dt layers)
+    const bool    ssm_dt_b_c_rms = hparams.ssm_dt_b_c_rms;
+
+    const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+
+    GGML_ASSERT(n_seqs != 0);
+    GGML_ASSERT(ubatch.equal_seqs());
+    GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
+
+    ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
+    ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);
+
+    ggml_tensor * conv = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
+    conv               = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner, n_seqs);
+
+    // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
+    cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
+
+    // {n_embd, 2*d_inner} @ {n_embd, n_seq_tokens, n_seqs} => {2*d_inner, n_seq_tokens, n_seqs}
+    ggml_tensor * xz = build_lora_mm(layer.ssm_in, cur);
+    // split the above in two
+    // => {d_inner, n_seq_tokens, n_seqs}
+    ggml_tensor * x  = ggml_view_3d(ctx0, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], 0);
+    ggml_tensor * z =
+        ggml_view_3d(ctx0, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], d_inner * ggml_element_size(xz));
+
+    // conv
+    {
+        // => {d_conv - 1 + n_seq_tokens, d_inner, n_seqs}
+        ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, x), 0);
+
+        // copy last (d_conv - 1) columns back into the state cache
+        ggml_tensor * last_conv = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner, n_seqs, conv_x->nb[1], conv_x->nb[2],
+                                               n_seq_tokens * (conv_x->nb[0]));
+
+        ggml_build_forward_expand(
+            gf, ggml_cpy(ctx0, last_conv,
+                         ggml_view_1d(ctx0, conv_states_all, (d_conv - 1) * (d_inner) * (n_seqs),
+                                      kv_head * (d_conv - 1) * (d_inner) *ggml_element_size(conv_states_all))));
+
+        // 1D convolution
+        // The equivalent is to make a self-overlapping view of conv_x
+        // over d_conv columns at each stride in the 3rd dimension,
+        // then element-wise multiply that with the conv1d weight,
+        // then sum the elements of each row,
+        // (the last two steps are a dot product over rows (also doable with mul_mat))
+        // then permute away the ne[0] dimension,
+        // and then you're left with the resulting x tensor.
+        // For simultaneous sequences, all sequences need to have the same length.
+        x = ggml_ssm_conv(ctx0, conv_x, layer.ssm_conv1d);
+
+        // bias
+        x = ggml_add(ctx0, x, layer.ssm_conv1d_b);
+
+        x = ggml_silu(ctx0, x);
+    }
+
+    // ssm
+    {
+        // {d_inner, dt_rank + 2*d_state} @ {d_inner, n_seq_tokens, n_seqs} => {dt_rank + 2*d_state, n_seq_tokens, n_seqs}
+        ggml_tensor * x_db = build_lora_mm(layer.ssm_x, x);
+        // split
+        ggml_tensor * dt   = ggml_view_3d(ctx0, x_db, dt_rank, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], 0);
+        ggml_tensor * B =
+            ggml_view_4d(ctx0, x_db, d_state, /* n_group */ 1, n_seq_tokens, n_seqs, d_state * x_db->nb[0], x_db->nb[1],
+                         x_db->nb[2], ggml_element_size(x_db) * dt_rank);
+        ggml_tensor * C =
+            ggml_view_4d(ctx0, x_db, d_state, /* n_group */ 1, n_seq_tokens, n_seqs, d_state * x_db->nb[0], x_db->nb[1],
+                         x_db->nb[2], ggml_element_size(x_db) * (dt_rank + d_state));
+
+        // Some Mamba variants (e.g. FalconMamba, Jamba) apply RMS norm in B, C & Dt layers
+        if (ssm_dt_b_c_rms || (layer.ssm_dt_norm && layer.ssm_b_norm && layer.ssm_c_norm)) {
+            dt = build_norm(dt, layer.ssm_dt_norm, NULL, LLM_NORM_RMS, il);
+            B  = build_norm(B, layer.ssm_b_norm, NULL, LLM_NORM_RMS, il);
+            C  = build_norm(C, layer.ssm_c_norm, NULL, LLM_NORM_RMS, il);
+        }
+
+        // {dt_rank, d_inner} @ {dt_rank, n_seq_tokens, n_seqs} => {d_inner, n_seq_tokens, n_seqs}
+        dt = build_lora_mm(layer.ssm_dt, dt);
+        dt = ggml_add(ctx0, dt, layer.ssm_dt_b);
+
+        cur = x;
+        x   = ggml_reshape_4d(ctx0, x, head_dim, n_head, n_seq_tokens, n_seqs);
+
+        ggml_tensor * A = layer.ssm_a;
+
+        // use the states and the indices provided by build_recurrent_state
+        // (this is necessary in order to properly use the states before they are overwritten,
+        //  while avoiding to make unnecessary copies of the states)
+        auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) {
+            ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_head, mctx_cur->get_size());
+
+            // Custom operator to optimize the parallel associative scan
+            // as described in the Annex D of the Mamba paper.
+            // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs}
+            return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids);
+        };
+
+        ggml_tensor * y_ssm = build_rs(inp, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows);
+
+        // store last states
+        ggml_build_forward_expand(
+            gf, ggml_cpy(ctx0, ggml_view_1d(ctx0, y_ssm, d_state * d_inner * n_seqs, x->nb[3] * x->ne[3]),
+                         ggml_view_1d(ctx0, ssm_states_all, d_state * d_inner * n_seqs,
+                                      kv_head * d_state * d_inner * ggml_element_size(ssm_states_all))));
+
+        ggml_tensor * y = ggml_view_3d(ctx0, y_ssm, d_inner, n_seq_tokens, n_seqs, x->nb[2], x->nb[3], 0);
+
+        // TODO: skip computing output earlier for unused tokens
+
+        y = ggml_add(ctx0, y, ggml_mul(ctx0, cur, layer.ssm_d));
+        y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y);
+
+        // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs}
+        cur = build_lora_mm(layer.ssm_out, y);
+    }
+
+    // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
+    cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs);
+
+    return cur;
+}
+
+ggml_tensor * llm_graph_context_mamba::build_mamba2_layer(llm_graph_input_rs * inp,
+                                                          ggml_tensor *        cur,
+                                                          const llama_model &  model,
+                                                          const llama_ubatch & ubatch,
+                                                          int                  il) const {
+    const auto * mctx_cur = inp->mctx;
+
+    const auto kv_head = mctx_cur->get_head();
+
+    const int64_t d_conv   = hparams.ssm_d_conv;
+    const int64_t d_inner  = hparams.ssm_d_inner;
+    const int64_t d_state  = hparams.ssm_d_state;
+    const int64_t n_head   = hparams.ssm_dt_rank;
+    const int64_t head_dim = d_inner / n_head;
+    const int64_t n_group  = hparams.ssm_n_group;
+    const int64_t n_seqs   = ubatch.n_seqs;
+
+    const int64_t n_seq_tokens = ubatch.n_seq_tokens;
+
+    GGML_ASSERT(n_seqs != 0);
+    GGML_ASSERT(ubatch.equal_seqs());
+    GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
+
+    ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
+    ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);
+
+    ggml_tensor * conv = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
+    conv               = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner + 2 * n_group * d_state, n_seqs);
+
+    // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
+    cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
+
+    // d_in_proj = 2 * self.d_inner + 2 * self.ngroups * self.d_state + self.nheads
+
+    // {n_embd, d_in_proj} @ {n_embd, n_seq_tokens, n_seqs} => {d_in_proj, n_seq_tokens, n_seqs}
+    ggml_tensor * zxBCdt = build_lora_mm(model.layers[il].ssm_in, cur);
+
+    // split the above in three
+    ggml_tensor * z   = ggml_view_4d(ctx0, zxBCdt, head_dim, n_head, n_seq_tokens, n_seqs, head_dim * zxBCdt->nb[0],
+                                     zxBCdt->nb[1], zxBCdt->nb[2], 0);
+    ggml_tensor * xBC = ggml_view_3d(ctx0, zxBCdt, d_inner + 2 * n_group * d_state, n_seq_tokens, n_seqs, zxBCdt->nb[1],
+                                     zxBCdt->nb[2], d_inner * ggml_element_size(zxBCdt));
+    ggml_tensor * dt  = ggml_view_3d(ctx0, zxBCdt, n_head, n_seq_tokens, n_seqs, zxBCdt->nb[1], zxBCdt->nb[2],
+                                     (2 * d_inner + 2 * n_group * d_state) * ggml_element_size(zxBCdt));
+
+    // conv
+    {
+        // => {d_conv - 1 + n_seq_tokens, d_inner + 2*n_group*d_state, n_seqs}
+        ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, xBC), 0);
+
+        // copy last (d_conv - 1) columns back into the state cache
+        ggml_tensor * last_conv = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner + 2 * n_group * d_state, n_seqs,
+                                               conv_x->nb[1], conv_x->nb[2], n_seq_tokens * (conv_x->nb[0]));
+
+        ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv,
+                                               ggml_view_1d(ctx0, conv_states_all,
+                                                            (d_conv - 1) * (d_inner + 2 * n_group * d_state) * (n_seqs),
+                                                            kv_head * (d_conv - 1) * (d_inner + 2 * n_group * d_state) *
+                                                                ggml_element_size(conv_states_all))));
+
+        // 1D convolution
+        // The equivalent is to make a self-overlapping view of conv_x
+        // over d_conv columns at each stride in the 3rd dimension,
+        // then element-wise multiply that with the conv1d weight,
+        // then sum the elements of each row,
+        // (the last two steps are a dot product over rows (also doable with mul_mat))
+        // then permute away the ne[0] dimension,
+        // and then you're left with the resulting x tensor.
+        // For simultaneous sequences, all sequences need to have the same length.
+        xBC = ggml_ssm_conv(ctx0, conv_x, model.layers[il].ssm_conv1d);
+
+        // bias
+        xBC = ggml_add(ctx0, xBC, model.layers[il].ssm_conv1d_b);
+
+        xBC = ggml_silu(ctx0, xBC);
+    }
+
+    // ssm
+    {
+        // These correspond to V K Q in SSM/attention duality
+        ggml_tensor * x = ggml_view_4d(ctx0, xBC, head_dim, n_head, n_seq_tokens, n_seqs, head_dim * xBC->nb[0],
+                                       xBC->nb[1], xBC->nb[2], 0);
+        ggml_tensor * B = ggml_view_4d(ctx0, xBC, d_state, n_group, n_seq_tokens, n_seqs, d_state * xBC->nb[0],
+                                       xBC->nb[1], xBC->nb[2], d_inner * ggml_element_size(xBC));
+        ggml_tensor * C = ggml_view_4d(ctx0, xBC, d_state, n_group, n_seq_tokens, n_seqs, d_state * xBC->nb[0],
+                                       xBC->nb[1], xBC->nb[2], (d_inner + n_group * d_state) * ggml_element_size(xBC));
+
+        // {n_head, n_seq_tokens, n_seqs}
+        dt = ggml_add(ctx0, ggml_cont(ctx0, dt), model.layers[il].ssm_dt_b);
+
+        ggml_tensor * A = model.layers[il].ssm_a;
+
+        // use the states and the indices provided by build_recurrent_state
+        // (this is necessary in order to properly use the states before they are overwritten,
+        //  while avoiding to make unnecessary copies of the states)
+        auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) {
+            ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_head, mctx_cur->get_size());
+
+            // TODO: use semistructured matrices to implement state-space duality
+            // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs}
+            return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids);
+        };
+
+        ggml_tensor * y_ssm = build_rs(inp, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows);
+
+        // store last states
+        ggml_build_forward_expand(
+            gf, ggml_cpy(ctx0, ggml_view_1d(ctx0, y_ssm, d_state * d_inner * n_seqs, ggml_nelements(x) * x->nb[0]),
+                         ggml_view_1d(ctx0, ssm_states_all, d_state * d_inner * n_seqs,
+                                      kv_head * d_state * d_inner * ggml_element_size(ssm_states_all))));
+
+        ggml_tensor * y = ggml_view_4d(ctx0, y_ssm, head_dim, n_head, n_seq_tokens, n_seqs, x->nb[1], n_head * x->nb[1],
+                                       n_seq_tokens * n_head * x->nb[1], 0);
+
+        // TODO: skip computing output earlier for unused tokens
+
+        y = ggml_add(ctx0, y, ggml_mul(ctx0, x, model.layers[il].ssm_d));
+        cb(y, "mamba2_y_add_d", il);
+        y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y);
+
+        // grouped RMS norm
+        if (model.layers[il].ssm_norm) {
+            y = ggml_reshape_4d(ctx0, y, d_inner / n_group, n_group, n_seq_tokens, n_seqs);
+            y = build_norm(y, model.layers[il].ssm_norm, NULL, LLM_NORM_RMS, il);
+        }
+
+        y = ggml_reshape_3d(ctx0, y, d_inner, n_seq_tokens, n_seqs);
+
+        // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs}
+        cur = build_lora_mm(model.layers[il].ssm_out, y);
+    }
+
+    // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
+    cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs);
+    cb(cur, "mamba_out", il);
+
+    return cur;
+}

src/models/grok.cpp

@@ -0,0 +1,160 @@
+#include "models.h"
+
+llm_build_grok::llm_build_grok(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv();
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
+            }
+            if (il == n_layer - 1 && inp_out_ids) {
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+            cur = build_norm(cur,
+                    model.layers[il].attn_out_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_out_norm", il);
+
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            // MoE branch
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_GELU, true,
+                    false, 0.0,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            if (model.layers[il].ffn_up) {
+                ggml_tensor * ffn_out = build_ffn(cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_PAR, il);
+                cb(ffn_out, "ffn_out", il);
+
+                cur = ggml_scale(ctx0, ggml_add(ctx0, ffn_out, moe_out), std::sqrt(2) / 2);
+                cb(cur, "ffn_out", il);
+            } else {
+                cur = moe_out;
+            }
+            cur = build_norm(cur,
+                    model.layers[il].ffn_post_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_post_norm", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        cur = ggml_scale(ctx0, cur, hparams.f_logit_scale);
+
+        // final logit soft-capping
+        if (hparams.f_final_logit_softcapping) {
+            cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
+            cur = ggml_tanh(ctx0, cur);
+            cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
+        }
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+}

src/models/grovemoe.cpp

@@ -0,0 +1,141 @@
+#include "models.h"
+
+
+
+llm_build_grovemoe::llm_build_grovemoe(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t n_embd_head    = hparams.n_embd_head_v;
+    const int64_t n_chunk_expert = n_expert / hparams.n_group_experts;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self_attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                                 ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // MoE branch
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        ggml_tensor * probs = build_lora_mm(model.layers[il].ffn_gate_inp, cur);  // [n_expert, n_tokens]
+        cb(probs, "ffn_moe_logits", il);
+
+        ggml_tensor * moe_out =
+            build_moe_ffn(cur,
+                nullptr,
+                model.layers[il].ffn_up_exps,
+                model.layers[il].ffn_gate_exps,
+                model.layers[il].ffn_down_exps,
+                nullptr,
+                n_expert, n_expert_used,
+                LLM_FFN_SILU, true,
+                false, 0.0,
+                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                il,
+                probs);
+        cb(moe_out, "ffn_moe_out", il);
+        cur = moe_out;
+
+        // TODO: Only do the expert selection and weights once
+        moe_out = build_moe_ffn(cur,
+                    nullptr,
+                    model.layers[il].ffn_up_chexps,
+                    model.layers[il].ffn_gate_chexps,
+                    model.layers[il].ffn_down_chexps,
+                    nullptr,
+                    n_chunk_expert, n_expert_used > n_chunk_expert ? n_chunk_expert : n_expert_used,
+                    LLM_FFN_SILU, true,
+                    false, 0.0,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il,
+                    probs);
+        cb(moe_out, "ffn_adj_moe_out", il);
+
+        cur = ggml_add(ctx0, cur, ggml_scale(ctx0, moe_out, hparams.expert_group_scale));
+        cb(cur, "ffn_final_moe_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/hunyuan-dense.cpp

@@ -0,0 +1,132 @@
+#include "models.h"
+
+llm_build_hunyuan_dense::llm_build_hunyuan_dense(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv();
+
+        const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+            // self-attention
+            {
+                // rope freq factors for llama3; may return nullptr for llama2 and other models
+                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+                // compute Q and K and RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_ext(
+                         ctx0, Qcur, inp_pos, rope_factors,
+                         n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                         ext_factor, attn_factor, beta_fast, beta_slow
+                         );
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Kcur = ggml_rope_ext(
+                         ctx0, Kcur, inp_pos, rope_factors,
+                         n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                         ext_factor, attn_factor, beta_fast, beta_slow
+                         );
+
+                Kcur = build_norm(Kcur,
+                         model.layers[il].attn_k_norm, nullptr,
+                         LLM_NORM_RMS, il);
+                cb(Kcur, "Kcur_norm", il);
+
+                Qcur = build_norm(Qcur,
+                         model.layers[il].attn_q_norm, nullptr,
+                         LLM_NORM_RMS, il);
+                cb(Qcur, "Qcur_norm", il);
+
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+                cb(cur, "attn_out", il);
+            }
+            if (il == n_layer - 1 && inp_out_ids) {
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+            // feed-forward network (non-MoE)
+            ggml_tensor * cur_mlp = build_ffn(cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur_mlp, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur_mlp, ffn_inp);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+}

src/models/hunyuan-moe.cpp

@@ -0,0 +1,154 @@
+#include "models.h"
+
+llm_build_hunyuan_moe::llm_build_hunyuan_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv();
+
+        const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // rope freq factors for llama3; may return nullptr for llama2 and other models
+                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+                // compute Q and K and RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, rope_factors,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = build_norm(Kcur,
+                        model.layers[il].attn_k_norm, nullptr,
+                        LLM_NORM_RMS, il);
+                cb(Kcur, "Kcur_norm", il);
+
+                Qcur = build_norm(Qcur,
+                        model.layers[il].attn_q_norm, nullptr,
+                        LLM_NORM_RMS, il);
+                cb(Qcur, "Qcur_norm", il);
+
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+                cb(cur, "attn_out", il);
+            }
+            if (il == n_layer - 1 && inp_out_ids) {
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            // feed-forward network (non-MoE)
+            ggml_tensor * cur_mlp = build_ffn(cur,
+                    model.layers[il].ffn_up_shexp,   NULL, NULL,
+                    model.layers[il].ffn_gate_shexp, NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur_mlp, "ffn_mlp", il);
+
+            // MoE branch
+            ggml_tensor * cur_moe = build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU,
+                    true, // norm_topk_prob
+                    false,
+                    0.0,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+            cb(cur_moe, "ffn_moe_out", il);
+
+            ggml_tensor * ffn_out = ggml_add(ctx0, cur_moe, cur_mlp);
+            cb(ffn_out, "ffn_out", il);
+
+            cur = ggml_add(ctx0, ffn_out, ffn_inp);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+}

src/models/internlm2.cpp

@@ -0,0 +1,121 @@
+#include "models.h"
+
+
+llm_build_internlm2::llm_build_internlm2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv();
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+            }
+            if (il == n_layer - 1 && inp_out_ids) {
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+}