llama : add Phi-4-mini support (supersede #12099 ) (#12108 )

* Added Phi-4-mini-instruct support * Update regex per ngxson * Change the vocab base to Xenova/gpt-4o * fix conversion update script * no need to check longrope * minor style fix * fix python style --------- Co-authored-by: Nicholas Sparks <nisparks@microsoft.com>
Update granite vision docs for 3.2 model (#12105 )
2026-10-08 13:57:34 -05:00 · 2025-02-28 12:44:11 +01:00 · 2025-02-28 11:31:47 +00:00 · 2025-02-28 09:42:52 +01:00
17 changed files with 734 additions and 77 deletions
@@ -699,6 +699,9 @@ class Model:
        if chkhsh == "b3f499bb4255f8ca19fccd664443283318f2fd2414d5e0b040fbdd0cc195d6c5":
            # ref: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
            res = "deepseek-r1-qwen"
+        if chkhsh == "ccc2ef013c104be7bae2965776d611e1d7a8a2a9c547dd93a682c9a9fc80352e":
+            # ref: https://huggingface.co/Xenova/gpt-4o
+            res = "gpt-4o"

        if res is None:
            logger.warning("\n")
@@ -2512,7 +2515,8 @@ class Phi3MiniModel(Model):
        rms_eps = self.find_hparam(["rms_norm_eps"])
        max_pos_embds = self.find_hparam(["n_positions", "max_position_embeddings"])
        orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"])
-        rope_dims = n_embd // n_head
+        rot_pct = self.hparams.get("partial_rotary_factor", 1.0)
+        rope_dims = int(rot_pct * n_embd) // n_head

        self.gguf_writer.add_context_length(max_pos_embds)
        self.gguf_writer.add_rope_scaling_orig_ctx_len(orig_max_pos_embds)
@@ -2536,7 +2540,8 @@ class Phi3MiniModel(Model):
        n_head = self.find_hparam(["num_attention_heads", "n_head"])
        max_pos_embds = self.find_hparam(["n_positions", "max_position_embeddings"])
        orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"])
-        rope_dims = n_embd // n_head
+        rot_pct = self.hparams.get("partial_rotary_factor", 1.0)
+        rope_dims = int(rot_pct * n_embd) // n_head

        # write rope scaling for long context (128k) model
        rope_scaling = self.find_hparam(['rope_scaling'], True)
@@ -2565,7 +2570,7 @@ class Phi3MiniModel(Model):
            raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor')

        if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2:
-            raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}')
+            raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}. long_factors = {len(long_factors)}, short_factors = {len(short_factors)}.')

        yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_LONG), torch.tensor(long_factors, dtype=torch.float32))
        yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT), torch.tensor(short_factors, dtype=torch.float32))
@@ -109,6 +109,7 @@ models = [
    {"name": "megrez",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Infinigence/Megrez-3B-Instruct"},
    {"name": "deepseek-v3",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-V3"},
    {"name": "deepseek-r1-qwen", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"},
+    {"name": "gpt-4o",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Xenova/gpt-4o", },
 ]


@@ -131,6 +132,10 @@ def download_model(model):

    files = ["config.json", "tokenizer.json", "tokenizer_config.json"]

+    if name == "gpt-4o":
+        # Xenova/gpt-4o is tokenizer-only, it does not contain config.json
+        files = ["tokenizer.json", "tokenizer_config.json"]
+
    if tokt == TOKENIZER_TYPE.SPM:
        files.append("tokenizer.model")

@@ -3,8 +3,8 @@
 Download the model and point your `GRANITE_MODEL` environment variable to the path.

 ```bash
-$ git clone https://huggingface.co/ibm-granite/granite-vision-3.1-2b-preview
-$ export GRANITE_MODEL=./granite-vision-3.1-2b-preview
+$ git clone https://huggingface.co/ibm-granite/granite-vision-3.2-2b
+$ export GRANITE_MODEL=./granite-vision-3.2-2b
 ```


@@ -41,10 +41,18 @@ If you actually inspect the `.keys()` of the loaded tensors, you should see a lo


 ### 2. Creating the Visual Component GGUF
-To create the GGUF for the visual components, we need to write a config for the visual encoder; make sure the config contains the correct `image_grid_pinpoints`
+Next, create a new directory to hold the visual components, and copy the llava.clip/projector files, as shown below.

+```bash
+$ ENCODER_PATH=$PWD/visual_encoder
+$ mkdir $ENCODER_PATH
+
+$ cp $GRANITE_MODEL/llava.clip $ENCODER_PATH/pytorch_model.bin
+$ cp $GRANITE_MODEL/llava.projector $ENCODER_PATH/
+```
+
+Now, we need to write a config for the visual encoder. In order to convert the model, be sure to use the correct `image_grid_pinpoints`, as these may vary based on the model. You can find the `image_grid_pinpoints` in `$GRANITE_MODEL/config.json`.

-Note: we refer to this file as `$VISION_CONFIG` later on.
 ```json
 {
    "_name_or_path": "siglip-model",
@@ -52,6 +60,7 @@ Note: we refer to this file as `$VISION_CONFIG` later on.
      "SiglipVisionModel"
    ],
    "image_grid_pinpoints": [
+        [384,384],
        [384,768],
        [384,1152],
        [384,1536],
@@ -94,24 +103,13 @@ Note: we refer to this file as `$VISION_CONFIG` later on.
 }
 ```

-Create a new directory to hold the visual components, and copy the llava.clip/projector files, as well as the vision config into it.
-
-```bash
-$ ENCODER_PATH=$PWD/visual_encoder
-$ mkdir $ENCODER_PATH
-
-$ cp $GRANITE_MODEL/llava.clip $ENCODER_PATH/pytorch_model.bin
-$ cp $GRANITE_MODEL/llava.projector $ENCODER_PATH/
-$ cp $VISION_CONFIG $ENCODER_PATH/config.json
-```
-
-At which point you should have something like this:
+At this point you should have something like this:
 ```bash
 $ ls $ENCODER_PATH
 config.json             llava.projector         pytorch_model.bin
 ```

-Now convert the components to GGUF; Note that we also override the image mean/std dev to `[.5,.5,.5]` since we use the siglip visual encoder - in the transformers model, you can find these numbers in the [preprocessor_config.json](https://huggingface.co/ibm-granite/granite-vision-3.1-2b-preview/blob/main/preprocessor_config.json).
+Now convert the components to GGUF; Note that we also override the image mean/std dev to `[.5,.5,.5]` since we use the SigLIP visual encoder - in the transformers model, you can find these numbers in the `preprocessor_config.json`.
 ```bash
 $ python convert_image_encoder_to_gguf.py \
    -m $ENCODER_PATH \
@@ -119,17 +117,18 @@ $ python convert_image_encoder_to_gguf.py \
    --output-dir $ENCODER_PATH \
    --clip-model-is-vision \
    --clip-model-is-siglip \
-    --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5
+    --image-mean 0.5 0.5 0.5 \
+    --image-std 0.5 0.5 0.5
 ```

-this will create the first GGUF file at `$ENCODER_PATH/mmproj-model-f16.gguf`; we will refer to the abs path of this file as the `$VISUAL_GGUF_PATH.`
+This will create the first GGUF file at `$ENCODER_PATH/mmproj-model-f16.gguf`; we will refer to the absolute path of this file as the `$VISUAL_GGUF_PATH.`


 ### 3. Creating the LLM GGUF.
 The granite vision model contains a granite LLM as its language model. For now, the easiest way to get the GGUF for LLM is by loading the composite model in `transformers` and exporting the LLM so that it can be directly converted with the normal conversion path.

 First, set the `LLM_EXPORT_PATH` to the path to export the `transformers` LLM to.
-```
+```bash
 $ export LLM_EXPORT_PATH=$PWD/granite_vision_llm
 ```

@@ -142,7 +141,7 @@ if not MODEL_PATH:
    raise ValueError("env var GRANITE_MODEL is unset!")

 LLM_EXPORT_PATH = os.getenv("LLM_EXPORT_PATH")
-if not MODEL_PATH:
+if not LLM_EXPORT_PATH:
    raise ValueError("env var LLM_EXPORT_PATH is unset!")

 tokenizer = transformers.AutoTokenizer.from_pretrained(MODEL_PATH)
@@ -166,18 +165,26 @@ $ python convert_hf_to_gguf.py --outfile $LLM_GGUF_PATH $LLM_EXPORT_PATH
 ```


-### 4. Running the Model in Llama cpp
-Build llama cpp normally; you should have a target binary named `llama-llava-cli`, which you can pass two binaries to. Sample usage:
+### 4. Quantization
+If you want to quantize the LLM, you can do so with `llama-quantize` as you would any other LLM. For example:
+```bash
+$ ./build/bin/llama-quantize $LLM_EXPORT_PATH/granite_llm.gguf $LLM_EXPORT_PATH/granite_llm_q4_k_m.gguf Q4_K_M
+$ LLM_GGUF_PATH=$LLM_EXPORT_PATH/granite_llm_q4_k_m.gguf
+```

-Note - the test image shown below can be found [here](https://github-production-user-asset-6210df.s3.amazonaws.com/10740300/415512792-d90d5562-8844-4f34-a0a5-77f62d5a58b5.jpg?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAVCODYLSA53PQK4ZA%2F20250221%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20250221T054145Z&X-Amz-Expires=300&X-Amz-Signature=86c60be490aa49ef7d53f25d6c973580a8273904fed11ed2453d0a38240ee40a&X-Amz-SignedHeaders=host).
+Note that currently you cannot quantize the visual encoder because granite vision models use SigLIP as the visual encoder, which has tensor dimensions that are not divisible by 32.
+
+
+### 5. Running the Model in Llama cpp
+Build llama cpp normally; you should have a target binary named `llama-llava-cli`, which you can pass two binaries to. As an example, we pass the the llama.cpp banner.

 ```bash
 $ ./build/bin/llama-llava-cli -m $LLM_GGUF_PATH \
    --mmproj $VISUAL_GGUF_PATH \
-    --image cherry_blossom.jpg \
+    --image ./media/llama0-banner.png \
    -c 16384 \
-    -p "<|system|>\nA chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.\n<|user|>\n\<image>\nWhat type of flowers are in this picture?\n<|assistant|>\n" \
+    -p "<|system|>\nA chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.\n<|user|>\n\<image>\nWhat does the text in this image say?\n<|assistant|>\n" \
    --temp 0
 ```

-Sample response: `The flowers in the picture are cherry blossoms, which are known for their delicate pink petals and are often associated with the beauty of spring.`
+Sample output: `The text in the image reads "LLAMA C++ Can it run DOOM Llama?"`
@@ -1992,6 +1992,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
        }
    } else if (device->vendor_id == VK_VENDOR_ID_INTEL)
        rm_stdq = 2;
+    uint32_t rm_iq = 2 * rm_kq;

    for (uint32_t i = 0; i < mul_mat_vec_max_cols; ++i) {
        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32_"+std::to_string(i+1),  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
@@ -2006,15 +2007,15 @@ static void ggml_vk_load_shaders(vk_device& device) {
        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_s_f32_f32_len,   mul_mat_vec_iq1_s_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_m_f32_f32_len,   mul_mat_vec_iq1_m_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xxs_f32_f32_len, mul_mat_vec_iq2_xxs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f32_f32_"+std::to_string(i+1),  mul_mat_vec_iq2_xs_f32_f32_len,  mul_mat_vec_iq2_xs_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq2_s_f32_f32_len,   mul_mat_vec_iq2_s_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq3_xxs_f32_f32_len, mul_mat_vec_iq3_xxs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq3_s_f32_f32_len,   mul_mat_vec_iq3_s_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f32_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_xs_f32_f32_len,  mul_mat_vec_iq4_xs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f32_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_nl_f32_f32_len,  mul_mat_vec_iq4_nl_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_s_f32_f32_len,   mul_mat_vec_iq1_s_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_m_f32_f32_len,   mul_mat_vec_iq1_m_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xxs_f32_f32_len, mul_mat_vec_iq2_xxs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f32_f32_"+std::to_string(i+1),  mul_mat_vec_iq2_xs_f32_f32_len,  mul_mat_vec_iq2_xs_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq2_s_f32_f32_len,   mul_mat_vec_iq2_s_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq3_xxs_f32_f32_len, mul_mat_vec_iq3_xxs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq3_s_f32_f32_len,   mul_mat_vec_iq3_s_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f32_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_xs_f32_f32_len,  mul_mat_vec_iq4_xs_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f32_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_nl_f32_f32_len,  mul_mat_vec_iq4_nl_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);

        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32_"+std::to_string(i+1),  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32_"+std::to_string(i+1),  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
@@ -2028,15 +2029,15 @@ static void ggml_vk_load_shaders(vk_device& device) {
        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_s_f16_f32_len,   mul_mat_vec_iq1_s_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_m_f16_f32_len,   mul_mat_vec_iq1_m_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xxs_f16_f32_len, mul_mat_vec_iq2_xxs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f16_f32_"+std::to_string(i+1),  mul_mat_vec_iq2_xs_f16_f32_len,  mul_mat_vec_iq2_xs_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq2_s_f16_f32_len,   mul_mat_vec_iq2_s_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq3_xxs_f16_f32_len, mul_mat_vec_iq3_xxs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq3_s_f16_f32_len,   mul_mat_vec_iq3_s_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f16_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_xs_f16_f32_len,  mul_mat_vec_iq4_xs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f16_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_nl_f16_f32_len,  mul_mat_vec_iq4_nl_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_s_f16_f32_len,   mul_mat_vec_iq1_s_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_m_f16_f32_len,   mul_mat_vec_iq1_m_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xxs_f16_f32_len, mul_mat_vec_iq2_xxs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f16_f32_"+std::to_string(i+1),  mul_mat_vec_iq2_xs_f16_f32_len,  mul_mat_vec_iq2_xs_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq2_s_f16_f32_len,   mul_mat_vec_iq2_s_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq3_xxs_f16_f32_len, mul_mat_vec_iq3_xxs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq3_s_f16_f32_len,   mul_mat_vec_iq3_s_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f16_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_xs_f16_f32_len,  mul_mat_vec_iq4_xs_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f16_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_nl_f16_f32_len,  mul_mat_vec_iq4_nl_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
    }

    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
@@ -2051,15 +2052,15 @@ static void ggml_vk_load_shaders(vk_device& device) {
    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ1_S],   "mul_mat_vec_id_iq1_s_f32",   mul_mat_vec_id_iq1_s_f32_len,   mul_mat_vec_id_iq1_s_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ1_M],   "mul_mat_vec_id_iq1_m_f32",   mul_mat_vec_id_iq1_m_f32_len,   mul_mat_vec_id_iq1_m_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_XXS], "mul_mat_vec_id_iq2_xxs_f32", mul_mat_vec_id_iq2_xxs_f32_len, mul_mat_vec_id_iq2_xxs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_XS],  "mul_mat_vec_id_iq2_xs_f32",  mul_mat_vec_id_iq2_xs_f32_len,  mul_mat_vec_id_iq2_xs_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_S],   "mul_mat_vec_id_iq2_s_f32",   mul_mat_vec_id_iq2_s_f32_len,   mul_mat_vec_id_iq2_s_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ3_XXS], "mul_mat_vec_id_iq3_xxs_f32", mul_mat_vec_id_iq3_xxs_f32_len, mul_mat_vec_id_iq3_xxs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ3_S],   "mul_mat_vec_id_iq3_s_f32",   mul_mat_vec_id_iq3_s_f32_len,   mul_mat_vec_id_iq3_s_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_XS],  "mul_mat_vec_id_iq4_xs_f32",  mul_mat_vec_id_iq4_xs_f32_len,  mul_mat_vec_id_iq4_xs_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL],  "mul_mat_vec_id_iq4_nl_f32",  mul_mat_vec_id_iq4_nl_f32_len,  mul_mat_vec_id_iq4_nl_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ1_S],   "mul_mat_vec_id_iq1_s_f32",   mul_mat_vec_id_iq1_s_f32_len,   mul_mat_vec_id_iq1_s_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ1_M],   "mul_mat_vec_id_iq1_m_f32",   mul_mat_vec_id_iq1_m_f32_len,   mul_mat_vec_id_iq1_m_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_XXS], "mul_mat_vec_id_iq2_xxs_f32", mul_mat_vec_id_iq2_xxs_f32_len, mul_mat_vec_id_iq2_xxs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_XS],  "mul_mat_vec_id_iq2_xs_f32",  mul_mat_vec_id_iq2_xs_f32_len,  mul_mat_vec_id_iq2_xs_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_S],   "mul_mat_vec_id_iq2_s_f32",   mul_mat_vec_id_iq2_s_f32_len,   mul_mat_vec_id_iq2_s_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ3_XXS], "mul_mat_vec_id_iq3_xxs_f32", mul_mat_vec_id_iq3_xxs_f32_len, mul_mat_vec_id_iq3_xxs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ3_S],   "mul_mat_vec_id_iq3_s_f32",   mul_mat_vec_id_iq3_s_f32_len,   mul_mat_vec_id_iq3_s_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_XS],  "mul_mat_vec_id_iq4_xs_f32",  mul_mat_vec_id_iq4_xs_f32_len,  mul_mat_vec_id_iq4_xs_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL],  "mul_mat_vec_id_iq4_nl_f32",  mul_mat_vec_id_iq4_nl_f32_len,  mul_mat_vec_id_iq4_nl_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);

    // dequant shaders
    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16",   dequant_f32_len,  dequant_f32_data,  "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
@@ -1,5 +1,7 @@
 #version 450

+#extension GL_EXT_control_flow_attributes : enable
+
 #include "types.comp"
 #include "generic_binary_head.comp"
 #include "dequant_funcs.comp"
@@ -0,0 +1,90 @@
+#version 450
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y_idx = i * QUANT_K + 16 * itid;
+    const uint nibble_shift = 4 * (itid & 1);
+    const uint ib32 = itid / 2; // 0..7
+
+    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const float d = float(data_a[ibi].d);
+        const uint scale = (data_a[ibi].scales[ib32] >> nibble_shift) & 0xF;
+        const float db = d * (0.5 + scale) * 0.25;
+
+        const uint qh = data_a[ibi].qh[ib32];
+        const u8vec2 qs16 = unpack8(data_a_packed16[ibi].qs[itid]);
+        const u8vec2 sign16 = unpack8(data_a_packed16[ibi].qs[QUANT_K / 16 + itid]);
+        [[unroll]] for (uint l = 0; l < 2; ++l) {
+            const uint8_t sign = sign16[l];
+            const uint qs = qs16[l] | ((qh << (8 - nibble_shift - 2 * l)) & 0x300);
+            const uvec2 grid = iq2s_grid[qs];
+            const vec4 grid0 = vec4(unpack8(grid.x));
+            const vec4 grid1 = vec4(unpack8(grid.y));
+
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
+                vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
+
+                FLOAT_TYPE sum =
+                      fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x),
+                      fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y),
+                      fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z),
+                      fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w),
+                      fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x),
+                      fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y),
+                      fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z),
+                      fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign & 128) != 0 ? -grid1.w : grid1.w),
+                      FLOAT_TYPE(0.0)))))))));
+                temp[j][n] = fma(db, sum, temp[j][n]);
+            }
+        }
+        ibi += num_blocks_per_row;
+    }
+}
+
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+
+    // 16 threads are used to process each block
+    const uint blocks_per_wg = gl_WorkGroupSize.x/16;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid % 16;  // 0...15
+    const uint ix = tid / 16;
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
+        calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}
@@ -0,0 +1,87 @@
+#version 450
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y_idx = i * QUANT_K + 16 * itid;
+    const uint nibble_shift = 4 * (itid & 1);
+    const uint ib32 = itid / 2; // 0..7
+
+    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const float d = float(data_a[ibi].d);
+        const uint scale = (data_a[ibi].scales[ib32] >> nibble_shift) & 0xF;
+        const float db = d * (0.5 + scale) * 0.25;
+
+        [[unroll]] for (uint l = 0; l < 2; ++l) {
+            const uint qs = data_a[ibi].qs[2 * itid + l];
+            const uint sign = qs >> 9;
+            const uint sign7 = bitCount(sign);
+            const vec4 grid0 = vec4(unpack8(iq2xs_grid[qs & 511].x));
+            const vec4 grid1 = vec4(unpack8(iq2xs_grid[qs & 511].y));
+
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
+                vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
+
+                FLOAT_TYPE sum =
+                      fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x),
+                      fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y),
+                      fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z),
+                      fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w),
+                      fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x),
+                      fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y),
+                      fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z),
+                      fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 &  1) != 0 ? -grid1.w : grid1.w),
+                      FLOAT_TYPE(0.0)))))))));
+                temp[j][n] = fma(db, sum, temp[j][n]);
+            }
+        }
+        ibi += num_blocks_per_row;
+    }
+}
+
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+
+    // 16 threads are used to process each block
+    const uint blocks_per_wg = gl_WorkGroupSize.x/16;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid % 16;  // 0...15
+    const uint ix = tid / 16;
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
+        calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}
@@ -0,0 +1,87 @@
+#version 450
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y_idx = i * QUANT_K + 16 * itid;
+    const uint ib32 = itid / 2; // 0..7
+
+    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const float d = float(data_a[ibi].d);
+        const uint signscale = pack32(u16vec2(
+            data_a_packed16[ibi].qs[4 * ib32 + 2],
+            data_a_packed16[ibi].qs[4 * ib32 + 3]));
+        const float db = d * 0.25 * (0.5 + (signscale >> 28));
+        [[unroll]] for (uint l = 0; l < 2; ++l) {
+            const uint qs = data_a[ibi].qs[8 * ib32 + 2 * (itid & 1) + l];
+            const uint sign = bitfieldExtract(signscale, 7 * int(2 * (itid & 1) + l), 7);
+            const uint sign7 = bitCount(sign);
+            const vec4 grid0 = vec4(unpack8(iq2xxs_grid[qs].x));
+            const vec4 grid1 = vec4(unpack8(iq2xxs_grid[qs].y));
+
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                const vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
+                const vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
+
+                FLOAT_TYPE sum =
+                      fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x),
+                      fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y),
+                      fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z),
+                      fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w),
+                      fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x),
+                      fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y),
+                      fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z),
+                      fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 &  1) != 0 ? -grid1.w : grid1.w),
+                      FLOAT_TYPE(0.0)))))))));
+                temp[j][n] = fma(db, sum, temp[j][n]);
+            }
+        }
+        ibi += num_blocks_per_row;
+    }
+}
+
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+
+    // 16 threads are used to process each block
+    const uint blocks_per_wg = gl_WorkGroupSize.x/16;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid % 16;  // 0...15
+    const uint ix = tid / 16;
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
+        calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}
@@ -0,0 +1,90 @@
+#version 450
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y_idx = i * QUANT_K + 32 * ib32;
+
+    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const float d = float(data_a[ibi].d);
+        const uint scale = (data_a[ibi].scales[ib32/2] >> (4 * (ib32 & 1))) & 0xF;
+        const float dscale = d * (1 + 2 * scale);
+        const uint qh = data_a[ibi].qh[ib32];
+        FLOAT_TYPE sum[NUM_COLS];
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            sum[j] = 0.0;
+        }
+        [[unroll]] for (uint l = 0; l < 4; ++l) {
+            const u8vec2 qs = unpack8(data_a_packed16[ibi].qs[4 * ib32 + l]);
+            const uint sign = data_a[ibi].signs[4 * ib32 + l];
+            const vec4 grid0 = vec4(unpack8(iq3s_grid[qs.x | ((qh << (8 - 2*l)) & 0x100)]));
+            const vec4 grid1 = vec4(unpack8(iq3s_grid[qs.y | ((qh << (7 - 2*l)) & 0x100)]));
+
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                const vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
+                const vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
+
+                sum[j] =
+                      fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x),
+                      fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y),
+                      fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z),
+                      fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w),
+                      fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x),
+                      fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y),
+                      fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z),
+                      fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign & 128) != 0 ? -grid1.w : grid1.w),
+                      sum[j]))))))));
+            }
+        }
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            temp[j][n] = fma(dscale, sum[j], temp[j][n]);
+        }
+        ibi += num_blocks_per_row;
+    }
+}
+
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+
+    // 8 threads are used to process each block
+    const uint blocks_per_wg = gl_WorkGroupSize.x/8;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid % 8;  // 0...7
+    const uint ix = tid / 8;
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
+        calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}
@@ -0,0 +1,88 @@
+#version 450
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y_idx = i * QUANT_K + 16 * itid;
+    const uint ib32 = itid / 2; // 0..7
+
+    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const float d = float(data_a[ibi].d);
+        const uint signscale = pack32(u16vec2(
+            data_a_packed16[ibi].qs[QUANT_K / 8 + 2 * ib32],
+            data_a_packed16[ibi].qs[QUANT_K / 8 + 2 * ib32 + 1]));
+        const float db = d * 0.5 * (0.5 + (signscale >> 28));
+        [[unroll]] for (uint l = 0; l < 2; ++l) {
+            const uint qs0 = data_a[ibi].qs[8 * ib32 + 4 * (itid & 1) + 2 * l];
+            const uint qs1 = data_a[ibi].qs[8 * ib32 + 4 * (itid & 1) + 2 * l + 1];
+            const uint sign = bitfieldExtract(signscale, 7 * int(2 * (itid & 1) + l), 7);
+            const uint sign7 = bitCount(sign);
+            const vec4 grid0 = vec4(unpack8(iq3xxs_grid[qs0]));
+            const vec4 grid1 = vec4(unpack8(iq3xxs_grid[qs1]));
+
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                const vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
+                const vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
+
+                FLOAT_TYPE sum =
+                      fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x),
+                      fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y),
+                      fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z),
+                      fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w),
+                      fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x),
+                      fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y),
+                      fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z),
+                      fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 &  1) != 0 ? -grid1.w : grid1.w),
+                      FLOAT_TYPE(0.0)))))))));
+                temp[j][n] = fma(db, sum, temp[j][n]);
+            }
+        }
+        ibi += num_blocks_per_row;
+    }
+}
+
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+
+    // 16 threads are used to process each block
+    const uint blocks_per_wg = gl_WorkGroupSize.x/16;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid % 16;  // 0...15
+    const uint ix = tid / 16;
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
+        calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}
@@ -466,10 +466,13 @@ shared uint16_t iq1s_grid[2048];
 void init_iq_shmem(uvec3 wgsize)
 {
    // copy the table into shared memory and sync
-    for (uint i = gl_LocalInvocationIndex.x; i < iq1s_grid_const.length(); i += wgsize.x) {
-        u16vec2 g = unpack16(iq1s_grid_const[i]);
-        iq1s_grid[2*i+0] = g.x;
-        iq1s_grid[2*i+1] = g.y;
+    [[unroll]] for (uint i = 0; i < iq1s_grid_const.length(); i += wgsize.x) {
+        uint idx = i + gl_LocalInvocationIndex.x;
+        if (iq1s_grid_const.length() % wgsize.x == 0 || idx < iq1s_grid_const.length()) {
+            u16vec2 g = unpack16(iq1s_grid_const[idx]);
+            iq1s_grid[2*idx+0] = g.x;
+            iq1s_grid[2*idx+1] = g.y;
+        }
    }
    barrier();
 }
@@ -565,8 +568,10 @@ shared uvec2 iq2xxs_grid[256];
 void init_iq_shmem(uvec3 wgsize)
 {
    // copy the table into shared memory and sync
-    for (uint i = gl_LocalInvocationIndex.x; i < iq2xxs_grid.length(); i += wgsize.x) {
-        iq2xxs_grid[i] = iq2xxs_grid_const[i];
+    [[unroll]] for (uint i = 0; i < iq2xxs_grid.length(); i += wgsize.x) {
+        if (iq2xxs_grid_const.length() % wgsize.x == 0 || i + gl_LocalInvocationIndex.x < iq2xxs_grid_const.length()) {
+            iq2xxs_grid[i + gl_LocalInvocationIndex.x] = iq2xxs_grid_const[i + gl_LocalInvocationIndex.x];
+        }
    }
    barrier();
 }
@@ -733,8 +738,10 @@ shared uvec2 iq2xs_grid[512];
 void init_iq_shmem(uvec3 wgsize)
 {
    // copy the table into shared memory and sync
-    for (uint i = gl_LocalInvocationIndex.x; i < iq2xs_grid.length(); i += wgsize.x) {
-        iq2xs_grid[i] = iq2xs_grid_const[i];
+    [[unroll]] for (uint i = 0; i < iq2xs_grid.length(); i += wgsize.x) {
+        if (iq2xs_grid.length() % wgsize.x == 0 || i + gl_LocalInvocationIndex.x < iq2xs_grid_const.length()) {
+            iq2xs_grid[i + gl_LocalInvocationIndex.x] = iq2xs_grid_const[i + gl_LocalInvocationIndex.x];
+        }
    }
    barrier();
 }
@@ -756,6 +763,14 @@ struct block_iq2_s
    uint8_t scales[QUANT_K_IQ2_S/32];
 };

+struct block_iq2_s_packed16
+{
+    float16_t d;
+    uint16_t qs[QUANT_K_IQ2_S/8];
+    uint16_t qh[QUANT_K_IQ2_S/64];
+    uint16_t scales[QUANT_K_IQ2_S/64];
+};
+
 #if defined(DATA_A_IQ2_S)

 const uvec2 iq2s_grid_const[1024] = {
@@ -1023,8 +1038,10 @@ shared uvec2 iq2s_grid[1024];
 void init_iq_shmem(uvec3 wgsize)
 {
    // copy the table into shared memory and sync
-    for (uint i = gl_LocalInvocationIndex.x; i < iq2s_grid.length(); i += wgsize.x) {
-        iq2s_grid[i] = iq2s_grid_const[i];
+    [[unroll]] for (uint i = 0; i < iq2s_grid.length(); i += wgsize.x) {
+        if (iq2s_grid.length() % wgsize.x == 0 || i + gl_LocalInvocationIndex.x < iq2s_grid_const.length()) {
+            iq2s_grid[i + gl_LocalInvocationIndex.x] = iq2s_grid_const[i + gl_LocalInvocationIndex.x];
+        }
    }
    barrier();
 }
@@ -1032,6 +1049,7 @@ void init_iq_shmem(uvec3 wgsize)
 #define QUANT_K QUANT_K_IQ2_S
 #define QUANT_R QUANT_R_IQ2_S
 #define A_TYPE block_iq2_s
+#define A_TYPE_PACKED16 block_iq2_s_packed16
 #endif

 #define QUANT_K_IQ3_XXS 256
@@ -1092,8 +1110,10 @@ shared uint32_t iq3xxs_grid[256];
 void init_iq_shmem(uvec3 wgsize)
 {
    // copy the table into shared memory and sync
-    for (uint i = gl_LocalInvocationIndex.x; i < iq3xxs_grid.length(); i += wgsize.x) {
-        iq3xxs_grid[i] = iq3xxs_grid_const[i];
+    [[unroll]] for (uint i = 0; i < iq3xxs_grid.length(); i += wgsize.x) {
+        if (iq3xxs_grid.length() % wgsize.x == 0 || i + gl_LocalInvocationIndex.x < iq3xxs_grid.length()) {
+            iq3xxs_grid[i + gl_LocalInvocationIndex.x] = iq3xxs_grid_const[i + gl_LocalInvocationIndex.x];
+        }
    }
    barrier();
 }
@@ -1200,8 +1220,10 @@ shared uint32_t iq3s_grid[512];
 void init_iq_shmem(uvec3 wgsize)
 {
    // copy the table into shared memory and sync
-    for (uint i = gl_LocalInvocationIndex.x; i < iq3s_grid.length(); i += wgsize.x) {
-        iq3s_grid[i] = iq3s_grid_const[i];
+    [[unroll]] for (uint i = 0; i < iq3s_grid.length(); i += wgsize.x) {
+        if (iq3s_grid.length() % wgsize.x == 0 || i + gl_LocalInvocationIndex.x < iq3s_grid.length()) {
+            iq3s_grid[i + gl_LocalInvocationIndex.x] = iq3s_grid_const[i + gl_LocalInvocationIndex.x];
+        }
    }
    barrier();
 }
@@ -402,7 +402,7 @@ void process_shaders() {
    for (const auto& tname : type_names) {
        // mul mat vec
        std::string data_a_key = "DATA_A_" + to_uppercase(tname);
-        std::string shader = (string_ends_with(tname, "_k") || string_starts_with(tname, "iq1_")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";
+        std::string shader = (string_ends_with(tname, "_k") || string_starts_with(tname, "iq1_") || string_starts_with(tname, "iq2_") || string_starts_with(tname, "iq3_")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";

        string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
        string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}}));
@@ -105,6 +105,7 @@ extern "C" {
        LLAMA_VOCAB_PRE_TYPE_CHAMELEON      = 26,
        LLAMA_VOCAB_PRE_TYPE_MINERVA        = 27,
        LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM  = 28,
+        LLAMA_VOCAB_PRE_TYPE_GPT4O          = 29,
    };

    enum llama_rope_type {
@@ -0,0 +1,112 @@
+ied 4 ½ months
+__ggml_vocab_test__
+Führer
+__ggml_vocab_test__
+
+__ggml_vocab_test__
+ 
+__ggml_vocab_test__
+  
+__ggml_vocab_test__
+   
+__ggml_vocab_test__
+	
+__ggml_vocab_test__
+
+
+__ggml_vocab_test__
+
+
+
+__ggml_vocab_test__
+
+
+
+
+__ggml_vocab_test__
+	
+
+__ggml_vocab_test__
+Hello world
+__ggml_vocab_test__
+ Hello world
+__ggml_vocab_test__
+Hello World
+__ggml_vocab_test__
+ Hello World
+__ggml_vocab_test__
+ Hello World!
+__ggml_vocab_test__
+Hello, world!
+__ggml_vocab_test__
+ Hello, world!
+__ggml_vocab_test__
+ this is 🦙.cpp
+__ggml_vocab_test__
+w048 7tuijk dsdfhu
+__ggml_vocab_test__
+нещо на Български
+__ggml_vocab_test__
+កាន់តែពិសេសអាចខលចេញ
+__ggml_vocab_test__
+🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)
+__ggml_vocab_test__
+Hello
+__ggml_vocab_test__
+ Hello
+__ggml_vocab_test__
+  Hello
+__ggml_vocab_test__
+   Hello
+__ggml_vocab_test__
+    Hello
+__ggml_vocab_test__
+    Hello
+    Hello
+__ggml_vocab_test__
+ (
+__ggml_vocab_test__
+
+ =
+__ggml_vocab_test__
+' era
+__ggml_vocab_test__
+Hello, y'all! How are you 😁 ?我想在apple工作1314151天～
+__ggml_vocab_test__
+!!!!!!
+__ggml_vocab_test__
+3
+__ggml_vocab_test__
+33
+__ggml_vocab_test__
+333
+__ggml_vocab_test__
+3333
+__ggml_vocab_test__
+33333
+__ggml_vocab_test__
+333333
+__ggml_vocab_test__
+3333333
+__ggml_vocab_test__
+33333333
+__ggml_vocab_test__
+333333333
+__ggml_vocab_test__
+Cửa Việt
+__ggml_vocab_test__
+ discards
+__ggml_vocab_test__
+
+ 
+
+ 
+
+
+ 	 		 	
+  
+   
+    
+     
+🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български ''''''```````""""......!!!!!!?????? I've been 'told he's there, 'RE you sure? 'M not sure I'll make it, 'D you like some tea? We'Ve a'lL
+__ggml_vocab_test__
@@ -0,0 +1,46 @@
+ 1165 220 19 220 27124 5503
+ 37 19194 259
+
+ 220
+ 256
+ 271
+ 197
+ 198
+ 279
+ 2499
+ 2775
+ 13225 2375
+ 32949 2375
+ 13225 5922
+ 32949 5922
+ 32949 5922 0
+ 13225 11 2375 0
+ 32949 11 2375 0
+ 495 382 9552 99 247 13 17159
+ 86 45404 220 22 10191 2852 22924 4750 6916
+ 3907 53641 1235 185386 8118
+ 11400 107516 15867 20804 22851 134178 77431 32010 104312 37984 16329 27751 89335
+ 112927 222 350 14559 8 22861 114 2524 64364 104 15148 350 76466 166700 121942 780 8 91349 350 7393 74471 484 853 1617 2316 6602 8
+ 13225
+ 32949
+ 220 32949
+ 256 32949
+ 271 32949
+ 271 32949 198 271 32949
+ 350
+ 198 314
+ 6 6837
+ 13225 11 342 70653 0 3253 553 481 22861 223 1423 7522 18165 2178 34058 22369 16412 32999 16 867 8208
+ 147475
+ 18
+ 2546
+ 15517
+ 15517 18
+ 15517 2546
+ 15517 15517
+ 15517 15517 18
+ 15517 15517 2546
+ 15517 15517 15517
+ 34 60213 53904
+ 2960 3098
+ 126470 25980 160432 16609 2775 4066 172261 19432 112927 222 350 14559 8 22861 114 2524 64364 104 15148 350 76466 166700 121942 780 8 91349 9552 99 247 4103 99 247 220 18 220 2546 220 15517 220 15517 18 220 15517 2546 220 15517 15517 220 15517 15517 18 220 15517 15517 2546 220 18 13 18 220 18 485 18 220 18 1008 18 44735 107516 15867 20804 22851 134178 77431 32010 104312 156437 1423 7522 18165 2178 34058 22369 16412 32999 16 867 8208 105024 106657 1967 53641 1235 185386 8118 22434 39336 26178 26178 168394 194663 27271 147475 25883 6961 9790 1339 461 83 1280 19016 1354 11 461 1099 481 3239 30 461 44 625 3239 17291 1520 480 11 461 35 481 1299 1236 17966 30 1416 6 27493 261 54602 43
@@ -2202,13 +2202,16 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                } break;
            case LLM_ARCH_PHI3:
                {
-                    const int64_t n_embd_head = n_embd / n_head;
-
                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);

                    // output
                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
-                    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab }, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }

                    for (int i = 0; i < n_layer; ++i) {
                        auto & layer = layers[i];
@@ -2223,8 +2226,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, 0);
                        layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, 2 * n_ff }, 0);

-                        layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), { n_embd_head/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
-                        layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), { n_embd_head/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+                        layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), { n_rot/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+                        layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), { n_rot/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
                    }
                } break;
            case LLM_ARCH_PHIMOE:
@@ -392,6 +392,13 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                    "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
                };
                break;
+            case LLAMA_VOCAB_PRE_TYPE_GPT4O:
+                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+",
+                    "[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))*((?=[\\p{L}])([^A-Z]))+(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))+((?=[\\p{L}])([^A-Z]))*(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+                };
+                break;
            default:
                // default regex for BPE tokenization pre-processing
                regex_exprs = {
@@ -1592,6 +1599,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
            } else if (
                tokenizer_pre == "megrez") {
                pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN2;
+            } else if (
+                tokenizer_pre == "gpt-4o") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_GPT4O;
+                clean_spaces = false;
            } else {
                throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
            }