diff --git a/conversion/qwen.py b/conversion/qwen.py index d1127f7431..7e3d8c0d12 100644 --- a/conversion/qwen.py +++ b/conversion/qwen.py @@ -268,8 +268,101 @@ class Qwen3MoeModel(Qwen2MoeModel): super().set_vocab() +class _QwenMtpMixin: + """Shared MTP wiring for Qwen3-Next and Qwen3.5/3.6 text variants. The HF + config carries the MTP block under `mtp_num_hidden_layers` (computed from + the checkpoint when absent, e.g. Qwen3-Next) and the tensors under + `mtp.*`; we extend block_count, emit the nextn metadata key, and remap + `mtp.*` to the standard layer-indexed nextn naming so the existing + tensor_map handles them.""" + + supports_mtp_export = True + hparams: dict[str, Any] + model_arch: gguf.MODEL_ARCH + gguf_writer: gguf.GGUFWriter + block_count: int + tensor_map: gguf.TensorNameMap + no_mtp: bool + mtp_only: bool + _original_block_count: int | None = None + opt_num_mtp_layers: int = 0 + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self.block_count = self.hparams["num_hidden_layers"] + if not self.no_mtp: + n_mtp = self.hparams.get("mtp_num_hidden_layers", 0) + # Qwen-3-Next doesn't include `mtp_num_hidden_layers` in config. + if n_mtp == 0: + assert self.opt_num_mtp_layers != 0 + n_mtp = self.opt_num_mtp_layers + self.block_count += n_mtp + self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count) + + def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]: + hparams = {**self.hparams, **self.hparams.get("text_config", {})} + key = next((k for k in ["n_layers", "num_hidden_layers", "n_layer", "num_layers"] if k in hparams), None) + type(self)._original_block_count = hparams.get(key) + type(self).opt_num_mtp_layers = 0 + return super().index_tensors(remote_hf_model_id=remote_hf_model_id) # ty: ignore[unresolved-attribute] + + @classmethod + def filter_tensors(cls, item): + assert cls._original_block_count is not None + # TODO: change TextModel to super() + if (titem := TextModel.filter_tensors(item)) is None: + return None + name, gen = titem + if name.startswith("model.mtp."): + name = name.replace("model.", "", 1) + if name.startswith("mtp."): + if cls.no_mtp: + return None + remapper = { + "fc": "eh_proj", + "pre_fc_norm_embedding": "enorm", + "pre_fc_norm_hidden": "hnorm", + "norm": "shared_head.norm", + } + parts = name.split(".", 3) + if len(parts) == 4 and parts[1] == "layers" and parts[2].isdecimal(): + mtp_idx = int(parts[2]) + name = f"model.layers.{cls._original_block_count + mtp_idx}.{parts[3]}" + cls.opt_num_mtp_layers = max(cls.opt_num_mtp_layers, mtp_idx + 1) + elif len(parts) == 3 and parts[1] in remapper: + name = f"model.layers.{cls._original_block_count}.{remapper[parts[1]]}.{parts[2]}" + elif cls.mtp_only: + keep = name in ( + "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight", + "embed_tokens.weight", "norm.weight", + ) + if not keep: + return None + return name, gen + + def set_gguf_parameters(self): + super().set_gguf_parameters() # ty: ignore[unresolved-attribute] + if self.no_mtp: + return + if (n := self.block_count - self.hparams["num_hidden_layers"]) > 0: + self.gguf_writer.add_nextn_predict_layers(n) + + def prepare_metadata(self, vocab_only: bool): + from_dir = self.fname_out.is_dir() + super().prepare_metadata(vocab_only=vocab_only) # ty: ignore[unresolved-attribute] + + if not self.mtp_only or not from_dir: + return + + output_type: str = self.ftype.name.partition("_")[2] # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute] + fname_default: str = gguf.naming_convention( + self.metadata.name, self.metadata.basename, self.metadata.finetune, # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute] + self.metadata.version, size_label=None, output_type=output_type, model_type=None) # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute] + self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf" + + @ModelBase.register("Qwen3NextForCausalLM") -class Qwen3NextModel(Qwen2MoeModel): +class Qwen3NextModel(_QwenMtpMixin, Qwen2MoeModel): model_arch = gguf.MODEL_ARCH.QWEN3NEXT def set_gguf_parameters(self): @@ -284,16 +377,6 @@ class Qwen3NextModel(Qwen2MoeModel): rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"] self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.rope_parameters.get("partial_rotary_factor", 0.25))) - @classmethod - def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None: - name, gen = item - - if name.startswith("mtp"): - # ignore MTP layers for now - return None - - return super().filter_tensors(item) - def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: if name.endswith(".A_log"): data_torch = -torch.exp(data_torch) @@ -536,97 +619,13 @@ class _Qwen35MRopeMixin: self.gguf_writer.add_rope_dimension_sections(self._QWEN35_DEFAULT_MROPE_SECTION) -class _Qwen35MtpMixin: - """Shared MTP wiring for Qwen3.5/3.6 text variants. The HF config carries - the MTP block under `mtp_num_hidden_layers` and the tensors under - `mtp.*`; we extend block_count, emit the nextn metadata key, and remap - `mtp.*` to the standard layer-indexed nextn naming so the existing - tensor_map handles them.""" - - supports_mtp_export = True - hparams: dict[str, Any] - model_arch: gguf.MODEL_ARCH - gguf_writer: gguf.GGUFWriter - block_count: int - tensor_map: gguf.TensorNameMap - no_mtp: bool - mtp_only: bool - _original_block_count: int | None = None - - def __init__(self, *args, **kwargs): - super().__init__(*args, **kwargs) - self.block_count = self.hparams["num_hidden_layers"] - if not self.no_mtp: - self.block_count += self.hparams.get("mtp_num_hidden_layers", 0) - self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count) - - def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]: - hparams = {**self.hparams, **self.hparams.get("text_config", {})} - key = next((k for k in ["n_layers", "num_hidden_layers", "n_layer", "num_layers"] if k in hparams), None) - type(self)._original_block_count = hparams.get(key) - return super().index_tensors(remote_hf_model_id=remote_hf_model_id) # ty: ignore[unresolved-attribute] - - @classmethod - def filter_tensors(cls, item): - assert cls._original_block_count is not None - # TODO: change TextModel to super() - if (titem := TextModel.filter_tensors(item)) is None: - return None - name, gen = titem - if name.startswith("model.mtp."): - name = name.replace("model.", "", 1) - if name.startswith("mtp."): - if cls.no_mtp: - return None - remapper = { - "fc": "eh_proj", - "pre_fc_norm_embedding": "enorm", - "pre_fc_norm_hidden": "hnorm", - "norm": "shared_head.norm", - } - parts = name.split(".", 3) - if len(parts) == 4 and parts[1] == "layers" and parts[2].isdecimal(): - mtp_idx = int(parts[2]) - name = f"model.layers.{cls._original_block_count + mtp_idx}.{parts[3]}" - elif len(parts) == 3 and parts[1] in remapper: - name = f"model.layers.{cls._original_block_count}.{remapper[parts[1]]}.{parts[2]}" - elif cls.mtp_only: - keep = name in ( - "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight", - "embed_tokens.weight", "norm.weight", - ) - if not keep: - return None - return name, gen - - def set_gguf_parameters(self): - super().set_gguf_parameters() # ty: ignore[unresolved-attribute] - if self.no_mtp: - return - if (n := self.hparams.get("mtp_num_hidden_layers", 0)) > 0: - self.gguf_writer.add_nextn_predict_layers(n) - - def prepare_metadata(self, vocab_only: bool): - from_dir = self.fname_out.is_dir() - super().prepare_metadata(vocab_only=vocab_only) # ty: ignore[unresolved-attribute] - - if not self.mtp_only or not from_dir: - return - - output_type: str = self.ftype.name.partition("_")[2] # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute] - fname_default: str = gguf.naming_convention( - self.metadata.name, self.metadata.basename, self.metadata.finetune, # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute] - self.metadata.version, size_label=None, output_type=output_type, model_type=None) # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute] - self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf" - - @ModelBase.register("Qwen3_5ForConditionalGeneration", "Qwen3_5ForCausalLM") -class Qwen3_5TextModel(_Qwen35MtpMixin, _Qwen35MRopeMixin, _LinearAttentionVReorderBase): +class Qwen3_5TextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase): model_arch = gguf.MODEL_ARCH.QWEN35 @ModelBase.register("Qwen3_5MoeForConditionalGeneration", "Qwen3_5MoeForCausalLM") -class Qwen3_5MoeTextModel(_Qwen35MtpMixin, _Qwen35MRopeMixin, _LinearAttentionVReorderBase): +class Qwen3_5MoeTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase): model_arch = gguf.MODEL_ARCH.QWEN35MOE diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index a0e55781ba..7df984432d 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -2329,7 +2329,13 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.SSM_NORM, MODEL_TENSOR.SSM_IN, MODEL_TENSOR.SSM_BETA_ALPHA, - MODEL_TENSOR.SSM_OUT + MODEL_TENSOR.SSM_OUT, + MODEL_TENSOR.NEXTN_EH_PROJ, + MODEL_TENSOR.NEXTN_EMBED_TOKENS, + MODEL_TENSOR.NEXTN_ENORM, + MODEL_TENSOR.NEXTN_HNORM, + MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD, + MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM, ], MODEL_ARCH.QWEN3VL: [ MODEL_TENSOR.TOKEN_EMBD, diff --git a/src/llama-model.cpp b/src/llama-model.cpp index 13023c643f..4b4fe4712c 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -2195,11 +2195,11 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, // checks default: { - // The MTP head is dense-attention only on hybrid Qwen3.5/3.6, so use a plain + // The MTP head is dense-attention only on hybrid Qwen3-Next/3.5/3.6, so use a plain // attention KV cache for the MTP context instead of the hybrid wrapper. - const bool mtp_on_hybrid_qwen35 = + const bool mtp_on_hybrid_qwen = params.ctx_type == LLAMA_CONTEXT_TYPE_MTP && - (arch == LLM_ARCH_QWEN35 || arch == LLM_ARCH_QWEN35MOE); + (arch == LLM_ARCH_QWEN3NEXT || arch == LLM_ARCH_QWEN35 || arch == LLM_ARCH_QWEN35MOE); if (llm_arch_is_recurrent(arch)) { res = new llama_memory_recurrent( @@ -2211,7 +2211,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, cparams.n_seq_max, cparams.n_rs_seq, nullptr); - } else if (llm_arch_is_hybrid(arch) && !mtp_on_hybrid_qwen35) { + } else if (llm_arch_is_hybrid(arch) && !mtp_on_hybrid_qwen) { // The main difference between hybrid architectures is the // layer filters, so pick the right one here llama_memory_hybrid::layer_filter_cb filter_attn = nullptr; @@ -2226,7 +2226,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, filter_recr = [&](uint32_t il) { return hparams.is_recr(il) && hparams.n_ff(il) == 0; }; - } else if (arch == LLM_ARCH_QWEN35 || arch == LLM_ARCH_QWEN35MOE) { + } else if (arch == LLM_ARCH_QWEN3NEXT || arch == LLM_ARCH_QWEN35 || arch == LLM_ARCH_QWEN35MOE) { filter_attn = [&](uint32_t il) { return il < hparams.n_layer() && !hparams.is_recr(il); }; @@ -2292,7 +2292,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, }; } - if (mtp_on_hybrid_qwen35) { + if (mtp_on_hybrid_qwen) { filter = [&](uint32_t il) { return il >= hparams.n_layer(); }; } diff --git a/src/models/models.h b/src/models/models.h index 55072f17e9..930cc31842 100644 --- a/src/models/models.h +++ b/src/models/models.h @@ -2037,6 +2037,10 @@ struct llama_model_qwen3next : public llama_model_base { const llama_model & model; }; + struct graph_mtp : public llm_graph_context { + graph_mtp(const llama_model & model, const llm_graph_params & params); + }; + std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; diff --git a/src/models/qwen3next.cpp b/src/models/qwen3next.cpp index 09b66423d5..0808fd87aa 100644 --- a/src/models/qwen3next.cpp +++ b/src/models/qwen3next.cpp @@ -13,7 +13,11 @@ void llama_model_qwen3next::load_arch_hparams(llama_model_loader & ml) { ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank); ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group); - // Mark recurrent layers (linear attention layers) + // NextN/MTP: extra decoder block appended beyond the main stack + ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.n_layer_nextn, false); + GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < n_layer_all"); + + // Mark recurrent layers (linear attention layers). if (!ml.get_key_or_arr(LLM_KV_ATTENTION_RECURRENT_LAYERS, hparams.is_recr_impl, hparams.n_layer_all, false)) { uint32_t full_attn_interval = 4; ml.get_key(LLM_KV_FULL_ATTENTION_INTERVAL, full_attn_interval, false); @@ -28,13 +32,17 @@ void llama_model_qwen3next::load_arch_hparams(llama_model_loader & ml) { } } -void llama_model_qwen3next::load_arch_tensors(llama_model_loader &) { +void llama_model_qwen3next::load_arch_tensors(llama_model_loader & ml) { LLAMA_LOAD_LOCALS; if (n_expert == 0) { throw std::runtime_error(arch_name() + " model cannot have zero experts"); } + const bool mtp_only = (hparams.n_layer_nextn > 0) && (ml.get_weight("blk.0.attn_norm.weight") == nullptr); + const int trunk_flags = mtp_only ? TENSOR_NOT_REQUIRED : 0; + int mtp_flags = !ml.load_mtp ? TENSOR_SKIP : 0; + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0); // output @@ -61,49 +69,73 @@ void llama_model_qwen3next::load_arch_tensors(llama_model_loader &) { const int64_t qkvz_dim = key_dim * 2 + value_dim * 2; const int64_t ba_dim = n_v_heads * 2; - for (int i = 0; i < n_layer; ++i) { - auto & layer = layers[i]; - const uint32_t n_ff_shexp = hparams.n_ff_shexp > 0 ? hparams.n_ff_shexp : hparams.n_ff(i); + auto load_block_trunk = [&](int il, int flags) { + auto & layer = layers[il]; + const uint32_t n_ff_shexp = hparams.n_ff_shexp > 0 ? hparams.n_ff_shexp : hparams.n_ff(il); - layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0); - layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), { n_embd }, 0); + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", il), { n_embd }, flags); + layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", il), { n_embd }, flags); - if (!hparams.is_recr(i)) { + if (!hparams.is_recr(il)) { // Attention layers - create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head * 2, n_embd_k_gqa, n_embd_v_gqa, 0); - layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0); - + create_tensor_qkv(layer, il, n_embd, n_embd_head_k * n_head * 2, n_embd_k_gqa, n_embd_v_gqa, flags); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", il), { n_embd_head_k * n_head, n_embd }, flags); // Q/K normalization for attention layers - layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, 0); - layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, 0); + layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", il), { n_embd_head_k }, flags); + layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", il), { n_embd_head_k }, flags); } else { // Linear attention (gated delta net) specific tensors // Create tensors with calculated dimensions // note: ssm_in is used by legacy GGUF - layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), { n_embd, qkvz_dim }, TENSOR_NOT_REQUIRED); - layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, key_dim * 2 + value_dim }, TENSOR_NOT_REQUIRED); - layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), { n_embd, value_dim }, TENSOR_NOT_REQUIRED); - layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), { hparams.ssm_d_conv, conv_dim }, 0); - layer.ssm_dt = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), { hparams.ssm_dt_rank }, 0); - layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN, i), { hparams.ssm_dt_rank }, 0); - layer.ssm_beta_alpha = create_tensor(tn(LLM_TENSOR_SSM_BETA_ALPHA, "weight", i), { n_embd, ba_dim }, 0); - layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), { head_v_dim }, 0); - layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), { value_dim, n_embd }, 0); + layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", il), { n_embd, qkvz_dim }, TENSOR_NOT_REQUIRED | flags); + layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", il), { n_embd, key_dim * 2 + value_dim }, TENSOR_NOT_REQUIRED | flags); + layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", il), { n_embd, value_dim }, TENSOR_NOT_REQUIRED | flags); + layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", il), { hparams.ssm_d_conv, conv_dim }, flags); + layer.ssm_dt = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", il), { hparams.ssm_dt_rank }, flags); + layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN, il), { hparams.ssm_dt_rank }, flags); + layer.ssm_beta_alpha = create_tensor(tn(LLM_TENSOR_SSM_BETA_ALPHA, "weight", il), { n_embd, ba_dim }, flags); + layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", il), { head_v_dim }, flags); + layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", il), { value_dim, n_embd }, flags); } - layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), { n_embd, n_expert }, 0); - layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, 0); - create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, 0); + layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", il), { n_embd, n_expert }, flags); + layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", il), { n_ff_exp, n_embd, n_expert }, flags); + create_tensor_gate_up_exps(layer, il, n_embd, n_ff_exp, n_expert, flags); // Shared experts - layer.ffn_gate_inp_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", i), { n_embd }, 0); - layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), { n_embd, n_ff_shexp }, 0); - layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), { n_embd, n_ff_shexp }, 0); - layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_shexp, n_embd }, 0); + layer.ffn_gate_inp_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", il), { n_embd }, flags); + layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", il), { n_embd, n_ff_shexp }, flags); + layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", il), { n_embd, n_ff_shexp }, flags); + layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", il), { n_ff_shexp, n_embd }, flags); + }; + + auto load_block_mtp = [&](int il) { + // MTP head is identical to the trunk block (full attention + FFN) + load_block_trunk(il, mtp_flags); + + auto & layer = layers[il]; + + // NextN-specific tensors that define the MTP block. + layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", il), { 2 * n_embd, n_embd }, mtp_flags); + layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", il), { n_embd }, mtp_flags); + layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", il), { n_embd }, mtp_flags); + layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", il), { n_embd, n_vocab }, mtp_flags | TENSOR_NOT_REQUIRED); + layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", il), { n_embd, n_vocab }, mtp_flags | TENSOR_NOT_REQUIRED); + layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", il), { n_embd }, mtp_flags | TENSOR_NOT_REQUIRED); + }; + + for (int i = 0; i < n_layer; i++) { + load_block_trunk(i, trunk_flags); + } + for (int i = n_layer; i < n_layer_all; i++) { + load_block_mtp(i); } } std::unique_ptr llama_model_qwen3next::build_arch_graph(const llm_graph_params & params) const { + if (params.gtype == LLM_GRAPH_TYPE_DECODER_MTP) { + return std::make_unique(*this, params); + } return std::make_unique(*this, params); } @@ -120,6 +152,7 @@ llama_model_qwen3next::graph::graph(const llama_model & model, const llm_graph_p ggml_tensor * inp_pos = build_inp_pos(); ggml_tensor * inp_out_ids = build_inp_out_ids(); + // MTP/NextN layers are loaded as extra decoder blocks but not executed in the main pass. for (int il = 0; il < n_layer; ++il) { res->t_layer_inp[il] = inpL; @@ -139,7 +172,7 @@ llama_model_qwen3next::graph::graph(const llama_model & model, const llm_graph_p cur = build_layer_attn(inp->get_attn(), cur, inp_pos, il); } - if (il == n_layer - 1 && inp_out_ids) { + if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -171,9 +204,16 @@ llama_model_qwen3next::graph::graph(const llama_model & model, const llm_graph_p } cur = inpL; - // Final norm + // post-norm hidden state is input to both the LM head and the MTP head cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1); + cb(cur, "h_nextn", -1); + res->t_h_nextn = cur; + + if (!cparams.embeddings_nextn_masked && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + } + cb(cur, "result_norm", -1); res->t_embd = cur; @@ -186,14 +226,6 @@ llama_model_qwen3next::graph::graph(const llama_model & model, const llm_graph_p ggml_build_forward_expand(gf, cur); } -// utility to get one slice from the third dimension -// input dim: [x, y, c, b] -// output dim: [x, y, 1, b] -static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t c) { - return ggml_view_4d(ctx0, t, t->ne[0], t->ne[1], 1, t->ne[3], - t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c); -} - ggml_tensor * llama_model_qwen3next::graph::build_norm_gated( ggml_tensor * input, ggml_tensor * weights, @@ -216,7 +248,7 @@ ggml_tensor * llama_model_qwen3next::graph::build_layer_attn( // Order: joint QG projection, QG split, Q norm, KV projection, K norm, RoPE, attention // Qwen3Next uses a single Q projection that outputs query + gate - ggml_tensor * Qcur_full = build_lora_mm(model.layers[il].wq, cur); + ggml_tensor * Qcur_full = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s); cb(Qcur_full, "Qcur_full", il); Qcur_full = ggml_reshape_4d(ctx0, Qcur_full, n_embd_head * 2, n_head, n_tokens, 1); @@ -232,10 +264,10 @@ ggml_tensor * llama_model_qwen3next::graph::build_layer_attn( Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], n_embd_head * ggml_element_size(Qcur_full)); cb(gate, "gate", il); - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s); cb(Kcur, "Kcur", il); - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s); cb(Vcur, "Vcur", il); Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); @@ -274,8 +306,6 @@ ggml_tensor * llama_model_qwen3next::graph::build_layer_attn( gate = ggml_sigmoid(ctx0, gate); cb(gate, "gate_sigmoid", il); - gate = ggml_reshape_2d(ctx0, gate, n_embd_head * n_head, n_tokens); - cur = ggml_mul(ctx0, cur, gate); cb(cur, "attn_gated", il); @@ -550,16 +580,19 @@ ggml_tensor * llama_model_qwen3next::graph::build_layer_ffn(ggml_tensor * cur, c LLM_FFN_SILU, true, hparams.expert_weights_scale, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il, - nullptr, model.layers[il].ffn_gate_up_exps); + nullptr, model.layers[il].ffn_gate_up_exps, + model.layers[il].ffn_up_exps_s, + model.layers[il].ffn_gate_exps_s, + model.layers[il].ffn_down_exps_s); cb(moe_out, "ffn_moe_out", il); // Add shared experts if present - following Qwen3Next reference implementation if (model.layers[il].ffn_up_shexp != nullptr) { 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, + model.layers[il].ffn_up_shexp, NULL, model.layers[il].ffn_up_shexp_s, + model.layers[il].ffn_gate_shexp, NULL, model.layers[il].ffn_gate_shexp_s, + model.layers[il].ffn_down_shexp, NULL, model.layers[il].ffn_down_shexp_s, NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); cb(ffn_shexp, "ffn_shexp", il); @@ -593,3 +626,198 @@ ggml_tensor * llama_model_qwen3next::graph::build_layer_ffn(ggml_tensor * cur, c } return cur; } + +// LLM_GRAPH_TYPE_DECODER_MTP draft head for Qwen3-Next +llama_model_qwen3next::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params) + : llm_graph_context(params) { + GGML_ASSERT(hparams.n_layer_nextn > 0 && "QWEN3NEXT MTP requires n_layer_nextn > 0"); + GGML_ASSERT(hparams.n_layer_nextn == 1 && "QWEN3NEXT MTP currently only supports a single MTP block"); + + const int64_t n_embd_head = hparams.n_embd_head_v(); + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k()); + + const int il = hparams.n_layer(); + const auto & layer = model.layers[il]; + + GGML_ASSERT(layer.nextn.eh_proj && "MTP block missing nextn.eh_proj"); + GGML_ASSERT(layer.nextn.enorm && "MTP block missing nextn.enorm"); + GGML_ASSERT(layer.nextn.hnorm && "MTP block missing nextn.hnorm"); + GGML_ASSERT(layer.ffn_gate_inp && "MTP block missing ffn_gate_inp"); + + // TODO: extract in a common llm_graph_context::build_inp_embd_h() + auto inp = std::make_unique(hparams.n_embd); + + inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + ggml_set_input(inp->tokens); + + inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd_inp(), n_tokens); + ggml_set_input(inp->embd); + + // TODO: make static using `ggml_build_forward_select()` + // see llm_graph_context::build_inp_embd() for reference + ggml_tensor * tok_embd; + if (ubatch.token) { + ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd; + + tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens); + } else { + tok_embd = inp->embd; + } + cb(tok_embd, "mtp_tok_embd", il); + + inp->h = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens); + ggml_set_input(inp->h); + ggml_set_name(inp->h, "mtp_h_input"); + + ggml_tensor * h_embd = inp->h; + + res->add_input(std::move(inp)); + + ggml_tensor * inp_pos = build_inp_pos(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * h_norm = build_norm(h_embd, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il); + cb(h_norm, "mtp_hnorm", il); + + ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il); + cb(e_norm, "mtp_enorm", il); + + ggml_tensor * concat = ggml_concat(ctx0, e_norm, h_norm, /*dim=*/ 0); + cb(concat, "mtp_concat", il); + + ggml_tensor * cur = build_lora_mm(layer.nextn.eh_proj, concat, layer.nextn.eh_proj_s); + cb(cur, "mtp_eh_proj", il); + + ggml_tensor * inpSA = cur; + + cur = build_norm(cur, layer.attn_norm, nullptr, LLM_NORM_RMS, il); + cb(cur, "mtp_attn_norm", il); + + ggml_tensor * Qcur_full = build_lora_mm(layer.wq, cur, layer.wq_s); + cb(Qcur_full, "mtp_Qcur_full", il); + + ggml_tensor * Qcur = ggml_view_3d(ctx0, Qcur_full, + n_embd_head, n_head, n_tokens, + ggml_element_size(Qcur_full) * n_embd_head * 2, + ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head, + 0); + Qcur = build_norm(Qcur, layer.attn_q_norm, nullptr, LLM_NORM_RMS, il); + cb(Qcur, "mtp_Qcur_normed", il); + + ggml_tensor * Kcur = build_lora_mm(layer.wk, cur, layer.wk_s); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Kcur = build_norm(Kcur, layer.attn_k_norm, nullptr, LLM_NORM_RMS, il); + cb(Kcur, "mtp_Kcur_normed", il); + + ggml_tensor * Vcur = build_lora_mm(layer.wv, cur, layer.wv_s); + 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, "mtp_Qcur", il); + cb(Kcur, "mtp_Kcur", il); + cb(Vcur, "mtp_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, + nullptr, nullptr, nullptr, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(cur, "mtp_attn_pregate", il); + + ggml_tensor * gate = ggml_view_3d(ctx0, Qcur_full, + n_embd_head, n_head, n_tokens, + ggml_element_size(Qcur_full) * n_embd_head * 2, + ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head, + ggml_element_size(Qcur_full) * n_embd_head); + + // TODO: CUDA is missing non-contiguous unary ops. when implemented: remove this cont + gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens); + cb(gate, "mtp_gate", il); + + cur = ggml_mul(ctx0, cur, ggml_sigmoid(ctx0, gate)); + cur = build_lora_mm(layer.wo, cur, layer.wo_s); + cb(cur, "mtp_attn_out", il); + + if (inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + cur = ggml_add(ctx0, cur, inpSA); + cb(cur, "mtp_attn_residual", il); + + ggml_tensor * ffn_residual = cur; + cur = build_norm(cur, layer.attn_post_norm, nullptr, LLM_NORM_RMS, il); + cb(cur, "mtp_attn_post_norm", il); + + // MoE FFN — routed experts plus gated shared expert (mirrors the trunk). + ggml_tensor * moe_out = + build_moe_ffn(cur, + layer.ffn_gate_inp, + layer.ffn_up_exps, + layer.ffn_gate_exps, + layer.ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + hparams.expert_weights_scale, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il, + nullptr, layer.ffn_gate_up_exps, + layer.ffn_up_exps_s, + layer.ffn_gate_exps_s, + layer.ffn_down_exps_s); + cb(moe_out, "mtp_ffn_moe_out", il); + + if (layer.ffn_up_shexp != nullptr) { + ggml_tensor * ffn_shexp = + build_ffn(cur, + layer.ffn_up_shexp, nullptr, layer.ffn_up_shexp_s, + layer.ffn_gate_shexp, nullptr, layer.ffn_gate_shexp_s, + layer.ffn_down_shexp, nullptr, layer.ffn_down_shexp_s, + nullptr, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(ffn_shexp, "mtp_ffn_shexp", il); + + ggml_tensor * shared_gate = build_lora_mm(layer.ffn_gate_inp_shexp, cur); + shared_gate = ggml_sigmoid(ctx0, shared_gate); + cb(shared_gate, "mtp_shared_expert_gate_sigmoid", il); + + ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate); + cb(ffn_shexp, "mtp_ffn_shexp_gated", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + } else { + cur = moe_out; + } + cb(cur, "mtp_ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_residual); + cb(cur, "mtp_post_ffn", il); + + ggml_tensor * head_norm_w = layer.nextn.shared_head_norm + ? layer.nextn.shared_head_norm + : model.output_norm; + GGML_ASSERT(head_norm_w && "QWEN3NEXT MTP: missing both nextn.shared_head_norm and output_norm"); + cur = build_norm(cur, head_norm_w, nullptr, LLM_NORM_RMS, -1); + + cb(cur, "h_nextn", -1); + res->t_h_nextn = cur; + + ggml_tensor * head_w = layer.nextn.shared_head_head ? layer.nextn.shared_head_head : model.output; + ggml_tensor * head_s = layer.nextn.shared_head_head ? layer.nextn.shared_head_head_s : model.output_s; + GGML_ASSERT(head_w && "QWEN3NEXT MTP: missing LM head (nextn.shared_head_head or model.output)"); + cur = build_lora_mm(head_w, cur, head_s); + cb(cur, "result_output", -1); + + res->t_logits = cur; + ggml_build_forward_expand(gf, cur); +}