diff --git a/conversion/step3.py b/conversion/step3.py index ba867fb831ba..59758ee0ad9a 100644 --- a/conversion/step3.py +++ b/conversion/step3.py @@ -99,6 +99,34 @@ class Step3VLTextModel(Qwen3Model): class Step35Model(TextModel): model_arch = gguf.MODEL_ARCH.STEP35 + # The --mtp / --no-mtp toggles are ModelBase.mtp_only / no_mtp (set in + # convert_hf_to_gguf.py main()). Unlike Qwen3.5, which stores MTP under a + # `mtp.*` namespace, Step3.5 appends MTP layers at + # `model.layers.{num_hidden_layers + i}`, so we filter them by layer index. + # The trunk layer count is captured before indexing so the classmethod + # filter_tensors can tell the appended MTP block(s) apart from the trunk. + _n_main_layers: int | None = None + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + # NextN/MTP layers are appended past num_hidden_layers; extend the + # tensor map to cover them so the MTP block's tensors get correctly + # indexed names. When --no-mtp drops the MTP blocks, fall back to the + # base num_hidden_layers so we don't reserve unused slots. + n_nextn = int(self.hparams.get("num_nextn_predict_layers", 0)) + if n_nextn > 0 and not self.no_mtp: + self.block_count += n_nextn + 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): + # filter_tensors is a classmethod and can't reach self.hparams; stash + # the trunk layer count here (before indexing runs) so it can detect + # the appended MTP layers by index. + 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)._n_main_layers = hparams.get(key) + return super().index_tensors(remote_hf_model_id=remote_hf_model_id) + def set_gguf_parameters(self): rope_theta = self.hparams.get("rope_theta") if isinstance(rope_theta, list): @@ -119,8 +147,25 @@ def set_gguf_parameters(self): n_head_swa = attn_other.get("num_attention_heads", n_head_base) n_kv_swa = attn_other.get("num_attention_groups", n_kv_base) - layer_types = layer_types[: self.block_count] - partial_rotary_factors = partial_rotary_factors[: self.block_count] + n_nextn = int(self.hparams.get("num_nextn_predict_layers", 0)) + + # The Step3p5 HF checkpoint stores layer_types/partial_rotary_factors + # entries for the MTP blocks past num_hidden_layers; preserve them so + # the MTP layer's attention shape, SWA flag, and partial RoPE dim are + # set correctly. Pad with full-attention defaults if the checkpoint + # truncated them. + def _pad(arr, n, default): + arr = list(arr) + if len(arr) < n: + arr = arr + [default] * (n - len(arr)) + return arr[:n] + + layer_types = _pad(layer_types, self.block_count, "full_attention") + partial_rotary_factors = _pad( + partial_rotary_factors, + self.block_count, + 0.5, # full_attention default for Step3p5 + ) assert [1.0 if lt == "sliding_attention" else 0.5 for lt in layer_types] == partial_rotary_factors head_arr = [n_head_swa if lt == "sliding_attention" else n_head_base for lt in layer_types] kv_arr = [n_kv_swa if lt == "sliding_attention" else n_kv_base for lt in layer_types] @@ -157,31 +202,61 @@ def set_gguf_parameters(self): self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("rms_norm_eps", 1e-5)) - # Optional per-layer SwiGLU clamps. + # Optional per-layer SwiGLU clamps. MTP layers default to no clamping (0.0). if (limits := self.hparams.get("swiglu_limits")) is not None: - limits_f = [0.0 if v is None else float(v) for v in limits[: self.block_count]] + limits_f = _pad( + [0.0 if v is None else float(v) for v in limits], + self.block_count, + 0.0, + ) self.gguf_writer.add_swiglu_clamp_exp(limits_f) if (limits_shared := self.hparams.get("swiglu_limits_shared")) is not None: - limits_shared_f = [0.0 if v is None else float(v) for v in limits_shared[: self.block_count]] + limits_shared_f = _pad( + [0.0 if v is None else float(v) for v in limits_shared], + self.block_count, + 0.0, + ) self.gguf_writer.add_swiglu_clamp_shexp(limits_shared_f) + if n_nextn > 0 and not self.no_mtp: + self.gguf_writer.add_nextn_predict_layers(n_nextn) + @classmethod def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None: - name, gen = item + if (titem := super().filter_tensors(item)) is None: + return None + name, gen = titem # Map router bias (expert selection bias) to a GGUF bias tensor if name.endswith(".moe.router_bias"): name += ".bias" - return super().filter_tensors((name, gen)) + # Step3.5 appends the MTP block(s) past num_hidden_layers. + assert cls._n_main_layers is not None + is_mtp = (m := re.match(r"model\.layers\.(\d+)\.", name)) is not None and int(m.group(1)) >= cls._n_main_layers + + # --no-mtp: drop the appended MTP block(s) entirely. + if is_mtp and cls.no_mtp: + return None + # --mtp: keep ONLY MTP-block tensors plus the shared embeddings/norm/ + # lm_head (so the resulting GGUF carries just the draft head). + if cls.mtp_only and not is_mtp and name not in ( + "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight", + ): + return None + + # The checkpoint nests the per-MTP-layer shared head under + # `model.layers.{N+i}.transformer.shared_head.{norm,output}.weight`; + # strip the `transformer.` infix and rename `output` → `head` so the + # existing NEXTN_SHARED_HEAD_{NORM,HEAD} tensor mapping picks them up. + # Mirrors vllm's `_rewrite_spec_layer_name` (step3p5_mtp.py). + if is_mtp: + name = name.replace(".transformer.", ".") + name = name.replace("shared_head.output", "shared_head.head") + + return name, gen def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None): - # remove mtp layers - if (m := re.match(r"model\.layers\.(\d+)\.", name)) is not None: - il = int(m.group(1)) - n_main = int(self.hparams.get("num_hidden_layers", self.block_count)) - if il >= n_main: - return if name.endswith("norm.weight"): data_torch += 1.0 @@ -190,6 +265,21 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None): yield from super().modify_tensors(data_torch, name, bid) + def prepare_metadata(self, vocab_only: bool): + from_dir = self.fname_out.is_dir() + super().prepare_metadata(vocab_only=vocab_only) + + # Mirror Qwen3.5's behavior: when emitting a draft-only file into a + # directory, prefix with "mtp-" so it doesn't collide with the trunk. + if not self.mtp_only or not from_dir: + return + + output_type: str = self.ftype.name.partition("_")[2] + fname_default: str = gguf.naming_convention( + self.metadata.name, self.metadata.basename, self.metadata.finetune, + self.metadata.version, size_label=None, output_type=output_type, model_type=None) + self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf" + def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: # Step35 can optionally use Llama-3 style RoPE scaling (HF: rope_scaling.rope_type == "llama3"). # llama.cpp represents this via a single extra tensor: "rope_freqs.weight" (aka MODEL_TENSOR.ROPE_FREQS). diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 85527553563d..cd19eebdfa34 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -251,8 +251,9 @@ def main() -> None: if args.mtp or args.no_mtp: from conversion.qwen import _Qwen35MtpMixin - if not issubclass(model_class, _Qwen35MtpMixin): - logger.error("--mtp / --no-mtp are only supported for Qwen3.5/3.6 text variants today") + from conversion.step3 import Step35Model + if not (issubclass(model_class, _Qwen35MtpMixin) or issubclass(model_class, Step35Model)): + logger.error("--mtp / --no-mtp are only supported for Qwen3.5/3.6 and Step3.5 text variants today") sys.exit(1) if args.no_mtp: model_class.no_mtp = True diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index b4dfd58382d5..cca5f79d52b3 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -3987,6 +3987,13 @@ class MODEL_TENSOR(IntEnum): MODEL_TENSOR.FFN_GATE_SHEXP, MODEL_TENSOR.FFN_DOWN_SHEXP, MODEL_TENSOR.FFN_EXP_PROBS_B, + # NextN/MTP tensors (Step3p5 draft head) + 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.LLAMA_EMBED: [ MODEL_TENSOR.TOKEN_EMBD, diff --git a/src/models/models.h b/src/models/models.h index 5251e2d82802..cbef040870b6 100644 --- a/src/models/models.h +++ b/src/models/models.h @@ -1913,5 +1913,9 @@ struct llama_model_step35 : public llama_model_base { graph(const llama_model & model, const llm_graph_params & params); }; + 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/step35.cpp b/src/models/step35.cpp index 3b68e68707ae..caf18c743ff4 100644 --- a/src/models/step35.cpp +++ b/src/models/step35.cpp @@ -26,20 +26,36 @@ void llama_model_step35::load_arch_hparams(llama_model_loader & ml) { ml.get_key_or_arr(LLM_KV_SWIGLU_CLAMP_EXP, hparams.swiglu_clamp_exp, hparams.n_layer, false); ml.get_key_or_arr(LLM_KV_SWIGLU_CLAMP_SHEXP, hparams.swiglu_clamp_shexp, hparams.n_layer, false); - switch (hparams.n_layer) { + // NextN/MTP (Step3p5): extra decoder block appended beyond the main stack. + ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.nextn_predict_layers, false); + GGML_ASSERT(hparams.nextn_predict_layers < hparams.n_layer && "nextn_predict_layers must be < n_layer"); + + switch (hparams.n_layer - hparams.nextn_predict_layers) { case 45: type = LLM_TYPE_196B_A11B; break; default: type = LLM_TYPE_UNKNOWN; } } -void llama_model_step35::load_arch_tensors(llama_model_loader &) { +void llama_model_step35::load_arch_tensors(llama_model_loader & ml) { LLAMA_LOAD_LOCALS; + const uint32_t n_main = n_layer - hparams.nextn_predict_layers; + const bool mtp_only = (hparams.nextn_predict_layers > 0) && + (ml.get_weight("blk.0.attn_norm.weight") == nullptr); + // Trunk-only: the GGUF declares MTP layers in metadata but the actual MTP + // tensors live in a separate file (e.g. user split target/draft). Mark + // MTP tensors NOT_REQUIRED so the trunk loads cleanly. + const std::string mtp_probe = "blk." + std::to_string(n_main) + ".nextn.eh_proj.weight"; + const bool trunk_only = (hparams.nextn_predict_layers > 0) && + (ml.get_weight(mtp_probe.c_str()) == nullptr); + const int trunk_flags = mtp_only ? TENSOR_NOT_REQUIRED : 0; + const int mtp_flags = trunk_only ? TENSOR_NOT_REQUIRED : 0; + 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}, trunk_flags); // STEP35 supports per-layer partial RoPE dims; rope factors are stored as a single shared tensor // ("rope_freqs.weight") and ggml uses only the first (n_rot_l/2) entries per layer. @@ -51,14 +67,14 @@ void llama_model_step35::load_arch_tensors(llama_model_loader &) { n_rot_max = n_rot; } - for (int i = 0; i < n_layer; ++i) { + auto load_block_trunk = [&](int i, int flags) { auto & layer = layers[i]; const uint32_t n_head_l = hparams.n_head(i); const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i); const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i); - layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags); layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, TENSOR_NOT_REQUIRED); layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, TENSOR_NOT_REQUIRED); @@ -70,13 +86,13 @@ void llama_model_step35::load_arch_tensors(llama_model_loader &) { layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot_max/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0)); } - create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head_l, n_embd_k_gqa, n_embd_v_gqa, 0); - layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_v * n_head_l, n_embd}, 0); + create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head_l, n_embd_k_gqa, n_embd_v_gqa, flags); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_v * n_head_l, n_embd}, flags); // head-wise attention gate (Step35 self_attn.g_proj) layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_head_l}, TENSOR_NOT_REQUIRED); - layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags); // dense MLP (leading dense blocks) layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED); @@ -95,10 +111,86 @@ void llama_model_step35::load_arch_tensors(llama_model_loader &) { layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, TENSOR_NOT_REQUIRED); layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, TENSOR_NOT_REQUIRED); layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, TENSOR_NOT_REQUIRED); + }; + + auto load_block_mtp = [&](int i, bool is_first_mtp) { + auto & layer = layers[i]; + + const uint32_t n_head_l = hparams.n_head(i); + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i); + + // The MTP block is a full Step3p5 decoder layer (mtp_block) plus the + // NextN-specific wiring (enorm/hnorm/eh_proj + optional shared head). + // `mtp_flags` becomes NOT_REQUIRED when the GGUF is trunk-only. + // + // Only the FIRST MTP block (i == n_main) is required for the + // single-block MTP runtime; trailing MTP blocks are always tolerated + // as missing so pruned GGUFs (block 0 only) load cleanly. Override + // mtp_flags to NOT_REQUIRED for those. + const int eff_mtp_flags = is_first_mtp ? mtp_flags : (mtp_flags | TENSOR_NOT_REQUIRED); + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, eff_mtp_flags); + layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, TENSOR_NOT_REQUIRED); + layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, TENSOR_NOT_REQUIRED); + + if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) { + layer.rope_long = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG, "weight", i), {n_rot_max/2}, TENSOR_NOT_REQUIRED | TENSOR_DUPLICATED); + layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot_max/2}, TENSOR_NOT_REQUIRED | TENSOR_DUPLICATED); + } else { + layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot_max/2}, TENSOR_NOT_REQUIRED | TENSOR_DUPLICATED); + } + + create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head_l, n_embd_k_gqa, n_embd_v_gqa, eff_mtp_flags); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_v * n_head_l, n_embd}, eff_mtp_flags); + + layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_head_l}, TENSOR_NOT_REQUIRED); + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, eff_mtp_flags); + + // dense MLP (leading dense blocks) — present if the MTP block isn't MoE + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, TENSOR_NOT_REQUIRED); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED); + + // MoE routed experts + selection bias (router_bias) + const int64_t n_ff_exp = hparams.n_ff_exp; + layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED); + layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED); + layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, TENSOR_NOT_REQUIRED); + layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED); + layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED); + + layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, TENSOR_NOT_REQUIRED); + layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, TENSOR_NOT_REQUIRED); + layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, TENSOR_NOT_REQUIRED); + + // NextN-specific tensors that define the MTP block. + layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, eff_mtp_flags); + layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, eff_mtp_flags); + layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, eff_mtp_flags); + layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED); + layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED); + layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, TENSOR_NOT_REQUIRED); + }; + + for (int i = 0; i < (int) n_main; ++i) { + load_block_trunk(i, trunk_flags); + } + // Only the first MTP block (i == n_main) is required at runtime — the + // single-block-MTP graph in build_arch_graph always uses that one. + // Trailing MTP blocks are loaded if present (so an un-pruned GGUF with + // all MTP layers still works) but tolerated when absent via the pruning + // path. See scripts/prune_step35_extra_mtp.py for the pruner. + for (int i = (int) n_main; i < n_layer; ++i) { + load_block_mtp(i, /*is_first_mtp=*/ i == (int) n_main); } } std::unique_ptr llama_model_step35::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); } @@ -111,7 +203,9 @@ llama_model_step35::graph::graph(const llama_model & model, const llm_graph_para auto * inp_attn = build_attn_inp_kv_iswa(); ggml_tensor * inp_out_ids = build_inp_out_ids(); - for (int il = 0; il < n_layer; ++il) { + // MTP/NextN layers are loaded as extra decoder blocks but not executed in the main pass. + const int n_transformer_layers = n_layer - (int) hparams.nextn_predict_layers; + for (int il = 0; il < n_transformer_layers; ++il) { ggml_tensor * inpSA = inpL; const uint32_t n_head_l = hparams.n_head(il); @@ -198,8 +292,8 @@ llama_model_step35::graph::graph(const llama_model & model, const llm_graph_para cb(cur, "attn_proj", il); } - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); + if (il == n_transformer_layers - 1 && inp_out_ids && cparams.embeddings_pre_norm_masked) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -257,6 +351,13 @@ llama_model_step35::graph::graph(const llama_model & model, const llm_graph_para cur = inpL; + cb(cur, "h_pre_norm", -1); + res->t_h_pre_norm = cur; + + if (!cparams.embeddings_pre_norm_masked && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + } + cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1); cb(cur, "result_norm", -1); res->t_embd = cur; @@ -267,3 +368,192 @@ llama_model_step35::graph::graph(const llama_model & model, const llm_graph_para ggml_build_forward_expand(gf, cur); } + +// LLM_GRAPH_TYPE_DECODER_MTP draft head for Step3p5 (MoE) +llama_model_step35::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params) + : llm_graph_context(params) { + GGML_ASSERT(hparams.nextn_predict_layers > 0 && "STEP35 MTP requires nextn_predict_layers > 0"); + + // Single-block MTP only: always run the first trained MTP block (Qwen + // MTP / vLLM single-MTP-layer style). Multi-block round-robin proved to + // be a much deeper refactor than this PR justifies; the trailing MTP + // blocks are loaded with TENSOR_NOT_REQUIRED so pruned GGUFs (with just + // block 0) also work — see load_arch_tensors below and + // scripts/prune_step35_extra_mtp.py. + const int il = (int) hparams.n_layer - (int) hparams.nextn_predict_layers; + 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"); + + const uint32_t n_head_l = hparams.n_head(il); + const uint32_t n_head_kv_l = hparams.n_head_kv(il); + + const float freq_base_l = model.get_rope_freq_base(cparams, il); + const float freq_scale_l = model.get_rope_freq_scale(cparams, il); + + 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, n_tokens); + ggml_set_input(inp->embd); + ggml_set_name(inp->embd, "mtp_h_input"); + + ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd; + + ggml_tensor * h_input = inp->embd; + ggml_tensor * tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens); + cb(tok_embd, "mtp_tok_embd", il); + + res->add_input(std::move(inp)); + + ggml_tensor * inp_pos = build_inp_pos(); + auto * inp_attn = build_attn_inp_kv_iswa(); + + ggml_tensor * h_norm = build_norm(h_input, 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); + cb(cur, "mtp_eh_proj", il); + + ggml_tensor * inpSA = cur; + + // mtp_block: full Step3p5 decoder layer (attention with optional head-wise gate, then MoE/dense FFN) + cur = build_norm(cur, layer.attn_norm, nullptr, LLM_NORM_RMS, il); + cb(cur, "mtp_attn_norm", il); + + ggml_tensor * Qcur = build_lora_mm(layer.wq, cur, layer.wq_s); + ggml_tensor * Kcur = build_lora_mm(layer.wk, cur, layer.wk_s); + ggml_tensor * Vcur = build_lora_mm(layer.wv, cur, layer.wv_s); + cb(Qcur, "mtp_Qcur", il); + cb(Kcur, "mtp_Kcur", il); + cb(Vcur, "mtp_Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head_l, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv_l, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head_v, n_head_kv_l, n_tokens); + + if (layer.attn_q_norm) { + Qcur = build_norm(Qcur, layer.attn_q_norm, nullptr, LLM_NORM_RMS, il); + cb(Qcur, "mtp_Qcur_normed", il); + } + if (layer.attn_k_norm) { + Kcur = build_norm(Kcur, layer.attn_k_norm, nullptr, LLM_NORM_RMS, il); + cb(Kcur, "mtp_Kcur_normed", il); + } + + const bool is_swa = hparams.is_swa(il); + ggml_tensor * rope_factors = is_swa ? nullptr : model.get_rope_factors(cparams, il); + const int64_t n_rot_l = hparams.n_rot(il); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow); + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow); + cb(Qcur, "mtp_Qcur_pos", il); + cb(Kcur, "mtp_Kcur_pos", il); + + const float kq_scale = 1.0f / sqrtf(float(n_embd_head_k)); + ggml_tensor * attn_out = build_attn(inp_attn, + nullptr, nullptr, nullptr, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(attn_out, "mtp_attn_out", il); + + // head-wise attention gate: sigmoid(g_proj(x)) + if (layer.wqkv_gate) { + ggml_tensor * gate = build_lora_mm(layer.wqkv_gate, cur); // [n_head_l, n_tokens] + cb(gate, "mtp_attn_gate", il); + + gate = ggml_sigmoid(ctx0, gate); + cb(gate, "mtp_attn_gate_sigmoid", il); + + ggml_tensor * attn_3d = ggml_reshape_3d(ctx0, attn_out, n_embd_head_v, n_head_l, n_tokens); + ggml_tensor * gate_3d = ggml_reshape_3d(ctx0, gate, 1, n_head_l, n_tokens); + cb(gate_3d, "mtp_attn_gate_3d", il); + + attn_3d = ggml_mul(ctx0, attn_3d, gate_3d); + cb(attn_3d, "mtp_attn_gated_3d", il); + + attn_out = ggml_reshape_2d(ctx0, attn_3d, n_embd_head_v * n_head_l, n_tokens); + cb(attn_out, "mtp_attn_gated", il); + } + + cur = build_lora_mm(layer.wo, attn_out, layer.wo_s); + cb(cur, "mtp_attn_proj", il); + + cur = ggml_add(ctx0, cur, inpSA); + cb(cur, "mtp_attn_residual", il); + + ggml_tensor * ffn_inp = cur; + cur = build_norm(cur, layer.ffn_norm, nullptr, LLM_NORM_RMS, il); + cb(cur, "mtp_ffn_norm", il); + + // FFN: dense MLP or MoE (mirrors trunk path) + if (layer.ffn_gate_inp == nullptr) { + cur = build_ffn(cur, + layer.ffn_up, layer.ffn_up_b, nullptr, + layer.ffn_gate, layer.ffn_gate_b, nullptr, + layer.ffn_down, layer.ffn_down_b, nullptr, + nullptr, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "mtp_ffn_out", il); + } else { + ggml_tensor * moe_out = build_moe_ffn(cur, + layer.ffn_gate_inp, + layer.ffn_up_exps, + layer.ffn_gate_exps, + layer.ffn_down_exps, + layer.ffn_exp_probs_b, + n_expert, n_expert_used, + LLM_FFN_SILU, hparams.expert_weights_norm, + hparams.expert_weights_scale, + (llama_expert_gating_func_type) hparams.expert_gating_func, + il); + cb(moe_out, "mtp_ffn_moe_out", il); + + ggml_tensor * sh_out = build_ffn(cur, + layer.ffn_up_shexp, nullptr, nullptr, + layer.ffn_gate_shexp, nullptr, nullptr, + layer.ffn_down_shexp, nullptr, nullptr, + nullptr, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(sh_out, "mtp_ffn_shared_out", il); + + cur = ggml_add(ctx0, moe_out, sh_out); + cb(cur, "mtp_ffn_out", il); + } + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "mtp_post_ffn", il); + + // Pre-norm hidden state: used by the AR draft loop to seed the next MTP step. + cb(cur, "h_pre_norm", -1); + res->t_h_pre_norm = cur; + + ggml_tensor * head_norm_w = layer.nextn.shared_head_norm + ? layer.nextn.shared_head_norm + : model.output_norm; + GGML_ASSERT(head_norm_w && "STEP35 MTP: missing both nextn.shared_head_norm and output_norm"); + cur = build_norm(cur, head_norm_w, nullptr, LLM_NORM_RMS, -1); + cb(cur, "mtp_shared_head_norm", -1); + + ggml_tensor * head_w = layer.nextn.shared_head_head ? layer.nextn.shared_head_head : model.output; + GGML_ASSERT(head_w && "STEP35 MTP: missing LM head (nextn.shared_head_head or model.output)"); + cur = build_lora_mm(head_w, cur); + cb(cur, "result_output", -1); + + res->t_logits = cur; + ggml_build_forward_expand(gf, cur); +}