From 60759183094792f4060cdc0556004562191f19f2 Mon Sep 17 00:00:00 2001 From: itigges22 Date: Tue, 17 Mar 2026 16:49:22 -0400 Subject: [PATCH] feat: add MTP (Multi-Token Prediction) support for dense Qwen 3.5 Add native MTP support for the dense Qwen 3.5 architecture (0.8B, 2B, 4B, 9B, 27B). What works: - MTP graph builder for dense qwen35 (build_mtp_head in qwen35.cpp) - MTP tensor loading and registration for QWEN35 arch - GGUF converter handles MTP tensors (mtp.fc, mtp.layers, mtp.norm, etc.) - Public API: llama_get_mtp_logits(), llama_model_n_mtp_layers() - Server auto-detects MTP from GGUF metadata - Speculative state machine for MTP draft token generation - PR #20075 applied: recurrent state checkpoint/restore for hybrid models - M-RoPE position check relaxed for speculative re-evaluation - Windows os.kill fix for gateway process detection What needs work: - Speculative verify loop conflicts with tool-calling requests (400 error) - The recommended fix: bypass the speculative framework entirely and implement MTP acceptance directly in the server generation loop (no seq_rm/rollback needed since MTP drafts are produced in-graph) - MTP attention skipped (projection + FFN path only) due to inp_out_ids token count mismatch Tested on: RTX 5060 8GB, Windows 11, CUDA 13.2 Model: Qwen3.5-9B with MTP tensors (Q4_K_M quantization) Base: llama.cpp b8388 --- common/arg.cpp | 7 +- common/common.h | 1 + common/speculative.cpp | 65 ++++++++++ convert_hf_to_gguf.py | 49 ++++++++ gguf-py/gguf/constants.py | 9 +- include/llama.h | 7 ++ src/llama-arch.cpp | 22 ++-- src/llama-batch.cpp | 5 +- src/llama-context.cpp | 30 +++++ src/llama-context.h | 5 + src/llama-graph.h | 4 + src/llama-memory-recurrent.cpp | 206 ++++++++++++++++++++++++++++---- src/llama-memory-recurrent.h | 4 + src/llama-model.cpp | 103 +++++++++++----- src/models/models.h | 3 + src/models/qwen35.cpp | 133 ++++++++++++++++++++- tools/server/server-context.cpp | 13 +- 17 files changed, 599 insertions(+), 67 deletions(-) diff --git a/common/arg.cpp b/common/arg.cpp index 10aa1b5e4f..ecfbc785b9 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -3463,8 +3463,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex } ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI})); add_opt(common_arg( - {"--spec-type"}, "[none|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v|ngram-mod]", - string_format("type of speculative decoding to use when no draft model is provided (default: %s)\n", + {"--spec-type"}, "[none|mtp|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v|ngram-mod]", + string_format("type of speculative decoding to use when no draft model is provided (default: %s)\n" + " mtp: use model's built-in Multi-Token Prediction head (requires MTP-capable model)\n", common_speculative_type_to_str(params.speculative.type).c_str()), [](common_params & params, const std::string & value) { if (value == "none") { @@ -3479,6 +3480,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V; } else if (value == "ngram-mod") { params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_MOD; + } else if (value == "mtp") { + params.speculative.type = COMMON_SPECULATIVE_TYPE_MTP; } else { throw std::invalid_argument("unknown speculative decoding type without draft model"); } diff --git a/common/common.h b/common/common.h index ee7a2d805e..452f45bc73 100644 --- a/common/common.h +++ b/common/common.h @@ -170,6 +170,7 @@ enum common_speculative_type { COMMON_SPECULATIVE_TYPE_NONE, // no speculative decoding COMMON_SPECULATIVE_TYPE_DRAFT, // draft model COMMON_SPECULATIVE_TYPE_EAGLE3, // eagle draft model + COMMON_SPECULATIVE_TYPE_MTP, // multi-token prediction (uses model's built-in MTP head) COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE, // simple self-speculative decoding COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K, // self-speculative decoding with n-gram keys only COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, // self-speculative decoding with n-gram keys and 4 m-gram values diff --git a/common/speculative.cpp b/common/speculative.cpp index 3e68c38e49..48a121fdef 100644 --- a/common/speculative.cpp +++ b/common/speculative.cpp @@ -21,6 +21,7 @@ const std::vector common_speculative_types = { COMMON_SPECULATIVE_TYPE_NONE, COMMON_SPECULATIVE_TYPE_DRAFT, COMMON_SPECULATIVE_TYPE_EAGLE3, + COMMON_SPECULATIVE_TYPE_MTP, COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE, COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K, COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, @@ -32,6 +33,7 @@ const std::map common_speculative_typ {"none", COMMON_SPECULATIVE_TYPE_NONE}, {"draft", COMMON_SPECULATIVE_TYPE_DRAFT}, {"eagle3", COMMON_SPECULATIVE_TYPE_EAGLE3}, + {"mtp", COMMON_SPECULATIVE_TYPE_MTP}, {"ngram_simple", COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE}, {"ngram_map_k", COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K}, {"ngram_map_k4v", COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V}, @@ -462,6 +464,60 @@ struct common_speculative_state_eagle3 : public common_speculative_state { } }; +// Multi-Token Prediction (MTP) speculative decoding state +struct common_speculative_state_mtp : public common_speculative_state { + llama_context * ctx_tgt; + + common_speculative_state_mtp( + enum common_speculative_type type, + llama_context * ctx_tgt) + : common_speculative_state(type) + , ctx_tgt(ctx_tgt) + { + } + + ~common_speculative_state_mtp() override = default; + + void begin(const llama_tokens & prompt) override { + GGML_UNUSED(prompt); + } + + void draft( + const common_params_speculative & params, + const llama_tokens & prompt_tgt, + llama_token id_last, + llama_tokens & result) override { + GGML_UNUSED(prompt_tgt); + + const float * mtp_logits = llama_get_mtp_logits(ctx_tgt); + if (mtp_logits == nullptr) { + return; + } + + const int n_vocab = llama_vocab_n_tokens(llama_model_get_vocab(llama_get_model(ctx_tgt))); + + llama_token best_token = 0; + float best_logit = mtp_logits[0]; + for (int i = 1; i < n_vocab; ++i) { + if (mtp_logits[i] > best_logit) { + best_logit = mtp_logits[i]; + best_token = i; + } + } + + if (best_token >= 0 && best_token < n_vocab) { + result.push_back(best_token); + } + + GGML_UNUSED(id_last); + GGML_UNUSED(params); + } + + void accept(uint16_t n_accepted) override { + GGML_UNUSED(n_accepted); + } +}; + // state of self-speculation (simple implementation, not ngram-map) struct common_speculative_state_ngram_simple : public common_speculative_state { common_ngram_simple_config config; @@ -781,6 +837,7 @@ std::string common_speculative_type_to_str(enum common_speculative_type type) { case COMMON_SPECULATIVE_TYPE_NONE: return "none"; case COMMON_SPECULATIVE_TYPE_DRAFT: return "draft"; case COMMON_SPECULATIVE_TYPE_EAGLE3: return "eagle3"; + case COMMON_SPECULATIVE_TYPE_MTP: return "mtp"; case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE: return "ngram_simple"; case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K: return "ngram_map_k"; case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V: return "ngram_map_k4v"; @@ -853,6 +910,7 @@ common_speculative * common_speculative_init( { bool has_draft = !params.mparams_dft.path.empty(); bool has_draft_eagle3 = false; // TODO PR-18039: if params.speculative.eagle3 + bool has_mtp = (params.type == COMMON_SPECULATIVE_TYPE_MTP); bool has_ngram_cache = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_CACHE); bool has_ngram_simple = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE); @@ -892,6 +950,9 @@ common_speculative * common_speculative_init( if (has_ngram_cache) { configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_CACHE, params)); } + if (has_mtp) { + configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_MTP, params)); + } if (has_draft) { configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT, params)); } @@ -919,6 +980,10 @@ common_speculative * common_speculative_init( impls.push_back(std::make_unique(config.type)); break; } + case COMMON_SPECULATIVE_TYPE_MTP: { + impls.push_back(std::make_unique(config.type, ctx_tgt)); + break; + } case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE: { common_ngram_map ngram_map = get_common_ngram_map(config); diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 46469c8620..cdbc8239df 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -5033,6 +5033,55 @@ class _LinearAttentionVReorderBase(Qwen3NextModel): class Qwen3_5TextModel(_LinearAttentionVReorderBase): model_arch = gguf.MODEL_ARCH.QWEN35 + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + # If model has MTP layers, include them in block_count + mtp_layers = self.hparams.get("mtp_num_hidden_layers", 0) + if mtp_layers > 0: + self.block_count = self.hparams["num_hidden_layers"] + mtp_layers + self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count) + + def set_gguf_parameters(self): + super().set_gguf_parameters() + mtp_layers = self.hparams.get("mtp_num_hidden_layers", 0) + if mtp_layers > 0: + self.gguf_writer.add_nextn_predict_layers(mtp_layers) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + if name.startswith("mtp."): + num_hidden = self.hparams["num_hidden_layers"] + + if "layers." in name: + # Remap MTP transformer block tensors to append after main layers + # mtp.layers.{k}.* -> model.layers.{k + num_hidden_layers}.* + new_bid = (bid or 0) + num_hidden + name = name.replace(f"mtp.layers.{bid}", f"model.layers.{new_bid}") + yield from super().modify_tensors(data_torch, name, new_bid) + else: + # Shared MTP weights -> nextn tensor slots + from pathlib import Path + remapper = { + "mtp.fc": "model.layers.{bid}.eh_proj", + "mtp.pre_fc_norm_embedding": "model.layers.{bid}.enorm", + "mtp.pre_fc_norm_hidden": "model.layers.{bid}.hnorm", + "mtp.norm": "model.layers.{bid}.shared_head.norm", + } + _n = Path(name) + matched = False + for prefix, template in remapper.items(): + if name.startswith(prefix): + suffix = name[len(prefix):] # e.g. ".weight" + for b in range(num_hidden, self.block_count): + new_name = template.format(bid=b) + suffix + yield from super().modify_tensors(data_torch, new_name, b) + matched = True + break + if not matched: + # Skip unknown MTP tensors (e.g. embed_tokens/lm_head if shared) + pass + return + yield from super().modify_tensors(data_torch, name, bid) + @ModelBase.register("Qwen3_5MoeForConditionalGeneration", "Qwen3_5MoeForCausalLM") class Qwen3_5MoeTextModel(_LinearAttentionVReorderBase): diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 0a032e9039..cbab348809 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -1898,7 +1898,14 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.SSM_NORM, MODEL_TENSOR.SSM_BETA, MODEL_TENSOR.SSM_ALPHA, - MODEL_TENSOR.SSM_OUT + MODEL_TENSOR.SSM_OUT, + # NextN/MTP tensors + 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.QWEN35MOE: [ MODEL_TENSOR.TOKEN_EMBD, diff --git a/include/llama.h b/include/llama.h index c6e102abe5..44584876f0 100644 --- a/include/llama.h +++ b/include/llama.h @@ -557,6 +557,9 @@ extern "C" { LLAMA_API int32_t llama_model_n_head_kv (const struct llama_model * model); LLAMA_API int32_t llama_model_n_swa (const struct llama_model * model); + // Returns the number of Multi-Token Prediction layers (0 if MTP is not available) + LLAMA_API int32_t llama_model_n_mtp_layers(const struct llama_model * model); + // Get the model's RoPE frequency scaling factor LLAMA_API float llama_model_rope_freq_scale_train(const struct llama_model * model); @@ -990,6 +993,10 @@ extern "C" { // returns NULL for invalid ids. LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i); + // Get MTP (Multi-Token Prediction) draft logits for the last output position. + // Returns a pointer to n_vocab floats, or NULL if MTP is not available. + LLAMA_API float * llama_get_mtp_logits(struct llama_context * ctx); + // Get all output token embeddings. // when pooling_type == LLAMA_POOLING_TYPE_NONE or when using a generative model, // the embeddings for which llama_batch.logits[i] != 0 are stored contiguously diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index 84dc6d8f1b..c2b66d408a 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -1051,6 +1051,13 @@ static std::set llm_get_tensor_names(llm_arch arch) { LLM_TENSOR_SSM_ALPHA, LLM_TENSOR_SSM_NORM, LLM_TENSOR_SSM_OUT, + // NextN/MTP tensors + LLM_TENSOR_NEXTN_EH_PROJ, + LLM_TENSOR_NEXTN_EMBED_TOKENS, + LLM_TENSOR_NEXTN_ENORM, + LLM_TENSOR_NEXTN_HNORM, + LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, + LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, }; case LLM_ARCH_QWEN35MOE: return { @@ -2753,14 +2760,13 @@ static const std::map LLM_TENSOR_INFOS = { {LLM_TENSOR_INDEXER_PROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_INDEXER_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_INDEXER_ATTN_Q_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, - // NextN/MTP tensors are currently ignored (reserved for future MTP support) - // These tensors only exist in the last layer(s) and are treated as output tensors - {LLM_TENSOR_NEXTN_EH_PROJ, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, - {LLM_TENSOR_NEXTN_EMBED_TOKENS, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}}, - {LLM_TENSOR_NEXTN_ENORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}}, - {LLM_TENSOR_NEXTN_HNORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}}, - {LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, - {LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}}, + // NextN/MTP tensors — per-layer (appended after main layers) + {LLM_TENSOR_NEXTN_EH_PROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_NEXTN_EMBED_TOKENS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_GET_ROWS}}, + {LLM_TENSOR_NEXTN_ENORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_NEXTN_HNORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, // Nemotron 3 Super {LLM_TENSOR_FFN_LATENT_DOWN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, {LLM_TENSOR_FFN_LATENT_UP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, diff --git a/src/llama-batch.cpp b/src/llama-batch.cpp index 6bf76939cd..c9fe274f0c 100644 --- a/src/llama-batch.cpp +++ b/src/llama-batch.cpp @@ -262,12 +262,13 @@ bool llama_batch_allocr::init( const llama_pos p0 = memory ? memory->seq_pos_max(s) : -1; if (batch.token) { - if (p0 >= 0 && p0 >= seq_pos_min(s)) { + // Allow X == Y for speculative decoding where seq_rm + re-eval at same position is valid + if (p0 >= 0 && p0 > seq_pos_min(s)) { LLAMA_LOG_ERROR( "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n" " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n" " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n" - " for M-RoPE, it is required that the position satisfies: X < Y\n", + " for M-RoPE, it is required that the position satisfies: X <= Y\n", __func__, s, s, p0, s, seq_pos_min(s)); return false; diff --git a/src/llama-context.cpp b/src/llama-context.cpp index 1f7a52d789..82e37034d2 100644 --- a/src/llama-context.cpp +++ b/src/llama-context.cpp @@ -777,6 +777,13 @@ float * llama_context::get_logits() { return logits.data; } +float * llama_context::get_mtp_logits() { + if (!mtp_logits_valid || mtp_logits_buf.empty()) { + return nullptr; + } + return mtp_logits_buf.data(); +} + int64_t llama_context::output_resolve_row(int32_t i) const { int64_t j = -1; @@ -1797,6 +1804,23 @@ int llama_context::decode(const llama_batch & batch_inp) { } } + // Extract MTP logits if available + if (res->t_logits_mtp != nullptr && n_outputs > 0) { + ggml_backend_t backend_mtp = ggml_backend_sched_get_tensor_backend(sched.get(), res->t_logits_mtp); + if (backend_mtp != nullptr) { + const int64_t mtp_n_vocab = res->t_logits_mtp->ne[0]; + const int64_t mtp_n_tokens = res->t_logits_mtp->ne[1]; + + mtp_logits_buf.resize(mtp_n_vocab); + const size_t offset = (mtp_n_tokens - 1) * mtp_n_vocab * sizeof(float); + ggml_backend_tensor_get_async(backend_mtp, res->t_logits_mtp, + mtp_logits_buf.data(), offset, mtp_n_vocab * sizeof(float)); + mtp_logits_valid = true; + } + } else { + mtp_logits_valid = false; + } + // Copy backend sampling output if this ubatch produced any sampling tensors. if (has_samplers && (!res->t_sampled.empty() || !res->t_sampled_probs.empty() || !res->t_sampled_logits.empty())) { const auto seq_to_output_row = build_seq_to_output_row(ubatch, n_outputs_prev); @@ -3079,6 +3103,12 @@ float * llama_get_logits_ith(llama_context * ctx, int32_t i) { return res; } +float * llama_get_mtp_logits(llama_context * ctx) { + ctx->synchronize(); + + return ctx->get_mtp_logits(); +} + float * llama_get_embeddings(llama_context * ctx) { ctx->synchronize(); diff --git a/src/llama-context.h b/src/llama-context.h index e0d0085c1c..faf1c67ec9 100644 --- a/src/llama-context.h +++ b/src/llama-context.h @@ -74,6 +74,7 @@ struct llama_context { float * get_logits(); float * get_logits_ith(int32_t i); + float * get_mtp_logits(); float * get_embeddings(); float * get_embeddings_ith(int32_t i); @@ -268,6 +269,10 @@ private: // decode output (2-dimensional array: [n_outputs][n_vocab]) buffer_view logits = {nullptr, 0}; + // MTP draft logits (1-dimensional array: [n_vocab]) + std::vector mtp_logits_buf; + bool mtp_logits_valid = false; + // embeddings output (2-dimensional array: [n_outputs][n_embd]) // populated only when pooling_type == LLAMA_POOLING_TYPE_NONE buffer_view embd = {nullptr, 0}; diff --git a/src/llama-graph.h b/src/llama-graph.h index 4855685ef7..cadcf56104 100644 --- a/src/llama-graph.h +++ b/src/llama-graph.h @@ -662,6 +662,10 @@ public: ggml_tensor * t_embd = nullptr; ggml_tensor * t_embd_pooled = nullptr; + // MTP (Multi-Token Prediction) output nodes + ggml_tensor * t_logits_mtp = nullptr; // [n_vocab, n_tokens] draft logits from MTP head + ggml_tensor * t_embd_mtp = nullptr; // [n_embd, n_tokens] hidden state from MTP head + std::map t_sampled_logits; std::map t_candidates; std::map t_sampled; diff --git a/src/llama-memory-recurrent.cpp b/src/llama-memory-recurrent.cpp index 6e8413f493..4d62ad9e08 100644 --- a/src/llama-memory-recurrent.cpp +++ b/src/llama-memory-recurrent.cpp @@ -163,12 +163,41 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos const auto & cell = cells[tail_id]; // partial intersection is invalid if it includes the final pos if (0 < p0 && p0 <= cell.pos && p1 > cell.pos) { - //printf("[DEBUG] inside `llama_memory_recurrent::seq_rm`: partial intersection is invalid, so returning false, p0 = %d, cell.pos = %d, p1 = %d\n", p0, cell.pos, p1); - return false; + // for speculative decoding, we search for a checkpoint in the history + int32_t best_cell = -1; + for (uint32_t i = 0; i < size; ++i) { + if (cells[i].has_seq_id(seq_id) && cells[i].pos == p0 - 1) { + best_cell = i; + break; + } + } + + if (best_cell >= 0) { + tail_id = best_cell; + } else { + // No checkpoint found at p0-1: SSM tensor state cannot be rolled back + // without re-evaluating the sequence. Signal failure to the caller. + return false; + } } // invalidate tails which will be cleared if (p0 <= cell.pos && cell.pos < p1) { - tail_id = -1; + if (p0 == 0) { + tail_id = -1; + } else { + // Search for the best remaining cell after removal + int32_t new_tail = -1; + llama_pos max_pos = -1; + for (uint32_t i = 0; i < size; ++i) { + if (cells[i].has_seq_id(seq_id) && cells[i].pos < p0) { + if (cells[i].pos > max_pos) { + max_pos = cells[i].pos; + new_tail = i; + } + } + } + tail_id = new_tail; + } } } } else { @@ -184,6 +213,11 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos if (seq_id < 0) { cells[i].seq_id.clear(); } else if (cells[i].has_seq_id(seq_id)) { + if (p0 > 0 && p1 == std::numeric_limits::max()) { + // partial removal: just move the position back + cells[i].pos = p0 - 1; + continue; + } cells[i].seq_id.erase(seq_id); } else { continue; @@ -224,25 +258,42 @@ void llama_memory_recurrent::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id } if ((uint32_t) seq_id_dst < size && (uint32_t) seq_id_src < size) { - auto & tail_src = cells[seq_id_src]; - auto & tail_dst = cells[seq_id_dst]; - if (tail_dst.tail >= 0) { + auto & tail_src_meta = cells[seq_id_src]; + auto & tail_dst_meta = cells[seq_id_dst]; + + if (tail_dst_meta.tail >= 0) { // clear destination seq_id if it wasn't empty - auto & cell_dst = cells[tail_dst.tail]; - - cell_dst.seq_id.erase(seq_id_dst); - tail_dst.tail = -1; - if (cell_dst.seq_id.empty()) { - cell_dst.pos = -1; - cell_dst.src = -1; - used -= 1; - } + seq_rm(seq_id_dst, -1, -1); } - if (tail_src.tail >= 0) { - auto & cell_src = cells[tail_src.tail]; - cell_src.seq_id.insert(seq_id_dst); - tail_dst.tail = tail_src.tail; + if (tail_src_meta.tail >= 0) { + auto & cell_src = cells[tail_src_meta.tail]; + + // For recurrent models, we must copy the state to a new cell + // Otherwise, both sequences would share the same mutable state + uint32_t next_empty_cell = size; + for (uint32_t i = head; i < head + size; ++i) { + uint32_t idx = i % size; + if (cells[idx].is_empty()) { + next_empty_cell = idx; + break; + } + } + + if (next_empty_cell != size) { + auto & empty_cell = cells[next_empty_cell]; + + // Copy tensors data + copy_cell(tail_src_meta.tail, next_empty_cell); + + empty_cell.pos = cell_src.pos; + empty_cell.src = next_empty_cell; // results in a copy in the graph if needed + empty_cell.seq_id.insert(seq_id_dst); + tail_dst_meta.tail = next_empty_cell; + used += 1; + } else { + LLAMA_LOG_ERROR("%s: failed to find available cell for copy\n", __func__); + } } } } @@ -367,6 +418,47 @@ llama_pos llama_memory_recurrent::seq_pos_max(llama_seq_id seq_id) const { return result; } +void llama_memory_recurrent::copy_cell(int32_t i_src, int32_t i_dst) { + if (i_src == i_dst || i_src < 0 || i_dst < 0) { + return; + } + + ggml_init_params params = { + /*.mem_size =*/ size_t(2*ggml_tensor_overhead()), + /*.mem_buffer =*/ NULL, + /*.no_alloc =*/ true, + }; + + for (uint32_t il = 0; il < hparams.n_layer; ++il) { + if (r_l[il]) { + ggml_context * ctx = ggml_init(params); + size_t r_row_size = ggml_row_size(r_l[il]->type, hparams.n_embd_r()); + ggml_tensor * src_v = ggml_view_1d(ctx, r_l[il], r_row_size, i_src * r_row_size); + ggml_tensor * dst_v = ggml_view_1d(ctx, r_l[il], r_row_size, i_dst * r_row_size); + ggml_backend_tensor_copy(src_v, dst_v); + ggml_free(ctx); + } + if (s_l[il]) { + ggml_context * ctx = ggml_init(params); + size_t s_row_size = ggml_row_size(s_l[il]->type, hparams.n_embd_s()); + ggml_tensor * src_v = ggml_view_1d(ctx, s_l[il], s_row_size, i_src * s_row_size); + ggml_tensor * dst_v = ggml_view_1d(ctx, s_l[il], s_row_size, i_dst * s_row_size); + ggml_backend_tensor_copy(src_v, dst_v); + ggml_free(ctx); + } + } +} + +int llama_memory_recurrent::get_cell_count(llama_seq_id seq_id) const { + int count = 0; + for (uint32_t i = 0; i < size; ++i) { + if (cells[i].has_seq_id(seq_id)) { + count++; + } + } + return count; +} + std::map llama_memory_recurrent::memory_breakdown() const { std::map ret; for (const auto & [_, buf] : ctxs_bufs) { @@ -551,10 +643,35 @@ bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) { if (seq_meta.tail >= 0) { auto & orig_cell = cells[seq_meta.tail]; empty_cell.pos = orig_cell.pos; - empty_cell.src = orig_cell.src; - orig_cell.seq_id.erase(seq_id); + empty_cell.src = seq_meta.tail; // the data should be copied from the previous tail + + // Copy state data + copy_cell(seq_meta.tail, next_empty_cell); + + // Keep history of previous states for rollback (up to 8 cells per sequence) + if (get_cell_count(seq_id) < 8 && used < size * 0.9) { + // Do not erase seq_id from orig_cell to keep it as a checkpoint + } else { + // Erase oldest history point for this sequence + int32_t oldest_cell = -1; + llama_pos min_pos = std::numeric_limits::max(); + for (uint32_t i = 0; i < size; ++i) { + if (cells[i].has_seq_id(seq_id) && cells[i].pos < min_pos) { + min_pos = cells[i].pos; + oldest_cell = i; + } + } + + if (oldest_cell >= 0) { + cells[oldest_cell].seq_id.erase(seq_id); + if (cells[oldest_cell].is_empty()) { + cells[oldest_cell].pos = -1; + cells[oldest_cell].src = -1; + used--; + } + } + } empty_cell.seq_id.insert(seq_id); // will be overwritten - GGML_ASSERT(!orig_cell.is_empty()); // has at least one remaining seq_id } seq_meta.tail = next_empty_cell; // find next empty cell @@ -566,6 +683,51 @@ bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) { if (cell.is_empty()) { break; } } } + } else { + // Sequence owns its cell. Save a checkpoint of the current state before it is + // overwritten by new tokens. This is required for speculative decoding rollback + // in recurrent/SSM models where tensor state cannot be partially rewound. + const int32_t cur_tail = seq_meta.tail; + if (cells[next_empty_cell].is_empty()) { + bool can_checkpoint = (get_cell_count(seq_id) < 8 && used < size * 0.9); + if (!can_checkpoint) { + // Try to evict the oldest checkpoint to make room + int32_t oldest = -1; + llama_pos min_pos = std::numeric_limits::max(); + for (uint32_t j = 0; j < size; ++j) { + if ((int32_t)j != cur_tail && cells[j].has_seq_id(seq_id) && cells[j].pos < min_pos) { + min_pos = cells[j].pos; + oldest = j; + } + } + if (oldest >= 0) { + cells[oldest].seq_id.erase(seq_id); + if (cells[oldest].is_empty()) { + cells[oldest].pos = -1; + cells[oldest].src = -1; + used--; + } + can_checkpoint = true; + } + } + if (can_checkpoint) { + auto & cp_cell = cells[next_empty_cell]; + copy_cell(cur_tail, next_empty_cell); + cp_cell.pos = cells[cur_tail].pos; + cp_cell.src = next_empty_cell; // independent copy, no further movement needed + cp_cell.seq_id.insert(seq_id); + used++; + // advance next_empty_cell for subsequent sequences in this batch + if (s + 1 < n_seqs) { + for (uint32_t j = 0; j < size; ++j) { + next_empty_cell += 1; + if (next_empty_cell >= size) { next_empty_cell -= size; } + if (cells[next_empty_cell].is_empty()) { break; } + } + } + } + } + // seq_meta.tail remains unchanged - sequence still owns its current cell } if (min > seq_meta.tail) { min = seq_meta.tail; } if (max < seq_meta.tail) { max = seq_meta.tail; } diff --git a/src/llama-memory-recurrent.h b/src/llama-memory-recurrent.h index 47f01d7391..b6b5d6cfbd 100644 --- a/src/llama-memory-recurrent.h +++ b/src/llama-memory-recurrent.h @@ -65,6 +65,10 @@ public: void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override; void state_read (llama_io_read_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override; + // cell management + void copy_cell(int32_t i_src, int32_t i_dst); + int get_cell_count(llama_seq_id seq_id) const; + uint32_t head = 0; // the location where the batch will be placed in the cache (see find_slot()) uint32_t size = 0; // total number of cells, shared across all sequences uint32_t used = 0; // used cells (i.e. at least one seq_id) diff --git a/src/llama-model.cpp b/src/llama-model.cpp index 85db938a7a..c29b466ba3 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -2403,16 +2403,29 @@ void llama_model::load_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 parameters + ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.nextn_predict_layers, false); + + // The total n_layer includes MTP layers appended after main layers. + // Determine the number of main transformer layers for type detection. + const uint32_t n_main_layers = hparams.n_layer - hparams.nextn_predict_layers; + + // Mark recurrent layers (linear attention layers) — main layers only + // MTP layers use full attention, so they are NOT recurrent { uint32_t full_attn_interval = 4; ml.get_key(LLM_KV_FULL_ATTENTION_INTERVAL, full_attn_interval, false); for (uint32_t i = 0; i < hparams.n_layer; ++i) { - hparams.recurrent_layer_arr[i] = ((i + 1) % full_attn_interval != 0); + if (i < n_main_layers) { + hparams.recurrent_layer_arr[i] = ((i + 1) % full_attn_interval != 0); + } else { + // MTP layers use full attention (not recurrent) + hparams.recurrent_layer_arr[i] = false; + } } } - switch (hparams.n_layer) { + switch (n_main_layers) { case 24: type = hparams.n_embd == 1024 ? LLM_TYPE_0_8B : LLM_TYPE_2B; break; case 32: type = hparams.n_embd == 2560 ? LLM_TYPE_4B : LLM_TYPE_9B; break; case 64: type = LLM_TYPE_27B; break; @@ -7272,39 +7285,67 @@ bool llama_model::load_tensors(llama_model_loader & ml) { const int64_t value_dim = head_v_dim * n_v_heads; const int64_t conv_dim = key_dim * 2 + value_dim; + const uint32_t n_main_layers = n_layer - hparams.nextn_predict_layers; + for (int i = 0; i < n_layer; ++i) { auto & layer = layers[i]; - 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); + const bool is_mtp_layer = (static_cast(i) >= n_main_layers); - if (!hparams.is_recurrent(i)) { - // Attention layers - layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head * 2 }, 0); - layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_k_gqa }, 0); - layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, 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); + if (is_mtp_layer) { + // MTP layer: nextn-specific tensors + standard attention + standard FFN + layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, 0); + layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, 0); + layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, 0); + layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, 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.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED); - // 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); + // MTP layer uses same gated attention as main model (joint QG projection) + 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.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head * 2 }, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_k_gqa }, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, 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); + 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); + + // MTP layer uses standard dense FFN + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); + 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, n_ff}, 0); } else { - // Linear attention (gated delta net) specific tensors - // Create tensors with calculated dimensions - 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 = create_tensor(tn(LLM_TENSOR_SSM_BETA, "weight", i), { n_embd, n_v_heads }, 0); - layer.ssm_alpha = create_tensor(tn(LLM_TENSOR_SSM_ALPHA, "weight", i), { n_embd, n_v_heads }, 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); - } + // Main transformer layers + 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.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); - 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, n_ff}, 0); + if (!hparams.is_recurrent(i)) { + // Full attention layers (joint QG projection + gated attention) + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head * 2 }, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_k_gqa }, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, 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); + + 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); + } else { + // Linear attention (gated delta net) specific tensors + 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 = create_tensor(tn(LLM_TENSOR_SSM_BETA, "weight", i), { n_embd, n_v_heads }, 0); + layer.ssm_alpha = create_tensor(tn(LLM_TENSOR_SSM_ALPHA, "weight", i), { n_embd, n_v_heads }, 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.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); + 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, n_ff}, 0); + } } } break; case LLM_ARCH_MIMO2: @@ -8755,6 +8796,10 @@ int32_t llama_model_n_swa(const llama_model * model) { return model->hparams.n_swa; } +int32_t llama_model_n_mtp_layers(const llama_model * model) { + return model->hparams.nextn_predict_layers; +} + uint32_t llama_model_n_cls_out(const struct llama_model * model) { return model->hparams.n_cls_out; } diff --git a/src/models/models.h b/src/models/models.h index a86b2b1ebd..ec6c14b5d4 100644 --- a/src/models/models.h +++ b/src/models/models.h @@ -597,6 +597,9 @@ private: ggml_tensor * input, int il); + // Build the MTP (Multi-Token Prediction) head with standard transformer block + void build_mtp_head(llm_graph_input_mem_hybrid * inp, ggml_tensor * inp_pos, int * sections); + const llama_model & model; }; diff --git a/src/models/qwen35.cpp b/src/models/qwen35.cpp index d07579ee87..def7ffb7ee 100644 --- a/src/models/qwen35.cpp +++ b/src/models/qwen35.cpp @@ -23,7 +23,10 @@ llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_pa ggml_tensor * inp_pos = build_inp_pos(); ggml_tensor * inp_out_ids = build_inp_out_ids(); - for (int il = 0; il < n_layer; ++il) { + // Only process main transformer layers (skip MTP layers appended at the end) + const int n_transformer_layers = n_layer - hparams.nextn_predict_layers; + + for (int il = 0; il < n_transformer_layers; ++il) { ggml_tensor * inpSA = inpL; cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il); @@ -40,7 +43,7 @@ llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_pa cur = build_layer_attn(inp->get_attn(), cur, inp_pos, sections, il); } - if (il == n_layer - 1 && inp_out_ids) { + if (il == n_transformer_layers - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -82,6 +85,11 @@ llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_pa res->t_logits = cur; ggml_build_forward_expand(gf, cur); + + // Build MTP head if nextn_predict_layers > 0 + if (hparams.nextn_predict_layers > 0) { + build_mtp_head(inp, inp_pos, sections); + } } std::pair llm_build_qwen35::build_qkvz( @@ -379,3 +387,124 @@ ggml_tensor * llm_build_qwen35::build_layer_ffn(ggml_tensor * cur, const int il) return cur; } + +void llm_build_qwen35::build_mtp_head( + llm_graph_input_mem_hybrid * inp, + ggml_tensor * inp_pos, + int * sections) { + // MTP (Multi-Token Prediction) head for dense Qwen 3.5 + // + // The MTP module takes the hidden state from the last main transformer layer + // and uses the model's built-in MTP head to produce draft logits. + // + // MTP forward pass: + // 1. sampled_token = argmax(main_logits) + // 2. emb = embed_tokens(sampled_token) + // 3. h_norm = RMSNorm(hidden_state, hnorm) + // 4. e_norm = RMSNorm(emb, enorm) + // 5. combined = eh_proj(concat(e_norm, h_norm)) + // 6. Standard self-attention (Q/K/V with Q/K norms + RoPE) + // 7. Standard FFN (gate_proj + up_proj → SiLU → down_proj) + // 8. logits = lm_head(RMSNorm(output, mtp_norm)) + + const int n_transformer_layers = n_layer - hparams.nextn_predict_layers; + const int64_t n_embd_head = hparams.n_embd_head_v(); + + // Main model logits and hidden state + ggml_tensor * main_logits = res->t_logits; // [n_vocab, n_tokens] + ggml_tensor * hidden_state = res->t_embd; // [n_embd, n_tokens] after final norm + GGML_ASSERT(main_logits != nullptr); + + // In-graph greedy token selection + ggml_tensor * greedy_tokens = ggml_argmax(ctx0, main_logits); // [n_tokens] + cb(greedy_tokens, "mtp_greedy_tokens", -1); + + ggml_tensor * mtp_hidden = hidden_state; + + for (uint32_t k = 0; k < hparams.nextn_predict_layers; ++k) { + const int il = n_transformer_layers + k; + const auto & layer = model.layers[il]; + + if (layer.nextn.eh_proj == nullptr) { + continue; + } + + // Step 1: Get token embedding (shared with main model) + ggml_tensor * tok_embd = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd; + ggml_tensor * emb = ggml_get_rows(ctx0, tok_embd, greedy_tokens); + cb(emb, "mtp_token_embd", il); + + // Step 2: Normalize hidden state and embedding + ggml_tensor * h_norm = build_norm(mtp_hidden, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il); + cb(h_norm, "mtp_hnorm", il); + + ggml_tensor * e_norm = build_norm(emb, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il); + cb(e_norm, "mtp_enorm", il); + + // Step 3: Concatenate and project + ggml_tensor * concat = ggml_concat(ctx0, e_norm, h_norm, 0); // [2*n_embd, n_tokens] + cb(concat, "mtp_concat", il); + + ggml_tensor * cur = build_lora_mm(layer.nextn.eh_proj, concat); + cb(cur, "mtp_projected", il); + + // Step 4: Skip attention for MTP head (MTP operates on already-contextualized hidden states + // from the main model's final layer, so the projection + FFN path is sufficient for draft + // token generation. Full MTP attention requires resolving inp_out_ids token count mismatch.) + // TODO: implement MTP-specific attention that handles filtered token counts + { + // The projection through eh_proj already combines embedding + hidden context + // Just pass through with a norm (attention weights are loaded but unused for now) + (void)inp_pos; + (void)sections; + } + + // Step 5: Post-attention norm + FFN + { + ggml_tensor * ffn_residual = cur; + + ggml_tensor * attn_post_norm = build_norm(cur, layer.attn_post_norm, nullptr, LLM_NORM_RMS, il); + cb(attn_post_norm, "mtp_attn_post_norm", il); + + // Standard dense FFN (same as main model FFN) + cur = build_ffn(attn_post_norm, + layer.ffn_up, NULL, layer.ffn_up_s, + layer.ffn_gate, NULL, layer.ffn_gate_s, + layer.ffn_down, NULL, layer.ffn_down_s, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "mtp_ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_residual); + cb(cur, "mtp_post_ffn", il); + } + + mtp_hidden = cur; + + // Step 6: Final norm + LM head for draft logits + ggml_tensor * mtp_normed; + if (layer.nextn.shared_head_norm != nullptr) { + mtp_normed = build_norm(mtp_hidden, layer.nextn.shared_head_norm, nullptr, LLM_NORM_RMS, il); + } else { + // Use main model's output norm + mtp_normed = build_norm(mtp_hidden, model.output_norm, nullptr, LLM_NORM_RMS, il); + } + cb(mtp_normed, "mtp_head_norm", il); + + ggml_tensor * lm_head = layer.nextn.shared_head_head ? layer.nextn.shared_head_head : model.output; + ggml_tensor * mtp_logits = build_lora_mm(lm_head, mtp_normed); + cb(mtp_logits, "mtp_logits", il); + + // Store MTP outputs in graph result + res->t_embd_mtp = mtp_hidden; + res->t_logits_mtp = mtp_logits; + + // For recursive MTP (multiple layers), feed greedy tokens forward + if (k + 1 < hparams.nextn_predict_layers) { + greedy_tokens = ggml_argmax(ctx0, mtp_logits); + cb(greedy_tokens, "mtp_greedy_next", il); + } + + ggml_build_forward_expand(gf, mtp_logits); + } +} diff --git a/tools/server/server-context.cpp b/tools/server/server-context.cpp index c47ad876cb..60fae14281 100644 --- a/tools/server/server-context.cpp +++ b/tools/server/server-context.cpp @@ -752,11 +752,22 @@ private: slots.clear(); - const bool can_spec = common_speculative_is_compat(ctx); + bool can_spec = common_speculative_is_compat(ctx); if (!can_spec) { SRV_WRN("%s", "speculative decoding not supported by this context\n"); } + // Auto-detect MTP capability — log presence but don't enable speculative + // decoding framework. The hybrid SSM + M-RoPE architecture is incompatible + // with the speculative verify loop when tool-calling is active. + // MTP tensors are still computed in the forward pass graph (build_mtp_head). + if (params_base.speculative.type == COMMON_SPECULATIVE_TYPE_NONE) { + const int32_t n_mtp = llama_model_n_mtp_layers(llama_get_model(ctx)); + if (n_mtp > 0) { + SRV_INF("model has %d MTP layer(s) (graph-only, speculative verify disabled for hybrid models)\n", n_mtp); + } + } + // initialize slots for (int i = 0; i < params_base.n_parallel; i++) { server_slot slot;