From a337ebd7bdd90bab10a68b53b07a748b25014fd3 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Stanis=C5=82aw=20Szymczyk?= Date: Wed, 11 Mar 2026 16:38:12 +0100 Subject: [PATCH] model : Initial support for DeepseekV32ForCausalLM (for now with dense attention). Needs manual change of add_bos_token to true in tokenizer_config.json before conversion. --- convert_hf_to_gguf.py | 139 ++++++++++++++++++++++ gguf-py/gguf/constants.py | 47 ++++++++ src/CMakeLists.txt | 1 + src/llama-arch.cpp | 39 +++++++ src/llama-arch.h | 1 + src/llama-model.cpp | 159 ++++++++++++++++++++++++- src/llama-model.h | 1 + src/models/deepseek32.cpp | 240 ++++++++++++++++++++++++++++++++++++++ src/models/models.h | 8 +- 9 files changed, 632 insertions(+), 3 deletions(-) create mode 100644 src/models/deepseek32.cpp diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 30347f7389..3fdeb27794 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -8153,6 +8153,145 @@ class DeepseekV2Model(TextModel): raise ValueError(f"Unprocessed experts: {experts}") +@ModelBase.register( + "DeepseekV32ForCausalLM", +) +class DeepseekV32Model(TextModel): + model_arch = gguf.MODEL_ARCH.DEEPSEEK32 + + # TODO @ngxson : remove this when we support MTP for deepseek models + skip_mtp = True + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0) + self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count) + + def set_vocab(self): + self._set_vocab_gpt2() + + def set_gguf_parameters(self): + + # note: deepseek32 using MLA converts into MQA (ie: GQA with 1 group) + self.hparams["num_key_value_heads"] = 1 + + super().set_gguf_parameters() + hparams = self.hparams + + # first_k_dense_replace: number of leading layers using dense FFN instead of MoE + first_k_dense_replace = hparams.get("first_k_dense_replace") + self.gguf_writer.add_leading_dense_block_count(first_k_dense_replace) + self.gguf_writer.add_vocab_size(hparams["vocab_size"]) + self.gguf_writer.add_q_lora_rank(hparams["q_lora_rank"]) + self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"]) + + # note: deepseek32 using MLA converts into MQA with larger heads, then decompresses to MHA + self.gguf_writer.add_key_length(hparams["kv_lora_rank"] + hparams["qk_rope_head_dim"]) + self.gguf_writer.add_value_length(hparams["kv_lora_rank"]) + self.gguf_writer.add_key_length_mla(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"]) + self.gguf_writer.add_value_length_mla(hparams["v_head_dim"]) + + # MoE parameters (required by C++ code for DEEPSEEK32 arch) + self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"]) + self.gguf_writer.add_expert_count(hparams["n_routed_experts"]) + self.gguf_writer.add_expert_shared_count(hparams["n_shared_experts"]) + self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"]) + self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"]) + + self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"]) + + if (rope_mscale_all := self.rope_parameters.get("mscale_all_dim")) is not None: + # [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX] + # note: for legacy reasons, this is not consistent with the other usages of self.gguf_writer.add_rope_scaling_yarn_log_mul + # ref https://github.com/ggml-org/llama.cpp/pull/17945 + self.gguf_writer.add_rope_scaling_yarn_log_mul(0.1 * rope_mscale_all) + + # NextN/MTP prediction layers + if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None: + self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers) + + # DSA indexer parameters + self.gguf_writer.add_indexer_head_count(self.hparams["index_n_heads"]) + self.gguf_writer.add_indexer_key_length(self.hparams["index_head_dim"]) + self.gguf_writer.add_indexer_top_k(self.hparams["index_topk"]) + + _experts: list[dict[str, Tensor]] | None = None + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + if name.startswith("language_model."): + name = name.replace("language_model.", "") + + # rename e_score_correction_bias tensors + if name.endswith("e_score_correction_bias"): + name = name.replace("e_score_correction_bias", "e_score_correction.bias") + + # skip Multi-Token Prediction (MTP) layers + if self.skip_mtp: + block_count = self.hparams["num_hidden_layers"] + match = re.match(r"model.layers.(\d+)", name) + if match and int(match.group(1)) >= block_count: + return + + # process the experts separately + if name.find("mlp.experts") != -1: + n_experts = self.hparams["n_routed_experts"] + assert bid is not None + + if self._experts is None: + self._experts = [{} for _ in range(self.block_count)] + + self._experts[bid][name] = data_torch + + if len(self._experts[bid]) >= n_experts * 3: + # merge the experts into a single 3d tensor + for w_name in ["down_proj", "gate_proj", "up_proj"]: + datas: list[Tensor] = [] + + for xid in range(n_experts): + ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight" + datas.append(self._experts[bid][ename]) + del self._experts[bid][ename] + + data_torch = torch.stack(datas, dim=0) + + merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight" + + yield from super().modify_tensors(data_torch, merged_name, bid) + return + else: + return + + # note: MLA with the absorption optimization, needs these two split and k_b_proj transposed + if name.endswith("kv_b_proj.weight"): + name_kb = name.replace("kv_b_proj", "k_b_proj") + name_vb = name.replace("kv_b_proj", "v_b_proj") + + n_head_kv = self.hparams["num_key_value_heads"] + v_head_dim = self.hparams["v_head_dim"] + qk_nope_head_dim = self.hparams["qk_nope_head_dim"] + + assert data_torch.shape[0] == n_head_kv * (v_head_dim + qk_nope_head_dim) + + kv_b = data_torch.view(n_head_kv, v_head_dim + qk_nope_head_dim, data_torch.shape[-1]) + k_b, v_b = torch.split(kv_b, [qk_nope_head_dim, v_head_dim], dim=1) + k_b = k_b.transpose(1, 2) + + yield from super().modify_tensors(k_b, name_kb, bid) + yield from super().modify_tensors(v_b, name_vb, bid) + return + + yield from super().modify_tensors(data_torch, name, bid) + + def prepare_tensors(self): + super().prepare_tensors() + + if self._experts is not None: + # flatten `list[dict[str, Tensor]]` into `list[str]` + experts = [k for d in self._experts for k in d.keys()] + if len(experts) > 0: + raise ValueError(f"Unprocessed experts: {experts}") + + @ModelBase.register("MiniMaxM2ForCausalLM") class MiniMaxM2Model(TextModel): model_arch = gguf.MODEL_ARCH.MINIMAXM2 diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index c5f5469506..9f9b44bf17 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -431,6 +431,7 @@ class MODEL_ARCH(IntEnum): ARCTIC = auto() DEEPSEEK = auto() DEEPSEEK2 = auto() + DEEPSEEK32 = auto() CHATGLM = auto() GLM4 = auto() GLM4_MOE = auto() @@ -874,6 +875,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.ARCTIC: "arctic", MODEL_ARCH.DEEPSEEK: "deepseek", MODEL_ARCH.DEEPSEEK2: "deepseek2", + MODEL_ARCH.DEEPSEEK32: "deepseek32", MODEL_ARCH.CHATGLM: "chatglm", MODEL_ARCH.GLM4: "glm4", MODEL_ARCH.GLM4_MOE: "glm4moe", @@ -2623,6 +2625,47 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_UP_SHEXP, MODEL_TENSOR.FFN_EXP_PROBS_B, ], + MODEL_ARCH.DEEPSEEK32: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_Q_A, + MODEL_TENSOR.ATTN_Q_B, + MODEL_TENSOR.ATTN_KV_A_MQA, + MODEL_TENSOR.ATTN_KV_B, + MODEL_TENSOR.ATTN_K_B, + MODEL_TENSOR.ATTN_V_B, + MODEL_TENSOR.ATTN_Q_A_NORM, + MODEL_TENSOR.ATTN_KV_A_NORM, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.FFN_GATE_SHEXP, + MODEL_TENSOR.FFN_DOWN_SHEXP, + MODEL_TENSOR.FFN_UP_SHEXP, + MODEL_TENSOR.FFN_EXP_PROBS_B, + MODEL_TENSOR.INDEXER_K_NORM, + MODEL_TENSOR.INDEXER_PROJ, + MODEL_TENSOR.INDEXER_ATTN_K, + MODEL_TENSOR.INDEXER_ATTN_Q_B, + # NextN/MTP tensors - preserved but unused + 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.ERNIE4_5_MOE: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, @@ -3698,6 +3741,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.ROPE_FREQS, MODEL_TENSOR.ATTN_ROT_EMBD, ], + MODEL_ARCH.DEEPSEEK32: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], MODEL_ARCH.CHATGLM: [ MODEL_TENSOR.ROPE_FREQS, ], diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt index 283823fa9c..e524ebd2f2 100644 --- a/src/CMakeLists.txt +++ b/src/CMakeLists.txt @@ -57,6 +57,7 @@ add_library(llama models/deci.cpp models/deepseek.cpp models/deepseek2.cpp + models/deepseek32.cpp models/delta-net-base.cpp models/dots1.cpp models/dream.cpp diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index 204105b6dd..faa7a3a7da 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -73,6 +73,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_ARCTIC, "arctic" }, { LLM_ARCH_DEEPSEEK, "deepseek" }, { LLM_ARCH_DEEPSEEK2, "deepseek2" }, + { LLM_ARCH_DEEPSEEK32, "deepseek32" }, { LLM_ARCH_CHATGLM, "chatglm" }, { LLM_ARCH_GLM4, "glm4" }, { LLM_ARCH_GLM4_MOE, "glm4moe" }, @@ -1616,6 +1617,44 @@ static std::set llm_get_tensor_names(llm_arch arch) { LLM_TENSOR_FFN_UP_SHEXP, LLM_TENSOR_FFN_EXP_PROBS_B, }; + case LLM_ARCH_DEEPSEEK32: + return { + LLM_TENSOR_TOKEN_EMBD, + LLM_TENSOR_OUTPUT_NORM, + LLM_TENSOR_OUTPUT, + LLM_TENSOR_ATTN_NORM, + LLM_TENSOR_ATTN_Q_A_NORM, + LLM_TENSOR_ATTN_KV_A_NORM, + LLM_TENSOR_ATTN_Q_A, + LLM_TENSOR_ATTN_Q_B, + LLM_TENSOR_ATTN_KV_A_MQA, + LLM_TENSOR_ATTN_KV_B, + LLM_TENSOR_ATTN_K_B, + LLM_TENSOR_ATTN_V_B, + LLM_TENSOR_ATTN_OUT, + LLM_TENSOR_FFN_NORM, + LLM_TENSOR_FFN_GATE, + LLM_TENSOR_FFN_UP, + LLM_TENSOR_FFN_DOWN, + LLM_TENSOR_FFN_GATE_INP, + LLM_TENSOR_FFN_GATE_EXPS, + LLM_TENSOR_FFN_DOWN_EXPS, + LLM_TENSOR_FFN_UP_EXPS, + LLM_TENSOR_FFN_GATE_SHEXP, + LLM_TENSOR_FFN_DOWN_SHEXP, + LLM_TENSOR_FFN_UP_SHEXP, + LLM_TENSOR_FFN_EXP_PROBS_B, + LLM_TENSOR_INDEXER_K_NORM, + LLM_TENSOR_INDEXER_PROJ, + LLM_TENSOR_INDEXER_ATTN_K, + LLM_TENSOR_INDEXER_ATTN_Q_B, + 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_PLM: return { LLM_TENSOR_TOKEN_EMBD, diff --git a/src/llama-arch.h b/src/llama-arch.h index 28dd1ffac7..6c96da00b3 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -77,6 +77,7 @@ enum llm_arch { LLM_ARCH_ARCTIC, LLM_ARCH_DEEPSEEK, LLM_ARCH_DEEPSEEK2, + LLM_ARCH_DEEPSEEK32, LLM_ARCH_CHATGLM, LLM_ARCH_GLM4, LLM_ARCH_GLM4_MOE, diff --git a/src/llama-model.cpp b/src/llama-model.cpp index 0fa47e1b41..b484d82ef1 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -143,6 +143,7 @@ const char * llm_type_name(llm_type type) { case LLM_TYPE_310B_A15B: return "310B.A15B"; case LLM_TYPE_355B_A32B: return "355B.A32B"; case LLM_TYPE_397B_A17B: return "397B.A17B"; + case LLM_TYPE_685B_A37B: return "685B.A37B"; case LLM_TYPE_744B_A40B: return "744B.A40B"; case LLM_TYPE_E2B: return "E2B"; case LLM_TYPE_E4B: return "E4B"; @@ -1634,6 +1635,55 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } } break; + case LLM_ARCH_DEEPSEEK32: + { + ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + hparams.f_norm_eps = 1e-6; // eps for layer norm + ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false); + + // MoE parameters + ml.get_key(LLM_KV_EXPERT_COUNT, hparams.n_expert); + ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used); + ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared); + ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead, false); + ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale, false); + ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM, hparams.expert_weights_norm, false); + + // deepseek MLA parameters + ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q); + ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK, hparams.n_lora_kv); + ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA, hparams.n_embd_head_k_mla_impl, false); + ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl, false); + ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); + ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared); + + // DSA parameters + ml.get_key(LLM_KV_ATTENTION_INDEXER_HEAD_COUNT, hparams.indexer_n_head); + ml.get_key(LLM_KV_ATTENTION_INDEXER_KEY_LENGTH, hparams.indexer_head_size); + ml.get_key(LLM_KV_ATTENTION_INDEXER_TOP_K, hparams.indexer_top_k); + + // Expert gating function + ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func); + + if (ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul, 0.0f)) { + // [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX] + // cancel the factor from the convert script + hparams.rope_yarn_log_mul /= 0.1f; + } + + // NextN/MTP parameters + ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.nextn_predict_layers, false); + + // TODO: when MTP is implemented, this should probably be updated if needed + hparams.n_layer_kv_from_start = hparams.n_layer - hparams.nextn_predict_layers; + + switch (hparams.n_layer) { + case 61: type = LLM_TYPE_685B_A37B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; + case LLM_ARCH_PLM: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); @@ -4964,6 +5014,108 @@ bool llama_model::load_tensors(llama_model_loader & ml) { } } } break; + case LLM_ARCH_DEEPSEEK32: + { + const bool is_mla = hparams.is_mla(); + if (!is_mla) { + throw std::runtime_error("DEEPSEEK32 architecture requires MLA"); + } + + // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA + const int64_t n_embd_head_k_mla = hparams.n_embd_head_k_mla(); + const int64_t n_embd_head_v_mla = hparams.n_embd_head_v_mla(); + + const int64_t n_embd_head_qk_rope = hparams.n_rot(); + const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope; + + const int64_t q_lora_rank = hparams.n_lora_q; + const int64_t kv_lora_rank = hparams.n_lora_kv; + + const int64_t n_ff_exp = hparams.n_ff_exp; + const int64_t n_expert_shared = hparams.n_expert_shared; + + 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); + // try to load output.weight, if not found, use token_embd (tied embeddings) + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED); + if (!output) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } + + for (int i = 0; i < n_layer; ++i) { + int flags = 0; + if (hparams.nextn_predict_layers > 0 && static_cast(i) >= n_layer - hparams.nextn_predict_layers) { + // skip all tensors in the NextN layers + // TODO @ngxson : TENSOR_NOT_REQUIRED was a hack, need to remove it later + flags |= TENSOR_SKIP | TENSOR_NOT_REQUIRED; + } + + auto & layer = layers[i]; + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags); + layer.attn_q_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}, flags); + layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, flags); + + layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, flags); + layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, flags); + + layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, flags); + + // note: only old legacy GGUF files will have the unsplit wkv_b tensor in + layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head}, flags); + layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head}, flags); + + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_embd_head_v_mla, n_embd}, flags); + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags); + + // DSA indexer + layer.indexer_k_norm = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM, "weight", i), {hparams.indexer_head_size}, flags); + layer.indexer_k_norm_b = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM, "bias", i), {hparams.indexer_head_size}, flags); + layer.indexer_proj = create_tensor(tn(LLM_TENSOR_INDEXER_PROJ, "weight", i), {n_embd, hparams.indexer_n_head}, flags); + layer.indexer_attn_k = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_K, "weight", i), {n_embd, hparams.indexer_head_size}, flags); + layer.indexer_attn_q_b = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_Q_B, "weight", i), {q_lora_rank, hparams.indexer_n_head * hparams.indexer_head_size}, flags); + if (i < (int) hparams.n_layer_dense_lead) { + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, flags); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, flags); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, flags); + } else { + layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, flags); + layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED); + + if (n_expert == 0) { + throw std::runtime_error("n_expert must be > 0"); + } + if (n_expert_used == 0) { + throw std::runtime_error("n_expert_used must be > 0"); + } + + // MoE branch + layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}, flags); + layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, flags); + layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}, flags); + + // Shared expert branch + layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags); + layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_exp * n_expert_shared, n_embd}, flags); + layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags); + } + + // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers + if (hparams.nextn_predict_layers > 0 && static_cast(i) >= n_layer - hparams.nextn_predict_layers) { + layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags); + layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags); + layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags); + + // Optional tensors + layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED); + layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED); + layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags | TENSOR_NOT_REQUIRED); + } + } + } break; case LLM_ARCH_PLM: { const int64_t n_embd_head_qk_rope = hparams.n_rot(); @@ -7772,7 +7924,7 @@ void llama_model::print_info() const { LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale); } - if (arch == LLM_ARCH_DEEPSEEK2 || arch == LLM_ARCH_GLM_DSA) { + if (arch == LLM_ARCH_DEEPSEEK2 || arch == LLM_ARCH_GLM_DSA || arch == LLM_ARCH_DEEPSEEK32) { LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead); LLAMA_LOG_INFO("%s: n_lora_q = %d\n", __func__, hparams.n_lora_q); LLAMA_LOG_INFO("%s: n_lora_kv = %d\n", __func__, hparams.n_lora_kv); @@ -8353,6 +8505,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const { { llm = std::make_unique(*this, params); } break; + case LLM_ARCH_DEEPSEEK32: + { + llm = std::make_unique(*this, params); + } break; case LLM_ARCH_CHATGLM: { llm = std::make_unique(*this, params); @@ -8748,6 +8904,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_ARCTIC: case LLM_ARCH_DEEPSEEK: case LLM_ARCH_DEEPSEEK2: + case LLM_ARCH_DEEPSEEK32: case LLM_ARCH_PLM: case LLM_ARCH_CHATGLM: case LLM_ARCH_GRANITE: diff --git a/src/llama-model.h b/src/llama-model.h index 5ecb8344a2..9431d338d7 100644 --- a/src/llama-model.h +++ b/src/llama-model.h @@ -134,6 +134,7 @@ enum llm_type { LLM_TYPE_310B_A15B, // /MiMo-V2-Flash LLM_TYPE_355B_A32B, // GLM-4.5 LLM_TYPE_397B_A17B, // Qwen3.5 + LLM_TYPE_685B_A37B, // DeepSeek V3.2 LLM_TYPE_744B_A40B, // GLM-5 LLM_TYPE_E2B, LLM_TYPE_E4B, diff --git a/src/models/deepseek32.cpp b/src/models/deepseek32.cpp new file mode 100644 index 0000000000..f843dbd41c --- /dev/null +++ b/src/models/deepseek32.cpp @@ -0,0 +1,240 @@ +#include "models.h" + +llm_build_deepseek32::llm_build_deepseek32(const llama_model & model, const llm_graph_params & params) : + llm_graph_context(params) { + const bool is_mla = hparams.is_mla(); + + // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA + const int64_t n_embd_head_k = hparams.n_embd_head_k_mla(); + const int64_t n_embd_head_v = hparams.n_embd_head_v_mla(); + + const int64_t n_embd_head_qk_rope = hparams.n_rot(); + const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope; + + const uint32_t kv_lora_rank = hparams.n_lora_kv; + + // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly. + // See https://github.com/ggml-org/llama.cpp/discussions/7416 for detailed explanation. + // And also: https://github.com/ggml-org/llama.cpp/pull/17945 [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX] + + // first cancel the adjustment from llama_hparams::yarn_attn_factor_adjust to get the original attn_factor + GGML_ASSERT(ext_factor >= 0.0f); + const float attn_factor_org = attn_factor * (1.0f + 0.1f * logf(1.0f / freq_scale)); + + // use the original attn_factor to pre-scale the kq_scale + const float mscale = attn_factor_org * (1.0f + 0.1f * hparams.rope_yarn_log_mul * logf(1.0f / freq_scale)); + const float kq_scale = 1.0f * mscale * mscale / sqrtf(float(n_embd_head_k)); + + ggml_tensor * cur; + ggml_tensor * inpL; + + // {n_embd, n_tokens} + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn_kv = !is_mla ? build_attn_inp_kv() : nullptr; + auto * inp_attn_k = is_mla ? build_attn_inp_k() : nullptr; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + int effective_n_layers = hparams.n_layer - hparams.nextn_predict_layers; + for (int il = 0; il < effective_n_layers; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + ggml_tensor * q = NULL; + + const bool is_lite = model.layers[il].wq; + + q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur); + cb(q, "q", il); + + q = build_norm(q, model.layers[il].attn_q_a_norm, nullptr, LLM_NORM_RMS, il); + cb(q, "q", il); + + q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q); + cb(q, "q", il); + + // split into {n_embd_head_qk_nope, n_head, n_tokens} + ggml_tensor * q_nope = + ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k), + ggml_row_size(q->type, n_embd_head_k) * n_head, 0); + cb(q_nope, "q_nope", il); + + // and {n_embd_head_qk_rope, n_head, n_tokens} + ggml_tensor * q_pe = ggml_view_3d( + ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k), + ggml_row_size(q->type, n_embd_head_k) * n_head, ggml_row_size(q->type, n_embd_head_qk_nope)); + cb(q_pe, "q_pe", il); + + ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur); + cb(kv_cmpr_pe, "kv_cmpr_pe", il); + + // split into {kv_lora_rank, n_tokens} + ggml_tensor * kv_cmpr = + ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens, + ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0); + cb(kv_cmpr, "kv_cmpr", il); + + // and {n_embd_head_qk_rope, 1, n_tokens} + ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens, + ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), + ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), + ggml_row_size(kv_cmpr_pe->type, kv_lora_rank)); + cb(k_pe, "k_pe", il); + + q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + cb(q_pe, "q_pe", il); + + k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + cb(k_pe, "k_pe", il); + + kv_cmpr = build_norm(kv_cmpr, model.layers[il].attn_kv_a_norm, nullptr, LLM_NORM_RMS, il); + cb(kv_cmpr, "kv_cmpr", il); + + if (is_mla) { + // {n_embd_head_qk_nope, n_tokens, n_head} + q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3); + cb(q_nope, "q_nope_perm", il); + + // {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head} + ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope); + cb(q_nope_absorbed, "q_nope_absorbed", il); + + // {kv_lora_rank, n_head, n_tokens} + q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3); + cb(q_nope_absorbed, "q_nope_absorbed_perm", il); + + // {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens} + // note: rope must go first for in-place context shifting in build_rope_shift() + ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0); + cb(Qcur, "Qcur", il); + + kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens); + cb(kv_cmpr, "kv_cmpr_reshape", il); + + // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens} + ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0); + cb(Kcur, "Kcur", il); + + // {kv_lora_rank, 1, n_tokens} + ggml_tensor * Vcur = kv_cmpr; + cb(Vcur, "Vcur", il); + + // note: MLA with the absorption optimization converts into MQA (ie: GQA with 1 group) + cur = build_attn(inp_attn_k, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il); + } else { + ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_cmpr); + cb(kv, "kv", il); + + // split into {n_embd_head_qk_nope, n_head, n_tokens} + ggml_tensor * k_nope = + ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v), + ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head, 0); + cb(k_nope, "k_nope_view", il); + + // and {n_embd_head_v, n_head, n_tokens} + ggml_tensor * Vcur = ggml_view_3d(ctx0, kv, n_embd_head_v, n_head, n_tokens, + ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v), + ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head, + ggml_row_size(kv->type, n_embd_head_qk_nope)); + cb(Vcur, "Vcur_view", il); + + Vcur = ggml_cont(ctx0, Vcur); + cb(Vcur, "Vcur_cont", il); + + ggml_tensor * Qcur = ggml_concat(ctx0, q_nope, q_pe, 0); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0); + cb(Kcur, "Kcur", il); + + // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups) + cur = build_attn(inp_attn_kv, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + } + } + if (il == effective_n_layers - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + if ((uint32_t) il < hparams.n_layer_dense_lead) { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } else { + // MoE branch + ggml_tensor * moe_out = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + model.layers[il].ffn_exp_probs_b, + n_expert, n_expert_used, + LLM_FFN_SILU, hparams.expert_weights_norm, + hparams.expert_weights_scale, + (llama_expert_gating_func_type) hparams.expert_gating_func, + il, + nullptr, + model.layers[il].ffn_gate_up_exps); + cb(moe_out, "ffn_moe_out", il); + + // FFN shared expert + { + ggml_tensor * ffn_shexp = + build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(ffn_shexp, "ffn_shexp", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } + } + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/src/models/models.h b/src/models/models.h index cf9ba04e7f..fd8c8d65c7 100644 --- a/src/models/models.h +++ b/src/models/models.h @@ -166,12 +166,16 @@ struct llm_build_deci : public llm_graph_context { llm_build_deci(const llama_model & model, const llm_graph_params & params); }; +struct llm_build_deepseek : public llm_graph_context { + llm_build_deepseek(const llama_model & model, const llm_graph_params & params); +}; + struct llm_build_deepseek2 : public llm_graph_context { llm_build_deepseek2(const llama_model & model, const llm_graph_params & params); }; -struct llm_build_deepseek : public llm_graph_context { - llm_build_deepseek(const llama_model & model, const llm_graph_params & params); +struct llm_build_deepseek32 : public llm_graph_context { + llm_build_deepseek32(const llama_model & model, const llm_graph_params & params); }; struct llm_build_dots1 : public llm_graph_context {