diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 010e209b22..26ffea568f 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -7387,7 +7387,7 @@ class FalconH1Model(Mamba2Model): @ModelBase.register("HunYuanMoEV1ForCausalLM") class HunYuanMoEModel(TextModel): - model_arch = gguf.MODEL_ARCH.HUNYUAN_MOE + model_arch = gguf.MODEL_ARCH.HUNYUAN_V1_MOE def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py index 6a653a8933..cac881c5d9 100755 --- a/convert_hf_to_gguf_update.py +++ b/convert_hf_to_gguf_update.py @@ -139,7 +139,7 @@ pre_computed_hashes = [ {"name": "chatglm-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-chat", "chkhsh": "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516"}, {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", "chkhsh": "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2"}, {"name": "minerva-7b", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", "chkhsh": "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35"}, - {"name": "hunyuan", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-A13B-Instruct", "chkhsh": "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664"}, + {"name": "hunyuan-v1-moe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-A13B-Instruct", "chkhsh": "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664"}, {"name": "hunyuan-v1-dense", "tokt": TOKENIZER_TYPE.BPE, "repo": "", "chkhsh": ""}, # TODO: update hunyuan-v1-dense repo # falcon-h1 series uses 4 different tokenizers across model sizes (0.5b - 34b), hence we need to define 4 different hashes {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-0.5B-Base", "chkhsh": "a6b57017d60e6edb4d88ecc2845188e0eb333a70357e45dcc9b53964a73bbae6"}, diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index d300c08253..70017f297f 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -372,7 +372,7 @@ class MODEL_ARCH(IntEnum): ARCEE = auto() ERNIE4_5 = auto() ERNIE4_5_MOE = auto() - HUNYUAN_MOE = auto() + HUNYUAN_V1_MOE = auto() HUNYUAN_V1_DENSE = auto() SMOLLM3 = auto() LFM2 = auto() @@ -692,7 +692,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.ERNIE4_5: "ernie4_5", MODEL_ARCH.ERNIE4_5_MOE: "ernie4_5-moe", MODEL_ARCH.FALCON_H1: "falcon-h1", - MODEL_ARCH.HUNYUAN_MOE: "hunyuan-moe", + MODEL_ARCH.HUNYUAN_V1_MOE: "hunyuan-v1-moe", MODEL_ARCH.HUNYUAN_V1_DENSE: "hunyuan-v1-dense", MODEL_ARCH.SMOLLM3: "smollm3", MODEL_ARCH.LFM2: "lfm2", @@ -2430,7 +2430,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.OUTPUT_NORM, # Final layer norm MODEL_TENSOR.OUTPUT, # Output projection (lm_head) ], - MODEL_ARCH.HUNYUAN_MOE: [ + MODEL_ARCH.HUNYUAN_V1_MOE: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, MODEL_TENSOR.OUTPUT, diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index 203d4606b7..f7062dc0cb 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -84,7 +84,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_ARCEE, "arcee" }, { LLM_ARCH_ERNIE4_5, "ernie4_5" }, { LLM_ARCH_ERNIE4_5_MOE, "ernie4_5-moe" }, - { LLM_ARCH_HUNYUAN_MOE, "hunyuan-moe" }, + { LLM_ARCH_HUNYUAN_V1_MOE, "hunyuan-v1-moe" }, { LLM_ARCH_HUNYUAN_V1_DENSE, "hunyuan-v1-dense" }, { LLM_ARCH_SMOLLM3, "smollm3" }, { LLM_ARCH_LFM2, "lfm2" }, @@ -1874,7 +1874,7 @@ static const std::map> LLM_TENSOR_N }, }, { - LLM_ARCH_HUNYUAN_MOE, + LLM_ARCH_HUNYUAN_V1_MOE, { { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, diff --git a/src/llama-arch.h b/src/llama-arch.h index c5c1455c27..d29b367878 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -88,7 +88,7 @@ enum llm_arch { LLM_ARCH_ARCEE, LLM_ARCH_ERNIE4_5, LLM_ARCH_ERNIE4_5_MOE, - LLM_ARCH_HUNYUAN_MOE, + LLM_ARCH_HUNYUAN_V1_MOE, LLM_ARCH_HUNYUAN_V1_DENSE, LLM_ARCH_SMOLLM3, LLM_ARCH_LFM2, diff --git a/src/llama-chat.cpp b/src/llama-chat.cpp index 0cda3215ed..882f08bb41 100644 --- a/src/llama-chat.cpp +++ b/src/llama-chat.cpp @@ -65,7 +65,7 @@ static const std::map LLM_CHAT_TEMPLATES = { { "bailing", LLM_CHAT_TEMPLATE_BAILING }, { "llama4", LLM_CHAT_TEMPLATE_LLAMA4 }, { "smolvlm", LLM_CHAT_TEMPLATE_SMOLVLM }, - { "hunyuan-moe", LLM_CHAT_TEMPLATE_HUNYUAN_MOE }, + { "hunyuan-v1-moe", LLM_CHAT_TEMPLATE_HUNYUAN_V1_MOE }, { "hunyuan-v1-dense", LLM_CHAT_TEMPLATE_HUNYUAN_V1_DENSE }, { "kimi-k2", LLM_CHAT_TEMPLATE_KIMI_K2 }, }; @@ -193,7 +193,7 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) { } else if (tmpl_contains("<|endofuserprompt|>")) { return LLM_CHAT_TEMPLATE_DOTS1; } else if (tmpl_contains("<|startoftext|>") && tmpl_contains("<|extra_4|>")) { - return LLM_CHAT_TEMPLATE_HUNYUAN_MOE; + return LLM_CHAT_TEMPLATE_HUNYUAN_V1_MOE; } else if (tmpl_contains("<｜hy_place▁holder▁no▁2｜>") && tmpl_contains("<｜hy_place▁holder▁no▁3｜>")) { return LLM_CHAT_TEMPLATE_HUNYUAN_V1_DENSE; } else if (tmpl_contains("<|im_assistant|>assistant<|im_middle|>")) { @@ -694,14 +694,14 @@ int32_t llm_chat_apply_template( if (add_ass) { ss << "<|response|>"; } - } else if (tmpl == LLM_CHAT_TEMPLATE_HUNYUAN_MOE) { + } else if (tmpl == LLM_CHAT_TEMPLATE_HUNYUAN_V1_MOE) { // tencent/Hunyuan-A13B-Instruct for (auto message : chat) { std::string role(message->role); if (role == "system") { ss << "<|startoftext|>" << message->content << "<|extra_4|>"; } else if (role == "assistant") { - ss << "<|startoftext|>" << message->content << "<|eos|>"; + ss << message->content << "<|eos|>"; } else { ss << "<|startoftext|>" << message->content << "<|extra_0|>"; } diff --git a/src/llama-chat.h b/src/llama-chat.h index ece18767e9..3e1f6a5843 100644 --- a/src/llama-chat.h +++ b/src/llama-chat.h @@ -45,7 +45,7 @@ enum llm_chat_template { LLM_CHAT_TEMPLATE_LLAMA4, LLM_CHAT_TEMPLATE_SMOLVLM, LLM_CHAT_TEMPLATE_DOTS1, - LLM_CHAT_TEMPLATE_HUNYUAN_MOE, + LLM_CHAT_TEMPLATE_HUNYUAN_V1_MOE, LLM_CHAT_TEMPLATE_HUNYUAN_V1_DENSE, LLM_CHAT_TEMPLATE_KIMI_K2, LLM_CHAT_TEMPLATE_UNKNOWN, diff --git a/src/llama-model.cpp b/src/llama-model.cpp index e3375c2903..62d3441be0 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -1733,7 +1733,7 @@ void llama_model::load_hparams(llama_model_loader & ml) { type = LLM_TYPE_UNKNOWN; } } break; - case LLM_ARCH_HUNYUAN_MOE: + case LLM_ARCH_HUNYUAN_V1_MOE: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); @@ -5078,7 +5078,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.ffn_up_b = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i), {ffn_intermediate_size}, TENSOR_NOT_REQUIRED); } } break; - case LLM_ARCH_HUNYUAN_MOE: + case LLM_ARCH_HUNYUAN_V1_MOE: { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -16580,8 +16580,8 @@ struct llm_build_arcee : public llm_graph_context { } }; -struct llm_build_hunyuan_moe : public llm_graph_context { - llm_build_hunyuan_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { +struct llm_build_hunyuan_v1_moe : public llm_graph_context { + llm_build_hunyuan_v1_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { const int64_t n_embd_head = hparams.n_embd_head_v; GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); @@ -17615,9 +17615,9 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const { { llm = std::make_unique(*this, params); } break; - case LLM_ARCH_HUNYUAN_MOE: + case LLM_ARCH_HUNYUAN_V1_MOE: { - llm = std::make_unique(*this, params); + llm = std::make_unique(*this, params); } break; case LLM_ARCH_HUNYUAN_V1_DENSE: { @@ -17831,7 +17831,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_EXAONE4: case LLM_ARCH_MINICPM3: case LLM_ARCH_DOTS1: - case LLM_ARCH_HUNYUAN_MOE: + case LLM_ARCH_HUNYUAN_V1_MOE: case LLM_ARCH_HUNYUAN_V1_DENSE: case LLM_ARCH_LFM2: return LLAMA_ROPE_TYPE_NEOX;