From 2c3f8b850a4a6cff0f5dda2135c03fc81d33ed8b Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Sun, 30 Mar 2025 22:21:03 +0200 Subject: [PATCH] llama : support BailingMoE (Ling) (#12634) --- README.md | 1 + convert_hf_to_gguf.py | 105 ++++++++++++++++ convert_hf_to_gguf_update.py | 1 + gguf-py/gguf/constants.py | 24 ++++ gguf-py/gguf/tensor_mapping.py | 1 + include/llama.h | 1 + src/llama-arch.cpp | 24 ++++ src/llama-arch.h | 1 + src/llama-chat.cpp | 20 ++++ src/llama-chat.h | 1 + src/llama-model.cpp | 213 +++++++++++++++++++++++++++++++++ src/llama-model.h | 1 + src/llama-vocab.cpp | 11 ++ 13 files changed, 404 insertions(+) diff --git a/README.md b/README.md index b637fe2ef..f8ed423c4 100644 --- a/README.md +++ b/README.md @@ -113,6 +113,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo - [x] [QRWKV-6](https://huggingface.co/recursal/QRWKV6-32B-Instruct-Preview-v0.1) - [x] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct) - [X] [Trillion-7B-preview](https://huggingface.co/trillionlabs/Trillion-7B-preview) +- [x] [Ling models](https://huggingface.co/collections/inclusionAI/ling-67c51c85b34a7ea0aba94c32) #### Multimodal diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index c322edc41..0919cd3f0 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -711,6 +711,9 @@ class Model: if chkhsh == "1994ffd01900cfb37395608534236ecd63f2bd5995d6cb1004dda1af50240f15": # ref: https://huggingface.co/trillionlabs/Trillion-7B-preview res = "trillion" + if chkhsh == "96a5f08be6259352137b512d4157e333e21df7edd3fcd152990608735a65b224": + # ref: https://huggingface.co/inclusionAI/Ling-lite + res = "bailingmoe" if res is None: logger.warning("\n") @@ -5133,6 +5136,108 @@ class GraniteMoeModel(GraniteModel): return super().modify_tensors(data_torch, name, bid) +@Model.register("BailingMoeForCausalLM") +class BailingMoeModel(Model): + model_arch = gguf.MODEL_ARCH.BAILINGMOE + + def set_vocab(self): + self._set_vocab_gpt2() + + def set_gguf_parameters(self): + super().set_gguf_parameters() + hparams = self.hparams + if "head_dim" in hparams: + rope_dim = hparams["head_dim"] + else: + rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"] + + self.gguf_writer.add_rope_dimension_count(rope_dim) + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE) + self.gguf_writer.add_leading_dense_block_count(hparams["first_k_dense_replace"]) + self.gguf_writer.add_vocab_size(hparams["vocab_size"]) + self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"]) + self.gguf_writer.add_expert_weights_scale(1.0) + self.gguf_writer.add_expert_count(hparams["num_experts"]) + self.gguf_writer.add_expert_shared_count(hparams["num_shared_experts"]) + self.gguf_writer.add_expert_weights_norm(hparams["norm_topk_prob"]) + + _experts: list[dict[str, Tensor]] | None = None + + @staticmethod + def permute(weights: Tensor, n_head: int, n_head_kv: int | None): + if n_head_kv is not None and n_head != n_head_kv: + n_head = n_head_kv + return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:]) + .swapaxes(1, 2) + .reshape(weights.shape)) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + n_head = self.hparams["num_attention_heads"] + n_kv_head = self.hparams.get("num_key_value_heads") + n_embd = self.hparams["hidden_size"] + head_dim = self.hparams.get("head_dim", n_embd // n_head) + + output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT) + + if name.endswith("attention.dense.weight"): + return [(self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_OUT, bid), data_torch)] + elif name.endswith("query_key_value.weight"): + q, k, v = data_torch.split([n_head * head_dim, n_kv_head * head_dim, n_kv_head * head_dim], dim=-2) + + return [ + (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), BailingMoeModel.permute(q, n_head, n_head)), + (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), BailingMoeModel.permute(k, n_head, n_kv_head)), + (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid), v) + ] + elif name.find("mlp.experts") != -1: + n_experts = self.hparams["num_experts"] + assert bid is not None + + tensors: list[tuple[str, Tensor]] = [] + + 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" + + new_name = self.map_tensor_name(merged_name) + + tensors.append((new_name, data_torch)) + + return tensors + + new_name = self.map_tensor_name(name) + + if new_name == output_name and self.hparams.get("norm_head"): + data_torch = data_torch.float() + data_torch /= torch.norm(data_torch, p=2, dim=0, keepdim=True) + 1e-7 + + return [(new_name, data_torch)] + + 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}") + + @Model.register("ChameleonForConditionalGeneration") @Model.register("ChameleonForCausalLM") # obsolete class ChameleonModel(Model): diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py index a3a647125..1b86f4c90 100755 --- a/convert_hf_to_gguf_update.py +++ b/convert_hf_to_gguf_update.py @@ -112,6 +112,7 @@ models = [ {"name": "gpt-4o", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Xenova/gpt-4o", }, {"name": "superbpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/UW/OLMo2-8B-SuperBPE-t180k", }, {"name": "trillion", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/trillionlabs/Trillion-7B-preview", }, + {"name": "bailingmoe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-lite", }, ] diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 1753dca4b..3a52cfd1e 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -287,6 +287,7 @@ class MODEL_ARCH(IntEnum): CHAMELEON = auto() WAVTOKENIZER_DEC = auto() PLM = auto() + BAILINGMOE = auto() class MODEL_TENSOR(IntEnum): @@ -490,6 +491,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.CHAMELEON: "chameleon", MODEL_ARCH.WAVTOKENIZER_DEC: "wavtokenizer-dec", MODEL_ARCH.PLM: "plm", + MODEL_ARCH.BAILINGMOE: "bailingmoe", } TENSOR_NAMES: dict[MODEL_TENSOR, str] = { @@ -1667,6 +1669,25 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.POSNET_ATTN_V, MODEL_TENSOR.POSNET_ATTN_OUT, ], + MODEL_ARCH.BAILINGMOE: [ + 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_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_NORM, + 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, + ], # TODO } @@ -1719,6 +1740,9 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.ROPE_FREQS, MODEL_TENSOR.ATTN_ROT_EMBD, ], + MODEL_ARCH.BAILINGMOE: [ + MODEL_TENSOR.ROPE_FREQS, + ], } # diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 8d4a2b032..50bef12e3 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -29,6 +29,7 @@ class TensorNameMap: "shared", # t5 "rwkv.embeddings", # rwkv6 "model.embeddings", # rwkv7 + "model.word_embeddings", # bailingmoe ), # Token type embeddings diff --git a/include/llama.h b/include/llama.h index 4eb70ec99..468ab1fa4 100644 --- a/include/llama.h +++ b/include/llama.h @@ -109,6 +109,7 @@ extern "C" { LLAMA_VOCAB_PRE_TYPE_GPT4O = 29, LLAMA_VOCAB_PRE_TYPE_SUPERBPE = 30, LLAMA_VOCAB_PRE_TYPE_TRILLION = 31, + LLAMA_VOCAB_PRE_TYPE_BAILINGMOE = 32, }; enum llama_rope_type { diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index 9e443d830..954ae65a3 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -66,6 +66,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_CHAMELEON, "chameleon" }, { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" }, { LLM_ARCH_PLM, "plm" }, + { LLM_ARCH_BAILINGMOE, "bailingmoe" }, { LLM_ARCH_UNKNOWN, "(unknown)" }, }; @@ -1409,6 +1410,29 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_POS_NET_ATTN_OUT, "posnet.%d.attn_output" }, }, }, + { + LLM_ARCH_BAILINGMOE, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + { LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" }, + { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" }, + { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" }, + { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" }, + }, + }, { LLM_ARCH_UNKNOWN, { diff --git a/src/llama-arch.h b/src/llama-arch.h index 39e3a2ce0..7e2a3dd29 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -70,6 +70,7 @@ enum llm_arch { LLM_ARCH_CHAMELEON, LLM_ARCH_WAVTOKENIZER_DEC, LLM_ARCH_PLM, + LLM_ARCH_BAILINGMOE, LLM_ARCH_UNKNOWN, }; diff --git a/src/llama-chat.cpp b/src/llama-chat.cpp index d6d781cba..dd27a3814 100644 --- a/src/llama-chat.cpp +++ b/src/llama-chat.cpp @@ -60,6 +60,7 @@ static const std::map LLM_CHAT_TEMPLATES = { { "gigachat", LLM_CHAT_TEMPLATE_GIGACHAT }, { "megrez", LLM_CHAT_TEMPLATE_MEGREZ }, { "yandex", LLM_CHAT_TEMPLATE_YANDEX }, + { "bailing", LLM_CHAT_TEMPLATE_BAILING }, }; llm_chat_template llm_chat_template_from_str(const std::string & name) { @@ -171,6 +172,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) { return LLM_CHAT_TEMPLATE_MEGREZ; } else if (tmpl_contains(" Ассистент:")) { return LLM_CHAT_TEMPLATE_YANDEX; + } else if (tmpl_contains("ASSISTANT") && tmpl_contains("'HUMAN'")) { + return LLM_CHAT_TEMPLATE_BAILING; } return LLM_CHAT_TEMPLATE_UNKNOWN; } @@ -588,6 +591,23 @@ int32_t llm_chat_apply_template( if (add_ass) { ss << " Ассистент:[SEP]"; } + } else if (tmpl == LLM_CHAT_TEMPLATE_BAILING) { + // Bailing (Ling) template + for (auto message : chat) { + std::string role(message->role); + + if (role == "user") { + role = "HUMAN"; + } else { + std::transform(role.begin(), role.end(), role.begin(), ::toupper); + } + + ss << "" << role << "" << message->content; + } + + if (add_ass) { + ss << "ASSISTANT"; + } } else { // template not supported return -1; diff --git a/src/llama-chat.h b/src/llama-chat.h index bbd5a6b72..0e0bd772c 100644 --- a/src/llama-chat.h +++ b/src/llama-chat.h @@ -39,6 +39,7 @@ enum llm_chat_template { LLM_CHAT_TEMPLATE_GIGACHAT, LLM_CHAT_TEMPLATE_MEGREZ, LLM_CHAT_TEMPLATE_YANDEX, + LLM_CHAT_TEMPLATE_BAILING, LLM_CHAT_TEMPLATE_UNKNOWN, }; diff --git a/src/llama-model.cpp b/src/llama-model.cpp index e712960f8..8d525e1be 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -88,6 +88,7 @@ const char * llm_type_name(llm_type type) { case LLM_TYPE_10B_128x3_66B: return "10B+128x3.66B"; case LLM_TYPE_57B_A14B: return "57B.A14B"; case LLM_TYPE_27B: return "27B"; + case LLM_TYPE_290B: return "290B"; default: return "?B"; } } @@ -1328,6 +1329,21 @@ void llama_model::load_hparams(llama_model_loader & ml) { ml.get_key(LLM_KV_ATTENTION_GROUPNORM_GROUPS, hparams.n_norm_groups); ml.get_key(LLM_KV_ATTENTION_CAUSAL, hparams.causal_attn); } break; + case LLM_ARCH_BAILINGMOE: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead); + 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); + ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale); + ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM, hparams.expert_weights_norm, false); + + switch (hparams.n_layer) { + case 28: type = LLM_TYPE_16B; break; + case 88: type = LLM_TYPE_290B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; default: throw std::runtime_error("unsupported model architecture"); } @@ -3739,6 +3755,46 @@ bool llama_model::load_tensors(llama_model_loader & ml) { output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {hparams.convnext.n_embd, n_embd}, 0); output_b = create_tensor(tn(LLM_TENSOR_OUTPUT, "bias"), {n_embd}, 0); } break; + case LLM_ARCH_BAILINGMOE: + { + 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); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0); + + 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.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_head * n_rot}, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_head_kv * n_rot}, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_head_kv * n_rot}, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_rot, n_embd}, 0); + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + + layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0); + + 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"); + } + + layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}, 0); + layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, 0); + layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}, 0); + + layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0); + layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_exp * n_expert_shared, n_embd}, 0); + layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0); + } + } break; default: throw std::runtime_error("unknown architecture"); } @@ -4026,6 +4082,14 @@ void llama_model::print_info() const { LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale); } + if (arch == LLM_ARCH_BAILINGMOE) { + LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead); + LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp); + LLAMA_LOG_INFO("%s: n_expert_shared = %d\n", __func__, hparams.n_expert_shared); + LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale); + LLAMA_LOG_INFO("%s: expert_weights_norm = %d\n", __func__, hparams.expert_weights_norm); + } + vocab.print_info(); } @@ -11814,6 +11878,150 @@ struct llm_build_plm : public llm_graph_context { } }; +struct llm_build_bailingmoe : public llm_graph_context { + llm_build_bailingmoe(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++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 + { + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_rot, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_rot, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_rot)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_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); + + 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, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, hparams.expert_weights_norm, + false, hparams.expert_weights_scale, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + 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 = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + llama_memory_i * llama_model::create_memory() const { llama_memory_i * res; @@ -12090,6 +12298,10 @@ llm_graph_result_ptr llama_model::build_graph( { llm = std::make_unique(*this, params, gf); } break; + case LLM_ARCH_BAILINGMOE: + { + llm = std::make_unique(*this, params, gf); + } break; default: GGML_ABORT("fatal error"); } @@ -12221,6 +12433,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_GRANITE: case LLM_ARCH_GRANITE_MOE: case LLM_ARCH_CHAMELEON: + case LLM_ARCH_BAILINGMOE: return LLAMA_ROPE_TYPE_NORM; // the pairs of head values are offset by n_rot/2 diff --git a/src/llama-model.h b/src/llama-model.h index 0064d597a..f1bf0df3a 100644 --- a/src/llama-model.h +++ b/src/llama-model.h @@ -85,6 +85,7 @@ enum llm_type { LLM_TYPE_10B_128x3_66B, LLM_TYPE_57B_A14B, LLM_TYPE_27B, + LLM_TYPE_290B, }; struct llama_layer_posnet { diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp index 5ace5e385..78072e17f 100644 --- a/src/llama-vocab.cpp +++ b/src/llama-vocab.cpp @@ -407,6 +407,13 @@ struct llm_tokenizer_bpe : llm_tokenizer { "(?=(\\d{3})+(?!\\d))", }; break; + case LLAMA_VOCAB_PRE_TYPE_BAILINGMOE: + regex_exprs = { + // original regex from tokenizer.json + // "'(?i:[sdmt]|ll|ve|re)|[^\\r\\n\\p{L}\\p{N}]?+\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]++[\\r\\n]*|\\s*[\\r\\n]|\\s+(?!\\S)|\\s+" + "'(?:[sSdDmMtT]|[lL][lL]|[vV][eE]|[rR][eE])|[^\\r\\n\\p{L}\\p{N}]?+\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]++[\\r\\n]*|\\s*[\\r\\n]|\\s+(?!\\S)|\\s+", + }; + break; default: // default regex for BPE tokenization pre-processing regex_exprs = { @@ -1619,6 +1626,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { tokenizer_pre == "trillion") { pre_type = LLAMA_VOCAB_PRE_TYPE_TRILLION; clean_spaces = false; + } else if ( + tokenizer_pre == "bailingmoe") { + pre_type = LLAMA_VOCAB_PRE_TYPE_BAILINGMOE; + clean_spaces = false; } else { throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str())); }