diff --git a/README.md b/README.md index 9a10ead83..4d24dd591 100644 --- a/README.md +++ b/README.md @@ -77,6 +77,7 @@ Typically finetunes of the base models below are supported as well. - [x] [SEA-LION](https://huggingface.co/models?search=sea-lion) - [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B) - [x] [OLMo](https://allenai.org/olmo) +- [x] [OLMoE](https://huggingface.co/allenai/OLMoE-1B-7B-0924) - [x] [Granite models](https://huggingface.co/collections/ibm-granite/granite-code-models-6624c5cec322e4c148c8b330) - [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia) - [x] [Snowflake-Arctic MoE](https://huggingface.co/collections/Snowflake/arctic-66290090abe542894a5ac520) diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index d995ed764..f026977e9 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -2998,6 +2998,66 @@ class OlmoModel(Model): return [(self.map_tensor_name(name), data_torch)] +@Model.register("OlmoeForCausalLM") +class OlmoeModel(Model): + model_arch = gguf.MODEL_ARCH.OLMOE + + def set_gguf_parameters(self): + super().set_gguf_parameters() + self.gguf_writer.add_layer_norm_rms_eps(1e-5) + if (n_experts := self.hparams.get("num_experts")) is not None: + self.gguf_writer.add_expert_count(n_experts) + + _experts: list[dict[str, Tensor]] | None = None + + # Copied from: Qwen2MoeModel + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + # process the experts separately + if name.find("experts") != -1: + n_experts = self.hparams["num_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: + tensors: list[tuple[str, Tensor]] = [] + + # 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 + else: + return [] + + return [(self.map_tensor_name(name), data_torch)] + + # Copied from: Qwen2MoeModel + 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("JinaBertModel", "JinaBertForMaskedLM") class JinaBertV2Model(BertModel): model_arch = gguf.MODEL_ARCH.JINA_BERT_V2 diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 2c8545455..0d88649d8 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -220,6 +220,7 @@ class MODEL_ARCH(IntEnum): COMMAND_R = auto() DBRX = auto() OLMO = auto() + OLMOE = auto() OPENELM = auto() ARCTIC = auto() DEEPSEEK2 = auto() @@ -375,6 +376,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.COMMAND_R: "command-r", MODEL_ARCH.DBRX: "dbrx", MODEL_ARCH.OLMO: "olmo", + MODEL_ARCH.OLMOE: "olmoe", MODEL_ARCH.OPENELM: "openelm", MODEL_ARCH.ARCTIC: "arctic", MODEL_ARCH.DEEPSEEK2: "deepseek2", @@ -1027,6 +1029,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, ], + MODEL_ARCH.OLMOE: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + ], MODEL_ARCH.OPENELM: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index bc9a13ee5..2ebfa2b43 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -13,7 +13,7 @@ class TensorNameMap: "transformer.wte", # gpt2 gpt-j mpt refact qwen dbrx jais exaone "transformer.word_embeddings", # falcon "word_embeddings", # bloom - "model.embed_tokens", # llama-hf nemotron + "model.embed_tokens", # llama-hf nemotron olmoe "tok_embeddings", # llama-pth "embeddings.word_embeddings", # bert nomic-bert "language_model.embedding.word_embeddings", # persimmon @@ -54,7 +54,7 @@ class TensorNameMap: # Output MODEL_TENSOR.OUTPUT: ( "embed_out", # gptneox - "lm_head", # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone + "lm_head", # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone olmoe "output", # llama-pth bloom internlm2 "word_embeddings_for_head", # persimmon "lm_head.linear", # phi2 @@ -66,7 +66,7 @@ class TensorNameMap: MODEL_TENSOR.OUTPUT_NORM: ( "gpt_neox.final_layer_norm", # gptneox "transformer.ln_f", # gpt2 gpt-j falcon jais exaone - "model.norm", # llama-hf baichuan internlm2 + "model.norm", # llama-hf baichuan internlm2 olmoe "norm", # llama-pth "transformer.norm_f", # mpt dbrx "ln_f", # refact bloom qwen gpt2 @@ -98,7 +98,7 @@ class TensorNameMap: "transformer.h.{bid}.input_layernorm", # falcon7b "h.{bid}.input_layernorm", # bloom "transformer.h.{bid}.ln_mlp", # falcon40b - "model.layers.{bid}.input_layernorm", # llama-hf nemotron + "model.layers.{bid}.input_layernorm", # llama-hf nemotron olmoe "layers.{bid}.attention_norm", # llama-pth "language_model.encoder.layers.{bid}.input_layernorm", # persimmon "model.layers.{bid}.ln1", # yi @@ -142,7 +142,7 @@ class TensorNameMap: # Attention query MODEL_TENSOR.ATTN_Q: ( - "model.layers.{bid}.self_attn.q_proj", # llama-hf nemotron + "model.layers.{bid}.self_attn.q_proj", # llama-hf nemotron olmoe "layers.{bid}.attention.wq", # llama-pth "encoder.layer.{bid}.attention.self.query", # bert "transformer.h.{bid}.attn.q_proj", # gpt-j @@ -154,7 +154,7 @@ class TensorNameMap: # Attention key MODEL_TENSOR.ATTN_K: ( - "model.layers.{bid}.self_attn.k_proj", # llama-hf nemotron + "model.layers.{bid}.self_attn.k_proj", # llama-hf nemotron olmoe "layers.{bid}.attention.wk", # llama-pth "encoder.layer.{bid}.attention.self.key", # bert "transformer.h.{bid}.attn.k_proj", # gpt-j @@ -167,7 +167,7 @@ class TensorNameMap: # Attention value MODEL_TENSOR.ATTN_V: ( - "model.layers.{bid}.self_attn.v_proj", # llama-hf nemotron + "model.layers.{bid}.self_attn.v_proj", # llama-hf nemotron olmoe "layers.{bid}.attention.wv", # llama-pth "encoder.layer.{bid}.attention.self.value", # bert "transformer.h.{bid}.attn.v_proj", # gpt-j @@ -185,7 +185,7 @@ class TensorNameMap: "transformer.blocks.{bid}.attn.out_proj", # mpt "transformer.h.{bid}.self_attention.dense", # falcon "h.{bid}.self_attention.dense", # bloom - "model.layers.{bid}.self_attn.o_proj", # llama-hf nemotron + "model.layers.{bid}.self_attn.o_proj", # llama-hf nemotron olmoe "layers.{bid}.attention.wo", # llama-pth "encoder.layer.{bid}.attention.output.dense", # bert "transformer.h.{bid}.attn.out_proj", # gpt-j @@ -229,7 +229,7 @@ class TensorNameMap: "transformer.h.{bid}.ln_2", # gpt2 refact qwen jais exaone "h.{bid}.post_attention_layernorm", # bloom "transformer.blocks.{bid}.norm_2", # mpt - "model.layers.{bid}.post_attention_layernorm", # llama-hf nemotron + "model.layers.{bid}.post_attention_layernorm", # llama-hf nemotron olmoe "layers.{bid}.ffn_norm", # llama-pth "language_model.encoder.layers.{bid}.post_attention_layernorm", # persimmon "model.layers.{bid}.ln2", # yi @@ -253,7 +253,7 @@ class TensorNameMap: MODEL_TENSOR.FFN_GATE_INP: ( "layers.{bid}.feed_forward.gate", # mixtral "model.layers.{bid}.block_sparse_moe.gate", # mixtral - "model.layers.{bid}.mlp.gate", # qwen2moe + "model.layers.{bid}.mlp.gate", # qwen2moe olmoe "transformer.decoder_layer.{bid}.router", # Grok "transformer.blocks.{bid}.ffn.router.layer", # dbrx ), @@ -295,7 +295,7 @@ class TensorNameMap: "layers.{bid}.feed_forward.experts.w3", # mixtral (merged) "transformer.decoder_layer.{bid}.moe.linear_v", # Grok (merged) "transformer.blocks.{bid}.ffn.experts.mlp.v1", # dbrx - "model.layers.{bid}.mlp.experts.up_proj", # qwen2moe (merged) + "model.layers.{bid}.mlp.experts.up_proj", # qwen2moe olmoe (merged) ), MODEL_TENSOR.FFN_UP_SHEXP: ( @@ -327,7 +327,7 @@ class TensorNameMap: "layers.{bid}.feed_forward.experts.w1", # mixtral (merged) "transformer.decoder_layer.{bid}.moe.linear", # Grok (merged) "transformer.blocks.{bid}.ffn.experts.mlp.w1", # dbrx - "model.layers.{bid}.mlp.experts.gate_proj", # qwen2moe (merged) + "model.layers.{bid}.mlp.experts.gate_proj", # qwen2moe olmoe (merged) ), MODEL_TENSOR.FFN_GATE_SHEXP: ( @@ -367,7 +367,7 @@ class TensorNameMap: "layers.{bid}.feed_forward.experts.w2", # mixtral (merged) "transformer.decoder_layer.{bid}.moe.linear_1", # Grok (merged) "transformer.blocks.{bid}.ffn.experts.mlp.w2", # dbrx - "model.layers.{bid}.mlp.experts.down_proj", # qwen2moe (merged) + "model.layers.{bid}.mlp.experts.down_proj", # qwen2moe olmoe (merged) ), MODEL_TENSOR.FFN_DOWN_SHEXP: ( @@ -378,7 +378,7 @@ class TensorNameMap: MODEL_TENSOR.ATTN_Q_NORM: ( "language_model.encoder.layers.{bid}.self_attention.q_layernorm", "model.layers.{bid}.self_attn.q_layernorm", # persimmon - "model.layers.{bid}.self_attn.q_norm", # cohere + "model.layers.{bid}.self_attn.q_norm", # cohere olmoe "transformer.blocks.{bid}.attn.q_ln", # sea-lion "encoder.layer.{bid}.attention.self.layer_norm_q", # jina-bert-v2 "transformer.layers.{bid}.attn.q_norm", # openelm @@ -387,7 +387,7 @@ class TensorNameMap: MODEL_TENSOR.ATTN_K_NORM: ( "language_model.encoder.layers.{bid}.self_attention.k_layernorm", "model.layers.{bid}.self_attn.k_layernorm", # persimmon - "model.layers.{bid}.self_attn.k_norm", # cohere + "model.layers.{bid}.self_attn.k_norm", # cohere olmoe "transformer.blocks.{bid}.attn.k_ln", # sea-lion "encoder.layer.{bid}.attention.self.layer_norm_k", # jina-bert-v2 "transformer.layers.{bid}.attn.k_norm", # openelm diff --git a/src/llama.cpp b/src/llama.cpp index aa565723b..30997bf15 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -202,6 +202,7 @@ enum llm_arch { LLM_ARCH_COMMAND_R, LLM_ARCH_DBRX, LLM_ARCH_OLMO, + LLM_ARCH_OLMOE, LLM_ARCH_OPENELM, LLM_ARCH_ARCTIC, LLM_ARCH_DEEPSEEK2, @@ -251,6 +252,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_COMMAND_R, "command-r" }, { LLM_ARCH_DBRX, "dbrx" }, { LLM_ARCH_OLMO, "olmo" }, + { LLM_ARCH_OLMOE, "olmoe" }, { LLM_ARCH_OPENELM, "openelm" }, { LLM_ARCH_ARCTIC, "arctic" }, { LLM_ARCH_DEEPSEEK2, "deepseek2" }, @@ -1193,6 +1195,26 @@ static const std::map> LLM_TENSOR_NA { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_OLMOE, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { 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_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { 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_ARCH_OPENELM, { @@ -2277,6 +2299,7 @@ enum e_model { MODEL_MEDIUM, MODEL_LARGE, MODEL_XL, + MODEL_A1_7B, MODEL_A2_7B, MODEL_8x7B, MODEL_8x22B, @@ -5241,6 +5264,7 @@ static const char * llama_model_type_name(e_model type) { case MODEL_MEDIUM: return "0.4B"; case MODEL_LARGE: return "0.8B"; case MODEL_XL: return "1.5B"; + case MODEL_A1_7B: return "A1.7B"; case MODEL_A2_7B: return "A2.7B"; case MODEL_8x7B: return "8x7B"; case MODEL_8x22B: return "8x22B"; @@ -5791,6 +5815,14 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_OLMOE: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + switch (hparams.n_layer) { + case 16: model.type = e_model::MODEL_A1_7B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; case LLM_ARCH_OPENELM: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); @@ -8018,6 +8050,44 @@ static bool llm_load_tensors( layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); } } break; + case LLM_ARCH_OLMOE: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + + // output + { + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); + } + + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); + layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); + layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}); + layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + + layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}); + + GGML_ASSERT(n_expert > 0); + GGML_ASSERT(n_expert_used > 0); + + // MoE branch + layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}); + layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff, n_embd, n_expert}); + layer.ffn_up_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff, n_expert}); + } + } break; case LLM_ARCH_OPENELM: { model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); @@ -13832,6 +13902,134 @@ struct llm_build_context { return gf; } + // based on the build_qwen2moe() function, changes: + // * removed shared experts + // * removed bias + // * added q, k norm + struct ggml_cgraph * build_olmoe() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); + + // mutable variable, needed during the last layer of the computation to skip unused tokens + int32_t n_tokens = this->n_tokens; + + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = build_inp_pos(); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + // self_attention + { + // compute Q and K and RoPE them + struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = llm_build_norm(ctx0, Qcur, hparams, model.layers[il].attn_q_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(Qcur, "Qcur_normed", il); + + Kcur = llm_build_norm(ctx0, Kcur, hparams, model.layers[il].attn_k_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(Kcur, "Kcur_normed", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur_rope", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur_rope", il); + + cur = llm_build_kv(ctx0, lctx, kv_self, gf, + model.layers[il].wo, NULL, + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + n_tokens = n_outputs; + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // MoE branch + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_moe_ffn(ctx0, lctx, 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, + n_expert, n_expert_used, + LLM_FFN_SILU, false, + false, 0.0, + cb, il); + cb(cur, "ffn_moe_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cur = lctx.cvec.apply_to(ctx0, cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, NULL, + LLM_NORM_RMS, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + struct ggml_cgraph * build_openelm() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); @@ -15712,6 +15910,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_olmo(); } break; + case LLM_ARCH_OLMOE: + { + result = llm.build_olmoe(); + } break; case LLM_ARCH_OPENELM: { result = llm.build_openelm(); @@ -18896,6 +19098,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_QWEN: case LLM_ARCH_QWEN2: case LLM_ARCH_QWEN2MOE: + case LLM_ARCH_OLMOE: case LLM_ARCH_PHI2: case LLM_ARCH_PHI3: case LLM_ARCH_GEMMA: