From c8c353f88d8892b5de23a4f622da761130840188 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Stanis=C5=82aw=20Szymczyk?= Date: Thu, 16 May 2024 20:16:35 +0200 Subject: [PATCH] Added initial support for DeepseekV2ForCausalLM. --- convert-hf-to-gguf.py | 74 +++++++ gguf-py/gguf/constants.py | 44 +++++ gguf-py/gguf/tensor_mapping.py | 29 ++- llama.cpp | 340 ++++++++++++++++++++++++++++++++- 4 files changed, 482 insertions(+), 5 deletions(-) diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index cd875fa4a..ef1d9a2bc 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -2389,6 +2389,80 @@ class JinaBertV2Model(BertModel): self.gguf_writer.add_add_eos_token(True) +@Model.register("DeepseekV2ForCausalLM") +class DeepseekV2Model(Model): + model_arch = gguf.MODEL_ARCH.DEEPSEEK2 + + def set_vocab(self): + self._set_vocab_gpt2() + + def set_gguf_parameters(self): + super().set_gguf_parameters() + hparams = self.hparams + self.gguf_writer.add_vocab_size(hparams["vocab_size"]) + self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"]) + + if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: + if self.hparams["rope_scaling"].get("type") == "yarn": + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN) + self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) + self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"]) + + self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"]) + self.gguf_writer.add_value_length(hparams["v_head_dim"]) + self.gguf_writer.add_expert_count(hparams["n_routed_experts"]) + + _experts: list[dict[str, Tensor]] | None = None + + 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") + + # 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: + 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)] + + def write_tensors(self): + super().write_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}") + + ###### CONVERSION LOGIC ###### diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 978fcada3..0a732022d 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -139,6 +139,7 @@ class MODEL_ARCH(IntEnum): COMMAND_R = auto() DBRX = auto() OLMO = auto() + DEEPSEEK2 = auto() class MODEL_TENSOR(IntEnum): @@ -181,6 +182,12 @@ class MODEL_TENSOR(IntEnum): SSM_A = auto() SSM_D = auto() SSM_OUT = auto() + ATTN_Q_A = auto() + ATTN_Q_B = auto() + ATTN_KV_A_MQA = auto() + ATTN_KV_B = auto() + ATTN_Q_A_NORM = auto() + ATTN_KV_A_NORM = auto() MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { @@ -217,6 +224,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.COMMAND_R: "command-r", MODEL_ARCH.DBRX: "dbrx", MODEL_ARCH.OLMO: "olmo", + MODEL_ARCH.DEEPSEEK2: "deepseek2", } TENSOR_NAMES: dict[MODEL_TENSOR, str] = { @@ -259,6 +267,12 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = { MODEL_TENSOR.SSM_A: "blk.{bid}.ssm_a", MODEL_TENSOR.SSM_D: "blk.{bid}.ssm_d", MODEL_TENSOR.SSM_OUT: "blk.{bid}.ssm_out", + MODEL_TENSOR.ATTN_Q_A: "blk.{bid}.attn_q_a", + MODEL_TENSOR.ATTN_Q_B: "blk.{bid}.attn_q_b", + MODEL_TENSOR.ATTN_KV_A_MQA: "blk.{bid}.attn_kv_a_mqa", + MODEL_TENSOR.ATTN_KV_B: "blk.{bid}.attn_kv_b", + MODEL_TENSOR.ATTN_Q_A_NORM: "blk.{bid}.attn_q_a_norm", + MODEL_TENSOR.ATTN_KV_A_NORM: "blk.{bid}.attn_kv_a_norm", } MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { @@ -743,6 +757,32 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, ], + MODEL_ARCH.DEEPSEEK2: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q_A, + MODEL_TENSOR.ATTN_Q_B, + MODEL_TENSOR.ATTN_KV_A_MQA, + MODEL_TENSOR.ATTN_KV_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, + ], # TODO } @@ -779,6 +819,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.ROPE_FREQS, MODEL_TENSOR.ATTN_ROT_EMBD, ], + MODEL_ARCH.DEEPSEEK2: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], } # diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 8e1cac915..383d8440b 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -255,6 +255,7 @@ class TensorNameMap: MODEL_TENSOR.FFN_UP_SHEXP: ( "model.layers.{bid}.mlp.shared_expert.up_proj", # qwen2moe + "model.layers.{bid}.mlp.shared_experts.up_proj", # deepseek2 ), # AWQ-activation gate @@ -283,6 +284,7 @@ class TensorNameMap: MODEL_TENSOR.FFN_GATE_SHEXP: ( "model.layers.{bid}.mlp.shared_expert.gate_proj", # qwen2moe + "model.layers.{bid}.mlp.shared_experts.gate_proj", # deepseek2 ), # Feed-forward down @@ -317,6 +319,7 @@ class TensorNameMap: MODEL_TENSOR.FFN_DOWN_SHEXP: ( "model.layers.{bid}.mlp.shared_expert.down_proj", # qwen2moe + "model.layers.{bid}.mlp.shared_experts.down_proj", # deepseek2 ), MODEL_TENSOR.ATTN_Q_NORM: ( @@ -380,6 +383,30 @@ class TensorNameMap: "model.layers.{bid}.out_proj", "backbone.layers.{bid}.mixer.out_proj", ), + + MODEL_TENSOR.ATTN_Q_A: ( + "model.layers.{bid}.self_attn.q_a_proj", # deepseek2 + ), + + MODEL_TENSOR.ATTN_Q_B: ( + "model.layers.{bid}.self_attn.q_b_proj", # deepseek2 + ), + + MODEL_TENSOR.ATTN_KV_A_MQA: ( + "model.layers.{bid}.self_attn.kv_a_proj_with_mqa", # deepseek2 + ), + + MODEL_TENSOR.ATTN_KV_B: ( + "model.layers.{bid}.self_attn.kv_b_proj", # deepseek2 + ), + + MODEL_TENSOR.ATTN_Q_A_NORM: ( + "model.layers.{bid}.self_attn.q_a_layernorm", # deepseek2 + ), + + MODEL_TENSOR.ATTN_KV_A_NORM: ( + "model.layers.{bid}.self_attn.kv_a_layernorm", # deepseek2 + ), } mapping: dict[str, tuple[MODEL_TENSOR, str]] @@ -398,7 +425,7 @@ class TensorNameMap: if tensor not in MODEL_TENSORS[arch]: continue # TODO: make this configurable - n_experts = 60 + n_experts = 160 for xid in range(n_experts): tensor_name = TENSOR_NAMES[tensor].format(bid = bid, xid = xid) self.mapping[tensor_name] = (tensor, tensor_name) diff --git a/llama.cpp b/llama.cpp index 7d26966e4..53959d83a 100644 --- a/llama.cpp +++ b/llama.cpp @@ -110,7 +110,7 @@ #endif #define LLAMA_MAX_NODES 8192 -#define LLAMA_MAX_EXPERTS 60 +#define LLAMA_MAX_EXPERTS 160 // // logging @@ -229,6 +229,7 @@ enum llm_arch { LLM_ARCH_COMMAND_R, LLM_ARCH_DBRX, LLM_ARCH_OLMO, + LLM_ARCH_DEEPSEEK2, LLM_ARCH_UNKNOWN, }; @@ -266,6 +267,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_COMMAND_R, "command-r" }, { LLM_ARCH_DBRX, "dbrx" }, { LLM_ARCH_OLMO, "olmo" }, + { LLM_ARCH_DEEPSEEK2, "deepseek2" }, { LLM_ARCH_UNKNOWN, "(unknown)" }, }; @@ -476,6 +478,12 @@ enum llm_tensor { LLM_TENSOR_SSM_A, LLM_TENSOR_SSM_D, LLM_TENSOR_SSM_OUT, + LLM_TENSOR_ATTN_Q_A, + LLM_TENSOR_ATTN_Q_B, + LLM_TENSOR_ATTN_KV_A_MQA, + LLM_TENSOR_ATTN_KV_B, + LLM_TENSOR_ATTN_Q_A_NORM, + LLM_TENSOR_ATTN_KV_A_NORM, }; static const std::map> LLM_TENSOR_NAMES = { @@ -1052,6 +1060,34 @@ static const std::map> LLM_TENSOR_NA { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_DEEPSEEK2, + { + { 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_A_NORM, "blk.%d.attn_q_a_norm" }, + { LLM_TENSOR_ATTN_KV_A_NORM, "blk.%d.attn_kv_a_norm" }, + { LLM_TENSOR_ATTN_Q_A, "blk.%d.attn_q_a" }, + { LLM_TENSOR_ATTN_Q_B, "blk.%d.attn_q_b" }, + { LLM_TENSOR_ATTN_KV_A_MQA, "blk.%d.attn_kv_a_mqa" }, + { LLM_TENSOR_ATTN_KV_B, "blk.%d.attn_kv_b" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { 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_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, { @@ -1895,6 +1931,8 @@ struct llama_layer { struct ggml_tensor * attn_k_norm_b; struct ggml_tensor * attn_out_norm; struct ggml_tensor * attn_out_norm_b; + struct ggml_tensor * attn_q_a_norm; + struct ggml_tensor * attn_kv_a_norm; // attention struct ggml_tensor * wq; @@ -1902,6 +1940,10 @@ struct llama_layer { struct ggml_tensor * wv; struct ggml_tensor * wo; struct ggml_tensor * wqkv; + struct ggml_tensor * wq_a; + struct ggml_tensor * wq_b; + struct ggml_tensor * wkv_a_mqa; + struct ggml_tensor * wkv_b; // attention bias struct ggml_tensor * bq; @@ -4261,6 +4303,11 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_DEEPSEEK2: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + model.type = e_model::MODEL_UNKNOWN; + } break; default: (void)0; } @@ -4920,8 +4967,6 @@ static bool llm_load_tensors( throw std::runtime_error("model has expert layers but no expert layers are used"); } - GGML_ASSERT(n_embd_gqa == n_embd_k_gqa); - ggml_context * ctx_input = ctx_map.at(model.buft_input.buft); ggml_context * ctx_output = ctx_map.at(model.buft_output.buft); ggml_context * ctx_output_split = ctx_map.at(model.buft_output.buft_matrix); @@ -6060,6 +6105,67 @@ 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_DEEPSEEK2: + { + // TODO maybe move some of these to hparams + const uint32_t n_shared_experts = 2; + const uint32_t moe_intermediate_size = 1536; + const uint32_t q_lora_rank = 1536; + const uint32_t kv_lora_rank = 512; + const uint32_t first_k_dense_replace = 1; + + // kept original names of these parameters from HF transformers code for clarity + const uint32_t qk_rope_head_dim = hparams.n_rot; + const uint32_t qk_nope_head_dim = hparams.n_embd_head_k - hparams.n_rot; + + 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.attn_q_a_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}); + layer.attn_kv_a_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}); + + layer.wq_a = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}); + layer.wq_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, hparams.n_head * hparams.n_embd_head_k}); + layer.wkv_a_mqa = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + qk_rope_head_dim}); + layer.wkv_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_KV_B, "weight", i), {kv_lora_rank, hparams.n_head * (qk_nope_head_dim + hparams.n_embd_head_v)}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {hparams.n_head * hparams.n_embd_head_v, n_embd}); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + + if ((uint32_t) i < first_k_dense_replace) { + layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + } else { + layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}); + + GGML_ASSERT(hparams.n_expert > 0); + GGML_ASSERT(hparams.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, moe_intermediate_size, n_expert}); + layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {moe_intermediate_size, n_embd, n_expert}); + layer.ffn_up_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, moe_intermediate_size, n_expert}); + + // Shared expert branch + layer.ffn_gate_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, moe_intermediate_size * n_shared_experts}); + layer.ffn_down_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { moe_intermediate_size * n_shared_experts, n_embd}); + layer.ffn_up_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, moe_intermediate_size * n_shared_experts}); + } + } + } break; default: throw std::runtime_error("unknown architecture"); } @@ -6700,7 +6806,7 @@ static struct ggml_tensor * llm_build_kqv( struct ggml_tensor * kqv_merged = ggml_permute(ctx, kqv, 0, 2, 1, 3); cb(kqv_merged, "kqv_merged", il); - cur = ggml_cont_2d(ctx, kqv_merged, n_embd_head_k*n_head, n_tokens); + cur = ggml_cont_2d(ctx, kqv_merged, n_embd_head_v*n_head, n_tokens); cb(cur, "kqv_merged_cont", il); } @@ -10779,6 +10885,227 @@ struct llm_build_context { return gf; } + + struct ggml_cgraph * build_deepseek2() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + // mutable variable, needed during the last layer of the computation to skip unused tokens + int32_t n_tokens = this->n_tokens; + + // TODO maybe move some of these to hparams + const uint32_t first_k_dense_replace = 1; + const uint32_t kv_lora_rank = 512; + + // kept original names of these parameters from HF transformers code for clarity + const uint32_t qk_rope_head_dim = hparams.n_rot; + const uint32_t qk_nope_head_dim = hparams.n_embd_head_k - hparams.n_rot; + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + // {n_embd, n_tokens} + 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 + { + // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens} + struct ggml_tensor * q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur); + cb(q, "q", il); + + q = llm_build_norm(ctx0, q, hparams, + model.layers[il].attn_q_a_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(q, "q", il); + + // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens} + q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q); + cb(q, "q", il); + + // split into {n_head * qk_nope_head_dim, n_tokens} + struct ggml_tensor * q_nope = ggml_view_3d(ctx0, q, qk_nope_head_dim, n_head, n_tokens, ggml_element_size(q) * hparams.n_embd_head_k, ggml_element_size(q) * hparams.n_embd_head_k * n_head, 0); + // and {n_head * qk_rope_head_dim, n_tokens} + struct ggml_tensor * q_pe = ggml_view_3d(ctx0, q, qk_rope_head_dim, n_head, n_tokens, ggml_element_size(q) * hparams.n_embd_head_k, ggml_element_size(q) * hparams.n_embd_head_k * n_head, ggml_element_size(q) * qk_nope_head_dim); + + q_nope = ggml_cont(ctx0, q_nope); + cb(q_nope, "q_nope", il); + + q_pe = ggml_cont(ctx0, q_pe); + cb(q_pe, "q_pe", il); + + // {n_embd, kv_lora_rank + qk_rope_head_dim} * {n_embd, n_tokens} -> {kv_lora_rank + qk_rope_head_dim, n_tokens} + struct ggml_tensor * compressed_kv_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur); + cb(compressed_kv_pe, "compressed_kv_pe", il); + + // split into {kv_lora_rank, n_tokens} + struct ggml_tensor * compressed_kv = ggml_view_2d(ctx0, compressed_kv_pe, kv_lora_rank, n_tokens, compressed_kv_pe->nb[1], 0); + // and {qk_rope_head_dim, n_tokens} + struct ggml_tensor * k_pe = ggml_view_2d(ctx0, compressed_kv_pe, qk_rope_head_dim, n_tokens, compressed_kv_pe->nb[1], ggml_element_size(compressed_kv_pe)*kv_lora_rank); + + k_pe = ggml_cont(ctx0, k_pe); + cb(k_pe, "k_pe", il); + + compressed_kv = llm_build_norm(ctx0, compressed_kv, hparams, + model.layers[il].attn_kv_a_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(compressed_kv, "compressed_kv", il); + + // {kv_lora_rank, n_head * (qk_nope_head_dim + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (qk_nope_head_dim + n_embd_head_v), n_tokens} + struct ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, compressed_kv); + cb(kv, "kv", il); + + // split into {n_head * qk_nope_head_dim, n_tokens} + struct ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, qk_nope_head_dim, n_head, n_tokens, ggml_element_size(kv) * (qk_nope_head_dim + hparams.n_embd_head_v), ggml_element_size(kv) * n_head * (qk_nope_head_dim + hparams.n_embd_head_v), 0); + // and {n_head * n_embd_head_v, n_tokens} + struct ggml_tensor * value_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens, ggml_element_size(kv) * (qk_nope_head_dim + hparams.n_embd_head_v), ggml_element_size(kv) * n_head * (qk_nope_head_dim + hparams.n_embd_head_v), ggml_element_size(kv) * qk_nope_head_dim); + + value_states = ggml_dup(ctx0, value_states); + cb(value_states, "value_states", il); + + value_states = ggml_reshape_2d(ctx0, value_states, hparams.n_embd_head_v * n_head, n_tokens); + + q_pe = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, q_pe, qk_rope_head_dim, n_head, n_tokens), inp_pos, + n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(q_pe, "q_pe", il); + + // shared RoPE key + k_pe = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, k_pe, qk_rope_head_dim, 1, n_tokens), inp_pos, + n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(k_pe, "k_pe", il); + + struct ggml_tensor * query_states = ggml_new_tensor_3d(ctx0, q_nope->type, hparams.n_embd_head_k, n_head, n_tokens); + cb(query_states, "query_states", il); + query_states = ggml_set_inplace(ctx0, query_states, q_nope, query_states->nb[1], query_states->nb[2], query_states->nb[3], 0); + query_states = ggml_set_inplace(ctx0, query_states, q_pe, query_states->nb[1], query_states->nb[2], query_states->nb[3], ggml_element_size(query_states) * qk_nope_head_dim); + + k_pe = ggml_repeat(ctx0, k_pe, q_pe); + cb(k_pe, "k_pe", il); + + struct ggml_tensor * key_states = ggml_new_tensor_3d(ctx0, q_nope->type, hparams.n_embd_head_k, n_head, n_tokens); + cb(key_states, "key_states", il); + key_states = ggml_set_inplace(ctx0, key_states, k_nope, key_states->nb[1], key_states->nb[2], key_states->nb[3], 0); + key_states = ggml_set_inplace(ctx0, key_states, k_pe, key_states->nb[1], key_states->nb[2], key_states->nb[3], ggml_element_size(key_states) * qk_nope_head_dim); + + // TODO see if we can avoid these operations by permuting + // rows/columns of some model tensors during model conversion + query_states = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_4d(ctx0, query_states, 2, hparams.n_embd_head_k / 2, n_head, n_tokens))); + cb(query_states, "query_states", il); + + query_states = ggml_reshape_3d(ctx0, query_states, hparams.n_embd_head_k, n_head, n_tokens); + cb(query_states, "query_states", il); + + key_states = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_4d(ctx0, key_states, 2, hparams.n_embd_head_k / 2, n_head, n_tokens))); + cb(key_states, "key_states", il); + + key_states = ggml_reshape_3d(ctx0, key_states, hparams.n_embd_head_k, n_head, n_tokens); + cb(key_states, "key_states", il); + + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, + model.layers[il].wo, model.layers[il].bo, + key_states, value_states, query_states, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(hparams.n_embd_head_k)), 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); + + if ((uint32_t) il < first_k_dense_replace) { + 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_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + } else { + // 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); + + ggml_tensor * moe_out = + llm_build_moe_ffn(ctx0, 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, true, + cb, il); + cb(moe_out, "ffn_moe_out", il); + + // FFN shared expert + { + ggml_tensor * ffn_shexp = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up_shexp, NULL, + model.layers[il].ffn_gate_shexp, NULL, + model.layers[il].ffn_down_shexp, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(ffn_shexp, "ffn_shexp", il); + + moe_out = ggml_add(ctx0, moe_out, ffn_shexp); + cb(moe_out, "ffn_out", il); + + cur = moe_out; + } + } + + cur = ggml_add(ctx0, cur, ffn_inp); + 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 = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } }; static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector & ids) { @@ -10993,6 +11320,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_olmo(); } break; + case LLM_ARCH_DEEPSEEK2: + { + result = llm.build_deepseek2(); + } break; default: GGML_ASSERT(false); } @@ -16008,6 +16339,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_XVERSE: case LLM_ARCH_COMMAND_R: case LLM_ARCH_OLMO: + case LLM_ARCH_DEEPSEEK2: return LLAMA_ROPE_TYPE_NORM; // the pairs of head values are offset by n_rot/2