diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 2c6d5d95b..d995ed764 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -1841,6 +1841,60 @@ class MiniCPMModel(Model): return [(self.map_tensor_name(name), data_torch)] +@Model.register("MiniCPM3ForCausalLM") +class MiniCPM3Model(Model): + model_arch = gguf.MODEL_ARCH.MINICPM3 + + def set_gguf_parameters(self): + hparams = self.hparams + + rope_dims = hparams["qk_rope_head_dim"] + + self.gguf_writer.add_file_type(self.ftype) + self.gguf_writer.add_context_length(hparams["max_position_embeddings"]) + self.gguf_writer.add_embedding_length(hparams["hidden_size"]) + self.gguf_writer.add_block_count(self.block_count) + self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"]) + self.gguf_writer.add_head_count(hparams["num_attention_heads"]) + self.gguf_writer.add_head_count_kv(hparams["num_key_value_heads"]) + self.gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"]) + self.gguf_writer.add_vocab_size(hparams["vocab_size"]) + if "q_lora_rank" in hparams and hparams["q_lora_rank"] is not None: + self.gguf_writer.add_q_lora_rank(hparams["q_lora_rank"]) + self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"]) + self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"]) + self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"]) + + rope_scaling = self.find_hparam(['rope_scaling'], True) + if rope_scaling is None: + return + + long_factors = rope_scaling.get('long_factor', None) + short_factors = rope_scaling.get('short_factor', None) + + if long_factors is None or short_factors is None: + raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor') + + if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2: + raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}') + + self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG] + ".weight", np.array(long_factors, dtype=np.float32)) + self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32)) + + def set_vocab(self): + self._set_vocab_llama_hf() + + def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor: + if n_kv_head is not None and n_head != n_kv_head: + n_head //= n_kv_head + + return ( + weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:]) + .swapaxes(1, 2) + .reshape(weights.shape) + ) + + @Model.register("QWenLMHeadModel") class QwenModel(Model): model_arch = gguf.MODEL_ARCH.QWEN diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index c87d08782..2c8545455 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -210,6 +210,7 @@ class MODEL_ARCH(IntEnum): ORION = auto() INTERNLM2 = auto() MINICPM = auto() + MINICPM3 = auto() GEMMA = auto() GEMMA2 = auto() STARCODER2 = auto() @@ -364,6 +365,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.ORION: "orion", MODEL_ARCH.INTERNLM2: "internlm2", MODEL_ARCH.MINICPM: "minicpm", + MODEL_ARCH.MINICPM3: "minicpm3", MODEL_ARCH.GEMMA: "gemma", MODEL_ARCH.GEMMA2: "gemma2", MODEL_ARCH.STARCODER2: "starcoder2", @@ -867,6 +869,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN_EXP, MODEL_TENSOR.FFN_UP_EXP, ], + MODEL_ARCH.MINICPM3: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + 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.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], MODEL_ARCH.GEMMA: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, diff --git a/src/llama.cpp b/src/llama.cpp index c917d1c7b..aa565723b 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -193,6 +193,7 @@ enum llm_arch { LLM_ARCH_ORION, LLM_ARCH_INTERNLM2, LLM_ARCH_MINICPM, + LLM_ARCH_MINICPM3, LLM_ARCH_GEMMA, LLM_ARCH_GEMMA2, LLM_ARCH_STARCODER2, @@ -241,6 +242,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_ORION, "orion" }, { LLM_ARCH_INTERNLM2, "internlm2" }, { LLM_ARCH_MINICPM, "minicpm" }, + { LLM_ARCH_MINICPM3, "minicpm3" }, { LLM_ARCH_GEMMA, "gemma" }, { LLM_ARCH_GEMMA2, "gemma2" }, { LLM_ARCH_STARCODER2, "starcoder2" }, @@ -1034,6 +1036,29 @@ static const std::map> LLM_TENSOR_NA { LLM_TENSOR_FFN_UP_EXP, "blk.%d.ffn_up.%d" }, }, }, + { + LLM_ARCH_MINICPM3, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FACTORS_LONG, "rope_factors_long" }, + { LLM_TENSOR_ROPE_FACTORS_SHORT, "rope_factors_short" }, + { 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, "blk.%d.attn_q" }, + { 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_ARCH_GEMMA, { @@ -5390,6 +5415,17 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_MINICPM3: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + 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); + + switch (hparams.n_layer) { + case 62: model.type = e_model::MODEL_4B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; case LLM_ARCH_GROK: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); @@ -6897,6 +6933,54 @@ static bool llm_load_tensors( } } } break; + case LLM_ARCH_MINICPM3: + { + const int64_t n_embd_head_qk_rope = hparams.n_rot; + const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot; + + const int64_t q_lora_rank = hparams.n_lora_q; + const int64_t kv_lora_rank = hparams.n_lora_kv; + 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}, llama_model_loader::TENSOR_NOT_REQUIRED); + + // if output is NULL, init from the input tok embed + if (model.output == NULL) { + model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); + } + } + + 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, n_head * 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 + (n_embd_head_qk_rope)}); + layer.wkv_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_KV_B, "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_head * ( n_embd_head_v), n_embd}); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + + 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}); + + layer.rope_long = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FACTORS_LONG, "weight"), { n_embd_head_qk_rope/2 }, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0)); + layer.rope_short = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight"), { n_embd_head_qk_rope/2 }, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0)); + } + } break; case LLM_ARCH_GROK: { if (n_expert == 0) { @@ -12843,6 +12927,215 @@ struct llm_build_context { return gf; } + struct ggml_cgraph * build_minicpm3() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); + + //TODO: if the model varies, these parameters need to be read from the model + const int64_t n_embd_base = 256; + const float scale_embd = 12.0f; + const float scale_depth = 1.4f; + const float kq_scale = 1.0f / sqrtf(float(hparams.n_embd_head_k)); + + const uint32_t n_embd_head_qk_rope = hparams.n_rot; + const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot; + const uint32_t kv_lora_rank = hparams.n_lora_kv; + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); + + // scale the input embeddings + inpL = ggml_scale(ctx0, inpL, scale_embd); + cb(inpL, "inp_scaled", -1); + + // 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; + + struct ggml_tensor * rope_factors = build_rope_factors(il); + // 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 + { + struct ggml_tensor * q = NULL; + // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens} + 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 * n_embd_head_qk_nope, n_tokens} + struct ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(q->type, hparams.n_embd_head_k), + ggml_row_size(q->type, hparams.n_embd_head_k * n_head), + 0); + cb(q_nope, "q_nope", il); + + // and {n_head * n_embd_head_qk_rope, n_tokens} + struct ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, + ggml_row_size(q->type, hparams.n_embd_head_k), + ggml_row_size(q->type, hparams.n_embd_head_k * n_head), + ggml_row_size(q->type, n_embd_head_qk_nope)); + cb(q_pe, "q_pe", il); + + // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens} + struct ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur); + cb(kv_pe_compresseed, "kv_pe_compresseed", il); + + // split into {kv_lora_rank, n_tokens} + struct ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens, + kv_pe_compresseed->nb[1], + 0); + cb(kv_compressed, "kv_compressed", il); + + // and {n_embd_head_qk_rope, n_tokens} + struct ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens, + kv_pe_compresseed->nb[1], + kv_pe_compresseed->nb[1], + ggml_row_size(kv_pe_compresseed->type, kv_lora_rank)); + cb(k_pe, "k_pe", il); + + kv_compressed = ggml_cont(ctx0, kv_compressed); // TODO: the CUDA backend does not support non-contiguous norm + kv_compressed = llm_build_norm(ctx0, kv_compressed, hparams, + model.layers[il].attn_kv_a_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(kv_compressed, "kv_compressed", il); + + // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens} + struct ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed); + cb(kv, "kv", il); + + // split into {n_head * n_embd_head_qk_nope, n_tokens} + struct 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 + hparams.n_embd_head_v), + ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)), + 0); + cb(k_nope, "k_nope", il); + + // and {n_head * n_embd_head_v, n_tokens} + struct ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens, + ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)), + ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head), + ggml_row_size(kv->type, (n_embd_head_qk_nope))); + cb(v_states, "v_states", il); + + v_states = ggml_cont(ctx0, v_states); + cb(v_states, "v_states", il); + + v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens, + ggml_row_size(kv->type, hparams.n_embd_head_v * n_head), + 0); + cb(v_states, "v_states", il); + + q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend does not support non-contiguous RoPE + q_pe = ggml_rope_ext( + ctx0, q_pe, inp_pos, rope_factors, + 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); + + // shared RoPE key + k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend does not support non-contiguous RoPE + k_pe = ggml_rope_ext( + ctx0, k_pe, inp_pos, rope_factors, + 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); + + struct ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0); + cb(q_states, "q_states", il); + + struct ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0); + cb(k_states, "k_states", il); + + cur = llm_build_kv(ctx0, lctx, kv_self, gf, + model.layers[il].wo, NULL, + k_states, v_states, q_states, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct 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); + } + + // scale_res - scale the hidden states for residual connection + const float scale_res = scale_depth/sqrtf(float(n_layer)); + cur = ggml_scale(ctx0, cur, scale_res); + cb(cur, "hidden_scaled", il); + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + 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, lctx, 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, cb, il); + cb(cur, "ffn_out", il); + } + + // scale the hidden states for residual connection + cur = ggml_scale(ctx0, cur, scale_res); + cb(cur, "hidden_scaled_ffn", 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 scaling + const float scale_lmhead = float(n_embd_base)/float(n_embd); + cur = ggml_scale(ctx0, cur, scale_lmhead); + cb(cur, "lmhead_scaling", -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_gemma() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); @@ -15383,6 +15676,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_minicpm(); } break; + case LLM_ARCH_MINICPM3: + { + result = llm.build_minicpm3(); + } break; case LLM_ARCH_GEMMA: { result = llm.build_gemma(); @@ -18609,6 +18906,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_CODESHELL: case LLM_ARCH_NEMOTRON: case LLM_ARCH_EXAONE: + case LLM_ARCH_MINICPM3: return LLAMA_ROPE_TYPE_NEOX; // all model arches should be listed explicitly here