diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 7a0a8c3db..6ab7f486e 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -201,6 +201,8 @@ class Model: return PlamoModel if model_architecture == "CodeShellForCausalLM": return CodeShellModel + if model_architecture == "OrionForCausalLM": + return OrionModel return Model def _is_model_safetensors(self) -> bool: @@ -250,6 +252,8 @@ class Model: return gguf.MODEL_ARCH.PLAMO if arch == "CodeShellForCausalLM": return gguf.MODEL_ARCH.CODESHELL + if arch == "OrionForCausalLM": + return gguf.MODEL_ARCH.ORION raise NotImplementedError(f'Architecture "{arch}" not supported!') @@ -572,6 +576,83 @@ class MPTModel(Model): self.gguf_writer.add_tensor("output.weight", data) +class OrionModel(Model): + def set_vocab(self): + self._set_vocab_sentencepiece() + + def set_gguf_parameters(self): + block_count = self.hparams["num_hidden_layers"] + head_count = self.hparams["num_attention_heads"] + head_count_kv = self.hparams.get("num_key_value_heads", head_count) + hf_repo = self.hparams.get("_name_or_path", "") + + ctx_length = 0 + if "max_sequence_length" in self.hparams: + ctx_length = self.hparams["max_sequence_length"] + elif "max_position_embeddings" in self.hparams: + ctx_length = self.hparams["max_position_embeddings"] + elif "model_max_length" in self.hparams: + ctx_length = self.hparams["model_max_length"] + else: + print("gguf: can not find ctx length parameter.") + sys.exit() + + self.gguf_writer.add_file_type(self.ftype) + self.gguf_writer.add_name(self.dir_model.name) + self.gguf_writer.add_source_hf_repo(hf_repo) + self.gguf_writer.add_tensor_data_layout("Meta AI original pth") + self.gguf_writer.add_context_length(ctx_length) + self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) + self.gguf_writer.add_block_count(block_count) + self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"]) + self.gguf_writer.add_head_count(head_count) + self.gguf_writer.add_head_count_kv(head_count_kv) + self.gguf_writer.add_layer_norm_eps(self.hparams["rms_norm_eps"]) + + def write_tensors(self): + # Collect tensors from generator object + model_kv = dict(self.get_tensors()) + block_count = self.hparams["num_hidden_layers"] + tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count) + + for name, data_torch in model_kv.items(): + # we don't need these + if name.endswith(".rotary_emb.inv_freq"): + continue + + old_dtype = data_torch.dtype + + # convert any unsupported data types to float32 + if data_torch.dtype not in (torch.float16, torch.float32): + data_torch = data_torch.to(torch.float32) + + data = data_torch.squeeze().numpy() + + # map tensor names + new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) + if new_name is None: + print(f"Can not map tensor {name!r}") + sys.exit() + + n_dims = len(data.shape) + data_dtype = data.dtype + + # if f32 desired, convert any float16 to float32 + if self.ftype == 0 and data_dtype == np.float16: + data = data.astype(np.float32) + + # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 + if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1: + data = data.astype(np.float32) + + # if f16 desired, convert any float32 2-dim weight tensors to float16 + if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: + data = data.astype(np.float16) + + print(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + self.gguf_writer.add_tensor(new_name, data) + + class BaichuanModel(Model): def set_vocab(self): self._set_vocab_sentencepiece() diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 2d9c33c7d..f5c933a41 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -101,6 +101,7 @@ class MODEL_ARCH(IntEnum): PHI2 = auto() PLAMO = auto() CODESHELL = auto() + ORION = auto() class MODEL_TENSOR(IntEnum): @@ -151,6 +152,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.PHI2: "phi2", MODEL_ARCH.PLAMO: "plamo", MODEL_ARCH.CODESHELL: "codeshell", + MODEL_ARCH.ORION: "orion", } TENSOR_NAMES: dict[MODEL_TENSOR, str] = { @@ -427,7 +429,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_NORM, MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, - ] + ], + MODEL_ARCH.ORION: [ + 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.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], # TODO } @@ -452,6 +470,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.ROPE_FREQS, MODEL_TENSOR.ATTN_ROT_EMBD, ], + MODEL_ARCH.ORION: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], } # diff --git a/llama.cpp b/llama.cpp index b03b67e16..4cd0f16eb 100644 --- a/llama.cpp +++ b/llama.cpp @@ -196,6 +196,7 @@ enum llm_arch { LLM_ARCH_PHI2, LLM_ARCH_PLAMO, LLM_ARCH_CODESHELL, + LLM_ARCH_ORION, LLM_ARCH_UNKNOWN, }; @@ -217,6 +218,7 @@ static std::map LLM_ARCH_NAMES = { { LLM_ARCH_PHI2, "phi2" }, { LLM_ARCH_PLAMO, "plamo" }, { LLM_ARCH_CODESHELL, "codeshell" }, + { LLM_ARCH_ORION, "orion" }, }; enum llm_kv { @@ -641,6 +643,25 @@ static std::map> LLM_TENSOR_NAMES = { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_ORION, + { + { 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_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, { LLM_ARCH_UNKNOWN, @@ -1332,6 +1353,7 @@ enum e_model { MODEL_7B, MODEL_8B, MODEL_13B, + MODEL_14B, MODEL_15B, MODEL_30B, MODEL_34B, @@ -2683,6 +2705,7 @@ static const char * llama_model_type_name(e_model type) { case MODEL_7B: return "7B"; case MODEL_8B: return "8B"; case MODEL_13B: return "13B"; + case MODEL_14B: return "14B"; case MODEL_15B: return "15B"; case MODEL_30B: return "30B"; case MODEL_34B: return "34B"; @@ -2950,7 +2973,15 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_ORION: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + switch (hparams.n_layer) { + case 40: model.type = e_model::MODEL_14B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; default: (void)0; } @@ -3933,6 +3964,38 @@ static bool llm_load_tensors( layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}); } } break; + case LLM_ARCH_ORION: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + { + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {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_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", 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.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", 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}); + } + } break; + + default: throw std::runtime_error("unknown architecture"); } @@ -4563,6 +4626,126 @@ struct llm_build_context { ctx0 = nullptr; } } + struct ggml_cgraph * build_orion() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + 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, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); + cb(inpL, "inp_embd", -1); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); + cb(inp_pos, "inp_pos", -1); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); + cb(KQ_mask, "KQ_mask", -1); + + // shift the entire K-cache if needed + if (do_rope_shift) { + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); + } + + 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, model.layers[il].attn_norm_b, + LLM_NORM, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, 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); + // } + + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, 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); + // } + + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, 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_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + model.layers[il].wo, NULL, + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + cb(cur, "kqv_out", 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, model.layers[il].ffn_norm_b, + LLM_NORM, 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); + + 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, model.output_norm_b, + LLM_NORM, 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; + } + + struct ggml_cgraph * build_llama() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); @@ -6520,6 +6703,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_codeshell(); } break; + case LLM_ARCH_ORION: + { + result = llm.build_orion(); + } break; default: GGML_ASSERT(false); }