From ea9c8e11436ad50719987fa23a289c74b7b40d40 Mon Sep 17 00:00:00 2001 From: Jared Van Bortel Date: Tue, 13 Feb 2024 12:03:53 -0500 Subject: [PATCH] llama : add support for Nomic Embed (#5468) --- convert-hf-to-gguf.py | 117 ++++++++++++------- gguf-py/gguf/constants.py | 56 +++++---- gguf-py/gguf/tensor_mapping.py | 12 +- llama.cpp | 201 ++++++++++++++++++++++++--------- 4 files changed, 273 insertions(+), 113 deletions(-) diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 5adfdc143a41f..ae471481d4a70 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -10,7 +10,7 @@ import sys from enum import IntEnum from pathlib import Path -from typing import TYPE_CHECKING, Any, ContextManager, Iterator, cast +from typing import TYPE_CHECKING, Any, ContextManager, Iterator, Sequence, cast import numpy as np import torch @@ -25,15 +25,6 @@ from convert import HfVocab -# check for any of the given keys in the dictionary and return the value of the first key found -def get_key_opts(d, keys): - for k in keys: - if k in d: - return d[k] - print(f"Could not find any of {keys}") - sys.exit() - - ###### MODEL DEFINITIONS ###### class SentencePieceTokenTypes(IntEnum): @@ -58,6 +49,15 @@ def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian: self.hparams = Model.load_hparams(self.dir_model) self.model_arch = self._get_model_architecture() self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=False) + self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"]) + + def find_hparam(self, keys: Sequence[str], optional: bool = False) -> Any: + key = next((k for k in keys if k in self.hparams), None) + if key is not None: + return self.hparams[key] + if optional: + return None + raise KeyError(f"could not find any of: {keys}") def set_vocab(self): self._set_vocab_gpt2() @@ -79,28 +79,33 @@ def get_tensors(self) -> Iterator[tuple[str, Tensor]]: def set_gguf_parameters(self): self.gguf_writer.add_name(self.dir_model.name) - self.gguf_writer.add_block_count(self.hparams.get( - "n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")), - )) - if (n_ctx := self.hparams.get("max_position_embeddings")) is not None: + self.gguf_writer.add_block_count(self.block_count) + + if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None: self.gguf_writer.add_context_length(n_ctx) - if (n_embd := self.hparams.get("hidden_size")) is not None: - self.gguf_writer.add_embedding_length(n_embd) - if (n_ff := self.hparams.get("intermediate_size")) is not None: + + n_embd = self.find_hparam(["hidden_size", "n_embd"]) + self.gguf_writer.add_embedding_length(n_embd) + + if (n_ff := self.find_hparam(["intermediate_size", "n_inner"], optional=True)) is not None: self.gguf_writer.add_feed_forward_length(n_ff) - if (n_head := self.hparams.get("num_attention_heads")) is not None: - self.gguf_writer.add_head_count(n_head) + + n_head = self.find_hparam(["num_attention_heads", "n_head"]) + self.gguf_writer.add_head_count(n_head) + if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None: self.gguf_writer.add_head_count_kv(n_head_kv) - if (n_rms_eps := self.hparams.get("rms_norm_eps")) is not None: - self.gguf_writer.add_layer_norm_rms_eps(n_rms_eps) + if (f_rms_eps := self.hparams.get("rms_norm_eps")) is not None: + self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps) + if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon"], optional=True)) is not None: + self.gguf_writer.add_layer_norm_eps(f_norm_eps) if (n_experts := self.hparams.get("num_local_experts")) is not None: self.gguf_writer.add_expert_count(n_experts) if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None: self.gguf_writer.add_expert_used_count(n_experts_used) - self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True)) + self.gguf_writer.add_file_type(self.ftype) def write_tensors(self): block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer"))) @@ -211,6 +216,8 @@ def from_model_architecture(model_architecture): return MiniCPMModel if model_architecture == "BertModel": return BertModel + if model_architecture == "NomicBertModel": + return NomicBertModel return Model def _is_model_safetensors(self) -> bool: @@ -268,6 +275,8 @@ def _get_model_architecture(self) -> gguf.MODEL_ARCH: return gguf.MODEL_ARCH.MINICPM if arch == "BertModel": return gguf.MODEL_ARCH.BERT + if arch == "NomicBertModel": + return gguf.MODEL_ARCH.NOMIC_BERT raise NotImplementedError(f'Architecture "{arch}" not supported!') @@ -1297,21 +1306,21 @@ def write_tensors(self): class Phi2Model(Model): def set_gguf_parameters(self): - block_count = get_key_opts(self.hparams, ["num_hidden_layers", "n_layer"]) + block_count = self.find_hparam(["num_hidden_layers", "n_layer"]) - rot_pct = get_key_opts(self.hparams, ["partial_rotary_factor"]) - n_embd = get_key_opts(self.hparams, ["hidden_size", "n_embd"]) - n_head = get_key_opts(self.hparams, ["num_attention_heads", "n_head"]) + rot_pct = self.find_hparam(["partial_rotary_factor"]) + n_embd = self.find_hparam(["hidden_size", "n_embd"]) + n_head = self.find_hparam(["num_attention_heads", "n_head"]) self.gguf_writer.add_name("Phi2") - self.gguf_writer.add_context_length(get_key_opts(self.hparams, ["n_positions", "max_position_embeddings"])) + self.gguf_writer.add_context_length(self.find_hparam(["n_positions", "max_position_embeddings"])) self.gguf_writer.add_embedding_length(n_embd) self.gguf_writer.add_feed_forward_length(4 * n_embd) self.gguf_writer.add_block_count(block_count) self.gguf_writer.add_head_count(n_head) self.gguf_writer.add_head_count_kv(n_head) - self.gguf_writer.add_layer_norm_eps(get_key_opts(self.hparams, ["layer_norm_epsilon", "layer_norm_eps"])) + self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_epsilon", "layer_norm_eps"])) self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head) self.gguf_writer.add_file_type(self.ftype) self.gguf_writer.add_add_bos_token(False) @@ -1636,20 +1645,12 @@ def write_tensors(self): class BertModel(Model): def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) - self.block_count = self.hparams["num_hidden_layers"] + self.vocab_size = None def set_gguf_parameters(self): - # TODO(cebtenzzre): merge with parent class - self.gguf_writer.add_name(self.dir_model.name) - self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"]) - self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) - self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"]) - self.gguf_writer.add_block_count(self.block_count) - self.gguf_writer.add_head_count(self.hparams["num_attention_heads"]) - self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"]) + super().set_gguf_parameters() self.gguf_writer.add_causal_attention(False) self.gguf_writer.add_pooling_layer(True) - self.gguf_writer.add_file_type(self.ftype) def set_vocab(self): path = self.dir_model @@ -1659,6 +1660,7 @@ def set_vocab(self): vocab = HfVocab(path, added_tokens_path) tokens, scores, toktypes = zip(*vocab.all_tokens()) assert len(tokens) == vocab.vocab_size + self.vocab_size = vocab.vocab_size # we need this to validate the size of the token_type embeddings # though currently we are passing all zeros to the token_type embeddings @@ -1672,7 +1674,7 @@ def phantom(tok, typ): if tok.startswith(b"##"): return tok[2:] return b"\xe2\x96\x81" + tok - tokens = [phantom(t, y) for t, y in zip(tokens, toktypes)] + tokens = tuple(phantom(t, y) for t, y in zip(tokens, toktypes)) # set up bos and eos tokens (cls and sep) self.gguf_writer.add_bos_token_id(vocab.tokenizer.cls_token_id) @@ -1724,6 +1726,43 @@ def write_tensors(self): self.gguf_writer.add_tensor(new_name, data) +class NomicBertModel(BertModel): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + # the HF config claims n_ctx=8192, but it uses RoPE scaling + self.hparams["n_ctx"] = 2048 + + # SwigLU activation + assert self.hparams["activation_function"] == "swiglu" + # this doesn't do anything in the HF version + assert self.hparams["causal"] is False + # no bias tensors + assert self.hparams["qkv_proj_bias"] is False + assert self.hparams["mlp_fc1_bias"] is False + assert self.hparams["mlp_fc2_bias"] is False + # norm at end of layer + assert self.hparams["prenorm"] is False + # standard RoPE + assert self.hparams["rotary_emb_fraction"] == 1.0 + assert self.hparams["rotary_emb_interleaved"] is False + assert self.hparams["rotary_emb_scale_base"] is None + + def set_gguf_parameters(self): + super().set_gguf_parameters() + self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"]) + + def get_tensors(self): + assert self.vocab_size is not None + for name, data in super().get_tensors(): + # Nomic Embed's token embeddings tensor is padded, but llama.cpp wants tensor sizes to match exactly. + if name == 'embeddings.word_embeddings.weight' and data.shape[1] != self.vocab_size: + rounded_vocab_size = (self.vocab_size + 63) // 64 * 64 + assert data.shape == (rounded_vocab_size, self.hparams["n_embd"]) + data = data[:self.vocab_size, :] + yield name, data + + ###### CONVERSION LOGIC ###### diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 644e1589c830d..5fba0171439bb 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -87,27 +87,28 @@ class Tokenizer: class MODEL_ARCH(IntEnum): - LLAMA = auto() - FALCON = auto() - BAICHUAN = auto() - GPT2 = auto() - GPTJ = auto() - GPTNEOX = auto() - MPT = auto() - STARCODER = auto() - PERSIMMON = auto() - REFACT = auto() - BERT = auto() - BLOOM = auto() - STABLELM = auto() - QWEN = auto() - QWEN2 = auto() - PHI2 = auto() - PLAMO = auto() - CODESHELL = auto() - ORION = auto() + LLAMA = auto() + FALCON = auto() + BAICHUAN = auto() + GPT2 = auto() + GPTJ = auto() + GPTNEOX = auto() + MPT = auto() + STARCODER = auto() + PERSIMMON = auto() + REFACT = auto() + BERT = auto() + NOMIC_BERT = auto() + BLOOM = auto() + STABLELM = auto() + QWEN = auto() + QWEN2 = auto() + PHI2 = auto() + PLAMO = auto() + CODESHELL = auto() + ORION = auto() INTERNLM2 = auto() - MINICPM = auto() + MINICPM = auto() class MODEL_TENSOR(IntEnum): @@ -153,6 +154,7 @@ class MODEL_TENSOR(IntEnum): MODEL_ARCH.PERSIMMON: "persimmon", MODEL_ARCH.REFACT: "refact", MODEL_ARCH.BERT: "bert", + MODEL_ARCH.NOMIC_BERT: "nomic-bert", MODEL_ARCH.BLOOM: "bloom", MODEL_ARCH.STABLELM: "stablelm", MODEL_ARCH.QWEN: "qwen", @@ -282,6 +284,20 @@ class MODEL_TENSOR(IntEnum): MODEL_TENSOR.FFN_UP, MODEL_TENSOR.LAYER_OUT_NORM, ], + MODEL_ARCH.NOMIC_BERT: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.TOKEN_EMBD_NORM, + MODEL_TENSOR.TOKEN_TYPES, + MODEL_TENSOR.POS_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_OUT_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.LAYER_OUT_NORM, + ], MODEL_ARCH.MPT: [ 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 c7ba1420e0453..8610037767fb6 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -15,7 +15,7 @@ class TensorNameMap: "word_embeddings", # bloom "model.embed_tokens", # llama-hf "tok_embeddings", # llama-pth - "embeddings.word_embeddings", # bert + "embeddings.word_embeddings", # bert nomic-bert "language_model.embedding.word_embeddings", # persimmon "wte", # gpt2 "transformer.embd.wte", # phi2 @@ -24,13 +24,14 @@ class TensorNameMap: # Token type embeddings MODEL_TENSOR.TOKEN_TYPES: ( - "embeddings.token_type_embeddings", # bert + "embeddings.token_type_embeddings", # bert nomic-bert ), # Normalization of token embeddings MODEL_TENSOR.TOKEN_EMBD_NORM: ( "word_embeddings_layernorm", # bloom "embeddings.LayerNorm", # bert + "emb_ln", # nomic-bert ), # Position embeddings @@ -103,6 +104,7 @@ class TensorNameMap: "model.layers.{bid}.self_attn.query_key_value", # persimmon "h.{bid}.attn.c_attn", # gpt2 "transformer.h.{bid}.mixer.Wqkv", # phi2 + "encoder.layers.{bid}.attn.Wqkv", # nomic-bert ), # Attention query @@ -152,11 +154,13 @@ class TensorNameMap: "transformer.h.{bid}.mixer.out_proj", # phi2 "model.layers.layers.{bid}.self_attn.o_proj", # plamo "model.layers.{bid}.attention.wo", # internlm2 + "encoder.layers.{bid}.attn.out_proj", # nomic-bert ), # Attention output norm MODEL_TENSOR.ATTN_OUT_NORM: ( "encoder.layer.{bid}.attention.output.LayerNorm", # bert + "encoder.layers.{bid}.norm1", # nomic-bert ), # Rotary embeddings @@ -205,6 +209,7 @@ class TensorNameMap: "model.layers.{bid}.mlp.fc1", # phi2 "model.layers.layers.{bid}.mlp.up_proj", # plamo "model.layers.{bid}.feed_forward.w3", # internlm2 + "encoder.layers.{bid}.mlp.fc11", # nomic-bert ), MODEL_TENSOR.FFN_UP_EXP: ( @@ -224,6 +229,7 @@ class TensorNameMap: "transformer.h.{bid}.mlp.w2", # qwen "model.layers.layers.{bid}.mlp.gate_proj", # plamo "model.layers.{bid}.feed_forward.w1", # internlm2 + "encoder.layers.{bid}.mlp.fc12", # nomic-bert ), MODEL_TENSOR.FFN_GATE_EXP: ( @@ -249,6 +255,7 @@ class TensorNameMap: "model.layers.{bid}.mlp.fc2", # phi2 "model.layers.layers.{bid}.mlp.down_proj", # plamo "model.layers.{bid}.feed_forward.w2", # internlm2 + "encoder.layers.{bid}.mlp.fc2", # nomic-bert ), MODEL_TENSOR.FFN_DOWN_EXP: ( @@ -272,6 +279,7 @@ class TensorNameMap: MODEL_TENSOR.LAYER_OUT_NORM: ( "encoder.layer.{bid}.output.LayerNorm", # bert + "encoder.layers.{bid}.norm2", # nomic-bert ) } diff --git a/llama.cpp b/llama.cpp index 8ebbf7628c1e4..14e8821cdf0e6 100644 --- a/llama.cpp +++ b/llama.cpp @@ -197,6 +197,7 @@ enum llm_arch { LLM_ARCH_PERSIMMON, LLM_ARCH_REFACT, LLM_ARCH_BERT, + LLM_ARCH_NOMIC_BERT, LLM_ARCH_BLOOM, LLM_ARCH_STABLELM, LLM_ARCH_QWEN, @@ -211,27 +212,28 @@ enum llm_arch { }; static std::map LLM_ARCH_NAMES = { - { LLM_ARCH_LLAMA, "llama" }, - { LLM_ARCH_FALCON, "falcon" }, - { LLM_ARCH_GPT2, "gpt2" }, - { LLM_ARCH_GPTJ, "gptj" }, - { LLM_ARCH_GPTNEOX, "gptneox" }, - { LLM_ARCH_MPT, "mpt" }, - { LLM_ARCH_BAICHUAN, "baichuan" }, - { LLM_ARCH_STARCODER, "starcoder" }, - { LLM_ARCH_PERSIMMON, "persimmon" }, - { LLM_ARCH_REFACT, "refact" }, - { LLM_ARCH_BERT, "bert" }, - { LLM_ARCH_BLOOM, "bloom" }, - { LLM_ARCH_STABLELM, "stablelm" }, - { LLM_ARCH_QWEN, "qwen" }, - { LLM_ARCH_QWEN2, "qwen2" }, - { LLM_ARCH_PHI2, "phi2" }, - { LLM_ARCH_PLAMO, "plamo" }, - { LLM_ARCH_CODESHELL, "codeshell" }, - { LLM_ARCH_ORION, "orion" }, - { LLM_ARCH_INTERNLM2, "internlm2" }, - { LLM_ARCH_MINICPM, "minicpm" }, + { LLM_ARCH_LLAMA, "llama" }, + { LLM_ARCH_FALCON, "falcon" }, + { LLM_ARCH_GPT2, "gpt2" }, + { LLM_ARCH_GPTJ, "gptj" }, + { LLM_ARCH_GPTNEOX, "gptneox" }, + { LLM_ARCH_MPT, "mpt" }, + { LLM_ARCH_BAICHUAN, "baichuan" }, + { LLM_ARCH_STARCODER, "starcoder" }, + { LLM_ARCH_PERSIMMON, "persimmon" }, + { LLM_ARCH_REFACT, "refact" }, + { LLM_ARCH_BERT, "bert" }, + { LLM_ARCH_NOMIC_BERT, "nomic-bert" }, + { LLM_ARCH_BLOOM, "bloom" }, + { LLM_ARCH_STABLELM, "stablelm" }, + { LLM_ARCH_QWEN, "qwen" }, + { LLM_ARCH_QWEN2, "qwen2" }, + { LLM_ARCH_PHI2, "phi2" }, + { LLM_ARCH_PLAMO, "plamo" }, + { LLM_ARCH_CODESHELL, "codeshell" }, + { LLM_ARCH_ORION, "orion" }, + { LLM_ARCH_INTERNLM2, "internlm2" }, + { LLM_ARCH_MINICPM, "minicpm" }, }; enum llm_kv { @@ -375,6 +377,7 @@ enum llm_tensor { LLM_TENSOR_ATTN_OUT, LLM_TENSOR_ATTN_NORM, LLM_TENSOR_ATTN_NORM_2, + LLM_TENSOR_ATTN_OUT_NORM, LLM_TENSOR_ATTN_ROT_EMBD, LLM_TENSOR_FFN_GATE_INP, LLM_TENSOR_FFN_NORM, @@ -387,6 +390,7 @@ enum llm_tensor { LLM_TENSOR_FFN_UP_EXP, LLM_TENSOR_ATTN_Q_NORM, LLM_TENSOR_ATTN_K_NORM, + LLM_TENSOR_LAYER_OUT_NORM, }; static std::map> LLM_TENSOR_NAMES = { @@ -552,12 +556,27 @@ static std::map> LLM_TENSOR_NAMES = { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" }, { LLM_TENSOR_TOKEN_TYPES, "token_types" }, { LLM_TENSOR_POS_EMBD, "position_embd" }, - { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_output_norm" }, + { LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_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_NORM, "blk.%d.layer_output_norm" }, + { LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_NOMIC_BERT, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" }, + { LLM_TENSOR_TOKEN_TYPES, "token_types" }, + { LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_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" }, }, @@ -1485,6 +1504,7 @@ enum e_model { MODEL_22M, MODEL_33M, MODEL_109M, + MODEL_137M, MODEL_335M, MODEL_0_5B, MODEL_1B, @@ -1620,6 +1640,8 @@ struct llama_layer { struct ggml_tensor * attn_q_norm_b; struct ggml_tensor * attn_k_norm; struct ggml_tensor * attn_k_norm_b; + struct ggml_tensor * attn_out_norm; + struct ggml_tensor * attn_out_norm_b; // attention struct ggml_tensor * wq; @@ -1638,6 +1660,8 @@ struct llama_layer { // normalization struct ggml_tensor * ffn_norm; struct ggml_tensor * ffn_norm_b; + struct ggml_tensor * layer_out_norm; + struct ggml_tensor * layer_out_norm_b; // ff struct ggml_tensor * ffn_gate; // w1 @@ -2855,6 +2879,11 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { static const char * llama_model_type_name(e_model type) { switch (type) { + case MODEL_22M: return "22M"; + case MODEL_33M: return "33M"; + case MODEL_109M: return "109M"; + case MODEL_137M: return "137M"; + case MODEL_0_5B: return "0.5B"; case MODEL_1B: return "1B"; case MODEL_2B: return "2B"; case MODEL_3B: return "3B"; @@ -3073,6 +3102,17 @@ static void llm_load_hparams( model.type = e_model::MODEL_335M; break; // bge-large } } break; + case LLM_ARCH_NOMIC_BERT: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + ml.get_key(LLM_KV_ATTENTION_CAUSAL, hparams.causal_attn); + ml.get_key(LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT, hparams.n_vocab_type); + ml.get_key(LLM_KV_POOLING_LAYER, hparams.pooling_layer); + + if (hparams.n_layer == 12 && hparams.n_embd == 768) { + model.type = e_model::MODEL_137M; + } + } break; case LLM_ARCH_BLOOM: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); @@ -3875,10 +3915,14 @@ static bool llm_load_tensors( } } break; case LLM_ARCH_BERT: + case LLM_ARCH_NOMIC_BERT: { - model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); - model.type_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_vocab_type}); - model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, hparams.n_ctx_train}); + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + model.type_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_vocab_type}); + if (model.arch == LLM_ARCH_BERT) { + model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, hparams.n_ctx_train}); + } + model.tok_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}); model.tok_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}); @@ -3888,29 +3932,38 @@ static bool llm_load_tensors( 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}); + if (model.arch == LLM_ARCH_BERT) { + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); + layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), {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.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); + layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}); - layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); - layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}); + layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); + layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}); + } else { + layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}); + } - layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); - layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); - layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); - layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}); + layer.attn_out_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}); + layer.attn_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i), {n_embd}); - layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); - layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "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.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}); + if (model.arch == LLM_ARCH_BERT) { + layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}); + layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}); - layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}); - layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}); + layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}); + } else { + layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); + } + + layer.layer_out_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}); + layer.layer_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_OUT_NORM, "bias", i), {n_embd}); } } break; case LLM_ARCH_BLOOM: @@ -5773,6 +5826,7 @@ struct llm_build_context { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); struct ggml_tensor * cur; @@ -5789,7 +5843,9 @@ struct llm_build_context { // token types are hardcoded to zero ("Sentence A") struct ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0); inpL = ggml_add(ctx0, inpL, type_row0); - inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL); + if (model.arch == LLM_ARCH_BERT) { + inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL); + } cb(inpL, "inp_embd", -1); // embed layer norm @@ -5805,7 +5861,7 @@ struct llm_build_context { struct ggml_tensor * cur = inpL; // self-attention - { + if (model.arch == LLM_ARCH_BERT) { struct ggml_tensor * Qcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, cur), model.layers[il].bq); cb(Qcur, "Qcur", il); @@ -5818,6 +5874,37 @@ struct llm_build_context { // seems like we just need to do this for Q? Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + model.layers[il].wo, model.layers[il].bo, + 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); + } else { + // compute Q and K and RoPE them + cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + 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, model.layers[il].bo, Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); @@ -5828,25 +5915,34 @@ struct llm_build_context { cur = ggml_add(ctx0, cur, inpL); // attention layer norm - cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].attn_norm, model.layers[il].attn_norm_b, LLM_NORM, cb, il); + cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].attn_out_norm, model.layers[il].attn_out_norm_b, LLM_NORM, cb, il); struct ggml_tensor * ffn_inp = cur; cb(ffn_inp, "ffn_inp", il); // feed-forward network - cur = llm_build_ffn(ctx0, cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, - NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); + if (model.arch == LLM_ARCH_BERT) { + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, + NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); + } else { + 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); // attentions bypass the intermediate layer cur = ggml_add(ctx0, cur, ffn_inp); // output layer norm - cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, LLM_NORM, cb, il); + cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].layer_out_norm, model.layers[il].layer_out_norm_b, LLM_NORM, cb, il); // input for next layer inpL = cur; @@ -7289,6 +7385,7 @@ static struct ggml_cgraph * llama_build_graph( result = llm.build_refact(); } break; case LLM_ARCH_BERT: + case LLM_ARCH_NOMIC_BERT: { result = llm.build_bert(); } break;