From 65de3281eaa40e54266b76cc15e5106d16237c93 Mon Sep 17 00:00:00 2001 From: root Date: Sat, 30 Sep 2023 00:31:41 +0200 Subject: [PATCH] GGUF : GPT2 Support --- convert-gpt2-hf-to-gguf.py | 245 +++++++++++++++++++++++++++++++++++++ gguf-py/gguf/gguf.py | 11 +- llama.cpp | 66 ++++++++++ 3 files changed, 321 insertions(+), 1 deletion(-) create mode 100755 convert-gpt2-hf-to-gguf.py diff --git a/convert-gpt2-hf-to-gguf.py b/convert-gpt2-hf-to-gguf.py new file mode 100755 index 000000000..475489240 --- /dev/null +++ b/convert-gpt2-hf-to-gguf.py @@ -0,0 +1,245 @@ +#!/usr/bin/env python3 +# HF gpt2 --> gguf conversion + +from __future__ import annotations + +import argparse +import json +import os +import sys +from pathlib import Path +from typing import Any, Dict + +import numpy as np +import torch +from transformers import AutoTokenizer # type: ignore[import] +import logging + +logging.basicConfig(level=logging.INFO, format='%(asctime)s %(levelname)s %(message)s') +logger = logging.getLogger(__name__) + +if 'NO_LOCAL_GGUF' not in os.environ: + sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) +import gguf + + +def bytes_to_unicode() -> Dict[int, str]: + # ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py + """ + Returns list of utf-8 byte and a corresponding list of unicode strings. + The reversible bpe codes work on unicode strings. + This means you need a large # of unicode characters in your vocab if you want to avoid UNKs. + When you're at something like a 10B token dataset you end up needing around 5K for decent coverage. + This is a significant percentage of your normal, say, 32K bpe vocab. + To avoid that, we want lookup tables between utf-8 bytes and unicode strings. + And avoids mapping to whitespace/control characters the bpe code barfs on. + """ + bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1)) + cs = bs[:] + n = 0 + for b in range(2**8): + if b not in bs: + bs.append(b) + cs.append(2**8+n) + n += 1 + return dict(zip(bs, (chr(n) for n in cs))) + + +def count_model_parts(dir_model: Path) -> int: + num_parts = 0 + for filename in os.listdir(dir_model): + if filename.startswith("pytorch_model-"): + num_parts += 1 + + if num_parts > 0: + logger.info("gguf: found " + str(num_parts) + " model parts") + return num_parts + + +def parse_args() -> argparse.Namespace: + parser = argparse.ArgumentParser(description="Convert a gpt2 model to a GGML compatible file") + parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") + parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") + parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.bin)") + parser.add_argument("ftype", type=int, help="output format - use 0 for float32, 1 for float16", choices=[0, 1], default = 1) + return parser.parse_args() + +args = parse_args() + +dir_model = args.model +ftype = args.ftype +arch_name = 'GPT2LMHeadModel' + +if not dir_model.is_dir(): + raise FileNotFoundError(f'{args.model} is not a directory') + +# possible tensor data types +# ftype == 0 -> float32 +# ftype == 1 -> float16 + +# map from ftype to string +ftype_str = ["f32", "f16"] + +if args.outfile is not None: + fname_out = args.outfile +else: + # output in the same directory as the model by default + fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf' + +logger.info(f"gguf: loading model {dir_model.name}") + +with open(dir_model / "config.json", "r", encoding="utf-8") as f: + hparams = json.load(f) + +if hparams["architectures"][0] != arch_name: + raise ValueError("Only GPT2LMHeadModel is supported") + +# get number of model parts +num_parts = count_model_parts(dir_model) + +ARCH = gguf.MODEL_ARCH.GPT2 +gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH]) + +logger.info("gguf: get model metadata") + +block_count = hparams["n_layer"] + +gguf_writer.add_name("gpt2") +gguf_writer.add_context_length(hparams["n_positions"]) +gguf_writer.add_embedding_length(hparams["n_embd"]) +gguf_writer.add_feed_forward_length(4 * hparams["n_embd"]) +gguf_writer.add_block_count(block_count) +gguf_writer.add_head_count(hparams["n_head"]) +gguf_writer.add_head_count_kv(1) +gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"]) +gguf_writer.add_file_type(ftype) + +# TOKENIZATION + +print("gguf: get tokenizer metadata") + +tokens: list[bytearray] = [] + +tokenizer_json_file = dir_model / 'tokenizer.json' +if not tokenizer_json_file.is_file(): + print(f'Error: Missing {tokenizer_json_file}', file = sys.stderr) + sys.exit(1) + +# gpt2 tokenizer +gguf_writer.add_tokenizer_model("gpt2") + +with open(tokenizer_json_file, "r", encoding="utf-8") as f: + tokenizer_json = json.load(f) + +print("gguf: get gpt2 tokenizer vocab") + +vocab_size = len(tokenizer_json["model"]["vocab"]) + +# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py +tokenizer = AutoTokenizer.from_pretrained(dir_model) + +reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()} +byte_encoder = bytes_to_unicode() +byte_decoder = {v: k for k, v in byte_encoder.items()} + +for i in range(vocab_size): + if i in reverse_vocab: + try: + text = bytearray([byte_decoder[c] for c in reverse_vocab[i]]) + except KeyError: + text = bytearray() + for c in reverse_vocab[i]: + if ord(c) < 256: # single byte character + text.append(byte_decoder[ord(c)]) + else: # multibyte special token character + text.extend(c.encode('utf-8')) + else: + print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.") + pad_token = f"[PAD{i}]".encode("utf8") + text = bytearray(pad_token) + + tokens.append(text) + +gguf_writer.add_token_list(tokens) + +special_vocab = gguf.SpecialVocab(dir_model, load_merges = True) +special_vocab.add_to_gguf(gguf_writer) + +# TENSORS + +tensor_map = gguf.get_tensor_name_map(ARCH,block_count) + +# params for qkv transform +n_head = hparams["n_head"] +n_head_kv = hparams["n_head_kv"] if "n_head_kv" in hparams else 1 + +head_dim = hparams["n_embd"] // n_head + +# tensor info +logger.info("gguf: get tensor metadata") + +if num_parts == 0: + part_names = iter(("pytorch_model.bin",)) +else: + part_names = ( + f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1) + ) + +for part_name in part_names: + if args.vocab_only: + break + logger.info("gguf: loading model part '" + part_name + "'") + model_part = torch.load(dir_model / part_name, map_location="cpu") + + for name in model_part.keys(): + data = model_part[name] + + if name.endswith("attn.masked_bias") or name.endswith(".attn.bias"): + logger.info(f"Skipping variable: {name}") + continue + + old_dtype = data.dtype + + # convert any unsupported data types to float32 + if data.dtype != torch.float16 and data.dtype != torch.float32: + data = data.to(torch.float32) + + data = data.squeeze().numpy() + + # map tensor names + new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias")) + if new_name is None: + raise ValueError(f"Cannot map tensor name: {name}") + + n_dims = len(data.shape) + data_dtype = data.dtype + + # if f32 desired, convert any float16 to float32 + if 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 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 ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: + data = data.astype(np.float16) + + logger.info(f"{name} => {new_name}, shape = {data.shape}, {old_dtype} --> {data.dtype}") + + gguf_writer.add_tensor(new_name, data) + + del model_part + +logger.info("gguf: write header") +gguf_writer.write_header_to_file() +logger.info("gguf: write metadata") +gguf_writer.write_kv_data_to_file() +if not args.vocab_only: + logger.info("gguf: write tensors") + gguf_writer.write_tensors_to_file() + +gguf_writer.close() + +logger.info(f"gguf: model successfully exported to '{fname_out}'\n") diff --git a/gguf-py/gguf/gguf.py b/gguf-py/gguf/gguf.py index 598cf8e59..0fc688616 100644 --- a/gguf-py/gguf/gguf.py +++ b/gguf-py/gguf/gguf.py @@ -186,7 +186,16 @@ MODEL_TENSOR_NAMES: dict[MODEL_ARCH, dict[MODEL_TENSOR, str]] = { MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up", }, MODEL_ARCH.GPT2: { - # TODO + MODEL_TENSOR.TOKEN_EMBD: "token_embd", + MODEL_TENSOR.POS_EMBD: "position_embd", + MODEL_TENSOR.OUTPUT_NORM: "output_norm", + MODEL_TENSOR.OUTPUT: "output", + MODEL_TENSOR.ATTN_NORM: "blk.{bid}.attn_norm", + MODEL_TENSOR.ATTN_QKV: "blk.{bid}.attn_qkv", + MODEL_TENSOR.ATTN_OUT: "blk.{bid}.attn_output", + MODEL_TENSOR.FFN_NORM: "blk.{bid}.ffn_norm", + MODEL_TENSOR.FFN_DOWN: "blk.{bid}.ffn_down", + MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up", }, # TODO } diff --git a/llama.cpp b/llama.cpp index bff17135b..c65539e60 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2436,6 +2436,72 @@ static void llm_load_tensors( } } } break; + case LLM_ARCH_GPT2: + { + model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + { + ggml_backend backend_norm; + ggml_backend backend_output; + + if (n_gpu_layers > int(n_layer)) { + // norm is not performance relevant on its own but keeping it in VRAM reduces data copying + // on Windows however this is detrimental unless everything is on the GPU +#ifndef _WIN32 + backend_norm = LLAMA_BACKEND_OFFLOAD; +#else + backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; +#endif // _WIN32 + + backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; + } else { + backend_norm = GGML_BACKEND_CPU; + backend_output = GGML_BACKEND_CPU; + } + + model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); + model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); + + if (backend_norm == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(model.output_norm); + } + if (backend_output == GGML_BACKEND_GPU_SPLIT) { + vram_weights += ggml_nbytes(model.output); + } + } + + const uint32_t n_ff = hparams.n_ff; + + const int i_gpu_start = n_layer - n_gpu_layers; + + model.layers.resize(n_layer); + + for (uint32_t i = 0; i < n_layer; ++i) { + const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT + const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); + + layer.wq = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, backend_split); + layer.wk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, backend_split); + layer.wv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, backend_split); + layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); + + layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); + + layer.w1 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); + layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + + if (backend == GGML_BACKEND_GPU) { + vram_weights += + ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) + + ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + + ggml_nbytes(layer.w1) + ggml_nbytes(layer.w2) + ggml_nbytes(layer.w3); + } + } + } break; default: throw std::runtime_error("unknown architecture"); }