From 344f6e373bff260ddfc0bdb4a846587054bec14c Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?M=2E=20Yusuf=20Sar=C4=B1g=C3=B6z?= Date: Thu, 24 Aug 2023 09:09:52 +0300 Subject: [PATCH] gguf: prepare as Pip package --- gguf-py/LICENSE | 21 ++ gguf-py/README.md | 34 ++ gguf-py/gguf/__init__.py | 3 + gguf-py/gguf/gguf.py | 723 +++++++++++++++++++++++++++++++++++++ gguf-py/pyproject.toml | 28 ++ gguf-py/tests/test_gguf.py | 7 + 6 files changed, 816 insertions(+) create mode 100644 gguf-py/LICENSE create mode 100644 gguf-py/README.md create mode 100644 gguf-py/gguf/__init__.py create mode 100644 gguf-py/gguf/gguf.py create mode 100644 gguf-py/pyproject.toml create mode 100644 gguf-py/tests/test_gguf.py diff --git a/gguf-py/LICENSE b/gguf-py/LICENSE new file mode 100644 index 000000000..4d1c4c326 --- /dev/null +++ b/gguf-py/LICENSE @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2023 Georgi Gerganov + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. \ No newline at end of file diff --git a/gguf-py/README.md b/gguf-py/README.md new file mode 100644 index 000000000..9b6870fa3 --- /dev/null +++ b/gguf-py/README.md @@ -0,0 +1,34 @@ +## gguf + +This is a Python package for writing binary files in the [GGUF](https://github.com/ggerganov/ggml/pull/302) +(GGML Universal File) format. + +See [convert-llama-hf-to-gguf.py](https://github.com/ggerganov/llama.cpp/blob/master/convert-llama-hf-to-gguf.py) +as an example for its usage. + +## Install +```sh +pip install gguf +``` + +## Development +Maintainers who participate in development of this package are advised to install it in editable mode: + + +```sh +cd /path/to/llama.cpp/gguf + +pip install --editable . +``` + +**Note**: This may require to upgrade your Pip installation, with a message saying that editable installation currently requires `setup.py`. +In this case, upgrade Pip to the latest: + +```sh +pip install --upgrade pip +``` + +## TODO + +- [ ] Add tests +- [ ] Include conversion scripts as command line entry points in this package. diff --git a/gguf-py/gguf/__init__.py b/gguf-py/gguf/__init__.py new file mode 100644 index 000000000..cfbad8ed0 --- /dev/null +++ b/gguf-py/gguf/__init__.py @@ -0,0 +1,3 @@ +from .gguf import GGUFWriter + +__version__ = '0.1.0' diff --git a/gguf-py/gguf/gguf.py b/gguf-py/gguf/gguf.py new file mode 100644 index 000000000..5c37f0f0b --- /dev/null +++ b/gguf-py/gguf/gguf.py @@ -0,0 +1,723 @@ +#!/usr/bin/env python3 +import shutil +import sys +import struct +import tempfile +import numpy as np + +from enum import IntEnum, auto +from typing import Any, IO, List, Optional + +# +# constants +# + +GGUF_MAGIC = 0x46554747 +GGUF_VERSION = 1 +GGUF_DEFAULT_ALIGNMENT = 32 + +# general +KEY_GENERAL_ARCHITECTURE = "general.architecture" +KEY_GENERAL_QUANTIZATION_VERSION = "general.quantization_version" +KEY_GENERAL_ALIGNMENT = "general.alignment" +KEY_GENERAL_NAME = "general.name" +KEY_GENERAL_AUTHOR = "general.author" +KEY_GENERAL_URL = "general.url" +KEY_GENERAL_DESCRIPTION = "general.description" +KEY_GENERAL_LICENSE = "general.license" +KEY_GENERAL_SOURCE_URL = "general.source.url" +KEY_GENERAL_SOURCE_HF_REPO = "general.source.hugginface.repository" +KEY_GENERAL_FILE_TYPE = "general.file_type" + +# LLM +KEY_CONTEXT_LENGTH = "{arch}.context_length" +KEY_EMBEDDING_LENGTH = "{arch}.embedding_length" +KEY_BLOCK_COUNT = "{arch}.block_count" +KEY_FEED_FORWARD_LENGTH = "{arch}.feed_forward_length" +KEY_USE_PARALLEL_RESIDUAL = "{arch}.use_parallel_residual" +KEY_TENSOR_DATA_LAYOUT = "{arch}.tensor_data_layout" + +# attention +KEY_ATTENTION_HEAD_COUNT = "{arch}.attention.head_count" +KEY_ATTENTION_HEAD_COUNT_KV = "{arch}.attention.head_count_kv" +KEY_ATTENTION_MAX_ALIBI_BIAS = "{arch}.attention.max_alibi_bias" +KEY_ATTENTION_CLAMP_KQV = "{arch}.attention.clamp_kqv" +KEY_ATTENTION_LAYERNORM_EPS = "{arch}.attention.layer_norm_epsilon" +KEY_ATTENTION_LAYERNORM_RMS_EPS = "{arch}.attention.layer_norm_rms_epsilon" + +# RoPE +KEY_ROPE_DIMENSION_COUNT = "{arch}.rope.dimension_count" +KEY_ROPE_SCALE_LINEAR = "{arch}.rope.scale_linear" + +# tokenization +KEY_TOKENIZER_MODEL = "tokenizer.ggml.model" +KEY_TOKENIZER_LIST = "tokenizer.ggml.tokens" +KEY_TOKENIZER_TOKEN_TYPE = "tokenizer.ggml.token_type" +KEY_TOKENIZER_SCORES = "tokenizer.ggml.scores" +KEY_TOKENIZER_MERGES = "tokenizer.ggml.merges" +KEY_TOKENIZER_BOS_ID = "tokenizer.ggml.bos_token_id" +KEY_TOKENIZER_EOS_ID = "tokenizer.ggml.eos_token_id" +KEY_TOKENIZER_UNK_ID = "tokenizer.ggml.unknown_token_id" +KEY_TOKENIZER_SEP_ID = "tokenizer.ggml.seperator_token_id" +KEY_TOKENIZER_PAD_ID = "tokenizer.ggml.padding_token_id" +KEY_TOKENIZER_HF_JSON = "tokenizer.huggingface.json" +KEY_TOKENIZER_RWKV = "tokenizer.rwkv.world" + + +# +# recommended mapping of model tensor names for storage in gguf +# + + +class MODEL_ARCH(IntEnum): + LLAMA = auto() + FALCON = auto() + GPT2 = auto() + GPTJ = auto() + GPTNEOX = auto() + MPT = auto() + + +class MODEL_TENSOR(IntEnum): + TOKEN_EMBD = auto() + POS_EMBD = auto() + OUTPUT = auto() + OUTPUT_NORM = auto() + ROPE_FREQS = auto() + ATTN_Q = auto() + ATTN_K = auto() + ATTN_V = auto() + ATTN_QKV = auto() + ATTN_OUT = auto() + ATTN_NORM = auto() + ATTN_NORM_2 = auto() + ATTN_ROT_EMBD = auto() + FFN_GATE = auto() + FFN_DOWN = auto() + FFN_UP = auto() + FFN_NORM = auto() + + +MODEL_ARCH_NAMES = { + MODEL_ARCH.LLAMA: "llama", + MODEL_ARCH.FALCON: "falcon", + MODEL_ARCH.GPT2: "gpt2", + MODEL_ARCH.GPTJ: "gptj", + MODEL_ARCH.GPTNEOX: "gptneox", + MODEL_ARCH.MPT: "mpt", +} + +MODEL_TENSOR_NAMES = { + MODEL_ARCH.LLAMA: { + MODEL_TENSOR.TOKEN_EMBD: "token_embd", + MODEL_TENSOR.OUTPUT_NORM: "output_norm", + MODEL_TENSOR.OUTPUT: "output", + MODEL_TENSOR.ROPE_FREQS: "rope_freqs", + MODEL_TENSOR.ATTN_NORM: "blk.{bid}.attn_norm", + MODEL_TENSOR.ATTN_Q: "blk.{bid}.attn_q", + MODEL_TENSOR.ATTN_K: "blk.{bid}.attn_k", + MODEL_TENSOR.ATTN_V: "blk.{bid}.attn_v", + MODEL_TENSOR.ATTN_OUT: "blk.{bid}.attn_output", + MODEL_TENSOR.ATTN_ROT_EMBD: "blk.{bid}.attn_rot_embd", + MODEL_TENSOR.FFN_NORM: "blk.{bid}.ffn_norm", + MODEL_TENSOR.FFN_GATE: "blk.{bid}.ffn_gate", + MODEL_TENSOR.FFN_DOWN: "blk.{bid}.ffn_down", + MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up", + }, + MODEL_ARCH.GPTNEOX: { + MODEL_TENSOR.TOKEN_EMBD: "token_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", + }, + MODEL_ARCH.FALCON: { + MODEL_TENSOR.TOKEN_EMBD: "token_embd", + MODEL_TENSOR.OUTPUT_NORM: "output_norm", + MODEL_TENSOR.OUTPUT: "output", + MODEL_TENSOR.ATTN_NORM: "blk.{bid}.attn_norm", + MODEL_TENSOR.ATTN_NORM_2: "blk.{bid}.attn_norm_2", + MODEL_TENSOR.ATTN_QKV: "blk.{bid}.attn_qkv", + MODEL_TENSOR.ATTN_OUT: "blk.{bid}.attn_output", + MODEL_TENSOR.FFN_DOWN: "blk.{bid}.ffn_down", + MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up", + }, + MODEL_ARCH.GPT2: { + # TODO + }, + # TODO +} + +# tensors that will not be serialized +MODEL_TENSOR_SKIP = { + MODEL_ARCH.LLAMA: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], +} + + +# TODO: the following helper functions should be removed +# instead, get_tensor_name_map should return tuples of (name, MODEL_TENSOR) +# however, my Python is very bad, and I couldn't figure out how to do this, hence these functions +# REMOVE +def should_skip_tensor_TMP(arch: MODEL_ARCH, n_blocks: int, name: str) -> bool: + for skip in MODEL_TENSOR_SKIP.get(arch, []): + for i in range(n_blocks): + if name == MODEL_TENSOR_NAMES[arch][skip].format(bid=i): + return True + + return False + + +def get_tensor_name_map(arch: MODEL_ARCH, n_blocks: int) -> dict: + tensor_map = {} + + # Token embeddings + mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.TOKEN_EMBD, None) + + tensor_map["gpt_neox.embed_in"] = mapped_to # gptneox + tensor_map["transformer.wte"] = mapped_to # gpt2 mpt + tensor_map["transformer.word_embeddings"] = mapped_to # falcon + tensor_map["model.embed_tokens"] = mapped_to # llama-hf + tensor_map["tok_embeddings"] = mapped_to # llama-pth + + # Position embeddings + mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.POS_EMBD, None) + + tensor_map["transformer.wpe"] = mapped_to # gpt2 + + # Output + mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.OUTPUT, None) + + tensor_map["embed_out"] = mapped_to # gptneox + tensor_map["lm_head"] = mapped_to # gpt2 mpt falcon llama-hf + tensor_map["output"] = mapped_to # llama-pth + + # Output norm + mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.OUTPUT_NORM, None) + + tensor_map["gpt_neox.final_layer_norm"] = mapped_to # gptneox + tensor_map["transformer.ln_f"] = mapped_to # gpt2 falcon + tensor_map["transformer.norm_f"] = mapped_to # mpt + tensor_map["model.norm"] = mapped_to # llama-hf + tensor_map["norm"] = mapped_to # llama-pth + + # Rope frequencies + mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ROPE_FREQS, None) + + tensor_map["rope.freqs"] = mapped_to # llama-pth + + # Attention and feed-forward blocks + for i in range(0, n_blocks): + # Attention norm + # TODO: is there are simpler way to write these 2 lines in Python? + mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_NORM, None) + mapped_to = mapped_to.format(bid=i) if mapped_to else None + + tensor_map["gpt_neox.layers."+str(i)+".input_layernorm"] = mapped_to # gptneox + tensor_map["transformer.h."+str(i)+".ln_1"] = mapped_to # gpt2 + tensor_map["transformer.blocks."+str(i)+".norm_1"] = mapped_to # mpt + tensor_map["transformer.h."+str(i)+".input_layernorm"] = mapped_to # falcon7b + tensor_map["transformer.h."+str(i)+".ln_mlp"] = mapped_to # falcon40b + tensor_map["model.layers."+str(i)+".input_layernorm"] = mapped_to # llama-hf + tensor_map["layers."+str(i)+".attention_norm"] = mapped_to # llama-pth + + # Attention norm 2 + mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_NORM_2, None) + mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None + + tensor_map["transformer.h."+str(i)+".ln_attn"] = mapped_to # falcon40b + + # Attention query-key-value + mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_QKV, None) + mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None + + tensor_map["gpt_neox.layers."+str(i)+".attention.query_key_value"] = mapped_to # gptneox + tensor_map["transformer.h."+str(i)+".attn.c_attn"] = mapped_to # gpt2 + tensor_map["transformer.blocks."+str(i)+".attn.Wqkv"] = mapped_to # mpt + tensor_map["transformer.h."+str(i)+".self_attention.query_key_value"] = mapped_to # falcon + + # Attention query + mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_Q, None) + mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None + + tensor_map["model.layers."+str(i)+".self_attn.q_proj"] = mapped_to # llama-hf + tensor_map["layers."+str(i)+".attention.wq"] = mapped_to # llama-pth + + # Attention key + mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_K, None) + mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None + + tensor_map["model.layers."+str(i)+".self_attn.k_proj"] = mapped_to # llama-hf + tensor_map["layers."+str(i)+".attention.wk"] = mapped_to # llama-pth + + # Attention value + mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_V, None) + mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None + + tensor_map["model.layers."+str(i)+".self_attn.v_proj"] = mapped_to # llama-hf + tensor_map["layers."+str(i)+".attention.wv"] = mapped_to # llama-pth + + # Attention output + mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_OUT, None) + mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None + + tensor_map["gpt_neox.layers."+str(i)+".attention.dense"] = mapped_to # gptneox + tensor_map["transformer.h."+str(i)+".attn.c_proj"] = mapped_to # gpt2 + tensor_map["transformer.blocks."+str(i)+".attn.out_proj"] = mapped_to # mpt + tensor_map["transformer.h."+str(i)+".self_attention.dense"] = mapped_to # falcon + tensor_map["model.layers."+str(i)+".self_attn.o_proj"] = mapped_to # llama-hf + tensor_map["layers."+str(i)+".attention.wo"] = mapped_to # llama-pth + + # Rotary embeddings + mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_ROT_EMBD, None) + mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None + + tensor_map["model.layers."+str(i)+".self_attn.rotary_emb.inv_freq"] = mapped_to # llama-hf + tensor_map["layers."+str(i)+".attention.inner_attention.rope.freqs"] = mapped_to # llama-pth + + # Feed-forward norm + mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.FFN_NORM, None) + mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None + + tensor_map["gpt_neox.layers."+str(i)+".post_attention_layernorm"] = mapped_to # gptneox + tensor_map["transformer.h."+str(i)+".ln_2"] = mapped_to # gpt2 + tensor_map["transformer.blocks."+str(i)+".norm_2"] = mapped_to # mpt + tensor_map["model.layers."+str(i)+".post_attention_layernorm"] = mapped_to # llama-hf + tensor_map["layers."+str(i)+".ffn_norm"] = mapped_to # llama-pth + + # Feed-forward up + mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.FFN_UP, None) + mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None + + tensor_map["gpt_neox.layers."+str(i)+".mlp.dense_h_to_4h"] = mapped_to # gptneox + tensor_map["transformer.h."+str(i)+".mlp.c_fc"] = mapped_to # gpt2 + tensor_map["transformer.blocks."+str(i)+".ffn.up_proj"] = mapped_to # mpt + tensor_map["transformer.h."+str(i)+".mlp.dense_h_to_4h"] = mapped_to # falcon + tensor_map["model.layers."+str(i)+".mlp.up_proj"] = mapped_to # llama-hf + tensor_map["layers."+str(i)+".feed_forward.w3"] = mapped_to # llama-pth + + # Feed-forward gate + mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.FFN_GATE, None) + mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None + + tensor_map["model.layers."+str(i)+".mlp.gate_proj"] = mapped_to # llama-hf + tensor_map["layers."+str(i)+".feed_forward.w1"] = mapped_to # llama-pth + + # Feed-forward down + mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.FFN_DOWN, None) + mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None + + tensor_map["gpt_neox.layers."+str(i)+".mlp.dense_4h_to_h"] = mapped_to # gptneox + tensor_map["transformer.h."+str(i)+".mlp.c_proj"] = mapped_to # gpt2 + tensor_map["transformer.blocks."+str(i)+".ffn.down_proj"] = mapped_to # mpt + tensor_map["transformer.h."+str(i)+".mlp.dense_4h_to_h"] = mapped_to # falcon + tensor_map["model.layers."+str(i)+".mlp.down_proj"] = mapped_to # llama-hf + tensor_map["layers."+str(i)+".feed_forward.w2"] = mapped_to # llama-pth + + return tensor_map + + +class TokenType(IntEnum): + NORMAL = 1 + UNKNOWN = 2 + CONTROL = 3 + USER_DEFINED = 4 + UNUSED = 5 + BYTE = 6 + +# +# implementation +# + + +class GGMLQuantizationType(IntEnum): + F32 = 0 + F16 = 1 + Q4_0 = 2 + Q4_1 = 3 + Q5_0 = 6 + Q5_1 = 7 + Q8_0 = 8 + Q8_1 = 9 + Q2_K = 10 + Q3_K = 11 + Q4_K = 12 + Q5_K = 13 + Q6_K = 14 + Q8_K = 15 + + +class GGUFValueType(IntEnum): + UINT8 = 0 + INT8 = 1 + UINT16 = 2 + INT16 = 3 + UINT32 = 4 + INT32 = 5 + FLOAT32 = 6 + BOOL = 7 + STRING = 8 + ARRAY = 9 + + @staticmethod + def get_type(val): + if isinstance(val, str) or isinstance(val, bytes) or isinstance(val, bytearray): + return GGUFValueType.STRING + elif isinstance(val, list): + return GGUFValueType.ARRAY + elif isinstance(val, float): + return GGUFValueType.FLOAT32 + elif isinstance(val, bool): + return GGUFValueType.BOOL + elif isinstance(val, int): + return GGUFValueType.INT32 + else: + print("Unknown type: "+str(type(val))) + sys.exit() + + +class GGUFWriter: + def __init__(self, path: str, arch: str, use_temp_file = True): + self.fout = open(path, "wb") + self.arch = arch + self.offset_tensor = 0 + self.data_alignment = GGUF_DEFAULT_ALIGNMENT + self.kv_data = b"" + self.kv_data_count = 0 + self.ti_data = b"" + self.ti_data_count = 0 + self.add_architecture() + self.use_temp_file = use_temp_file + self.tensors = [] + + def write_header_to_file(self): + self.fout.write(struct.pack(" int: + return ((x + n - 1) // n) * n + + def add_tensor_info(self, name: str, tensor_shape: np.ndarray, tensor_dtype: np.dtype, tensor_nbytes: int, raw_dtype: Optional[GGMLQuantizationType] = None): + assert raw_dtype is not None or tensor_dtype in (np.float32, np.float16), "Only F32 and F16 tensors are supported for now" + + encoded_name = name.encode("utf8") + self.ti_data += struct.pack(""] +packages = [ + {include = "gguf"}, +] +readme = "README.md" +homepage = "https://ggml.ai" +repository = "https://github.com/ggerganov/llama.cpp" +keywords = ["ggml", "gguf", "llama.cpp"] +classifiers = [ + "Programming Language :: Python :: 3", + "License :: OSI Approved :: MIT License", + "Operating System :: OS Independent", +] + +[tool.poetry.dependencies] +python = ">=3.8" +numpy = ">=1.17" + +[tool.poetry.dev-dependencies] +pytest = "^5.2" + +[build-system] +requires = ["poetry-core>=1.0.0"] +build-backend = "poetry.core.masonry.api" diff --git a/gguf-py/tests/test_gguf.py b/gguf-py/tests/test_gguf.py new file mode 100644 index 000000000..2296ab599 --- /dev/null +++ b/gguf-py/tests/test_gguf.py @@ -0,0 +1,7 @@ +from gguf import __version__ + +# TODO: add tests + + +def test_version(): + assert __version__ == '0.1.0'