ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-30 19:31:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge mainline llama.cpp (#3)

Browse Source 
* Merging mainline - WIP

* Merging mainline - WIP

AVX2 and CUDA appear to work.
CUDA performance seems slightly (~1-2%) lower as it is so often
the case with llama.cpp/ggml after some "improvements" have been made.

* Merging mainline - fix Metal

* Remove check

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2024-07-27 07:55:01 +02:00
committed by GitHub
parent afd9fd274e
commit 0ceeb11721
612 changed files with 50817 additions and 165936 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
gguf-py/README.md
View File

@@ -3,7 +3,7 @@
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-hf-to-gguf.py)
See [convert_hf_to_gguf.py](https://github.com/ggerganov/llama.cpp/blob/master/convert_hf_to_gguf.py)
as an example for its usage.
## Installation
@@ -15,13 +15,13 @@ pip install gguf
[examples/writer.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/examples/writer.py) — Generates `example.gguf` in the current directory to demonstrate generating a GGUF file. Note that this file cannot be used as a model.
[scripts/gguf-dump.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf-dump.py) — Dumps a GGUF file's metadata to the console.
[scripts/gguf_dump.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf_dump.py) — Dumps a GGUF file's metadata to the console.
[scripts/gguf-set-metadata.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf-set-metadata.py) — Allows changing simple metadata values in a GGUF file by key.
[scripts/gguf_set_metadata.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf_set_metadata.py) — Allows changing simple metadata values in a GGUF file by key.
[scripts/gguf-convert-endian.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf-convert-endian.py) — Allows converting the endianness of GGUF files.
[scripts/gguf_convert_endian.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf_convert_endian.py) — Allows converting the endianness of GGUF files.
[scripts/gguf-new-metadata.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf-new-metadata.py) — Copies a GGUF file with added/modified/removed metadata values.
[scripts/gguf_new_metadata.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf_new_metadata.py) — Copies a GGUF file with added/modified/removed metadata values.
## Development
Maintainers who participate in development of this package are advised to install it in editable mode:
@@ -78,6 +78,13 @@ python -m build
python -m twine upload dist/*
```
## Run Unit Tests
From root of this repository you can run this command to run all the unit tests
```bash
python -m unittest discover ./gguf-py -v
```
## TODO
- [ ] Add tests
- [ ] Include conversion scripts as command line entry points in this package.

2
gguf-py/gguf/__init__.py
View File

@@ -5,3 +5,5 @@ from .gguf_writer import *
from .quants import *
from .tensor_mapping import *
from .vocab import *
from .utility import *
from .metadata import *

569
gguf-py/gguf/constants.py
View File

@@ -19,18 +19,60 @@ GGML_QUANT_VERSION = 2 # GGML_QNT_VERSION from ggml.h
class Keys:
class General:
ARCHITECTURE = "general.architecture"
QUANTIZATION_VERSION = "general.quantization_version"
ALIGNMENT = "general.alignment"
NAME = "general.name"
AUTHOR = "general.author"
VERSION = "general.version"
URL = "general.url"
DESCRIPTION = "general.description"
LICENSE = "general.license"
SOURCE_URL = "general.source.url"
SOURCE_HF_REPO = "general.source.huggingface.repository"
FILE_TYPE = "general.file_type"
TYPE = "general.type"
ARCHITECTURE = "general.architecture"
QUANTIZATION_VERSION = "general.quantization_version"
ALIGNMENT = "general.alignment"
FILE_TYPE = "general.file_type"
# Authorship Metadata
NAME = "general.name"
AUTHOR = "general.author"
VERSION = "general.version"
ORGANIZATION = "general.organization"
FINETUNE = "general.finetune"
BASENAME = "general.basename"
DESCRIPTION = "general.description"
QUANTIZED_BY = "general.quantized_by"
SIZE_LABEL = "general.size_label"
# Licensing details
LICENSE = "general.license"
LICENSE_NAME = "general.license.name"
LICENSE_LINK = "general.license.link"
# Typically represents the converted GGUF repo (Unless native)
URL = "general.url" # Model Website/Paper
DOI = "general.doi"
UUID = "general.uuid"
REPO_URL = "general.repo_url" # Model Source Repository (git/svn/etc...)
# Model Source during conversion
SOURCE_URL = "general.source.url" # Model Website/Paper
SOURCE_DOI = "general.source.doi"
SOURCE_UUID = "general.source.uuid"
SOURCE_REPO_URL = "general.source.repo_url" # Model Source Repository (git/svn/etc...)
# Base Model Source. There can be more than one source if it's a merged
# model like with 'Mistral-7B-Merge-14-v0.1'. This will assist in
# tracing linage of models as it is finetuned or merged over time.
BASE_MODEL_COUNT = "general.base_model.count"
BASE_MODEL_NAME = "general.base_model.{id}.name"
BASE_MODEL_AUTHOR = "general.base_model.{id}.author"
BASE_MODEL_VERSION = "general.base_model.{id}.version"
BASE_MODEL_ORGANIZATION = "general.base_model.{id}.organization"
BASE_MODEL_URL = "general.base_model.{id}.url" # Model Website/Paper
BASE_MODEL_DOI = "general.base_model.{id}.doi"
BASE_MODEL_UUID = "general.base_model.{id}.uuid"
BASE_MODEL_REPO_URL = "general.base_model.{id}.repo_url" # Model Source Repository (git/svn/etc...)
# Array based KV stores
TAGS = "general.tags"
LANGUAGES = "general.languages"
DATASETS = "general.datasets"
class LLM:
VOCAB_SIZE = "{arch}.vocab_size"
@@ -49,6 +91,9 @@ class Keys:
EXPERT_WEIGHTS_SCALE = "{arch}.expert_weights_scale"
POOLING_TYPE = "{arch}.pooling_type"
LOGIT_SCALE = "{arch}.logit_scale"
DECODER_START_TOKEN_ID = "{arch}.decoder_start_token_id"
ATTN_LOGIT_SOFTCAPPING = "{arch}.attn_logit_softcapping"
FINAL_LOGIT_SOFTCAPPING = "{arch}.final_logit_softcapping"
class Attention:
HEAD_COUNT = "{arch}.attention.head_count"
@@ -62,6 +107,8 @@ class Keys:
CAUSAL = "{arch}.attention.causal"
Q_LORA_RANK = "{arch}.attention.q_lora_rank"
KV_LORA_RANK = "{arch}.attention.kv_lora_rank"
REL_BUCKETS_COUNT = "{arch}.attention.relative_buckets_count"
SLIDING_WINDOW = "{arch}.attention.sliding_window"
class Rope:
DIMENSION_COUNT = "{arch}.rope.dimension_count"
@@ -73,6 +120,11 @@ class Keys:
SCALING_FINETUNED = "{arch}.rope.scaling.finetuned"
SCALING_YARN_LOG_MUL = "{arch}.rope.scaling.yarn_log_multiplier"
class Split:
LLM_KV_SPLIT_NO = "split.no"
LLM_KV_SPLIT_COUNT = "split.count"
LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count"
class SSM:
CONV_KERNEL = "{arch}.ssm.conv_kernel"
INNER_SIZE = "{arch}.ssm.inner_size"
@@ -80,129 +132,175 @@ class Keys:
TIME_STEP_RANK = "{arch}.ssm.time_step_rank"
class Tokenizer:
MODEL = "tokenizer.ggml.model"
PRE = "tokenizer.ggml.pre"
LIST = "tokenizer.ggml.tokens"
TOKEN_TYPE = "tokenizer.ggml.token_type"
TOKEN_TYPE_COUNT = "tokenizer.ggml.token_type_count" # for BERT-style token types
SCORES = "tokenizer.ggml.scores"
MERGES = "tokenizer.ggml.merges"
BOS_ID = "tokenizer.ggml.bos_token_id"
EOS_ID = "tokenizer.ggml.eos_token_id"
UNK_ID = "tokenizer.ggml.unknown_token_id"
SEP_ID = "tokenizer.ggml.seperator_token_id"
PAD_ID = "tokenizer.ggml.padding_token_id"
CLS_ID = "tokenizer.ggml.cls_token_id"
MASK_ID = "tokenizer.ggml.mask_token_id"
ADD_BOS = "tokenizer.ggml.add_bos_token"
ADD_EOS = "tokenizer.ggml.add_eos_token"
ADD_PREFIX = "tokenizer.ggml.add_space_prefix"
HF_JSON = "tokenizer.huggingface.json"
RWKV = "tokenizer.rwkv.world"
CHAT_TEMPLATE = "tokenizer.chat_template"
CHAT_TEMPLATE_N = "tokenizer.chat_template.{name}"
CHAT_TEMPLATES = "tokenizer.chat_templates"
MODEL = "tokenizer.ggml.model"
PRE = "tokenizer.ggml.pre"
LIST = "tokenizer.ggml.tokens"
TOKEN_TYPE = "tokenizer.ggml.token_type"
TOKEN_TYPE_COUNT = "tokenizer.ggml.token_type_count" # for BERT-style token types
SCORES = "tokenizer.ggml.scores"
MERGES = "tokenizer.ggml.merges"
BOS_ID = "tokenizer.ggml.bos_token_id"
EOS_ID = "tokenizer.ggml.eos_token_id"
UNK_ID = "tokenizer.ggml.unknown_token_id"
SEP_ID = "tokenizer.ggml.seperator_token_id"
PAD_ID = "tokenizer.ggml.padding_token_id"
CLS_ID = "tokenizer.ggml.cls_token_id"
MASK_ID = "tokenizer.ggml.mask_token_id"
ADD_BOS = "tokenizer.ggml.add_bos_token"
ADD_EOS = "tokenizer.ggml.add_eos_token"
ADD_PREFIX = "tokenizer.ggml.add_space_prefix"
REMOVE_EXTRA_WS = "tokenizer.ggml.remove_extra_whitespaces"
PRECOMPILED_CHARSMAP = "tokenizer.ggml.precompiled_charsmap"
HF_JSON = "tokenizer.huggingface.json"
RWKV = "tokenizer.rwkv.world"
CHAT_TEMPLATE = "tokenizer.chat_template"
CHAT_TEMPLATE_N = "tokenizer.chat_template.{name}"
CHAT_TEMPLATES = "tokenizer.chat_templates"
# FIM/Infill special tokens constants
PREFIX_ID = "tokenizer.ggml.prefix_token_id"
SUFFIX_ID = "tokenizer.ggml.suffix_token_id"
MIDDLE_ID = "tokenizer.ggml.middle_token_id"
EOT_ID = "tokenizer.ggml.eot_token_id"
PREFIX_ID = "tokenizer.ggml.prefix_token_id"
SUFFIX_ID = "tokenizer.ggml.suffix_token_id"
MIDDLE_ID = "tokenizer.ggml.middle_token_id"
EOT_ID = "tokenizer.ggml.eot_token_id"
class Adapter:
TYPE = "adapter.type"
LORA_ALPHA = "adapter.lora.alpha"
#
# recommended mapping of model tensor names for storage in gguf
#
class GGUFType:
MODEL = "model"
ADAPTER = "adapter"
class MODEL_ARCH(IntEnum):
LLAMA = auto()
FALCON = auto()
BAICHUAN = auto()
GROK = auto()
GPT2 = auto()
GPTJ = auto()
GPTNEOX = auto()
MPT = auto()
STARCODER = auto()
REFACT = auto()
BERT = auto()
NOMIC_BERT = auto()
LLAMA = auto()
FALCON = auto()
BAICHUAN = auto()
GROK = auto()
GPT2 = auto()
GPTJ = auto()
GPTNEOX = auto()
MPT = auto()
STARCODER = auto()
REFACT = auto()
BERT = auto()
NOMIC_BERT = auto()
JINA_BERT_V2 = auto()
BLOOM = auto()
STABLELM = auto()
QWEN = auto()
QWEN2 = auto()
QWEN2MOE = auto()
PHI2 = auto()
PHI3 = auto()
PLAMO = auto()
CODESHELL = auto()
ORION = auto()
INTERNLM2 = auto()
MINICPM = auto()
GEMMA = auto()
STARCODER2 = auto()
MAMBA = auto()
XVERSE = auto()
COMMAND_R = auto()
DBRX = auto()
OLMO = auto()
ARCTIC = auto()
DEEPSEEK2 = auto()
BITNET = auto()
BLOOM = auto()
STABLELM = auto()
QWEN = auto()
QWEN2 = auto()
QWEN2MOE = auto()
PHI2 = auto()
PHI3 = auto()
PLAMO = auto()
CODESHELL = auto()
ORION = auto()
INTERNLM2 = auto()
MINICPM = auto()
GEMMA = auto()
GEMMA2 = auto()
STARCODER2 = auto()
MAMBA = auto()
XVERSE = auto()
COMMAND_R = auto()
DBRX = auto()
OLMO = auto()
OPENELM = auto()
ARCTIC = auto()
DEEPSEEK2 = auto()
CHATGLM = auto()
BITNET = auto()
T5 = auto()
JAIS = auto()
class MODEL_TENSOR(IntEnum):
TOKEN_EMBD = auto()
TOKEN_EMBD_NORM = auto()
TOKEN_TYPES = auto()
POS_EMBD = auto()
OUTPUT = auto()
OUTPUT_NORM = auto()
ROPE_FREQS = auto()
ROPE_FACTORS_LONG = auto()
ROPE_FACTORS_SHORT = auto()
ATTN_Q = auto()
ATTN_K = auto()
ATTN_V = auto()
ATTN_QKV = auto()
ATTN_OUT = auto()
ATTN_NORM = auto()
ATTN_NORM_2 = auto()
ATTN_OUT_NORM = auto()
ATTN_ROT_EMBD = auto()
FFN_GATE_INP = auto()
FFN_GATE_INP_SHEXP = auto()
FFN_NORM = auto()
FFN_GATE = auto()
FFN_DOWN = auto()
FFN_UP = auto()
FFN_ACT = auto()
FFN_NORM_EXP = auto()
FFN_GATE_EXP = auto()
FFN_DOWN_EXP = auto()
FFN_UP_EXP = auto()
FFN_GATE_SHEXP = auto()
FFN_DOWN_SHEXP = auto()
FFN_UP_SHEXP = auto()
ATTN_Q_NORM = auto()
ATTN_K_NORM = auto()
LAYER_OUT_NORM = auto()
SSM_IN = auto()
SSM_CONV1D = auto()
SSM_X = auto()
SSM_DT = auto()
SSM_A = auto()
SSM_D = auto()
SSM_OUT = auto()
ATTN_Q_A = auto()
ATTN_Q_B = auto()
ATTN_KV_A_MQA = auto()
ATTN_KV_B = auto()
ATTN_Q_A_NORM = auto()
ATTN_KV_A_NORM = auto()
FFN_SUB_NORM = auto()
ATTN_SUB_NORM = auto()
TOKEN_EMBD = auto()
TOKEN_EMBD_NORM = auto()
TOKEN_TYPES = auto()
POS_EMBD = auto()
OUTPUT = auto()
OUTPUT_NORM = auto()
ROPE_FREQS = auto()
ROPE_FACTORS_LONG = auto()
ROPE_FACTORS_SHORT = auto()
ATTN_Q = auto()
ATTN_K = auto()
ATTN_V = auto()
ATTN_QKV = auto()
ATTN_OUT = auto()
ATTN_NORM = auto()
ATTN_NORM_2 = auto()
ATTN_OUT_NORM = auto()
ATTN_POST_NORM = auto()
ATTN_ROT_EMBD = auto()
FFN_GATE_INP = auto()
FFN_GATE_INP_SHEXP = auto()
FFN_NORM = auto()
FFN_PRE_NORM = auto()
FFN_POST_NORM = auto()
FFN_GATE = auto()
FFN_DOWN = auto()
FFN_UP = auto()
FFN_ACT = auto()
FFN_NORM_EXP = auto()
FFN_GATE_EXP = auto()
FFN_DOWN_EXP = auto()
FFN_UP_EXP = auto()
FFN_GATE_SHEXP = auto()
FFN_DOWN_SHEXP = auto()
FFN_UP_SHEXP = auto()
ATTN_Q_NORM = auto()
ATTN_K_NORM = auto()
LAYER_OUT_NORM = auto()
SSM_IN = auto()
SSM_CONV1D = auto()
SSM_X = auto()
SSM_DT = auto()
SSM_A = auto()
SSM_D = auto()
SSM_OUT = auto()
ATTN_Q_A = auto()
ATTN_Q_B = auto()
ATTN_KV_A_MQA = auto()
ATTN_KV_B = auto()
ATTN_Q_A_NORM = auto()
ATTN_KV_A_NORM = auto()
FFN_SUB_NORM = auto()
ATTN_SUB_NORM = auto()
DEC_ATTN_NORM = auto()
DEC_ATTN_Q = auto()
DEC_ATTN_K = auto()
DEC_ATTN_V = auto()
DEC_ATTN_OUT = auto()
DEC_ATTN_REL_B = auto()
DEC_CROSS_ATTN_NORM = auto()
DEC_CROSS_ATTN_Q = auto()
DEC_CROSS_ATTN_K = auto()
DEC_CROSS_ATTN_V = auto()
DEC_CROSS_ATTN_OUT = auto()
DEC_CROSS_ATTN_REL_B = auto()
DEC_FFN_NORM = auto()
DEC_FFN_GATE = auto()
DEC_FFN_DOWN = auto()
DEC_FFN_UP = auto()
DEC_OUTPUT_NORM = auto()
ENC_ATTN_NORM = auto()
ENC_ATTN_Q = auto()
ENC_ATTN_K = auto()
ENC_ATTN_V = auto()
ENC_ATTN_OUT = auto()
ENC_ATTN_REL_B = auto()
ENC_FFN_NORM = auto()
ENC_FFN_GATE = auto()
ENC_FFN_DOWN = auto()
ENC_FFN_UP = auto()
ENC_OUTPUT_NORM = auto()
MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
@@ -232,68 +330,104 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.INTERNLM2: "internlm2",
MODEL_ARCH.MINICPM: "minicpm",
MODEL_ARCH.GEMMA: "gemma",
MODEL_ARCH.GEMMA2: "gemma2",
MODEL_ARCH.STARCODER2: "starcoder2",
MODEL_ARCH.MAMBA: "mamba",
MODEL_ARCH.XVERSE: "xverse",
MODEL_ARCH.COMMAND_R: "command-r",
MODEL_ARCH.DBRX: "dbrx",
MODEL_ARCH.OLMO: "olmo",
MODEL_ARCH.OPENELM: "openelm",
MODEL_ARCH.ARCTIC: "arctic",
MODEL_ARCH.DEEPSEEK2: "deepseek2",
MODEL_ARCH.CHATGLM: "chatglm",
MODEL_ARCH.BITNET: "bitnet",
MODEL_ARCH.T5: "t5",
MODEL_ARCH.JAIS: "jais",
}
TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.TOKEN_EMBD: "token_embd",
MODEL_TENSOR.TOKEN_EMBD_NORM: "token_embd_norm",
MODEL_TENSOR.TOKEN_TYPES: "token_types",
MODEL_TENSOR.POS_EMBD: "position_embd",
MODEL_TENSOR.OUTPUT_NORM: "output_norm",
MODEL_TENSOR.OUTPUT: "output",
MODEL_TENSOR.ROPE_FREQS: "rope_freqs",
MODEL_TENSOR.ROPE_FACTORS_LONG: "rope_factors_long",
MODEL_TENSOR.ROPE_FACTORS_SHORT: "rope_factors_short",
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_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.ATTN_Q_NORM: "blk.{bid}.attn_q_norm",
MODEL_TENSOR.ATTN_K_NORM: "blk.{bid}.attn_k_norm",
MODEL_TENSOR.ATTN_OUT_NORM: "blk.{bid}.attn_output_norm",
MODEL_TENSOR.FFN_GATE_INP: "blk.{bid}.ffn_gate_inp",
MODEL_TENSOR.FFN_GATE_INP_SHEXP: "blk.{bid}.ffn_gate_inp_shexp",
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_TENSOR.FFN_GATE_SHEXP: "blk.{bid}.ffn_gate_shexp",
MODEL_TENSOR.FFN_DOWN_SHEXP: "blk.{bid}.ffn_down_shexp",
MODEL_TENSOR.FFN_UP_SHEXP: "blk.{bid}.ffn_up_shexp",
MODEL_TENSOR.FFN_ACT: "blk.{bid}.ffn",
MODEL_TENSOR.FFN_NORM_EXP: "blk.{bid}.ffn_norm_exps",
MODEL_TENSOR.FFN_GATE_EXP: "blk.{bid}.ffn_gate_exps",
MODEL_TENSOR.FFN_DOWN_EXP: "blk.{bid}.ffn_down_exps",
MODEL_TENSOR.FFN_UP_EXP: "blk.{bid}.ffn_up_exps",
MODEL_TENSOR.LAYER_OUT_NORM: "blk.{bid}.layer_output_norm",
MODEL_TENSOR.SSM_IN: "blk.{bid}.ssm_in",
MODEL_TENSOR.SSM_CONV1D: "blk.{bid}.ssm_conv1d",
MODEL_TENSOR.SSM_X: "blk.{bid}.ssm_x",
MODEL_TENSOR.SSM_DT: "blk.{bid}.ssm_dt",
MODEL_TENSOR.SSM_A: "blk.{bid}.ssm_a",
MODEL_TENSOR.SSM_D: "blk.{bid}.ssm_d",
MODEL_TENSOR.SSM_OUT: "blk.{bid}.ssm_out",
MODEL_TENSOR.ATTN_Q_A: "blk.{bid}.attn_q_a",
MODEL_TENSOR.ATTN_Q_B: "blk.{bid}.attn_q_b",
MODEL_TENSOR.ATTN_KV_A_MQA: "blk.{bid}.attn_kv_a_mqa",
MODEL_TENSOR.ATTN_KV_B: "blk.{bid}.attn_kv_b",
MODEL_TENSOR.ATTN_Q_A_NORM: "blk.{bid}.attn_q_a_norm",
MODEL_TENSOR.ATTN_KV_A_NORM: "blk.{bid}.attn_kv_a_norm",
MODEL_TENSOR.ATTN_SUB_NORM: "blk.{bid}.attn_sub_norm",
MODEL_TENSOR.FFN_SUB_NORM: "blk.{bid}.ffn_sub_norm",
MODEL_TENSOR.TOKEN_EMBD: "token_embd",
MODEL_TENSOR.TOKEN_EMBD_NORM: "token_embd_norm",
MODEL_TENSOR.TOKEN_TYPES: "token_types",
MODEL_TENSOR.POS_EMBD: "position_embd",
MODEL_TENSOR.OUTPUT_NORM: "output_norm",
MODEL_TENSOR.OUTPUT: "output",
MODEL_TENSOR.ROPE_FREQS: "rope_freqs",
MODEL_TENSOR.ROPE_FACTORS_LONG: "rope_factors_long",
MODEL_TENSOR.ROPE_FACTORS_SHORT: "rope_factors_short",
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_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.ATTN_Q_NORM: "blk.{bid}.attn_q_norm",
MODEL_TENSOR.ATTN_K_NORM: "blk.{bid}.attn_k_norm",
MODEL_TENSOR.ATTN_OUT_NORM: "blk.{bid}.attn_output_norm",
MODEL_TENSOR.ATTN_POST_NORM: "blk.{bid}.post_attention_norm",
MODEL_TENSOR.FFN_GATE_INP: "blk.{bid}.ffn_gate_inp",
MODEL_TENSOR.FFN_GATE_INP_SHEXP: "blk.{bid}.ffn_gate_inp_shexp",
MODEL_TENSOR.FFN_NORM: "blk.{bid}.ffn_norm",
MODEL_TENSOR.FFN_PRE_NORM: "blk.{bid}.ffn_norm",
MODEL_TENSOR.FFN_POST_NORM: "blk.{bid}.post_ffw_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_TENSOR.FFN_GATE_SHEXP: "blk.{bid}.ffn_gate_shexp",
MODEL_TENSOR.FFN_DOWN_SHEXP: "blk.{bid}.ffn_down_shexp",
MODEL_TENSOR.FFN_UP_SHEXP: "blk.{bid}.ffn_up_shexp",
MODEL_TENSOR.FFN_ACT: "blk.{bid}.ffn",
MODEL_TENSOR.FFN_NORM_EXP: "blk.{bid}.ffn_norm_exps",
MODEL_TENSOR.FFN_GATE_EXP: "blk.{bid}.ffn_gate_exps",
MODEL_TENSOR.FFN_DOWN_EXP: "blk.{bid}.ffn_down_exps",
MODEL_TENSOR.FFN_UP_EXP: "blk.{bid}.ffn_up_exps",
MODEL_TENSOR.LAYER_OUT_NORM: "blk.{bid}.layer_output_norm",
MODEL_TENSOR.SSM_IN: "blk.{bid}.ssm_in",
MODEL_TENSOR.SSM_CONV1D: "blk.{bid}.ssm_conv1d",
MODEL_TENSOR.SSM_X: "blk.{bid}.ssm_x",
MODEL_TENSOR.SSM_DT: "blk.{bid}.ssm_dt",
MODEL_TENSOR.SSM_A: "blk.{bid}.ssm_a",
MODEL_TENSOR.SSM_D: "blk.{bid}.ssm_d",
MODEL_TENSOR.SSM_OUT: "blk.{bid}.ssm_out",
MODEL_TENSOR.ATTN_Q_A: "blk.{bid}.attn_q_a",
MODEL_TENSOR.ATTN_Q_B: "blk.{bid}.attn_q_b",
MODEL_TENSOR.ATTN_KV_A_MQA: "blk.{bid}.attn_kv_a_mqa",
MODEL_TENSOR.ATTN_KV_B: "blk.{bid}.attn_kv_b",
MODEL_TENSOR.ATTN_Q_A_NORM: "blk.{bid}.attn_q_a_norm",
MODEL_TENSOR.ATTN_KV_A_NORM: "blk.{bid}.attn_kv_a_norm",
MODEL_TENSOR.ATTN_SUB_NORM: "blk.{bid}.attn_sub_norm",
MODEL_TENSOR.FFN_SUB_NORM: "blk.{bid}.ffn_sub_norm",
MODEL_TENSOR.DEC_ATTN_NORM: "dec.blk.{bid}.attn_norm",
MODEL_TENSOR.DEC_ATTN_Q: "dec.blk.{bid}.attn_q",
MODEL_TENSOR.DEC_ATTN_K: "dec.blk.{bid}.attn_k",
MODEL_TENSOR.DEC_ATTN_V: "dec.blk.{bid}.attn_v",
MODEL_TENSOR.DEC_ATTN_OUT: "dec.blk.{bid}.attn_o",
MODEL_TENSOR.DEC_ATTN_REL_B: "dec.blk.{bid}.attn_rel_b",
MODEL_TENSOR.DEC_CROSS_ATTN_NORM: "dec.blk.{bid}.cross_attn_norm",
MODEL_TENSOR.DEC_CROSS_ATTN_Q: "dec.blk.{bid}.cross_attn_q",
MODEL_TENSOR.DEC_CROSS_ATTN_K: "dec.blk.{bid}.cross_attn_k",
MODEL_TENSOR.DEC_CROSS_ATTN_V: "dec.blk.{bid}.cross_attn_v",
MODEL_TENSOR.DEC_CROSS_ATTN_OUT: "dec.blk.{bid}.cross_attn_o",
MODEL_TENSOR.DEC_CROSS_ATTN_REL_B: "dec.blk.{bid}.cross_attn_rel_b",
MODEL_TENSOR.DEC_FFN_NORM: "dec.blk.{bid}.ffn_norm",
MODEL_TENSOR.DEC_FFN_GATE: "dec.blk.{bid}.ffn_gate",
MODEL_TENSOR.DEC_FFN_DOWN: "dec.blk.{bid}.ffn_down",
MODEL_TENSOR.DEC_FFN_UP: "dec.blk.{bid}.ffn_up",
MODEL_TENSOR.DEC_OUTPUT_NORM: "dec.output_norm",
MODEL_TENSOR.ENC_ATTN_NORM: "enc.blk.{bid}.attn_norm",
MODEL_TENSOR.ENC_ATTN_Q: "enc.blk.{bid}.attn_q",
MODEL_TENSOR.ENC_ATTN_K: "enc.blk.{bid}.attn_k",
MODEL_TENSOR.ENC_ATTN_V: "enc.blk.{bid}.attn_v",
MODEL_TENSOR.ENC_ATTN_OUT: "enc.blk.{bid}.attn_o",
MODEL_TENSOR.ENC_ATTN_REL_B: "enc.blk.{bid}.attn_rel_b",
MODEL_TENSOR.ENC_FFN_NORM: "enc.blk.{bid}.ffn_norm",
MODEL_TENSOR.ENC_FFN_GATE: "enc.blk.{bid}.ffn_gate",
MODEL_TENSOR.ENC_FFN_DOWN: "enc.blk.{bid}.ffn_down",
MODEL_TENSOR.ENC_FFN_UP: "enc.blk.{bid}.ffn_up",
MODEL_TENSOR.ENC_OUTPUT_NORM: "enc.output_norm",
}
MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@@ -684,6 +818,21 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.FFN_NORM,
],
MODEL_ARCH.GEMMA2: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_POST_NORM,
MODEL_TENSOR.FFN_PRE_NORM,
MODEL_TENSOR.FFN_POST_NORM,
],
MODEL_ARCH.STARCODER2: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
@@ -766,6 +915,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.OPENELM: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_Q_NORM,
MODEL_TENSOR.ATTN_K_NORM,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.ARCTIC: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
@@ -814,17 +976,26 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN_SHEXP,
MODEL_TENSOR.FFN_UP_SHEXP,
],
MODEL_ARCH.CHATGLM : [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.BITNET: [
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_ROT_EMBD,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
@@ -832,6 +1003,50 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.ATTN_SUB_NORM,
MODEL_TENSOR.FFN_SUB_NORM,
],
MODEL_ARCH.T5: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.DEC_ATTN_NORM,
MODEL_TENSOR.DEC_ATTN_Q,
MODEL_TENSOR.DEC_ATTN_K,
MODEL_TENSOR.DEC_ATTN_V,
MODEL_TENSOR.DEC_ATTN_OUT,
MODEL_TENSOR.DEC_ATTN_REL_B,
MODEL_TENSOR.DEC_CROSS_ATTN_NORM,
MODEL_TENSOR.DEC_CROSS_ATTN_Q,
MODEL_TENSOR.DEC_CROSS_ATTN_K,
MODEL_TENSOR.DEC_CROSS_ATTN_V,
MODEL_TENSOR.DEC_CROSS_ATTN_OUT,
MODEL_TENSOR.DEC_CROSS_ATTN_REL_B,
MODEL_TENSOR.DEC_FFN_NORM,
MODEL_TENSOR.DEC_FFN_GATE,
MODEL_TENSOR.DEC_FFN_DOWN,
MODEL_TENSOR.DEC_FFN_UP,
MODEL_TENSOR.DEC_OUTPUT_NORM,
MODEL_TENSOR.ENC_ATTN_NORM,
MODEL_TENSOR.ENC_ATTN_Q,
MODEL_TENSOR.ENC_ATTN_K,
MODEL_TENSOR.ENC_ATTN_V,
MODEL_TENSOR.ENC_ATTN_OUT,
MODEL_TENSOR.ENC_ATTN_REL_B,
MODEL_TENSOR.ENC_FFN_NORM,
MODEL_TENSOR.ENC_FFN_GATE,
MODEL_TENSOR.ENC_FFN_DOWN,
MODEL_TENSOR.ENC_FFN_UP,
MODEL_TENSOR.ENC_OUTPUT_NORM,
],
MODEL_ARCH.JAIS: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_UP,
],
# TODO
}
@@ -869,6 +1084,9 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_ROT_EMBD,
],
MODEL_ARCH.CHATGLM: [
MODEL_TENSOR.ROPE_FREQS,
],
}
#
@@ -1056,7 +1274,6 @@ KEY_GENERAL_URL = Keys.General.URL
KEY_GENERAL_DESCRIPTION = Keys.General.DESCRIPTION
KEY_GENERAL_LICENSE = Keys.General.LICENSE
KEY_GENERAL_SOURCE_URL = Keys.General.SOURCE_URL
KEY_GENERAL_SOURCE_HF_REPO = Keys.General.SOURCE_HF_REPO
KEY_GENERAL_FILE_TYPE = Keys.General.FILE_TYPE
# LLM

35
gguf-py/gguf/gguf_reader.py
View File

@@ -67,8 +67,9 @@ class ReaderTensor(NamedTuple):
class GGUFReader:
# I - same as host, S - swapped
byte_order: Literal['I'] | Literal['S'] = 'I'
byte_order: Literal['I', 'S'] = 'I'
alignment: int = GGUF_DEFAULT_ALIGNMENT
data_offset: int
# Note: Internal helper, API may change.
gguf_scalar_to_np: dict[GGUFValueType, type[np.generic]] = {
@@ -85,12 +86,16 @@ class GGUFReader:
GGUFValueType.BOOL: np.bool_,
}
def __init__(self, path: os.PathLike[str] | str, mode: Literal['r'] | Literal['r+'] | Literal['c'] = 'r'):
def __init__(self, path: os.PathLike[str] | str, mode: Literal['r', 'r+', 'c'] = 'r'):
self.data = np.memmap(path, mode = mode)
offs = 0
# Check for GGUF magic
if self._get(offs, np.uint32, override_order = '<')[0] != GGUF_MAGIC:
raise ValueError('GGUF magic invalid')
offs += 4
# Check GGUF version
temp_version = self._get(offs, np.uint32)
if temp_version[0] & 65535 == 0:
# If we get 0 here that means it's (probably) a GGUF file created for
@@ -103,12 +108,16 @@ class GGUFReader:
self.fields: OrderedDict[str, ReaderField] = OrderedDict()
self.tensors: list[ReaderTensor] = []
offs += self._push_field(ReaderField(offs, 'GGUF.version', [temp_version], [0], [GGUFValueType.UINT32]))
# Check tensor count and kv count
temp_counts = self._get(offs, np.uint64, 2)
offs += self._push_field(ReaderField(offs, 'GGUF.tensor_count', [temp_counts[:1]], [0], [GGUFValueType.UINT64]))
offs += self._push_field(ReaderField(offs, 'GGUF.kv_count', [temp_counts[1:]], [0], [GGUFValueType.UINT64]))
tensor_count, kv_count = temp_counts
offs = self._build_fields(offs, kv_count)
offs, tensors_fields = self._build_tensors_fields(offs, tensor_count)
# Build Tensor Info Fields
offs, tensors_fields = self._build_tensor_info(offs, tensor_count)
new_align = self.fields.get('general.alignment')
if new_align is not None:
if new_align.types != [GGUFValueType.UINT32]:
@@ -117,6 +126,7 @@ class GGUFReader:
padding = offs % self.alignment
if padding != 0:
offs += self.alignment - padding
self.data_offset = offs
self._build_tensors(offs, tensors_fields)
_DT = TypeVar('_DT', bound = npt.DTypeLike)
@@ -130,7 +140,7 @@ class GGUFReader:
return self.tensors[idx]
def _get(
self, offset: int, dtype: npt.DTypeLike, count: int = 1, override_order: None | Literal['I'] | Literal['S'] | Literal['<'] = None,
self, offset: int, dtype: npt.DTypeLike, count: int = 1, override_order: None | Literal['I', 'S', '<'] = None,
) -> npt.NDArray[Any]:
count = int(count)
itemsize = int(np.empty([], dtype = dtype).itemsize)
@@ -193,18 +203,29 @@ class GGUFReader:
# We can't deal with this one.
raise ValueError('Unknown/unhandled field type {gtype}')
def _get_tensor(self, orig_offs: int) -> ReaderField:
def _get_tensor_info_field(self, orig_offs: int) -> ReaderField:
offs = orig_offs
# Get Tensor Name
name_len, name_data = self._get_str(offs)
offs += int(name_len.nbytes + name_data.nbytes)
# Get Tensor Dimensions Count
n_dims = self._get(offs, np.uint32)
offs += int(n_dims.nbytes)
# Get Tensor Dimension Array
dims = self._get(offs, np.uint64, n_dims[0])
offs += int(dims.nbytes)
# Get Tensor Encoding Scheme Type
raw_dtype = self._get(offs, np.uint32)
offs += int(raw_dtype.nbytes)
# Get Tensor Offset
offset_tensor = self._get(offs, np.uint64)
offs += int(offset_tensor.nbytes)
return ReaderField(
orig_offs,
str(bytes(name_data), encoding = 'utf-8'),
@@ -233,10 +254,10 @@ class GGUFReader:
offs += field_size
return offs
def _build_tensors_fields(self, offs: int, count: int) -> tuple[int, list[ReaderField]]:
def _build_tensor_info(self, offs: int, count: int) -> tuple[int, list[ReaderField]]:
tensor_fields = []
for _ in range(count):
field = self._get_tensor(offs)
field = self._get_tensor_info_field(offs)
offs += sum(int(part.nbytes) for part in field.parts)
tensor_fields.append(field)
return offs, tensor_fields

461
gguf-py/gguf/gguf_writer.py
View File

@@ -7,6 +7,8 @@ import struct
import tempfile
from dataclasses import dataclass
from enum import Enum, auto
from math import prod
from pathlib import Path
from io import BufferedWriter
from typing import IO, Any, Sequence, Mapping
from string import ascii_letters, digits
@@ -31,6 +33,9 @@ from .quants import quant_shape_from_byte_shape
logger = logging.getLogger(__name__)
SHARD_NAME_FORMAT = "{:s}-{:05d}-of-{:05d}.gguf"
@dataclass
class TensorInfo:
shape: Sequence[int]
@@ -55,11 +60,11 @@ class WriterState(Enum):
class GGUFWriter:
fout: BufferedWriter | None
path: os.PathLike[str] | str | None
fout: list[BufferedWriter] | None
path: Path | None
temp_file: tempfile.SpooledTemporaryFile[bytes] | None
tensors: dict[str, TensorInfo]
kv_data: dict[str, GGUFValue]
tensors: list[dict[str, TensorInfo]]
kv_data: list[dict[str, GGUFValue]]
state: WriterState
_simple_value_packing = {
GGUFValueType.UINT8: "B",
@@ -76,29 +81,89 @@ class GGUFWriter:
}
def __init__(
self, path: os.PathLike[str] | str | None, arch: str, use_temp_file: bool = False,
endianess: GGUFEndian = GGUFEndian.LITTLE,
self, path: os.PathLike[str] | str | None, arch: str, use_temp_file: bool = False, endianess: GGUFEndian = GGUFEndian.LITTLE,
split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False
):
self.fout = None
self.path = path
self.path = Path(path) if path else None
self.arch = arch
self.endianess = endianess
self.data_alignment = GGUF_DEFAULT_ALIGNMENT
self.use_temp_file = use_temp_file
self.temp_file = None
self.tensors = dict()
self.kv_data = dict()
self.tensors = [{}]
self.kv_data = [{}]
self.split_max_tensors = split_max_tensors
self.split_max_size = split_max_size
self.dry_run = dry_run
self.small_first_shard = small_first_shard
logger.info("gguf: This GGUF file is for {0} Endian only".format(
"Big" if self.endianess == GGUFEndian.BIG else "Little",
))
self.state = WriterState.NO_FILE
if self.small_first_shard:
self.tensors.append({})
self.add_architecture()
def open_output_file(self, path: os.PathLike[str] | str | None = None) -> None:
def get_total_parameter_count(self) -> tuple[int, int, int, int]:
total_params = 0
shared_params = 0
expert_params = 0
expert_sum = 0
n_expert_tensors = 0
last_lora_a: tuple[str, TensorInfo] | None = None
for tensors in self.tensors:
for name, info in tensors.items():
shape = info.shape
if name.endswith(".lora_a"):
last_lora_a = (name, info)
continue
elif name.endswith(".lora_b"):
if last_lora_a is None or last_lora_a[0] != name[:-1] + "a":
# Bail when the LoRA pair can't be found trivially
logger.warning("can't measure LoRA size correctly, tensor order is unusual")
return 0, 0, 0, 0
else:
shape = (*shape[:-1], last_lora_a[1].shape[-1])
size = prod(shape)
if "_exps." in name:
expert_params += (size // shape[-3])
expert_sum += shape[-3]
n_expert_tensors += 1
else:
shared_params += size
total_params += size
# Hopefully this should work even for variable-expert-count models
expert_count = (expert_sum // n_expert_tensors) if n_expert_tensors > 0 else 0
# Negate the total to signal it's likely not exact
if last_lora_a is not None:
total_params = -total_params
# NOTE: keep the output in the same order as accepted by 'size_label' in gguf-py/gguf/utility.py
return total_params, shared_params, expert_params, expert_count
def format_shard_names(self, path: Path) -> list[Path]:
if len(self.tensors) == 1:
return [path]
return [path.with_name(SHARD_NAME_FORMAT.format(path.stem, i + 1, len(self.tensors))) for i in range(len(self.tensors))]
def open_output_file(self, path: Path | None = None) -> None:
if self.state is WriterState.EMPTY and self.fout is not None and (path is None or path == self.path):
# allow calling this multiple times as long as the path is the same
return
if self.state is not WriterState.NO_FILE:
raise ValueError(f'Expected output file to be not yet opened, got {self.state}')
@@ -106,22 +171,60 @@ class GGUFWriter:
self.path = path
if self.path is not None:
if self.fout is not None:
self.fout.close()
self.fout = open(self.path, "wb")
filenames = self.print_plan()
self.fout = [open(filename, "wb") for filename in filenames]
self.state = WriterState.EMPTY
def write_header_to_file(self, path: os.PathLike[str] | str | None = None) -> None:
def print_plan(self) -> list[Path]:
logger.info("Writing the following files:")
assert self.path is not None
filenames = self.format_shard_names(self.path)
assert len(filenames) == len(self.tensors)
for name, tensors in zip(filenames, self.tensors):
logger.info(f"{name}: n_tensors = {len(tensors)}, total_size = {GGUFWriter.format_n_bytes_to_str(sum(ti.nbytes for ti in tensors.values()))}")
if self.dry_run:
logger.info("Dry run, not writing files")
for name in filenames:
print(name) # noqa: NP100
exit()
return filenames
def add_shard_kv_data(self) -> None:
if len(self.tensors) == 1:
return
total_tensors = sum(len(t) for t in self.tensors)
assert self.fout is not None
total_splits = len(self.fout)
self.kv_data.extend({} for _ in range(len(self.kv_data), total_splits))
for i, kv_data in enumerate(self.kv_data):
kv_data[Keys.Split.LLM_KV_SPLIT_NO] = GGUFValue(i, GGUFValueType.UINT16)
kv_data[Keys.Split.LLM_KV_SPLIT_COUNT] = GGUFValue(total_splits, GGUFValueType.UINT16)
kv_data[Keys.Split.LLM_KV_SPLIT_TENSORS_COUNT] = GGUFValue(total_tensors, GGUFValueType.INT32)
def write_header_to_file(self, path: Path | None = None) -> None:
if len(self.tensors) == 1 and (self.split_max_tensors != 0 or self.split_max_size != 0):
logger.warning("Model fails split requirements, not splitting")
self.open_output_file(path)
if self.state is not WriterState.EMPTY:
raise ValueError(f'Expected output file to be empty, got {self.state}')
self._write_packed("<I", GGUF_MAGIC, skip_pack_prefix = True)
self._write_packed("I", GGUF_VERSION)
self._write_packed("Q", len(self.tensors))
self._write_packed("Q", len(self.kv_data))
self.flush()
assert self.fout is not None
assert len(self.fout) == len(self.tensors)
assert len(self.kv_data) == 1
self.add_shard_kv_data()
for fout, tensors, kv_data in zip(self.fout, self.tensors, self.kv_data):
fout.write(self._pack("<I", GGUF_MAGIC, skip_pack_prefix = True))
fout.write(self._pack("I", GGUF_VERSION))
fout.write(self._pack("Q", len(tensors)))
fout.write(self._pack("Q", len(kv_data)))
fout.flush()
self.state = WriterState.HEADER
def write_kv_data_to_file(self) -> None:
@@ -129,13 +232,15 @@ class GGUFWriter:
raise ValueError(f'Expected output file to contain the header, got {self.state}')
assert self.fout is not None
kv_data = bytearray()
for fout, kv_data in zip(self.fout, self.kv_data):
kv_bytes = bytearray()
for key, val in self.kv_data.items():
kv_data += self._pack_val(key, GGUFValueType.STRING, add_vtype=False)
kv_data += self._pack_val(val.value, val.type, add_vtype=True)
for key, val in kv_data.items():
kv_bytes += self._pack_val(key, GGUFValueType.STRING, add_vtype=False)
kv_bytes += self._pack_val(val.value, val.type, add_vtype=True)
fout.write(kv_bytes)
self.fout.write(kv_data)
self.flush()
self.state = WriterState.KV_DATA
@@ -144,28 +249,29 @@ class GGUFWriter:
raise ValueError(f'Expected output file to contain KV data, got {self.state}')
assert self.fout is not None
ti_data = bytearray()
offset_tensor = 0
for fout, tensors in zip(self.fout, self.tensors):
ti_data = bytearray()
offset_tensor = 0
for name, ti in self.tensors.items():
ti_data += self._pack_val(name, GGUFValueType.STRING, add_vtype=False)
n_dims = len(ti.shape)
ti_data += self._pack("I", n_dims)
for i in range(n_dims):
ti_data += self._pack("Q", ti.shape[n_dims - 1 - i])
ti_data += self._pack("I", ti.dtype)
ti_data += self._pack("Q", offset_tensor)
offset_tensor += GGUFWriter.ggml_pad(ti.nbytes, self.data_alignment)
for name, ti in tensors.items():
ti_data += self._pack_val(name, GGUFValueType.STRING, add_vtype=False)
n_dims = len(ti.shape)
ti_data += self._pack("I", n_dims)
for j in range(n_dims):
ti_data += self._pack("Q", ti.shape[n_dims - 1 - j])
ti_data += self._pack("I", ti.dtype)
ti_data += self._pack("Q", offset_tensor)
offset_tensor += GGUFWriter.ggml_pad(ti.nbytes, self.data_alignment)
self.fout.write(ti_data)
self.flush()
fout.write(ti_data)
fout.flush()
self.state = WriterState.TI_DATA
def add_key_value(self, key: str, val: Any, vtype: GGUFValueType) -> None:
if key in self.kv_data:
if any(key in kv_data for kv_data in self.kv_data):
raise ValueError(f'Duplicated key name {key!r}')
self.kv_data[key] = GGUFValue(value=val, type=vtype)
self.kv_data[0][key] = GGUFValue(value=val, type=vtype)
def add_uint8(self, key: str, val: int) -> None:
self.add_key_value(key,val, GGUFValueType.UINT8)
@@ -206,9 +312,6 @@ class GGUFWriter:
self.add_key_value(key, val, GGUFValueType.STRING)
def add_array(self, key: str, val: Sequence[Any]) -> None:
if not isinstance(val, Sequence):
raise ValueError("Value must be a sequence for array type")
self.add_key_value(key, val, GGUFValueType.ARRAY)
@staticmethod
@@ -222,7 +325,7 @@ class GGUFWriter:
if self.state is not WriterState.NO_FILE:
raise ValueError(f'Expected output file to be not yet opened, got {self.state}')
if name in self.tensors:
if any(name in tensors for tensors in self.tensors):
raise ValueError(f'Duplicated tensor name {name!r}')
if raw_dtype is None:
@@ -247,7 +350,18 @@ class GGUFWriter:
if tensor_dtype == np.uint8:
tensor_shape = quant_shape_from_byte_shape(tensor_shape, raw_dtype)
self.tensors[name] = TensorInfo(shape=tensor_shape, dtype=dtype, nbytes=tensor_nbytes)
# make sure there is at least one tensor before splitting
if len(self.tensors[-1]) > 0:
if ( # split when over tensor limit
self.split_max_tensors != 0
and len(self.tensors[-1]) >= self.split_max_tensors
) or ( # split when over size limit
self.split_max_size != 0
and sum(ti.nbytes for ti in self.tensors[-1].values()) + tensor_nbytes > self.split_max_size
):
self.tensors.append({})
self.tensors[-1][name] = TensorInfo(shape=tensor_shape, dtype=dtype, nbytes=tensor_nbytes)
def add_tensor(
self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Sequence[int] | None = None,
@@ -264,7 +378,7 @@ class GGUFWriter:
self.add_tensor_info(name, shape, tensor.dtype, tensor.nbytes, raw_dtype=raw_dtype)
if self.temp_file is None:
self.tensors[name].tensor = tensor
self.tensors[-1][name].tensor = tensor
return
tensor.tofile(self.temp_file)
@@ -282,9 +396,24 @@ class GGUFWriter:
if self.endianess == GGUFEndian.BIG:
tensor.byteswap(inplace=True)
self.write_padding(self.fout, self.fout.tell())
tensor.tofile(self.fout)
self.write_padding(self.fout, tensor.nbytes)
file_id = -1
for i, tensors in enumerate(self.tensors):
if len(tensors) > 0:
file_id = i
break
fout = self.fout[file_id]
# pop the first tensor info
# TODO: cleaner way to get the first key
first_tensor_name = [name for name, _ in zip(self.tensors[file_id].keys(), range(1))][0]
ti = self.tensors[file_id].pop(first_tensor_name)
assert ti.nbytes == tensor.nbytes
self.write_padding(fout, fout.tell())
tensor.tofile(fout)
self.write_padding(fout, tensor.nbytes)
self.state = WriterState.WEIGHTS
@@ -293,31 +422,43 @@ class GGUFWriter:
assert self.fout is not None
self.write_padding(self.fout, self.fout.tell())
for fout in self.fout:
self.write_padding(fout, fout.tell())
if self.temp_file is None:
shard_bar = None
bar = None
if progress:
from tqdm import tqdm
total_bytes = sum(t.nbytes for t in self.tensors.values())
total_bytes = sum(ti.nbytes for t in self.tensors for ti in t.values())
if len(self.fout) > 1:
shard_bar = tqdm(desc=f"Shard (0/{len(self.fout)})", total=None, unit="byte", unit_scale=True)
bar = tqdm(desc="Writing", total=total_bytes, unit="byte", unit_scale=True)
# relying on the fact that Python dicts preserve insertion order (since 3.7)
for ti in self.tensors.values():
assert ti.tensor is not None # can only iterate once over the tensors
assert ti.tensor.nbytes == ti.nbytes
ti.tensor.tofile(self.fout)
if bar is not None:
bar.update(ti.nbytes)
self.write_padding(self.fout, ti.nbytes)
ti.tensor = None
for i, (fout, tensors) in enumerate(zip(self.fout, self.tensors)):
if shard_bar is not None:
shard_bar.set_description(f"Shard ({i + 1}/{len(self.fout)})")
total = sum(ti.nbytes for ti in tensors.values())
shard_bar.reset(total=(total if total > 0 else None))
# relying on the fact that Python dicts preserve insertion order (since 3.7)
for ti in tensors.values():
assert ti.tensor is not None # can only iterate once over the tensors
assert ti.tensor.nbytes == ti.nbytes
ti.tensor.tofile(fout)
if shard_bar is not None:
shard_bar.update(ti.nbytes)
if bar is not None:
bar.update(ti.nbytes)
self.write_padding(fout, ti.nbytes)
ti.tensor = None
else:
self.temp_file.seek(0)
shutil.copyfileobj(self.temp_file, self.fout)
shutil.copyfileobj(self.temp_file, self.fout[0 if not self.small_first_shard else 1])
self.flush()
self.temp_file.close()
@@ -325,39 +466,27 @@ class GGUFWriter:
def flush(self) -> None:
assert self.fout is not None
self.fout.flush()
for fout in self.fout:
fout.flush()
def close(self) -> None:
if self.fout is not None:
self.fout.close()
for fout in self.fout:
fout.close()
self.fout = None
def add_type(self, type_name: str) -> None:
self.add_string(Keys.General.TYPE, type_name)
def add_architecture(self) -> None:
self.add_string(Keys.General.ARCHITECTURE, self.arch)
def add_author(self, author: str) -> None:
self.add_string(Keys.General.AUTHOR, author)
def add_quantization_version(self, quantization_version: int) -> None:
self.add_uint32(Keys.General.QUANTIZATION_VERSION, quantization_version)
def add_version(self, version: str) -> None:
self.add_string(Keys.General.VERSION, version)
def add_tensor_data_layout(self, layout: str) -> None:
self.add_string(Keys.LLM.TENSOR_DATA_LAYOUT.format(arch=self.arch), layout)
def add_url(self, url: str) -> None:
self.add_string(Keys.General.URL, url)
def add_description(self, description: str) -> None:
self.add_string(Keys.General.DESCRIPTION, description)
def add_licence(self, licence: str) -> None:
self.add_string(Keys.General.LICENSE, licence)
def add_source_url(self, url: str) -> None:
self.add_string(Keys.General.SOURCE_URL, url)
def add_source_hf_repo(self, repo: str) -> None:
self.add_string(Keys.General.SOURCE_HF_REPO, repo)
def add_custom_alignment(self, alignment: int) -> None:
self.data_alignment = alignment
self.add_uint32(Keys.General.ALIGNMENT, alignment)
def add_file_type(self, ftype: int) -> None:
self.add_uint32(Keys.General.FILE_TYPE, ftype)
@@ -365,13 +494,101 @@ class GGUFWriter:
def add_name(self, name: str) -> None:
self.add_string(Keys.General.NAME, name)
def add_quantization_version(self, quantization_version: int) -> None:
self.add_uint32(
Keys.General.QUANTIZATION_VERSION, quantization_version)
def add_author(self, author: str) -> None:
self.add_string(Keys.General.AUTHOR, author)
def add_custom_alignment(self, alignment: int) -> None:
self.data_alignment = alignment
self.add_uint32(Keys.General.ALIGNMENT, alignment)
def add_version(self, version: str) -> None:
self.add_string(Keys.General.VERSION, version)
def add_organization(self, organization: str) -> None:
self.add_string(Keys.General.ORGANIZATION, organization)
def add_finetune(self, finetune: str) -> None:
self.add_string(Keys.General.FINETUNE, finetune)
def add_basename(self, basename: str) -> None:
self.add_string(Keys.General.BASENAME, basename)
def add_description(self, description: str) -> None:
self.add_string(Keys.General.DESCRIPTION, description)
def add_quantized_by(self, quantized: str) -> None:
self.add_string(Keys.General.QUANTIZED_BY, quantized)
def add_size_label(self, size_label: str) -> None:
self.add_string(Keys.General.SIZE_LABEL, size_label)
def add_license(self, license: str) -> None:
self.add_string(Keys.General.LICENSE, license)
def add_license_name(self, license: str) -> None:
self.add_string(Keys.General.LICENSE_NAME, license)
def add_license_link(self, license: str) -> None:
self.add_string(Keys.General.LICENSE_LINK, license)
def add_url(self, url: str) -> None:
self.add_string(Keys.General.URL, url)
def add_doi(self, doi: str) -> None:
self.add_string(Keys.General.DOI, doi)
def add_uuid(self, uuid: str) -> None:
self.add_string(Keys.General.UUID, uuid)
def add_repo_url(self, repo_url: str) -> None:
self.add_string(Keys.General.REPO_URL, repo_url)
def add_source_url(self, url: str) -> None:
self.add_string(Keys.General.SOURCE_URL, url)
def add_source_doi(self, doi: str) -> None:
self.add_string(Keys.General.SOURCE_DOI, doi)
def add_source_uuid(self, uuid: str) -> None:
self.add_string(Keys.General.SOURCE_UUID, uuid)
def add_source_repo_url(self, repo_url: str) -> None:
self.add_string(Keys.General.SOURCE_REPO_URL, repo_url)
def add_base_model_count(self, source_count: int) -> None:
self.add_uint32(Keys.General.BASE_MODEL_COUNT, source_count)
def add_base_model_name(self, source_id: int, name: str) -> None:
self.add_string(Keys.General.BASE_MODEL_NAME.format(id=source_id), name)
def add_base_model_author(self, source_id: int, author: str) -> None:
self.add_string(Keys.General.BASE_MODEL_AUTHOR.format(id=source_id), author)
def add_base_model_version(self, source_id: int, version: str) -> None:
self.add_string(Keys.General.BASE_MODEL_VERSION.format(id=source_id), version)
def add_base_model_organization(self, source_id: int, organization: str) -> None:
self.add_string(Keys.General.BASE_MODEL_ORGANIZATION.format(id=source_id), organization)
def add_base_model_url(self, source_id: int, url: str) -> None:
self.add_string(Keys.General.BASE_MODEL_URL.format(id=source_id), url)
def add_base_model_doi(self, source_id: int, doi: str) -> None:
self.add_string(Keys.General.BASE_MODEL_DOI.format(id=source_id), doi)
def add_base_model_uuid(self, source_id: int, uuid: str) -> None:
self.add_string(Keys.General.BASE_MODEL_UUID.format(id=source_id), uuid)
def add_base_model_repo_url(self, source_id: int, repo_url: str) -> None:
self.add_string(Keys.General.BASE_MODEL_REPO_URL.format(id=source_id), repo_url)
def add_tags(self, tags: Sequence[str]) -> None:
self.add_array(Keys.General.TAGS, tags)
def add_languages(self, languages: Sequence[str]) -> None:
self.add_array(Keys.General.LANGUAGES, languages)
def add_datasets(self, datasets: Sequence[str]) -> None:
self.add_array(Keys.General.DATASETS, datasets)
def add_tensor_data_layout(self, layout: str) -> None:
self.add_string(Keys.LLM.TENSOR_DATA_LAYOUT.format(arch=self.arch), layout)
def add_vocab_size(self, size: int) -> None:
self.add_uint32(Keys.LLM.VOCAB_SIZE.format(arch=self.arch), size)
@@ -388,8 +605,11 @@ class GGUFWriter:
def add_leading_dense_block_count(self, length: int) -> None:
self.add_uint32(Keys.LLM.LEADING_DENSE_BLOCK_COUNT.format(arch=self.arch), length)
def add_feed_forward_length(self, length: int) -> None:
self.add_uint32(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length)
def add_feed_forward_length(self, length: int | Sequence[int]) -> None:
if isinstance(length, int):
self.add_uint32(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length)
else:
self.add_array(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length)
def add_expert_feed_forward_length(self, length: int) -> None:
self.add_uint32(Keys.LLM.EXPERT_FEED_FORWARD_LENGTH.format(arch=self.arch), length)
@@ -400,11 +620,20 @@ class GGUFWriter:
def add_parallel_residual(self, use: bool) -> None:
self.add_bool(Keys.LLM.USE_PARALLEL_RESIDUAL.format(arch=self.arch), use)
def add_head_count(self, count: int) -> None:
self.add_uint32(Keys.Attention.HEAD_COUNT.format(arch=self.arch), count)
def add_decoder_start_token_id(self, id: int) -> None:
self.add_uint32(Keys.LLM.DECODER_START_TOKEN_ID.format(arch=self.arch), id)
def add_head_count_kv(self, count: int) -> None:
self.add_uint32(Keys.Attention.HEAD_COUNT_KV.format(arch=self.arch), count)
def add_head_count(self, count: int | Sequence[int]) -> None:
if isinstance(count, int):
self.add_uint32(Keys.Attention.HEAD_COUNT.format(arch=self.arch), count)
else:
self.add_array(Keys.Attention.HEAD_COUNT.format(arch=self.arch), count)
def add_head_count_kv(self, count: int | Sequence[int]) -> None:
if isinstance(count, int):
self.add_uint32(Keys.Attention.HEAD_COUNT_KV.format(arch=self.arch), count)
else:
self.add_array(Keys.Attention.HEAD_COUNT_KV.format(arch=self.arch), count)
def add_key_length(self, length: int) -> None:
self.add_uint32(Keys.Attention.KEY_LENGTH.format(arch=self.arch), length)
@@ -421,6 +650,12 @@ class GGUFWriter:
def add_logit_scale(self, value: float) -> None:
self.add_float32(Keys.LLM.LOGIT_SCALE.format(arch=self.arch), value)
def add_attn_logit_softcapping(self, value: float) -> None:
self.add_float32(Keys.LLM.ATTN_LOGIT_SOFTCAPPING.format(arch=self.arch), value)
def add_final_logit_softcapping(self, value: float) -> None:
self.add_float32(Keys.LLM.FINAL_LOGIT_SOFTCAPPING.format(arch=self.arch), value)
def add_expert_count(self, count: int) -> None:
self.add_uint32(Keys.LLM.EXPERT_COUNT.format(arch=self.arch), count)
@@ -448,6 +683,12 @@ class GGUFWriter:
def add_kv_lora_rank(self, length: int) -> None:
self.add_uint32(Keys.Attention.KV_LORA_RANK.format(arch=self.arch), length)
def add_relative_attn_buckets_count(self, value: int) -> None:
self.add_uint32(Keys.Attention.REL_BUCKETS_COUNT.format(arch=self.arch), value)
def add_sliding_window(self, value: int) -> None:
self.add_uint32(Keys.Attention.SLIDING_WINDOW.format(arch=self.arch), value)
def add_pooling_type(self, value: PoolingType) -> None:
self.add_uint32(Keys.LLM.POOLING_TYPE.format(arch=self.arch), value.value)
@@ -538,6 +779,12 @@ class GGUFWriter:
def add_add_space_prefix(self, value: bool) -> None:
self.add_bool(Keys.Tokenizer.ADD_PREFIX, value)
def add_remove_extra_whitespaces(self, value: bool) -> None:
self.add_bool(Keys.Tokenizer.REMOVE_EXTRA_WS, value)
def add_precompiled_charsmap(self, charsmap: Sequence[bytes]) -> None:
self.add_array(Keys.Tokenizer.PRECOMPILED_CHARSMAP, charsmap)
def add_chat_template(self, value: str | Sequence[Mapping[str, str]]) -> None:
if not isinstance(value, str):
template_default = None
@@ -599,9 +846,12 @@ class GGUFWriter:
kv_data += self._pack("Q", len(encoded_val))
kv_data += encoded_val
elif vtype == GGUFValueType.ARRAY and isinstance(val, Sequence) and val:
ltype = GGUFValueType.get_type(val[0])
if not all(GGUFValueType.get_type(i) is ltype for i in val[1:]):
raise ValueError("All items in a GGUF array should be of the same type")
if isinstance(val, bytes):
ltype = GGUFValueType.UINT8
else:
ltype = GGUFValueType.get_type(val[0])
if not all(GGUFValueType.get_type(i) is ltype for i in val[1:]):
raise ValueError("All items in a GGUF array should be of the same type")
kv_data += self._pack("I", ltype)
kv_data += self._pack("Q", len(val))
for item in val:
@@ -611,6 +861,13 @@ class GGUFWriter:
return kv_data
def _write_packed(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> None:
assert self.fout is not None
self.fout.write(self._pack(fmt, value, skip_pack_prefix))
@staticmethod
def format_n_bytes_to_str(num: int) -> str:
if num == 0:
return "negligible - metadata only"
fnum = float(num)
for unit in ("", "K", "M", "G"):
if abs(fnum) < 1000.0:
return f"{fnum:3.1f}{unit}"
fnum /= 1000.0
return f"{fnum:.1f}T - over 1TB, split recommended"

71
gguf-py/gguf/lazy.py
View File

@@ -3,10 +3,8 @@ from abc import ABC, ABCMeta, abstractmethod
import logging
from typing import Any, Callable
from collections import deque
import numpy as np
from numpy._typing import _Shape
from numpy.typing import DTypeLike
@@ -16,16 +14,16 @@ logger = logging.getLogger(__name__)
class LazyMeta(ABCMeta):
def __new__(cls, name: str, bases: tuple[type, ...], namespace: dict[str, Any], **kwargs):
def __getattr__(self, __name: str) -> Any:
meta_attr = getattr(self._meta, __name)
def __getattr__(self, name: str) -> Any:
meta_attr = getattr(self._meta, name)
if callable(meta_attr):
return type(self)._wrap_fn(
(lambda s, *args, **kwargs: getattr(s, __name)(*args, **kwargs)),
(lambda s, *args, **kwargs: getattr(s, name)(*args, **kwargs)),
use_self=self,
)
elif isinstance(meta_attr, self._tensor_type):
# e.g. self.T with torch.Tensor should still be wrapped
return type(self)._wrap_fn(lambda s: getattr(s, __name))(self)
return type(self)._wrap_fn(lambda s: getattr(s, name))(self)
else:
# no need to wrap non-tensor properties,
# and they likely don't depend on the actual contents of the tensor
@@ -75,20 +73,18 @@ class LazyBase(ABC, metaclass=LazyMeta):
_tensor_type: type
_meta: Any
_data: Any | None
_lazy: deque[LazyBase] # shared within a graph, to avoid deep recursion when making eager
_args: tuple
_func: Callable[[tuple], Any] | None
_kwargs: dict[str, Any]
_func: Callable[[Any], Any] | None
def __init__(self, *, meta: Any, data: Any | None = None, lazy: deque[LazyBase] | None = None, args: tuple = (), func: Callable[[tuple], Any] | None = None):
def __init__(self, *, meta: Any, data: Any | None = None, args: tuple = (), kwargs: dict[str, Any] | None = None, func: Callable[[Any], Any] | None = None):
super().__init__()
self._meta = meta
self._data = data
self._lazy = lazy if lazy is not None else deque()
self._args = args
self._kwargs = kwargs if kwargs is not None else {}
self._func = func
assert self._func is not None or self._data is not None
if self._data is None:
self._lazy.append(self)
def __init_subclass__(cls) -> None:
if "_tensor_type" not in cls.__dict__:
@@ -118,6 +114,7 @@ class LazyBase(ABC, metaclass=LazyMeta):
args = ((use_self,) if use_self is not None else ()) + args
meta_args = LazyBase._recurse_apply(args, lambda t: t._meta)
# TODO: maybe handle tensors in kwargs too
if isinstance(meta_noop, bool) and not meta_noop:
try:
@@ -141,21 +138,7 @@ class LazyBase(ABC, metaclass=LazyMeta):
res = cls.meta_with_dtype_and_shape(meta_noop, res.shape)
if isinstance(res, cls._tensor_type):
def collect_replace(t: LazyBase):
if collect_replace.shared_lazy is None:
collect_replace.shared_lazy = t._lazy
else:
collect_replace.shared_lazy.extend(t._lazy)
t._lazy = collect_replace.shared_lazy
# emulating a static variable
collect_replace.shared_lazy = None
LazyBase._recurse_apply(args, collect_replace)
shared_lazy = collect_replace.shared_lazy
return cls(meta=cls.eager_to_meta(res), lazy=shared_lazy, args=args, func=lambda a: fn(*a, **kwargs))
return cls(meta=cls.eager_to_meta(res), args=args, kwargs=kwargs, func=fn)
else:
del res # not needed
# non-tensor return likely relies on the contents of the args
@@ -167,25 +150,18 @@ class LazyBase(ABC, metaclass=LazyMeta):
@classmethod
def to_eager(cls, t: Any) -> Any:
def simple_to_eager(_t: LazyBase) -> Any:
def already_eager_to_eager(_t: LazyBase) -> Any:
assert _t._data is not None
if _t._data is not None:
return _t._data
while _t._data is None:
lt = _t._lazy.popleft()
if lt._data is not None:
# Lazy tensor did not belong in the lazy queue.
# Weirdly only happens with Bloom models...
# likely because tensors aren't unique in the queue.
# The final output is still the same as in eager mode,
# so it's safe to ignore this.
continue
assert lt._func is not None
lt._args = cls._recurse_apply(lt._args, already_eager_to_eager)
lt._data = lt._func(lt._args)
# sanity check
assert lt._data.dtype == lt._meta.dtype
assert lt._data.shape == lt._meta.shape
# NOTE: there's a recursion limit in Python (usually 1000)
assert _t._func is not None
_t._args = cls._recurse_apply(_t._args, simple_to_eager)
_t._data = _t._func(*_t._args, **_t._kwargs)
# sanity check
assert _t._data is not None
assert _t._data.dtype == _t._meta.dtype
assert _t._data.shape == _t._meta.shape
return _t._data
@@ -204,7 +180,7 @@ class LazyBase(ABC, metaclass=LazyMeta):
@classmethod
def from_eager(cls, t: Any) -> Any:
if type(t) is cls:
# already eager
# already lazy
return t
elif isinstance(t, cls._tensor_type):
return cls(meta=cls.eager_to_meta(t), data=t)
@@ -216,7 +192,7 @@ class LazyNumpyTensor(LazyBase):
_tensor_type = np.ndarray
@classmethod
def meta_with_dtype_and_shape(cls, dtype: DTypeLike, shape: _Shape) -> np.ndarray[Any, Any]:
def meta_with_dtype_and_shape(cls, dtype: DTypeLike, shape: tuple[int, ...]) -> np.ndarray[Any, Any]:
# The initial idea was to use np.nan as the fill value,
# but non-float types like np.int16 can't use that.
# So zero it is.
@@ -226,8 +202,7 @@ class LazyNumpyTensor(LazyBase):
def astype(self, dtype, *args, **kwargs):
meta = type(self).meta_with_dtype_and_shape(dtype, self._meta.shape)
full_args = (self, dtype,) + args
# very important to pass the shared _lazy deque, or else there's an infinite loop somewhere.
return type(self)(meta=meta, args=full_args, lazy=self._lazy, func=(lambda a: a[0].astype(*a[1:], **kwargs)))
return type(self)(meta=meta, args=full_args, kwargs=kwargs, func=(lambda a, *args, **kwargs: a.astype(*args, **kwargs)))
def tofile(self, *args, **kwargs):
eager = LazyNumpyTensor.to_eager(self)

503
gguf-py/gguf/metadata.py Normal file
View File

@@ -0,0 +1,503 @@
from __future__ import annotations
import re
import json
import yaml
import logging
from pathlib import Path
from typing import Any, Literal, Optional
from dataclasses import dataclass
from .constants import Keys
import gguf
logger = logging.getLogger("metadata")
@dataclass
class Metadata:
# Authorship Metadata to be written to GGUF KV Store
name: Optional[str] = None
author: Optional[str] = None
version: Optional[str] = None
organization: Optional[str] = None
finetune: Optional[str] = None
basename: Optional[str] = None
description: Optional[str] = None
quantized_by: Optional[str] = None
size_label: Optional[str] = None
url: Optional[str] = None
doi: Optional[str] = None
uuid: Optional[str] = None
repo_url: Optional[str] = None
source_url: Optional[str] = None
source_doi: Optional[str] = None
source_uuid: Optional[str] = None
source_repo_url: Optional[str] = None
license: Optional[str] = None
license_name: Optional[str] = None
license_link: Optional[str] = None
base_models: Optional[list[dict]] = None
tags: Optional[list[str]] = None
languages: Optional[list[str]] = None
datasets: Optional[list[str]] = None
@staticmethod
def load(metadata_override_path: Optional[Path] = None, model_path: Optional[Path] = None, model_name: Optional[str] = None, total_params: int = 0) -> Metadata:
# This grabs as many contextual authorship metadata as possible from the model repository
# making any conversion as required to match the gguf kv store metadata format
# as well as giving users the ability to override any authorship metadata that may be incorrect
# Create a new Metadata instance
metadata = Metadata()
model_card = Metadata.load_model_card(model_path)
hf_params = Metadata.load_hf_parameters(model_path)
# TODO: load adapter_config.json when possible, it usually contains the base model of the LoRA adapter
# heuristics
metadata = Metadata.apply_metadata_heuristic(metadata, model_card, hf_params, model_path, total_params)
# Metadata Override File Provided
# This is based on LLM_KV_NAMES mapping in llama.cpp
metadata_override = Metadata.load_metadata_override(metadata_override_path)
metadata.name = metadata_override.get(Keys.General.NAME, metadata.name)
metadata.author = metadata_override.get(Keys.General.AUTHOR, metadata.author)
metadata.version = metadata_override.get(Keys.General.VERSION, metadata.version)
metadata.organization = metadata_override.get(Keys.General.ORGANIZATION, metadata.organization)
metadata.finetune = metadata_override.get(Keys.General.FINETUNE, metadata.finetune)
metadata.basename = metadata_override.get(Keys.General.BASENAME, metadata.basename)
metadata.description = metadata_override.get(Keys.General.DESCRIPTION, metadata.description)
metadata.quantized_by = metadata_override.get(Keys.General.QUANTIZED_BY, metadata.quantized_by)
metadata.size_label = metadata_override.get(Keys.General.SIZE_LABEL, metadata.size_label)
metadata.license_name = metadata_override.get(Keys.General.LICENSE_NAME, metadata.license_name)
metadata.license_link = metadata_override.get(Keys.General.LICENSE_LINK, metadata.license_link)
metadata.url = metadata_override.get(Keys.General.URL, metadata.url)
metadata.doi = metadata_override.get(Keys.General.DOI, metadata.doi)
metadata.uuid = metadata_override.get(Keys.General.UUID, metadata.uuid)
metadata.repo_url = metadata_override.get(Keys.General.REPO_URL, metadata.repo_url)
metadata.source_url = metadata_override.get(Keys.General.SOURCE_URL, metadata.source_url)
metadata.source_doi = metadata_override.get(Keys.General.SOURCE_DOI, metadata.source_doi)
metadata.source_uuid = metadata_override.get(Keys.General.SOURCE_UUID, metadata.source_uuid)
metadata.source_repo_url = metadata_override.get(Keys.General.SOURCE_REPO_URL, metadata.source_repo_url)
# Base Models is received here as an array of models
metadata.base_models = metadata_override.get("general.base_models", metadata.base_models)
metadata.tags = metadata_override.get(Keys.General.TAGS, metadata.tags)
metadata.languages = metadata_override.get(Keys.General.LANGUAGES, metadata.languages)
metadata.datasets = metadata_override.get(Keys.General.DATASETS, metadata.datasets)
# Direct Metadata Override (via direct cli argument)
if model_name is not None:
metadata.name = model_name
return metadata
@staticmethod
def load_metadata_override(metadata_override_path: Optional[Path] = None) -> dict[str, Any]:
if metadata_override_path is None or not metadata_override_path.is_file():
return {}
with open(metadata_override_path, "r", encoding="utf-8") as f:
return json.load(f)
@staticmethod
def load_model_card(model_path: Optional[Path] = None) -> dict[str, Any]:
if model_path is None or not model_path.is_dir():
return {}
model_card_path = model_path / "README.md"
if not model_card_path.is_file():
return {}
# The model card metadata is assumed to always be in YAML
# ref: https://github.com/huggingface/transformers/blob/a5c642fe7a1f25d3bdcd76991443ba6ff7ee34b2/src/transformers/modelcard.py#L468-L473
with open(model_card_path, "r", encoding="utf-8") as f:
if f.readline() == "---\n":
raw = f.read().partition("---\n")[0]
data = yaml.safe_load(raw)
if isinstance(data, dict):
return data
else:
logger.error(f"while reading YAML model card frontmatter, data is {type(data)} instead of dict")
return {}
else:
return {}
@staticmethod
def load_hf_parameters(model_path: Optional[Path] = None) -> dict[str, Any]:
if model_path is None or not model_path.is_dir():
return {}
config_path = model_path / "config.json"
if not config_path.is_file():
return {}
with open(config_path, "r", encoding="utf-8") as f:
return json.load(f)
@staticmethod
def id_to_title(string):
# Convert capitalization into title form unless acronym or version number
return ' '.join([w.title() if w.islower() and not re.match(r'^(v\d+(?:\.\d+)*|\d.*)$', w) else w for w in string.strip().replace('-', ' ').split()])
@staticmethod
def get_model_id_components(model_id: Optional[str] = None, total_params: int = 0) -> tuple[str | None, str | None, str | None, str | None, str | None, str | None]:
# Huggingface often store model id as '<org>/<model name>'
# so let's parse it and apply some heuristics if possible for model name components
if model_id is None:
# model ID missing
return None, None, None, None, None, None
if ' ' in model_id:
# model ID is actually a normal human sentence
# which means its most likely a normal model name only
# not part of the hugging face naming standard, but whatever
return model_id, None, None, None, None, None
if '/' in model_id:
# model ID (huggingface style)
org_component, model_full_name_component = model_id.split('/', 1)
else:
# model ID but missing org components
org_component, model_full_name_component = None, model_id
# Check if we erroneously matched against './' or '../' etc...
if org_component is not None and org_component[0] == '.':
org_component = None
name_parts: list[str] = model_full_name_component.split('-')
# Remove empty parts
for i in reversed(range(len(name_parts))):
if len(name_parts[i]) == 0:
del name_parts[i]
name_types: list[
set[Literal["basename", "size_label", "finetune", "version", "type"]]
] = [set() for _ in name_parts]
# Annotate the name
for i, part in enumerate(name_parts):
# Version
if re.fullmatch(r'(v|iter)?\d+([.]\d+)*', part, re.IGNORECASE):
name_types[i].add("version")
# Quant type (should not be there for base models, but still annotated)
elif re.fullmatch(r'i?q\d(_\w)*|b?fp?(16|32)', part, re.IGNORECASE):
name_types[i].add("type")
name_parts[i] = part.upper()
# Model size
elif i > 0 and re.fullmatch(r'(([A]|\d+[x])?\d+([._]\d+)?[KMBT][\d]?|small|mini|medium|large|x?xl)', part, re.IGNORECASE):
part = part.replace("_", ".")
# Handle weird bloom-7b1 notation
if part[-1].isdecimal():
part = part[:-2] + "." + part[-1] + part[-2]
# Normalize the size suffixes
if len(part) > 1 and part[-2].isdecimal():
if part[-1] in "kmbt":
part = part[:-1] + part[-1].upper()
if total_params != 0:
try:
label_params = float(part[:-1]) * pow(1000, " KMBT".find(part[-1]))
# Only use it as a size label if it's close or bigger than the model size
# Note that LoRA adapters don't necessarily include all layers,
# so this is why bigger label sizes are accepted.
# Do not use the size label when it's smaller than 1/8 of the model size
if (total_params < 0 and label_params < abs(total_params) // 8) or (
# Check both directions when the current model isn't a LoRA adapter
total_params > 0 and abs(label_params - total_params) > 7 * total_params // 8
):
# Likely a context length
name_types[i].add("finetune")
# Lowercase the size when it's a context length
part = part[:-1] + part[-1].lower()
except ValueError:
# Failed to convert the size label to float, use it anyway
pass
if len(name_types[i]) == 0:
name_types[i].add("size_label")
name_parts[i] = part
# Some easy to recognize finetune names
elif i > 0 and re.fullmatch(r'chat|instruct|vision|lora', part, re.IGNORECASE):
if total_params < 0 and part.lower() == "lora":
# ignore redundant "lora" in the finetune part when the output is a lora adapter
name_types[i].add("type")
else:
name_types[i].add("finetune")
# Ignore word-based size labels when there is at least a number-based one present
# TODO: should word-based size labels always be removed instead?
if any(c.isdecimal() for n, t in zip(name_parts, name_types) if "size_label" in t for c in n):
for n, t in zip(name_parts, name_types):
if "size_label" in t:
if all(c.isalpha() for c in n):
t.remove("size_label")
at_start = True
# Find the basename through the annotated name
for part, t in zip(name_parts, name_types):
if at_start and ((len(t) == 0 and part[0].isalpha()) or "version" in t):
t.add("basename")
else:
if at_start:
at_start = False
if len(t) == 0:
t.add("finetune")
# Remove the basename annotation from trailing version
for part, t in zip(reversed(name_parts), reversed(name_types)):
if "basename" in t and len(t) > 1:
t.remove("basename")
else:
break
basename = "-".join(n for n, t in zip(name_parts, name_types) if "basename" in t) or None
# Deduplicate size labels using order-preserving 'dict' ('set' seems to sort the keys)
size_label = "-".join(dict.fromkeys(s for s, t in zip(name_parts, name_types) if "size_label" in t).keys()) or None
finetune = "-".join(f for f, t in zip(name_parts, name_types) if "finetune" in t) or None
# TODO: should the basename version always be excluded?
# NOTE: multiple finetune versions are joined together
version = "-".join(v for v, t, in zip(name_parts, name_types) if "version" in t and "basename" not in t) or None
if size_label is None and finetune is None and version is None:
# Too ambiguous, output nothing
basename = None
return model_full_name_component, org_component, basename, finetune, version, size_label
@staticmethod
def apply_metadata_heuristic(metadata: Metadata, model_card: Optional[dict] = None, hf_params: Optional[dict] = None, model_path: Optional[Path] = None, total_params: int = 0) -> Metadata:
# Reference Model Card Metadata: https://github.com/huggingface/hub-docs/blob/main/modelcard.md?plain=1
# Model Card Heuristics
########################
if model_card is not None:
if "model_name" in model_card and metadata.name is None:
# Not part of huggingface model card standard but notice some model creator using it
# such as TheBloke in 'TheBloke/Mistral-7B-Instruct-v0.2-GGUF'
metadata.name = model_card.get("model_name")
if "model_creator" in model_card and metadata.author is None:
# Not part of huggingface model card standard but notice some model creator using it
# such as TheBloke in 'TheBloke/Mistral-7B-Instruct-v0.2-GGUF'
metadata.author = model_card.get("model_creator")
if "model_type" in model_card and metadata.basename is None:
# Not part of huggingface model card standard but notice some model creator using it
# such as TheBloke in 'TheBloke/Mistral-7B-Instruct-v0.2-GGUF'
metadata.basename = model_card.get("model_type")
if "base_model" in model_card:
# This represents the parent models that this is based on
# Example: stabilityai/stable-diffusion-xl-base-1.0. Can also be a list (for merges)
# Example of merges: https://huggingface.co/EmbeddedLLM/Mistral-7B-Merge-14-v0.1/blob/main/README.md
metadata_base_models = []
base_model_value = model_card.get("base_model", None)
if base_model_value is not None:
if isinstance(base_model_value, str):
metadata_base_models.append(base_model_value)
elif isinstance(base_model_value, list):
metadata_base_models.extend(base_model_value)
if metadata.base_models is None:
metadata.base_models = []
for model_id in metadata_base_models:
# NOTE: model size of base model is assumed to be similar to the size of the current model
model_full_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(model_id, total_params)
base_model = {}
if model_full_name_component is not None:
base_model["name"] = Metadata.id_to_title(model_full_name_component)
if org_component is not None:
base_model["organization"] = Metadata.id_to_title(org_component)
if version is not None:
base_model["version"] = version
if org_component is not None and model_full_name_component is not None:
base_model["repo_url"] = f"https://huggingface.co/{org_component}/{model_full_name_component}"
metadata.base_models.append(base_model)
if "license" in model_card and metadata.license is None:
metadata.license = model_card.get("license")
if "license_name" in model_card and metadata.license_name is None:
metadata.license_name = model_card.get("license_name")
if "license_link" in model_card and metadata.license_link is None:
metadata.license_link = model_card.get("license_link")
tags_value = model_card.get("tags", None)
if tags_value is not None:
if metadata.tags is None:
metadata.tags = []
if isinstance(tags_value, str):
metadata.tags.append(tags_value)
elif isinstance(tags_value, list):
metadata.tags.extend(tags_value)
pipeline_tags_value = model_card.get("pipeline_tag", None)
if pipeline_tags_value is not None:
if metadata.tags is None:
metadata.tags = []
if isinstance(pipeline_tags_value, str):
metadata.tags.append(pipeline_tags_value)
elif isinstance(pipeline_tags_value, list):
metadata.tags.extend(pipeline_tags_value)
language_value = model_card.get("languages", model_card.get("language", None))
if language_value is not None:
if metadata.languages is None:
metadata.languages = []
if isinstance(language_value, str):
metadata.languages.append(language_value)
elif isinstance(language_value, list):
metadata.languages.extend(language_value)
dataset_value = model_card.get("datasets", model_card.get("dataset", None))
if dataset_value is not None:
if metadata.datasets is None:
metadata.datasets = []
if isinstance(dataset_value, str):
metadata.datasets.append(dataset_value)
elif isinstance(dataset_value, list):
metadata.datasets.extend(dataset_value)
# Hugging Face Parameter Heuristics
####################################
if hf_params is not None:
hf_name_or_path = hf_params.get("_name_or_path")
if hf_name_or_path is not None and hf_name_or_path.count('/') <= 1:
# Use _name_or_path only if its actually a model name and not some computer path
# e.g. 'meta-llama/Llama-2-7b-hf'
model_id = hf_name_or_path
model_full_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(model_id, total_params)
if metadata.name is None and model_full_name_component is not None:
metadata.name = Metadata.id_to_title(model_full_name_component)
if metadata.organization is None and org_component is not None:
metadata.organization = Metadata.id_to_title(org_component)
if metadata.basename is None and basename is not None:
metadata.basename = basename
if metadata.finetune is None and finetune is not None:
metadata.finetune = finetune
if metadata.version is None and version is not None:
metadata.version = version
if metadata.size_label is None and size_label is not None:
metadata.size_label = size_label
# Directory Folder Name Fallback Heuristics
############################################
if model_path is not None:
model_id = model_path.name
model_full_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(model_id, total_params)
if metadata.name is None and model_full_name_component is not None:
metadata.name = Metadata.id_to_title(model_full_name_component)
if metadata.organization is None and org_component is not None:
metadata.organization = Metadata.id_to_title(org_component)
if metadata.basename is None and basename is not None:
metadata.basename = basename
if metadata.finetune is None and finetune is not None:
metadata.finetune = finetune
if metadata.version is None and version is not None:
metadata.version = version
if metadata.size_label is None and size_label is not None:
metadata.size_label = size_label
return metadata
def set_gguf_meta_model(self, gguf_writer: gguf.GGUFWriter):
assert self.name is not None
gguf_writer.add_name(self.name)
if self.author is not None:
gguf_writer.add_author(self.author)
if self.version is not None:
gguf_writer.add_version(self.version)
if self.organization is not None:
gguf_writer.add_organization(self.organization)
if self.finetune is not None:
gguf_writer.add_finetune(self.finetune)
if self.basename is not None:
gguf_writer.add_basename(self.basename)
if self.description is not None:
gguf_writer.add_description(self.description)
if self.quantized_by is not None:
gguf_writer.add_quantized_by(self.quantized_by)
if self.size_label is not None:
gguf_writer.add_size_label(self.size_label)
if self.license is not None:
gguf_writer.add_license(self.license)
if self.license_name is not None:
gguf_writer.add_license_name(self.license_name)
if self.license_link is not None:
gguf_writer.add_license_link(self.license_link)
if self.url is not None:
gguf_writer.add_url(self.url)
if self.doi is not None:
gguf_writer.add_doi(self.doi)
if self.uuid is not None:
gguf_writer.add_uuid(self.uuid)
if self.repo_url is not None:
gguf_writer.add_repo_url(self.repo_url)
if self.source_url is not None:
gguf_writer.add_source_url(self.source_url)
if self.source_doi is not None:
gguf_writer.add_source_doi(self.source_doi)
if self.source_uuid is not None:
gguf_writer.add_source_uuid(self.source_uuid)
if self.source_repo_url is not None:
gguf_writer.add_source_repo_url(self.source_repo_url)
if self.base_models is not None:
gguf_writer.add_base_model_count(len(self.base_models))
for key, base_model_entry in enumerate(self.base_models):
if "name" in base_model_entry:
gguf_writer.add_base_model_name(key, base_model_entry["name"])
if "author" in base_model_entry:
gguf_writer.add_base_model_author(key, base_model_entry["author"])
if "version" in base_model_entry:
gguf_writer.add_base_model_version(key, base_model_entry["version"])
if "organization" in base_model_entry:
gguf_writer.add_base_model_organization(key, base_model_entry["organization"])
if "url" in base_model_entry:
gguf_writer.add_base_model_url(key, base_model_entry["url"])
if "doi" in base_model_entry:
gguf_writer.add_base_model_doi(key, base_model_entry["doi"])
if "uuid" in base_model_entry:
gguf_writer.add_base_model_uuid(key, base_model_entry["uuid"])
if "repo_url" in base_model_entry:
gguf_writer.add_base_model_repo_url(key, base_model_entry["repo_url"])
if self.tags is not None:
gguf_writer.add_tags(self.tags)
if self.languages is not None:
gguf_writer.add_languages(self.languages)
if self.datasets is not None:
gguf_writer.add_datasets(self.datasets)

2
gguf-py/gguf/quants.py
View File

@@ -43,7 +43,7 @@ def __apply_over_grouped_rows(func: Callable[[np.ndarray], np.ndarray], arr: np.
osize *= dim
out = np.empty(shape=osize, dtype=otype)
# compute over groups of 16 rows (arbitrary, but seems good for performance)
n_groups = rows.shape[0] // 16
n_groups = (rows.shape[0] // 16) or 1
np.concatenate([func(group).ravel() for group in np.array_split(rows, n_groups)], axis=0, out=out)
return out.reshape(oshape)

192
gguf-py/gguf/tensor_mapping.py
View File

@@ -10,7 +10,7 @@ class TensorNameMap:
# Token embeddings
MODEL_TENSOR.TOKEN_EMBD: (
"gpt_neox.embed_in", # gptneox
"transformer.wte", # gpt2 gpt-j mpt refact qwen dbrx
"transformer.wte", # gpt2 gpt-j mpt refact qwen dbrx jais
"transformer.word_embeddings", # falcon
"word_embeddings", # bloom
"model.embed_tokens", # llama-hf
@@ -24,6 +24,9 @@ class TensorNameMap:
"backbone.embedding", # mamba
"backbone.embeddings", # mamba-hf
"transformer.in_out_embed", # Grok
"embedding.word_embeddings", # chatglm
"transformer.token_embeddings", # openelm
"shared", # t5
),
# Token type embeddings
@@ -36,6 +39,7 @@ class TensorNameMap:
"word_embeddings_layernorm", # bloom
"embeddings.LayerNorm", # bert
"emb_ln", # nomic-bert
"transformer.norm", # openelm
),
# Position embeddings
@@ -48,16 +52,17 @@ class TensorNameMap:
# Output
MODEL_TENSOR.OUTPUT: (
"embed_out", # gptneox
"lm_head", # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx
"lm_head", # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais
"output", # llama-pth bloom internlm2
"word_embeddings_for_head", # persimmon
"lm_head.linear", # phi2
"output_layer", # chatglm
),
# Output norm
MODEL_TENSOR.OUTPUT_NORM: (
"gpt_neox.final_layer_norm", # gptneox
"transformer.ln_f", # gpt2 gpt-j falcon
"transformer.ln_f", # gpt2 gpt-j falcon jais
"model.norm", # llama-hf baichuan internlm2
"norm", # llama-pth
"transformer.norm_f", # mpt dbrx
@@ -68,11 +73,14 @@ class TensorNameMap:
"model.norm_f", # mamba-qbert
"backbone.norm_f", # mamba
"transformer.rms_norm", # Grok
"encoder.final_layernorm", # chatglm
"transformer.norm", # openelm
),
# Rope frequencies
MODEL_TENSOR.ROPE_FREQS: (
"rope.freqs", # llama-pth
"rotary_pos_emb.inv_freq", # chatglm
),
}
@@ -80,7 +88,7 @@ class TensorNameMap:
# Attention norm
MODEL_TENSOR.ATTN_NORM: (
"gpt_neox.layers.{bid}.input_layernorm", # gptneox
"transformer.h.{bid}.ln_1", # gpt2 gpt-j refact qwen
"transformer.h.{bid}.ln_1", # gpt2 gpt-j refact qwen jais
"transformer.blocks.{bid}.norm_1", # mpt
"transformer.h.{bid}.input_layernorm", # falcon7b
"h.{bid}.input_layernorm", # bloom
@@ -97,6 +105,8 @@ class TensorNameMap:
"backbone.layers.{bid}.norm", # mamba
"transformer.decoder_layer.{bid}.rms_norm", # Grok
"transformer.blocks.{bid}.norm_attn_norm.norm_1", # dbrx
"encoder.layers.{bid}.input_layernorm", # chatglm
"transformer.layers.{bid}.attn_norm", # openelm
),
# Attention norm 2
@@ -108,7 +118,7 @@ class TensorNameMap:
# Attention query-key-value
MODEL_TENSOR.ATTN_QKV: (
"gpt_neox.layers.{bid}.attention.query_key_value", # gptneox
"transformer.h.{bid}.attn.c_attn", # gpt2 qwen
"transformer.h.{bid}.attn.c_attn", # gpt2 qwen jais
"transformer.blocks.{bid}.attn.Wqkv", # mpt
"transformer.blocks.{bid}.norm_attn_norm.attn.Wqkv", # dbrx
"transformer.h.{bid}.self_attention.query_key_value", # falcon
@@ -118,7 +128,9 @@ class TensorNameMap:
"h.{bid}.attn.c_attn", # gpt2
"transformer.h.{bid}.mixer.Wqkv", # phi2
"encoder.layers.{bid}.attn.Wqkv", # nomic-bert
"model.layers.{bid}.self_attn.qkv_proj" # phi3
"model.layers.{bid}.self_attn.qkv_proj", # phi3
"encoder.layers.{bid}.self_attention.query_key_value", # chatglm
"transformer.layers.{bid}.attn.qkv_proj", # openelm
),
# Attention query
@@ -129,7 +141,7 @@ class TensorNameMap:
"transformer.h.{bid}.attn.q_proj", # gpt-j
"model.layers.layers.{bid}.self_attn.q_proj", # plamo
"model.layers.{bid}.attention.wq", # internlm2
"transformer.decoder_layer.{bid}.multi_head_attention.query" # Grok
"transformer.decoder_layer.{bid}.multi_head_attention.query",# Grok
),
# Attention key
@@ -141,7 +153,7 @@ class TensorNameMap:
"transformer.h.{bid}.attn.k", # refact
"model.layers.layers.{bid}.self_attn.k_proj", # plamo
"model.layers.{bid}.attention.wk", # internlm2
"transformer.decoder_layer.{bid}.multi_head_attention.key" # Grok
"transformer.decoder_layer.{bid}.multi_head_attention.key",# Grok
),
# Attention value
@@ -159,7 +171,7 @@ class TensorNameMap:
# Attention output
MODEL_TENSOR.ATTN_OUT: (
"gpt_neox.layers.{bid}.attention.dense", # gptneox
"transformer.h.{bid}.attn.c_proj", # gpt2 refact qwen
"transformer.h.{bid}.attn.c_proj", # gpt2 refact qwen jais
"transformer.blocks.{bid}.attn.out_proj", # mpt
"transformer.h.{bid}.self_attention.dense", # falcon
"h.{bid}.self_attention.dense", # bloom
@@ -176,6 +188,8 @@ class TensorNameMap:
"encoder.layers.{bid}.attn.out_proj", # nomic-bert
"transformer.decoder_layer.{bid}.multi_head_attention.linear", # Grok
"transformer.blocks.{bid}.norm_attn_norm.attn.out_proj", # dbrx
"encoder.layers.{bid}.self_attention.dense", # chatglm
"transformer.layers.{bid}.attn.out_proj", # openelm
),
# Attention output norm
@@ -186,6 +200,10 @@ class TensorNameMap:
"transformer.blocks.{bid}.norm_attn_norm.norm_2", # dbrx
),
MODEL_TENSOR.ATTN_POST_NORM: (
"model.layers.{bid}.post_attention_layernorm", # gemma2
),
# Rotary embeddings
MODEL_TENSOR.ATTN_ROT_EMBD: (
"model.layers.{bid}.self_attn.rotary_emb.inv_freq", # llama-hf
@@ -197,7 +215,7 @@ class TensorNameMap:
# Feed-forward norm
MODEL_TENSOR.FFN_NORM: (
"gpt_neox.layers.{bid}.post_attention_layernorm", # gptneox
"transformer.h.{bid}.ln_2", # gpt2 refact qwen
"transformer.h.{bid}.ln_2", # gpt2 refact qwen jais
"h.{bid}.post_attention_layernorm", # bloom
"transformer.blocks.{bid}.norm_2", # mpt
"model.layers.{bid}.post_attention_layernorm", # llama-hf
@@ -207,6 +225,18 @@ class TensorNameMap:
"h.{bid}.ln_2", # gpt2
"model.layers.{bid}.ffn_norm", # internlm2
"transformer.decoder_layer.{bid}.rms_norm_2", # Grok
"encoder.layers.{bid}.post_attention_layernorm", # chatglm
"transformer.layers.{bid}.ffn_norm", # openelm
),
# Post feed-forward norm
MODEL_TENSOR.FFN_PRE_NORM: (
"model.layers.{bid}.pre_feedforward_layernorm", # gemma2
),
# Post feed-forward norm
MODEL_TENSOR.FFN_POST_NORM: (
"model.layers.{bid}.post_feedforward_layernorm", # gemma2
),
MODEL_TENSOR.FFN_GATE_INP: (
@@ -224,7 +254,7 @@ class TensorNameMap:
# Feed-forward up
MODEL_TENSOR.FFN_UP: (
"gpt_neox.layers.{bid}.mlp.dense_h_to_4h", # gptneox
"transformer.h.{bid}.mlp.c_fc", # gpt2
"transformer.h.{bid}.mlp.c_fc", # gpt2 jais
"transformer.blocks.{bid}.ffn.up_proj", # mpt
"transformer.h.{bid}.mlp.dense_h_to_4h", # falcon
"h.{bid}.mlp.dense_h_to_4h", # bloom
@@ -246,6 +276,7 @@ class TensorNameMap:
"model.layers.{bid}.mlp.c_fc", # starcoder2
"encoder.layer.{bid}.mlp.gated_layers_v", # jina-bert-v2
"model.layers.{bid}.residual_mlp.w3", # arctic
"encoder.layers.{bid}.mlp.dense_h_to_4h", # chatglm
),
MODEL_TENSOR.FFN_UP_EXP: (
@@ -270,6 +301,7 @@ class TensorNameMap:
"model.layers.{bid}.mlp.gate_proj", # llama-hf refact
"layers.{bid}.feed_forward.w1", # llama-pth
"transformer.h.{bid}.mlp.w2", # qwen
"transformer.h.{bid}.mlp.c_fc2", # jais
"model.layers.layers.{bid}.mlp.gate_proj", # plamo
"model.layers.{bid}.feed_forward.w1", # internlm2
"encoder.layers.{bid}.mlp.fc12", # nomic-bert
@@ -293,7 +325,7 @@ class TensorNameMap:
# Feed-forward down
MODEL_TENSOR.FFN_DOWN: (
"gpt_neox.layers.{bid}.mlp.dense_4h_to_h", # gptneox
"transformer.h.{bid}.mlp.c_proj", # gpt2 refact qwen
"transformer.h.{bid}.mlp.c_proj", # gpt2 refact qwen jais
"transformer.blocks.{bid}.ffn.down_proj", # mpt
"transformer.h.{bid}.mlp.dense_4h_to_h", # falcon
"h.{bid}.mlp.dense_4h_to_h", # bloom
@@ -311,8 +343,10 @@ class TensorNameMap:
"encoder.layers.{bid}.mlp.fc2", # nomic-bert
"model.layers.{bid}.mlp.c_proj", # starcoder2
"encoder.layer.{bid}.mlp.wo", # jina-bert-v2
"transformer.layers.{bid}.ffn.proj_2", # openelm
"model.layers.{bid}.residual_mlp.w2", # arctic
"encoder.layer.{bid}.mlp.down_layer", # jina-bert-v2
"encoder.layers.{bid}.mlp.dense_4h_to_h", # chatglm
),
MODEL_TENSOR.FFN_DOWN_EXP: (
@@ -332,7 +366,8 @@ class TensorNameMap:
"model.layers.{bid}.self_attn.q_layernorm", # persimmon
"model.layers.{bid}.self_attn.q_norm", # cohere
"transformer.blocks.{bid}.attn.q_ln", # sea-lion
"encoder.layer.{bid}.attention.self.layer_norm_q" # jina-bert-v2
"encoder.layer.{bid}.attention.self.layer_norm_q", # jina-bert-v2
"transformer.layers.{bid}.attn.q_norm", # openelm
),
MODEL_TENSOR.ATTN_K_NORM: (
@@ -340,7 +375,8 @@ class TensorNameMap:
"model.layers.{bid}.self_attn.k_layernorm", # persimmon
"model.layers.{bid}.self_attn.k_norm", # cohere
"transformer.blocks.{bid}.attn.k_ln", # sea-lion
"encoder.layer.{bid}.attention.self.layer_norm_k" # jina-bert-v2
"encoder.layer.{bid}.attention.self.layer_norm_k", # jina-bert-v2
"transformer.layers.{bid}.attn.k_norm", # openelm
),
MODEL_TENSOR.ROPE_FREQS: (
@@ -421,6 +457,120 @@ class TensorNameMap:
MODEL_TENSOR.FFN_SUB_NORM: (
"model.layers.{bid}.mlp.ffn_layernorm", # bitnet
),
MODEL_TENSOR.DEC_ATTN_NORM: (
"decoder.block.{bid}.layer.0.layer_norm", # t5
),
MODEL_TENSOR.DEC_ATTN_Q: (
"decoder.block.{bid}.layer.0.SelfAttention.q", # t5
),
MODEL_TENSOR.DEC_ATTN_K: (
"decoder.block.{bid}.layer.0.SelfAttention.k", # t5
),
MODEL_TENSOR.DEC_ATTN_V: (
"decoder.block.{bid}.layer.0.SelfAttention.v", # t5
),
MODEL_TENSOR.DEC_ATTN_OUT: (
"decoder.block.{bid}.layer.0.SelfAttention.o", # t5
),
MODEL_TENSOR.DEC_ATTN_REL_B: (
"decoder.block.{bid}.layer.0.SelfAttention.relative_attention_bias", # t5
),
MODEL_TENSOR.DEC_CROSS_ATTN_NORM: (
"decoder.block.{bid}.layer.1.layer_norm", # t5
),
MODEL_TENSOR.DEC_CROSS_ATTN_Q: (
"decoder.block.{bid}.layer.1.EncDecAttention.q", # t5
),
MODEL_TENSOR.DEC_CROSS_ATTN_K: (
"decoder.block.{bid}.layer.1.EncDecAttention.k", # t5
),
MODEL_TENSOR.DEC_CROSS_ATTN_V: (
"decoder.block.{bid}.layer.1.EncDecAttention.v", # t5
),
MODEL_TENSOR.DEC_CROSS_ATTN_OUT: (
"decoder.block.{bid}.layer.1.EncDecAttention.o", # t5
),
MODEL_TENSOR.DEC_CROSS_ATTN_REL_B: (
"decoder.block.{bid}.layer.1.EncDecAttention.relative_attention_bias", # t5
),
MODEL_TENSOR.DEC_FFN_NORM: (
"decoder.block.{bid}.layer.2.layer_norm", # t5
),
MODEL_TENSOR.DEC_FFN_GATE: (
"decoder.block.{bid}.layer.2.DenseReluDense.wi_0", # flan-t5
),
MODEL_TENSOR.DEC_FFN_UP: (
"decoder.block.{bid}.layer.2.DenseReluDense.wi", # t5
"decoder.block.{bid}.layer.2.DenseReluDense.wi_1", # flan-t5
),
MODEL_TENSOR.DEC_FFN_DOWN: (
"decoder.block.{bid}.layer.2.DenseReluDense.wo", # t5
),
MODEL_TENSOR.DEC_OUTPUT_NORM: (
"decoder.final_layer_norm", # t5
),
MODEL_TENSOR.ENC_ATTN_NORM: (
"encoder.block.{bid}.layer.0.layer_norm", # t5
),
MODEL_TENSOR.ENC_ATTN_Q: (
"encoder.block.{bid}.layer.0.SelfAttention.q", # t5
),
MODEL_TENSOR.ENC_ATTN_K: (
"encoder.block.{bid}.layer.0.SelfAttention.k", # t5
),
MODEL_TENSOR.ENC_ATTN_V: (
"encoder.block.{bid}.layer.0.SelfAttention.v", # t5
),
MODEL_TENSOR.ENC_ATTN_OUT: (
"encoder.block.{bid}.layer.0.SelfAttention.o", # t5
),
MODEL_TENSOR.ENC_ATTN_REL_B: (
"encoder.block.{bid}.layer.0.SelfAttention.relative_attention_bias", # t5
),
MODEL_TENSOR.ENC_FFN_NORM: (
"encoder.block.{bid}.layer.1.layer_norm", # t5
),
MODEL_TENSOR.ENC_FFN_GATE: (
"encoder.block.{bid}.layer.1.DenseReluDense.wi_0", # flan-t5
),
MODEL_TENSOR.ENC_FFN_UP: (
"encoder.block.{bid}.layer.1.DenseReluDense.wi", # t5
"encoder.block.{bid}.layer.1.DenseReluDense.wi_1", # flan-t5
),
MODEL_TENSOR.ENC_FFN_DOWN: (
"encoder.block.{bid}.layer.1.DenseReluDense.wo", # t5
),
MODEL_TENSOR.ENC_OUTPUT_NORM: (
"encoder.final_layer_norm", # t5
),
}
# architecture-specific block mappings
@@ -452,14 +602,12 @@ class TensorNameMap:
for tensor, keys in self.block_mappings_cfg.items():
if tensor not in MODEL_TENSORS[arch]:
continue
# TODO: make this configurable
n_experts = 160
for xid in range(n_experts):
tensor_name = TENSOR_NAMES[tensor].format(bid = bid, xid = xid)
self.mapping[tensor_name] = (tensor, tensor_name)
for key in keys:
key = key.format(bid = bid, xid = xid)
self.mapping[key] = (tensor, tensor_name)
tensor_name = TENSOR_NAMES[tensor].format(bid = bid)
self.mapping[tensor_name] = (tensor, tensor_name)
for key in keys:
key = key.format(bid = bid)
self.mapping[key] = (tensor, tensor_name)
def get_type_and_name(self, key: str, try_suffixes: Sequence[str] = ()) -> tuple[MODEL_TENSOR, str] | None:
result = self.mapping.get(key)

69
gguf-py/gguf/utility.py Normal file
View File

@@ -0,0 +1,69 @@
from __future__ import annotations
from typing import Literal
def fill_templated_filename(filename: str, output_type: str | None) -> str:
# Given a file name fill in any type templates e.g. 'some-model-name.{ftype}.gguf'
ftype_lowercase: str = output_type.lower() if output_type is not None else ""
ftype_uppercase: str = output_type.upper() if output_type is not None else ""
return filename.format(ftype_lowercase,
outtype=ftype_lowercase, ftype=ftype_lowercase,
OUTTYPE=ftype_uppercase, FTYPE=ftype_uppercase)
def model_weight_count_rounded_notation(model_params_count: int, min_digits: int = 2) -> str:
if model_params_count > 1e12 :
# Trillions Of Parameters
scaled_model_params = model_params_count * 1e-12
scale_suffix = "T"
elif model_params_count > 1e9 :
# Billions Of Parameters
scaled_model_params = model_params_count * 1e-9
scale_suffix = "B"
elif model_params_count > 1e6 :
# Millions Of Parameters
scaled_model_params = model_params_count * 1e-6
scale_suffix = "M"
else:
# Thousands Of Parameters
scaled_model_params = model_params_count * 1e-3
scale_suffix = "K"
fix = max(min_digits - len(str(round(scaled_model_params)).lstrip('0')), 0)
return f"{scaled_model_params:.{fix}f}{scale_suffix}"
def size_label(total_params: int, shared_params: int, expert_params: int, expert_count: int) -> str:
if expert_count > 0:
pretty_size = model_weight_count_rounded_notation(abs(shared_params) + abs(expert_params), min_digits=2)
size_class = f"{expert_count}x{pretty_size}"
else:
size_class = model_weight_count_rounded_notation(abs(total_params), min_digits=2)
return size_class
def naming_convention(model_name: str | None, base_name: str | None, finetune_string: str | None, version_string: str | None, size_label: str | None, output_type: str | None, model_type: Literal['vocab', 'LoRA'] | None = None) -> str:
# Reference: https://github.com/ggerganov/ggml/blob/master/docs/gguf.md#gguf-naming-convention
if base_name is not None:
name = base_name.strip().replace(' ', '-').replace('/', '-')
elif model_name is not None:
name = model_name.strip().replace(' ', '-').replace('/', '-')
else:
name = "ggml-model"
parameters = f"-{size_label}" if size_label is not None else ""
finetune = f"-{finetune_string.strip().replace(' ', '-')}" if finetune_string is not None else ""
version = f"-{version_string.strip().replace(' ', '-')}" if version_string is not None else ""
encoding = f"-{output_type.strip().replace(' ', '-').upper()}" if output_type is not None else ""
kind = f"-{model_type.strip().replace(' ', '-')}" if model_type is not None else ""
return f"{name}{parameters}{finetune}{version}{encoding}{kind}"

3
gguf-py/pyproject.toml
View File

@@ -1,6 +1,6 @@
[tool.poetry]
name = "gguf"
version = "0.9.0"
version = "0.9.1"
description = "Read and write ML models in GGUF for GGML"
authors = ["GGML <ggml@ggml.ai>"]
packages = [
@@ -22,6 +22,7 @@ classifiers = [
python = ">=3.8"
numpy = ">=1.17"
tqdm = ">=4.27"
pyyaml = ">=5.1"
[tool.poetry.dev-dependencies]
pytest = "^5.2"

17
gguf-py/scripts/__init__.py
View File

@@ -1,13 +1,6 @@
import os
# pyright: reportUnusedImport=false
from importlib import import_module
os.environ["NO_LOCAL_GGUF"] = "TRUE"
gguf_convert_endian_entrypoint = import_module("scripts.gguf-convert-endian").main
gguf_dump_entrypoint = import_module("scripts.gguf-dump").main
gguf_set_metadata_entrypoint = import_module("scripts.gguf-set-metadata").main
gguf_new_metadata_entrypoint = import_module("scripts.gguf-new-metadata").main
del import_module, os
from .gguf_convert_endian import main as gguf_convert_endian_entrypoint
from .gguf_dump import main as gguf_dump_entrypoint
from .gguf_set_metadata import main as gguf_set_metadata_entrypoint
from .gguf_new_metadata import main as gguf_new_metadata_entrypoint

0
gguf-py/scripts/gguf-convert-endian.py → gguf-py/scripts/gguf_convert_endian.py
View File

80
gguf-py/scripts/gguf-dump.py → gguf-py/scripts/gguf_dump.py
View File

@@ -4,6 +4,7 @@ from __future__ import annotations
import logging
import argparse
import os
import re
import sys
from pathlib import Path
from typing import Any
@@ -208,7 +209,9 @@ def translate_tensor_name(name):
'ssm_d': 'State space model skip connection',
'ssm_dt': 'State space model time step',
'ssm_out': 'State space model output projection',
'blk': 'Block'
'blk': 'Block',
'enc': 'Encoder',
'dec': 'Decoder',
}
expanded_words = []
@@ -242,26 +245,58 @@ def dump_markdown_metadata(reader: GGUFReader, args: argparse.Namespace) -> None
else:
pretty_type = str(field.types[-1].name)
def escape_markdown_inline_code(value_string):
# Find the longest contiguous sequence of backticks in the string then
# wrap string with appropriate number of backticks required to escape it
max_backticks = max((len(match.group(0)) for match in re.finditer(r'`+', value_string)), default=0)
inline_code_marker = '`' * (max_backticks + 1)
# If the string starts or ends with a backtick, add a space at the beginning and end
if value_string.startswith('`') or value_string.endswith('`'):
value_string = f" {value_string} "
return f"{inline_code_marker}{value_string}{inline_code_marker}"
total_elements = len(field.data)
value = ""
if len(field.types) == 1:
curr_type = field.types[0]
if curr_type == GGUFValueType.STRING:
value = repr(str(bytes(field.parts[-1]), encoding='utf-8')[:60])
truncate_length = 60
value_string = str(bytes(field.parts[-1]), encoding='utf-8')
if len(value_string) > truncate_length:
head = escape_markdown_inline_code(value_string[:truncate_length // 2])
tail = escape_markdown_inline_code(value_string[-truncate_length // 2:])
value = "{head}...{tail}".format(head=head, tail=tail)
else:
value = escape_markdown_inline_code(value_string)
elif curr_type in reader.gguf_scalar_to_np:
value = str(field.parts[-1][0])
else:
if field.types[0] == GGUFValueType.ARRAY:
curr_type = field.types[1]
array_elements = []
if curr_type == GGUFValueType.STRING:
render_element = min(5, total_elements)
for element_pos in range(render_element):
value += repr(str(bytes(field.parts[-1 - element_pos]), encoding='utf-8')[:5]) + (", " if total_elements > 1 else "")
truncate_length = 30
value_string = str(bytes(field.parts[-1 - (total_elements - element_pos - 1) * 2]), encoding='utf-8')
if len(value_string) > truncate_length:
head = escape_markdown_inline_code(value_string[:truncate_length // 2])
tail = escape_markdown_inline_code(value_string[-truncate_length // 2:])
value = "{head}...{tail}".format(head=head, tail=tail)
else:
value = escape_markdown_inline_code(value_string)
array_elements.append(value)
elif curr_type in reader.gguf_scalar_to_np:
render_element = min(7, total_elements)
for element_pos in range(render_element):
value += str(field.parts[-1 - element_pos][0]) + (", " if total_elements > 1 else "")
value = f'[ {value}{" ..." if total_elements > 1 else ""} ]'
array_elements.append(str(field.parts[-1 - (total_elements - element_pos - 1)][0]))
value = f'[ {", ".join(array_elements).strip()}{", ..." if total_elements > len(array_elements) else ""} ]'
kv_dump_table.append({"n":n, "pretty_type":pretty_type, "total_elements":total_elements, "field_name":field.name, "value":value})
kv_dump_table_header_map = [
@@ -291,6 +326,10 @@ def dump_markdown_metadata(reader: GGUFReader, args: argparse.Namespace) -> None
tensor_group_name = "base"
if tensor_components[0] == 'blk':
tensor_group_name = f"{tensor_components[0]}.{tensor_components[1]}"
elif tensor_components[0] in ['enc', 'dec'] and tensor_components[1] == 'blk':
tensor_group_name = f"{tensor_components[0]}.{tensor_components[1]}.{tensor_components[2]}"
elif tensor_components[0] in ['enc', 'dec']:
tensor_group_name = f"{tensor_components[0]}"
# Check if new Tensor Group
if tensor_group_name not in tensor_groups:
@@ -313,6 +352,27 @@ def dump_markdown_metadata(reader: GGUFReader, args: argparse.Namespace) -> None
markdown_content += "\n"
markdown_content += "### Tensor Data Offset\n"
markdown_content += '\n'
markdown_content += 'This table contains the offset and data segment relative to start of file\n'
markdown_content += '\n'
tensor_mapping_table: list[dict[str, str | int]] = []
for key, tensor in enumerate(reader.tensors):
data_offset_pretty = '{0:#16x}'.format(tensor.data_offset)
data_size_pretty = '{0:#16x}'.format(tensor.n_bytes)
tensor_mapping_table.append({"t_id":key, "layer_name":tensor.name, "data_offset":data_offset_pretty, "data_size":data_size_pretty})
tensors_mapping_table_header_map = [
{'key_name':'t_id', 'header_name':'T_ID', 'align':'right'},
{'key_name':'layer_name', 'header_name':'Tensor Layer Name', 'align':'left'},
{'key_name':'data_offset', 'header_name':'Data Offset (B)', 'align':'right'},
{'key_name':'data_size', 'header_name':'Data Size (B)', 'align':'right'},
]
markdown_content += markdown_table_with_alignment_support(tensors_mapping_table_header_map, tensor_mapping_table)
markdown_content += "\n"
for group in tensor_prefix_order:
tensors = tensor_groups[group]
group_elements = sum(tensor.n_elements for tensor in tensors)
@@ -355,7 +415,7 @@ def dump_markdown_metadata(reader: GGUFReader, args: argparse.Namespace) -> None
markdown_content += f"- Percentage of total elements: {group_percentage:.2f}%\n"
markdown_content += "\n\n"
print(markdown_content) # noqa: NP100
print(markdown_content) # noqa: NP100
def main() -> None:
@@ -364,6 +424,8 @@ def main() -> None:
parser.add_argument("--no-tensors", action="store_true", help="Don't dump tensor metadata")
parser.add_argument("--json", action="store_true", help="Produce JSON output")
parser.add_argument("--json-array", action="store_true", help="Include full array values in JSON output (long)")
parser.add_argument("--data-offset", action="store_true", help="Start of data offset")
parser.add_argument("--data-alignment", action="store_true", help="Data alignment applied globally to data field")
parser.add_argument("--markdown", action="store_true", help="Produce markdown output")
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
@@ -371,7 +433,7 @@ def main() -> None:
logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
if not args.json and not args.markdown:
if not args.json and not args.markdown and not args.data_offset and not args.data_alignment:
logger.info(f'* Loading: {args.model}')
reader = GGUFReader(args.model, 'r')
@@ -380,6 +442,10 @@ def main() -> None:
dump_metadata_json(reader, args)
elif args.markdown:
dump_markdown_metadata(reader, args)
elif args.data_offset:
print(reader.data_offset) # noqa: NP100
elif args.data_alignment:
print(reader.alignment) # noqa: NP100
else:
dump_metadata(reader, args)

102
gguf-py/scripts/gguf_hash.py Executable file
View File

@@ -0,0 +1,102 @@
#!/usr/bin/env python3
from __future__ import annotations
import uuid
import hashlib
import logging
import argparse
import os
import sys
from pathlib import Path
from tqdm import tqdm
# Necessary to load the local gguf package
if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent / 'gguf-py').exists():
sys.path.insert(0, str(Path(__file__).parent.parent))
from gguf import GGUFReader # noqa: E402
logger = logging.getLogger("gguf-hash")
# UUID_NAMESPACE_LLAMA_CPP = uuid.uuid5(uuid.NAMESPACE_URL, 'en.wikipedia.org/wiki/Llama.cpp')
UUID_NAMESPACE_LLAMA_CPP = uuid.UUID('ef001206-dadc-5f6d-a15f-3359e577d4e5')
# For more information about what field.parts and field.data represent,
# please see the comments in the modify_gguf.py example.
def gguf_hash(reader: GGUFReader, filename: str, disable_progress_bar: bool, no_layer: bool) -> None:
sha1 = hashlib.sha1()
sha256 = hashlib.sha256()
uuidv5_sha1 = hashlib.sha1()
uuidv5_sha1.update(UUID_NAMESPACE_LLAMA_CPP.bytes)
# Total Weight Calculation For Progress Bar
total_weights = 0
for n, tensor in enumerate(reader.tensors, 1):
# We don't need these
if tensor.name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")):
continue
# Calculate Tensor Volume
sum_weights_in_tensor = 1
for dim in tensor.shape:
sum_weights_in_tensor *= dim
total_weights += sum_weights_in_tensor
# Hash Progress Bar
bar = tqdm(desc="Hashing", total=total_weights, unit="weights", unit_scale=True, disable=disable_progress_bar)
# Hashing Process
for tensor in reader.tensors:
# We don't need these
if tensor.name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")):
continue
# Progressbar
sum_weights_in_tensor = 1
for dim in tensor.shape:
sum_weights_in_tensor *= dim
bar.update(sum_weights_in_tensor)
if not no_layer:
sha1_layer = hashlib.sha1()
sha1_layer.update(tensor.data.data)
print("sha1 {0} {1}:{2}".format(sha1_layer.hexdigest(), filename, tensor.name)) # noqa: NP100
sha256_layer = hashlib.sha256()
sha256_layer.update(tensor.data.data)
print("sha256 {0} {1}:{2}".format(sha256_layer.hexdigest(), filename, tensor.name)) # noqa: NP100
sha1.update(tensor.data.data)
sha256.update(tensor.data.data)
uuidv5_sha1.update(tensor.data.data)
# Flush Hash Progress Bar
bar.close()
# Display Hash Output
print("sha1 {0} {1}".format(sha1.hexdigest(), filename)) # noqa: NP100
print("sha256 {0} {1}".format(sha256.hexdigest(), filename)) # noqa: NP100
print("uuid {0} {1}".format(uuid.UUID(bytes=uuidv5_sha1.digest()[:16], version=5), filename)) # noqa: NP100
def main() -> None:
parser = argparse.ArgumentParser(description="Dump GGUF file metadata")
parser.add_argument("model", type=str, help="GGUF format model filename")
parser.add_argument("--no-layer", action="store_true", help="exclude per layer hash")
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
parser.add_argument("--progressbar", action="store_true", help="enable progressbar")
args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"])
logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
reader = GGUFReader(args.model, 'r')
gguf_hash(reader, args.model, not args.progressbar, args.no_layer)
if __name__ == '__main__':
main()

2
gguf-py/scripts/gguf-new-metadata.py → gguf-py/scripts/gguf_new_metadata.py
View File

@@ -1,4 +1,6 @@
#!/usr/bin/env python3
from __future__ import annotations
import logging
import argparse
import os

0
gguf-py/scripts/gguf-set-metadata.py → gguf-py/scripts/gguf_set_metadata.py
View File

1
gguf-py/tests/__init__.py Normal file
View File

@@ -0,0 +1 @@
from .test_metadata import *

7
gguf-py/tests/test_gguf.py
View File

@@ -1,7 +0,0 @@
import gguf # noqa: F401
# TODO: add tests
def test_write_gguf() -> None:
pass

203
gguf-py/tests/test_metadata.py Executable file
View File

@@ -0,0 +1,203 @@
#!/usr/bin/env python3
import unittest
from pathlib import Path
import os
import sys
# Necessary to load the local gguf package
if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent / 'gguf-py').exists():
sys.path.insert(0, str(Path(__file__).parent.parent))
import gguf
class TestMetadataMethod(unittest.TestCase):
def test_id_to_title(self):
self.assertEqual(gguf.Metadata.id_to_title("Mixtral-8x7B-Instruct-v0.1"), "Mixtral 8x7B Instruct v0.1")
self.assertEqual(gguf.Metadata.id_to_title("Meta-Llama-3-8B"), "Meta Llama 3 8B")
self.assertEqual(gguf.Metadata.id_to_title("hermes-2-pro-llama-3-8b-DPO"), "Hermes 2 Pro Llama 3 8b DPO")
def test_get_model_id_components(self):
# This is the basic standard form with organization marker
self.assertEqual(gguf.Metadata.get_model_id_components("Mistral/Mixtral-8x7B-Instruct-v0.1"),
('Mixtral-8x7B-Instruct-v0.1', "Mistral", 'Mixtral', 'Instruct', 'v0.1', '8x7B'))
# Similar to basic standard form but without organization marker
self.assertEqual(gguf.Metadata.get_model_id_components("Mixtral-8x7B-Instruct-v0.1"),
('Mixtral-8x7B-Instruct-v0.1', None, 'Mixtral', 'Instruct', 'v0.1', '8x7B'))
# Missing version
self.assertEqual(gguf.Metadata.get_model_id_components("Mixtral-8x7B-Instruct"),
('Mixtral-8x7B-Instruct', None, 'Mixtral', 'Instruct', None, '8x7B'))
# Missing finetune
self.assertEqual(gguf.Metadata.get_model_id_components("Mixtral-8x7B-v0.1"),
('Mixtral-8x7B-v0.1', None, 'Mixtral', None, 'v0.1', '8x7B'))
# Base name and size label only
self.assertEqual(gguf.Metadata.get_model_id_components("Mixtral-8x7B"),
('Mixtral-8x7B', None, 'Mixtral', None, None, '8x7B'))
# Base name and version only
self.assertEqual(gguf.Metadata.get_model_id_components("Mixtral-v0.1"),
('Mixtral-v0.1', None, 'Mixtral', None, 'v0.1', None))
## Edge Cases ##
# This is too ambiguous... best to err on caution and output nothing
self.assertEqual(gguf.Metadata.get_model_id_components("Mixtral"),
('Mixtral', None, None, None, None, None))
# Basename has numbers mixed in and also size label provided. Must avoid capturing number in basename
self.assertEqual(gguf.Metadata.get_model_id_components("NousResearch/Meta-Llama-3-8B"),
('Meta-Llama-3-8B', "NousResearch", 'Meta-Llama-3', None, None, '8B'))
# Non standard naming
self.assertEqual(gguf.Metadata.get_model_id_components("Qwen1.5-MoE-A2.7B-Chat"),
('Qwen1.5-MoE-A2.7B-Chat', None, 'Qwen1.5-MoE', 'Chat', None, 'A2.7B'))
# Capture 'sub size labels' e.g. A14B in '57B-A14B' usually refers to activated params/weight count
self.assertEqual(gguf.Metadata.get_model_id_components("Qwen2-57B-A14B-Instruct"),
('Qwen2-57B-A14B-Instruct', None, 'Qwen2', 'Instruct', None, '57B-A14B'))
# Check that it can handle a real model id with no version code
# Note that 4k in this string is non standard and microsoft were referring to context length rather than weight count
self.assertEqual(gguf.Metadata.get_model_id_components("microsoft/Phi-3-mini-4k-instruct", 4 * 10**9),
('Phi-3-mini-4k-instruct', 'microsoft', 'Phi-3', '4k-instruct', None, 'mini'))
# There is some legitimate models with only thousands of parameters
self.assertEqual(gguf.Metadata.get_model_id_components("delphi-suite/stories-llama2-50k", 50 * 10**3),
('stories-llama2-50k', 'delphi-suite', 'stories-llama2', None, None, '50K'))
# Non standard and not easy to disambiguate
self.assertEqual(gguf.Metadata.get_model_id_components("DeepSeek-Coder-V2-Lite-Instruct"),
('DeepSeek-Coder-V2-Lite-Instruct', None, 'DeepSeek-Coder-V2-Lite', 'Instruct', None, None))
# This is a real model_id where they append 2DPO to refer to Direct Preference Optimization
self.assertEqual(gguf.Metadata.get_model_id_components("crestf411/daybreak-kunoichi-2dpo-7b"),
('daybreak-kunoichi-2dpo-7b', 'crestf411', 'daybreak-kunoichi', '2dpo', None, '7B'))
# This is a real model id where the weight size has a decimal point
self.assertEqual(gguf.Metadata.get_model_id_components("Qwen2-0.5B-Instruct"),
('Qwen2-0.5B-Instruct', None, 'Qwen2', 'Instruct', None, '0.5B'))
# Uses an underscore in the size label
self.assertEqual(gguf.Metadata.get_model_id_components("smallcloudai/Refact-1_6B-fim"),
('Refact-1_6B-fim', 'smallcloudai', 'Refact', 'fim', None, '1.6B'))
# Uses Iter3 for the version
self.assertEqual(gguf.Metadata.get_model_id_components("UCLA-AGI/Gemma-2-9B-It-SPPO-Iter3"),
('Gemma-2-9B-It-SPPO-Iter3', 'UCLA-AGI', 'Gemma-2', 'It-SPPO', 'Iter3', '9B'))
# Has two potential versions in the basename
self.assertEqual(gguf.Metadata.get_model_id_components("NousResearch/Hermes-2-Theta-Llama-3-8B"),
('Hermes-2-Theta-Llama-3-8B', 'NousResearch', 'Hermes-2-Theta-Llama-3', None, None, '8B'))
# Potential version in the basename
self.assertEqual(gguf.Metadata.get_model_id_components("SeaLLMs/SeaLLMs-v3-7B-Chat"),
('SeaLLMs-v3-7B-Chat', 'SeaLLMs', 'SeaLLMs-v3', 'Chat', None, '7B'))
# Underscore in the basename, and 1m for the context size
self.assertEqual(gguf.Metadata.get_model_id_components("internlm/internlm2_5-7b-chat-1m", 7 * 10**9),
('internlm2_5-7b-chat-1m', 'internlm', 'internlm2_5', 'chat-1m', None, '7B'))
# Version before the finetune name
self.assertEqual(gguf.Metadata.get_model_id_components("pszemraj/jamba-900M-v0.13-KIx2"),
('jamba-900M-v0.13-KIx2', 'pszemraj', 'jamba', 'KIx2', 'v0.13', '900M'))
# TODO: hf suffix which could be ignored but isn't
self.assertEqual(gguf.Metadata.get_model_id_components("state-spaces/mamba-2.8b-hf"),
('mamba-2.8b-hf', 'state-spaces', 'mamba', 'hf', None, '2.8B'))
# Two sizes, don't merge them, the other is the number of tokens on which it was trained
self.assertEqual(gguf.Metadata.get_model_id_components("abacaj/llama-161M-100B", 161 * 10**6),
('llama-161M-100B', 'abacaj', 'llama', '100b', None, '161M'))
# It's a trap, there is no size label
self.assertEqual(gguf.Metadata.get_model_id_components("SparseLLM/relu-100B", 1340 * 10**6),
('relu-100B', 'SparseLLM', 'relu', '100b', None, None))
# Weird size notation
self.assertEqual(gguf.Metadata.get_model_id_components("bigscience/bloom-7b1-petals"),
('bloom-7b1-petals', 'bigscience', 'bloom', 'petals', None, '7.1B'))
# Ignore full-text size labels when there are number-based ones, and deduplicate size labels
self.assertEqual(gguf.Metadata.get_model_id_components("MaziyarPanahi/GreenNode-mini-7B-multilingual-v1olet-Mistral-7B-Instruct-v0.1"),
('GreenNode-mini-7B-multilingual-v1olet-Mistral-7B-Instruct-v0.1', 'MaziyarPanahi', 'GreenNode-mini', 'multilingual-v1olet-Mistral-Instruct', 'v0.1', '7B'))
# Instruct in a name without a size label
self.assertEqual(gguf.Metadata.get_model_id_components("mistralai/Mistral-Nemo-Instruct-2407"),
('Mistral-Nemo-Instruct-2407', 'mistralai', 'Mistral-Nemo', 'Instruct', '2407', None))
# Non-obvious splitting relying on 'chat' keyword
self.assertEqual(gguf.Metadata.get_model_id_components("deepseek-ai/DeepSeek-V2-Chat-0628"),
('DeepSeek-V2-Chat-0628', 'deepseek-ai', 'DeepSeek-V2', 'Chat', '0628', None))
# Multiple versions
self.assertEqual(gguf.Metadata.get_model_id_components("OpenGVLab/Mini-InternVL-Chat-2B-V1-5"),
('Mini-InternVL-Chat-2B-V1-5', 'OpenGVLab', 'Mini-InternVL', 'Chat', 'V1-5', '2B'))
# TODO: DPO in the name
self.assertEqual(gguf.Metadata.get_model_id_components("jondurbin/bagel-dpo-2.8b-v0.2"),
('bagel-dpo-2.8b-v0.2', 'jondurbin', 'bagel-dpo', None, 'v0.2', '2.8B'))
# DPO in name, but can't be used for the finetune to keep 'LLaMA-3' in the basename
self.assertEqual(gguf.Metadata.get_model_id_components("voxmenthe/SFR-Iterative-DPO-LLaMA-3-8B-R-unquantized"),
('SFR-Iterative-DPO-LLaMA-3-8B-R-unquantized', 'voxmenthe', 'SFR-Iterative-DPO-LLaMA-3', 'R-unquantized', None, '8B'))
# Too ambiguous
# TODO: should "base" be a 'finetune' or 'size_label'?
# (in this case it should be a size label, but other models use it to signal that they are not finetuned)
self.assertEqual(gguf.Metadata.get_model_id_components("microsoft/Florence-2-base"),
('Florence-2-base', 'microsoft', None, None, None, None))
## Invalid cases ##
# Start with a dash and has dashes in rows
self.assertEqual(gguf.Metadata.get_model_id_components("mistralai/-Mistral--Nemo-Base-2407-"),
('-Mistral--Nemo-Base-2407-', 'mistralai', 'Mistral-Nemo-Base', None, '2407', None))
## LoRA ##
self.assertEqual(gguf.Metadata.get_model_id_components("Llama-3-Instruct-abliteration-LoRA-8B"),
('Llama-3-Instruct-abliteration-LoRA-8B', None, 'Llama-3', 'Instruct-abliteration-LoRA', None, '8B'))
# Negative size --> output is a LoRA adaper --> prune "LoRA" out of the name to avoid redundancy with the suffix
self.assertEqual(gguf.Metadata.get_model_id_components("Llama-3-Instruct-abliteration-LoRA-8B", -1234),
('Llama-3-Instruct-abliteration-LoRA-8B', None, 'Llama-3', 'Instruct-abliteration', None, '8B'))
def test_apply_metadata_heuristic_from_model_card(self):
model_card = {
'tags': ['Llama-3', 'instruct', 'finetune', 'chatml', 'DPO', 'RLHF', 'gpt4', 'synthetic data', 'distillation', 'function calling', 'json mode', 'axolotl'],
'model-index': [{'name': 'Mixtral-8x7B-Instruct-v0.1', 'results': []}],
'language': ['en'],
'datasets': ['teknium/OpenHermes-2.5'],
'widget': [{'example_title': 'Hermes 2 Pro', 'messages': [{'role': 'system', 'content': 'You are a sentient, superintelligent artificial general intelligence, here to teach and assist me.'}, {'role': 'user', 'content': 'Write a short story about Goku discovering kirby has teamed up with Majin Buu to destroy the world.'}]}],
'base_model': ["EmbeddedLLM/Mistral-7B-Merge-14-v0", "janai-hq/trinity-v1"]
}
got = gguf.Metadata.apply_metadata_heuristic(gguf.Metadata(), model_card, None, None)
expect = gguf.Metadata()
expect.base_models=[{'name': 'Mistral 7B Merge 14 v0', 'organization': 'EmbeddedLLM', 'version': '14-v0', 'repo_url': 'https://huggingface.co/EmbeddedLLM/Mistral-7B-Merge-14-v0'}, {'name': 'Trinity v1', 'organization': 'Janai Hq', 'version': 'v1', 'repo_url': 'https://huggingface.co/janai-hq/trinity-v1'}]
expect.tags=['Llama-3', 'instruct', 'finetune', 'chatml', 'DPO', 'RLHF', 'gpt4', 'synthetic data', 'distillation', 'function calling', 'json mode', 'axolotl']
expect.languages=['en']
expect.datasets=['teknium/OpenHermes-2.5']
self.assertEqual(got, expect)
def test_apply_metadata_heuristic_from_hf_parameters(self):
hf_params = {"_name_or_path": "./hermes-2-pro-llama-3-8b-DPO"}
got = gguf.Metadata.apply_metadata_heuristic(gguf.Metadata(), model_card=None, hf_params=hf_params, model_path=None)
expect = gguf.Metadata(name='Hermes 2 Pro Llama 3 8b DPO', finetune='DPO', basename='hermes-2-pro-llama-3', size_label='8B')
self.assertEqual(got, expect)
def test_apply_metadata_heuristic_from_model_dir(self):
model_dir_path = Path("./hermes-2-pro-llama-3-8b-DPO")
got = gguf.Metadata.apply_metadata_heuristic(gguf.Metadata(), model_card=None, hf_params=None, model_path=model_dir_path)
expect = gguf.Metadata(name='Hermes 2 Pro Llama 3 8b DPO', finetune='DPO', basename='hermes-2-pro-llama-3', size_label='8B')
self.assertEqual(got, expect)
if __name__ == "__main__":
unittest.main()