vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into compilade/imatrix-batched-chunks

Browse source
This commit is contained in:
Francis Couture-Harpin 2025-02-09 12:06:15 -05:00
commit db502ddd0e
762 changed files with 149462 additions and 91773 deletions
Download patch file Download diff file Expand all files Collapse all files

10
gguf-py/README.md
View file

@ -15,13 +15,15 @@ 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.
[examples/reader.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/examples/reader.py) — Extracts and displays key-value pairs and tensor details from a GGUF file in a readable format.
[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.
[gguf/scripts/gguf_dump.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/gguf/scripts/gguf_dump.py) — Dumps a GGUF file's metadata to the console.
[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.
[gguf/scripts/gguf_set_metadata.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/gguf/scripts/gguf_set_metadata.py) — Allows changing simple metadata values in a GGUF file by key.
[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.
[gguf/scripts/gguf_convert_endian.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/gguf/scripts/gguf_convert_endian.py) — Allows converting the endianness of GGUF files.
[gguf/scripts/gguf_new_metadata.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/gguf/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:

580
gguf-py/gguf/constants.py
View file

@ -64,20 +64,33 @@ class Keys:
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_DESCRIPTION = "general.base_model.{id}.description"
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...)
# Dataset Source
DATASET_COUNT = "general.dataset.count"
DATASET_NAME = "general.dataset.{id}.name"
DATASET_AUTHOR = "general.dataset.{id}.author"
DATASET_VERSION = "general.dataset.{id}.version"
DATASET_ORGANIZATION = "general.dataset.{id}.organization"
DATASET_DESCRIPTION = "general.dataset.{id}.description"
DATASET_URL = "general.dataset.{id}.url" # Model Website/Paper
DATASET_DOI = "general.dataset.{id}.doi"
DATASET_UUID = "general.dataset.{id}.uuid"
DATASET_REPO_URL = "general.dataset.{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"
CONTEXT_LENGTH = "{arch}.context_length"
EMBEDDING_LENGTH = "{arch}.embedding_length"
FEATURES_LENGTH = "{arch}.features_length"
BLOCK_COUNT = "{arch}.block_count"
LEADING_DENSE_BLOCK_COUNT = "{arch}.leading_dense_block_count"
FEED_FORWARD_LENGTH = "{arch}.feed_forward_length"
@ -89,14 +102,20 @@ class Keys:
EXPERT_USED_COUNT = "{arch}.expert_used_count"
EXPERT_SHARED_COUNT = "{arch}.expert_shared_count"
EXPERT_WEIGHTS_SCALE = "{arch}.expert_weights_scale"
EXPERT_WEIGHTS_NORM = "{arch}.expert_weights_norm"
EXPERT_GATING_FUNC = "{arch}.expert_gating_func"
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"
SWIN_NORM = "{arch}.swin_norm"
RESCALE_EVERY_N_LAYERS = "{arch}.rescale_every_n_layers"
TIME_MIX_EXTRA_DIM = "{arch}.time_mix_extra_dim"
TIME_DECAY_EXTRA_DIM = "{arch}.time_decay_extra_dim"
RESIDUAL_SCALE = "{arch}.residual_scale"
EMBEDDING_SCALE = "{arch}.embedding_scale"
TOKEN_SHIFT_COUNT = "{arch}.token_shift_count"
class Attention:
HEAD_COUNT = "{arch}.attention.head_count"
@ -107,14 +126,18 @@ class Keys:
VALUE_LENGTH = "{arch}.attention.value_length"
LAYERNORM_EPS = "{arch}.attention.layer_norm_epsilon"
LAYERNORM_RMS_EPS = "{arch}.attention.layer_norm_rms_epsilon"
GROUPNORM_EPS = "{arch}.attention.group_norm_epsilon"
GROUPNORM_GROUPS = "{arch}.attention.group_norm_groups"
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"
SCALE = "{arch}.attention.scale"
class Rope:
DIMENSION_COUNT = "{arch}.rope.dimension_count"
DIMENSION_SECTIONS = "{arch}.rope.dimension_sections"
FREQ_BASE = "{arch}.rope.freq_base"
SCALING_TYPE = "{arch}.rope.scaling.type"
SCALING_FACTOR = "{arch}.rope.scaling.factor"
@ -138,6 +161,14 @@ class Keys:
class WKV:
HEAD_SIZE = "{arch}.wkv.head_size"
class PosNet:
EMBEDDING_LENGTH = "{arch}.posnet.embedding_length"
BLOCK_COUNT = "{arch}.posnet.block_count"
class ConvNext:
EMBEDDING_LENGTH = "{arch}.convnext.embedding_length"
BLOCK_COUNT = "{arch}.convnext.block_count"
class Tokenizer:
MODEL = "tokenizer.ggml.model"
PRE = "tokenizer.ggml.pre"
@ -148,10 +179,11 @@ class Keys:
MERGES = "tokenizer.ggml.merges"
BOS_ID = "tokenizer.ggml.bos_token_id"
EOS_ID = "tokenizer.ggml.eos_token_id"
EOT_ID = "tokenizer.ggml.eot_token_id"
EOM_ID = "tokenizer.ggml.eom_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"
@ -164,11 +196,16 @@ class Keys:
CHAT_TEMPLATE_N = "tokenizer.chat_template.{name}"
CHAT_TEMPLATES = "tokenizer.chat_templates"
# FIM/Infill special tokens constants
FIM_PRE_ID = "tokenizer.ggml.fim_pre_token_id"
FIM_SUF_ID = "tokenizer.ggml.fim_suf_token_id"
FIM_MID_ID = "tokenizer.ggml.fim_mid_token_id"
FIM_PAD_ID = "tokenizer.ggml.fim_pad_token_id"
FIM_REP_ID = "tokenizer.ggml.fim_rep_token_id"
FIM_SEP_ID = "tokenizer.ggml.fim_sep_token_id"
# deprecated:
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"
EOM_ID = "tokenizer.ggml.eom_token_id"
class Adapter:
TYPE = "adapter.type"
@ -192,50 +229,63 @@ class GGUFType:
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()
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()
GEMMA2 = auto()
STARCODER2 = auto()
RWKV6 = auto()
MAMBA = auto()
XVERSE = auto()
COMMAND_R = auto()
DBRX = auto()
OLMO = auto()
OPENELM = auto()
ARCTIC = auto()
DEEPSEEK2 = auto()
CHATGLM = auto()
BITNET = auto()
T5 = auto()
T5ENCODER = auto()
JAIS = auto()
NEMOTRON = auto()
EXAONE = auto()
LLAMA = auto()
DECI = 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()
QWEN2VL = auto()
PHI2 = auto()
PHI3 = auto()
PHIMOE = auto()
PLAMO = auto()
CODESHELL = auto()
ORION = auto()
INTERNLM2 = auto()
MINICPM = auto()
MINICPM3 = auto()
GEMMA = auto()
GEMMA2 = auto()
STARCODER2 = auto()
RWKV6 = auto()
RWKV6QWEN2 = auto()
MAMBA = auto()
XVERSE = auto()
COMMAND_R = auto()
COHERE2 = auto()
DBRX = auto()
OLMO = auto()
OLMO2 = auto()
OLMOE = auto()
OPENELM = auto()
ARCTIC = auto()
DEEPSEEK = auto()
DEEPSEEK2 = auto()
CHATGLM = auto()
BITNET = auto()
T5 = auto()
T5ENCODER = auto()
JAIS = auto()
NEMOTRON = auto()
EXAONE = auto()
GRANITE = auto()
GRANITE_MOE = auto()
CHAMELEON = auto()
WAVTOKENIZER_DEC = auto()
class MODEL_TENSOR(IntEnum):
@ -274,6 +324,7 @@ class MODEL_TENSOR(IntEnum):
FFN_GATE_SHEXP = auto()
FFN_DOWN_SHEXP = auto()
FFN_UP_SHEXP = auto()
FFN_EXP_PROBS_B = auto()
ATTN_Q_NORM = auto()
ATTN_K_NORM = auto()
LAYER_OUT_NORM = auto()
@ -291,6 +342,7 @@ class MODEL_TENSOR(IntEnum):
TIME_MIX_LERP_V = auto()
TIME_MIX_LERP_R = auto()
TIME_MIX_LERP_G = auto()
TIME_MIX_LERP_FUSED = auto()
TIME_MIX_LERP_W = auto()
TIME_MIX_FIRST = auto()
TIME_MIX_DECAY = auto()
@ -343,53 +395,84 @@ class MODEL_TENSOR(IntEnum):
ENC_FFN_DOWN = auto()
ENC_FFN_UP = auto()
ENC_OUTPUT_NORM = auto()
CLS = auto() # classifier
CLS_OUT = auto() # classifier output projection
CONV1D = auto()
CONVNEXT_DW = auto()
CONVNEXT_NORM = auto()
CONVNEXT_PW1 = auto()
CONVNEXT_PW2 = auto()
CONVNEXT_GAMMA = auto()
POSNET_CONV1 = auto()
POSNET_CONV2 = auto()
POSNET_NORM = auto()
POSNET_NORM1 = auto()
POSNET_NORM2 = auto()
POSNET_ATTN_NORM = auto()
POSNET_ATTN_Q = auto()
POSNET_ATTN_K = auto()
POSNET_ATTN_V = auto()
POSNET_ATTN_OUT = auto()
MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.LLAMA: "llama",
MODEL_ARCH.FALCON: "falcon",
MODEL_ARCH.BAICHUAN: "baichuan",
MODEL_ARCH.GROK: "grok",
MODEL_ARCH.GPT2: "gpt2",
MODEL_ARCH.GPTJ: "gptj",
MODEL_ARCH.GPTNEOX: "gptneox",
MODEL_ARCH.MPT: "mpt",
MODEL_ARCH.STARCODER: "starcoder",
MODEL_ARCH.REFACT: "refact",
MODEL_ARCH.BERT: "bert",
MODEL_ARCH.NOMIC_BERT: "nomic-bert",
MODEL_ARCH.JINA_BERT_V2: "jina-bert-v2",
MODEL_ARCH.BLOOM: "bloom",
MODEL_ARCH.STABLELM: "stablelm",
MODEL_ARCH.QWEN: "qwen",
MODEL_ARCH.QWEN2: "qwen2",
MODEL_ARCH.QWEN2MOE: "qwen2moe",
MODEL_ARCH.PHI2: "phi2",
MODEL_ARCH.PHI3: "phi3",
MODEL_ARCH.PLAMO: "plamo",
MODEL_ARCH.CODESHELL: "codeshell",
MODEL_ARCH.ORION: "orion",
MODEL_ARCH.INTERNLM2: "internlm2",
MODEL_ARCH.MINICPM: "minicpm",
MODEL_ARCH.GEMMA: "gemma",
MODEL_ARCH.GEMMA2: "gemma2",
MODEL_ARCH.STARCODER2: "starcoder2",
MODEL_ARCH.RWKV6: "rwkv6",
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.T5ENCODER: "t5encoder",
MODEL_ARCH.JAIS: "jais",
MODEL_ARCH.NEMOTRON: "nemotron",
MODEL_ARCH.EXAONE: "exaone",
MODEL_ARCH.LLAMA: "llama",
MODEL_ARCH.DECI: "deci",
MODEL_ARCH.FALCON: "falcon",
MODEL_ARCH.BAICHUAN: "baichuan",
MODEL_ARCH.GROK: "grok",
MODEL_ARCH.GPT2: "gpt2",
MODEL_ARCH.GPTJ: "gptj",
MODEL_ARCH.GPTNEOX: "gptneox",
MODEL_ARCH.MPT: "mpt",
MODEL_ARCH.STARCODER: "starcoder",
MODEL_ARCH.REFACT: "refact",
MODEL_ARCH.BERT: "bert",
MODEL_ARCH.NOMIC_BERT: "nomic-bert",
MODEL_ARCH.JINA_BERT_V2: "jina-bert-v2",
MODEL_ARCH.BLOOM: "bloom",
MODEL_ARCH.STABLELM: "stablelm",
MODEL_ARCH.QWEN: "qwen",
MODEL_ARCH.QWEN2: "qwen2",
MODEL_ARCH.QWEN2MOE: "qwen2moe",
MODEL_ARCH.QWEN2VL: "qwen2vl",
MODEL_ARCH.PHI2: "phi2",
MODEL_ARCH.PHI3: "phi3",
MODEL_ARCH.PHIMOE: "phimoe",
MODEL_ARCH.PLAMO: "plamo",
MODEL_ARCH.CODESHELL: "codeshell",
MODEL_ARCH.ORION: "orion",
MODEL_ARCH.INTERNLM2: "internlm2",
MODEL_ARCH.MINICPM: "minicpm",
MODEL_ARCH.MINICPM3: "minicpm3",
MODEL_ARCH.GEMMA: "gemma",
MODEL_ARCH.GEMMA2: "gemma2",
MODEL_ARCH.STARCODER2: "starcoder2",
MODEL_ARCH.RWKV6: "rwkv6",
MODEL_ARCH.RWKV6QWEN2: "rwkv6qwen2",
MODEL_ARCH.MAMBA: "mamba",
MODEL_ARCH.XVERSE: "xverse",
MODEL_ARCH.COMMAND_R: "command-r",
MODEL_ARCH.COHERE2: "cohere2",
MODEL_ARCH.DBRX: "dbrx",
MODEL_ARCH.OLMO: "olmo",
MODEL_ARCH.OLMO2: "olmo2",
MODEL_ARCH.OLMOE: "olmoe",
MODEL_ARCH.OPENELM: "openelm",
MODEL_ARCH.ARCTIC: "arctic",
MODEL_ARCH.DEEPSEEK: "deepseek",
MODEL_ARCH.DEEPSEEK2: "deepseek2",
MODEL_ARCH.CHATGLM: "chatglm",
MODEL_ARCH.BITNET: "bitnet",
MODEL_ARCH.T5: "t5",
MODEL_ARCH.T5ENCODER: "t5encoder",
MODEL_ARCH.JAIS: "jais",
MODEL_ARCH.NEMOTRON: "nemotron",
MODEL_ARCH.EXAONE: "exaone",
MODEL_ARCH.GRANITE: "granite",
MODEL_ARCH.GRANITE_MOE: "granitemoe",
MODEL_ARCH.CHAMELEON: "chameleon",
MODEL_ARCH.WAVTOKENIZER_DEC: "wavtokenizer-dec",
}
TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@ -430,6 +513,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
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.FFN_EXP_PROBS_B: "blk.{bid}.exp_probs_b",
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",
@ -445,6 +529,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.TIME_MIX_LERP_V: "blk.{bid}.time_mix_lerp_v",
MODEL_TENSOR.TIME_MIX_LERP_R: "blk.{bid}.time_mix_lerp_r",
MODEL_TENSOR.TIME_MIX_LERP_G: "blk.{bid}.time_mix_lerp_g",
MODEL_TENSOR.TIME_MIX_LERP_FUSED: "blk.{bid}.time_mix_lerp_fused",
MODEL_TENSOR.TIME_MIX_LERP_W: "blk.{bid}.time_mix_lerp_w",
MODEL_TENSOR.TIME_MIX_FIRST: "blk.{bid}.time_mix_first",
MODEL_TENSOR.TIME_MIX_DECAY: "blk.{bid}.time_mix_decay",
@ -497,6 +582,24 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
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_TENSOR.CLS: "cls",
MODEL_TENSOR.CLS_OUT: "cls.output",
MODEL_TENSOR.CONV1D: "conv1d",
MODEL_TENSOR.CONVNEXT_DW: "convnext.{bid}.dw",
MODEL_TENSOR.CONVNEXT_NORM: "convnext.{bid}.norm",
MODEL_TENSOR.CONVNEXT_PW1: "convnext.{bid}.pw1",
MODEL_TENSOR.CONVNEXT_PW2: "convnext.{bid}.pw2",
MODEL_TENSOR.CONVNEXT_GAMMA: "convnext.{bid}.gamma",
MODEL_TENSOR.POSNET_CONV1: "posnet.{bid}.conv1",
MODEL_TENSOR.POSNET_CONV2: "posnet.{bid}.conv2",
MODEL_TENSOR.POSNET_NORM: "posnet.{bid}.norm",
MODEL_TENSOR.POSNET_NORM1: "posnet.{bid}.norm1",
MODEL_TENSOR.POSNET_NORM2: "posnet.{bid}.norm2",
MODEL_TENSOR.POSNET_ATTN_NORM: "posnet.{bid}.attn_norm",
MODEL_TENSOR.POSNET_ATTN_Q: "posnet.{bid}.attn_q",
MODEL_TENSOR.POSNET_ATTN_K: "posnet.{bid}.attn_k",
MODEL_TENSOR.POSNET_ATTN_V: "posnet.{bid}.attn_v",
MODEL_TENSOR.POSNET_ATTN_OUT: "posnet.{bid}.attn_output",
}
MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@ -520,6 +623,26 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
],
MODEL_ARCH.DECI: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_ROT_EMBD,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.FFN_GATE_EXP,
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
],
MODEL_ARCH.GROK: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
@ -606,6 +729,8 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.LAYER_OUT_NORM,
MODEL_TENSOR.CLS,
MODEL_TENSOR.CLS_OUT,
],
MODEL_ARCH.NOMIC_BERT: [
MODEL_TENSOR.TOKEN_EMBD,
@ -637,6 +762,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.LAYER_OUT_NORM,
MODEL_TENSOR.CLS,
],
MODEL_ARCH.MPT: [
MODEL_TENSOR.TOKEN_EMBD,
@ -723,6 +849,21 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.QWEN2: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.QWEN2VL: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
@ -800,6 +941,8 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ROPE_FACTORS_LONG,
MODEL_TENSOR.ROPE_FACTORS_SHORT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_Q,
@ -810,6 +953,24 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.PHIMOE: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ROPE_FACTORS_LONG,
MODEL_TENSOR.ROPE_FACTORS_SHORT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_GATE_EXP,
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
],
MODEL_ARCH.CODESHELL: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.POS_EMBD,
@ -859,6 +1020,8 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ROPE_FACTORS_LONG,
MODEL_TENSOR.ROPE_FACTORS_SHORT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
@ -874,6 +1037,25 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
],
MODEL_ARCH.MINICPM3: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ROPE_FACTORS_LONG,
MODEL_TENSOR.ROPE_FACTORS_SHORT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q_A,
MODEL_TENSOR.ATTN_Q_B,
MODEL_TENSOR.ATTN_KV_A_MQA,
MODEL_TENSOR.ATTN_KV_B,
MODEL_TENSOR.ATTN_Q_A_NORM,
MODEL_TENSOR.ATTN_KV_A_NORM,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.GEMMA: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
@ -932,6 +1114,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.TIME_MIX_LERP_R,
MODEL_TENSOR.TIME_MIX_LERP_G,
MODEL_TENSOR.TIME_MIX_LERP_W,
MODEL_TENSOR.TIME_MIX_LERP_FUSED,
MODEL_TENSOR.TIME_MIX_FIRST,
MODEL_TENSOR.TIME_MIX_DECAY,
MODEL_TENSOR.TIME_MIX_DECAY_W1,
@ -948,6 +1131,35 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.CHANNEL_MIX_RECEPTANCE,
MODEL_TENSOR.CHANNEL_MIX_VALUE,
],
MODEL_ARCH.RWKV6QWEN2: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.TIME_MIX_W1,
MODEL_TENSOR.TIME_MIX_W2,
MODEL_TENSOR.TIME_MIX_LERP_X,
MODEL_TENSOR.TIME_MIX_LERP_K,
MODEL_TENSOR.TIME_MIX_LERP_V,
MODEL_TENSOR.TIME_MIX_LERP_R,
MODEL_TENSOR.TIME_MIX_LERP_G,
MODEL_TENSOR.TIME_MIX_LERP_W,
MODEL_TENSOR.TIME_MIX_LERP_FUSED,
MODEL_TENSOR.TIME_MIX_FIRST,
MODEL_TENSOR.TIME_MIX_DECAY,
MODEL_TENSOR.TIME_MIX_DECAY_W1,
MODEL_TENSOR.TIME_MIX_DECAY_W2,
MODEL_TENSOR.TIME_MIX_KEY,
MODEL_TENSOR.TIME_MIX_VALUE,
MODEL_TENSOR.TIME_MIX_RECEPTANCE,
MODEL_TENSOR.TIME_MIX_GATE,
MODEL_TENSOR.TIME_MIX_LN,
MODEL_TENSOR.TIME_MIX_OUTPUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.MAMBA: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
@ -991,6 +1203,18 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.ATTN_K_NORM,
MODEL_TENSOR.ATTN_Q_NORM,
],
MODEL_ARCH.COHERE2: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.ATTN_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_ARCH.DBRX: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
@ -1015,6 +1239,39 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.OLMO2: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_POST_NORM,
MODEL_TENSOR.ATTN_Q_NORM,
MODEL_TENSOR.ATTN_K_NORM,
MODEL_TENSOR.FFN_POST_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.OLMOE: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q_NORM,
MODEL_TENSOR.ATTN_K_NORM,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_GATE_EXP,
MODEL_TENSOR.FFN_UP_EXP,
MODEL_TENSOR.FFN_DOWN_EXP,
],
MODEL_ARCH.OPENELM: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
@ -1049,6 +1306,29 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
],
MODEL_ARCH.DEEPSEEK: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_ROT_EMBD,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.FFN_GATE_EXP,
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
MODEL_TENSOR.FFN_GATE_SHEXP,
MODEL_TENSOR.FFN_DOWN_SHEXP,
MODEL_TENSOR.FFN_UP_SHEXP,
],
MODEL_ARCH.DEEPSEEK2: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
@ -1075,6 +1355,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_GATE_SHEXP,
MODEL_TENSOR.FFN_DOWN_SHEXP,
MODEL_TENSOR.FFN_UP_SHEXP,
MODEL_TENSOR.FFN_EXP_PROBS_B,
],
MODEL_ARCH.CHATGLM : [
MODEL_TENSOR.TOKEN_EMBD,
@ -1193,6 +1474,73 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.GRANITE: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.GRANITE_MOE: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_GATE_EXP,
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
],
MODEL_ARCH.CHAMELEON: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_Q_NORM,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_K_NORM,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.WAVTOKENIZER_DEC: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.TOKEN_EMBD_NORM,
MODEL_TENSOR.CONV1D,
MODEL_TENSOR.CONVNEXT_DW,
MODEL_TENSOR.CONVNEXT_NORM,
MODEL_TENSOR.CONVNEXT_PW1,
MODEL_TENSOR.CONVNEXT_PW2,
MODEL_TENSOR.CONVNEXT_GAMMA,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.POSNET_CONV1,
MODEL_TENSOR.POSNET_CONV2,
MODEL_TENSOR.POSNET_NORM,
MODEL_TENSOR.POSNET_NORM1,
MODEL_TENSOR.POSNET_NORM2,
MODEL_TENSOR.POSNET_ATTN_NORM,
MODEL_TENSOR.POSNET_ATTN_Q,
MODEL_TENSOR.POSNET_ATTN_K,
MODEL_TENSOR.POSNET_ATTN_V,
MODEL_TENSOR.POSNET_ATTN_OUT,
],
# TODO
}
@ -1202,6 +1550,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_ROT_EMBD,
],
MODEL_ARCH.DECI: [
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_ROT_EMBD,
],
MODEL_ARCH.BAICHUAN: [
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_ROT_EMBD,
@ -1226,6 +1578,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_ROT_EMBD,
],
MODEL_ARCH.DEEPSEEK: [
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_ROT_EMBD,
],
MODEL_ARCH.DEEPSEEK2: [
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_ROT_EMBD,
@ -1254,9 +1610,10 @@ class TokenType(IntEnum):
class RopeScalingType(Enum):
NONE = 'none'
LINEAR = 'linear'
YARN = 'yarn'
NONE = 'none'
LINEAR = 'linear'
YARN = 'yarn'
LONGROPE = 'longrope'
class PoolingType(IntEnum):
@ -1295,13 +1652,15 @@ class GGMLQuantizationType(IntEnum):
F64 = 28
IQ1_M = 29
BF16 = 30
Q4_0_4_4 = 31
Q4_0_4_8 = 32
Q4_0_8_8 = 33
TQ1_0 = 34
TQ2_0 = 35
class ExpertGatingFuncType(IntEnum):
SOFTMAX = 1
SIGMOID = 2
# TODO: add GGMLFileType from ggml_ftype in ggml.h
@ -1341,9 +1700,9 @@ class LlamaFileType(IntEnum):
MOSTLY_IQ4_XS = 30 # except 1d tensors
MOSTLY_IQ1_M = 31 # except 1d tensors
MOSTLY_BF16 = 32 # except 1d tensors
MOSTLY_Q4_0_4_4 = 33 # except 1d tensors
MOSTLY_Q4_0_4_8 = 34 # except 1d tensors
MOSTLY_Q4_0_8_8 = 35 # except 1d tensors
# MOSTLY_Q4_0_4_4 = 33 # removed from gguf files, use Q4_0 and runtime repack
# MOSTLY_Q4_0_4_8 = 34 # removed from gguf files, use Q4_0 and runtime repack
# MOSTLY_Q4_0_8_8 = 35 # removed from gguf files, use Q4_0 and runtime repack
MOSTLY_TQ1_0 = 36 # except 1d tensors
MOSTLY_TQ2_0 = 37 # except 1d tensors
@ -1419,9 +1778,6 @@ GGML_QUANT_SIZES: dict[GGMLQuantizationType, tuple[int, int]] = {
GGMLQuantizationType.F64: (1, 8),
GGMLQuantizationType.IQ1_M: (256, QK_K // 8 + QK_K // 16 + QK_K // 32),
GGMLQuantizationType.BF16: (1, 2),
GGMLQuantizationType.Q4_0_4_4:(32, 2 + 16),
GGMLQuantizationType.Q4_0_4_8:(32, 2 + 16),
GGMLQuantizationType.Q4_0_8_8:(32, 2 + 16),
GGMLQuantizationType.TQ1_0: (256, 2 + 4 * 13),
GGMLQuantizationType.TQ2_0: (256, 2 + 64),
}
@ -1482,15 +1838,23 @@ KEY_TOKENIZER_SCORES = Keys.Tokenizer.SCORES
KEY_TOKENIZER_MERGES = Keys.Tokenizer.MERGES
KEY_TOKENIZER_BOS_ID = Keys.Tokenizer.BOS_ID
KEY_TOKENIZER_EOS_ID = Keys.Tokenizer.EOS_ID
KEY_TOKENIZER_EOT_ID = Keys.Tokenizer.EOT_ID
KEY_TOKENIZER_EOM_ID = Keys.Tokenizer.EOM_ID
KEY_TOKENIZER_UNK_ID = Keys.Tokenizer.UNK_ID
KEY_TOKENIZER_SEP_ID = Keys.Tokenizer.SEP_ID
KEY_TOKENIZER_PAD_ID = Keys.Tokenizer.PAD_ID
KEY_TOKENIZER_CLS_ID = Keys.Tokenizer.CLS_ID
KEY_TOKENIZER_MASK_ID = Keys.Tokenizer.MASK_ID
KEY_TOKENIZER_HF_JSON = Keys.Tokenizer.HF_JSON
KEY_TOKENIZER_RWKV = Keys.Tokenizer.RWKV
KEY_TOKENIZER_PRIFIX_ID = Keys.Tokenizer.PREFIX_ID
KEY_TOKENIZER_FIM_PRE_ID = Keys.Tokenizer.FIM_PRE_ID
KEY_TOKENIZER_FIM_SUF_ID = Keys.Tokenizer.FIM_SUF_ID
KEY_TOKENIZER_FIM_MID_ID = Keys.Tokenizer.FIM_MID_ID
KEY_TOKENIZER_FIM_PAD_ID = Keys.Tokenizer.FIM_PAD_ID
KEY_TOKENIZER_FIM_REP_ID = Keys.Tokenizer.FIM_REP_ID
KEY_TOKENIZER_FIM_SEP_ID = Keys.Tokenizer.FIM_SEP_ID
# deprecated
KEY_TOKENIZER_PREFIX_ID = Keys.Tokenizer.PREFIX_ID
KEY_TOKENIZER_SUFFIX_ID = Keys.Tokenizer.SUFFIX_ID
KEY_TOKENIZER_MIDDLE_ID = Keys.Tokenizer.MIDDLE_ID
KEY_TOKENIZER_EOT_ID = Keys.Tokenizer.EOT_ID
KEY_TOKENIZER_EOM_ID = Keys.Tokenizer.EOM_ID

9
gguf-py/gguf/gguf_reader.py
View file

@ -145,11 +145,10 @@ class GGUFReader:
count = int(count)
itemsize = int(np.empty([], dtype = dtype).itemsize)
end_offs = offset + itemsize * count
return (
self.data[offset:end_offs]
.view(dtype = dtype)[:count]
.newbyteorder(override_order or self.byte_order)
)
arr = self.data[offset:end_offs].view(dtype=dtype)[:count]
if override_order is None:
return arr
return arr.view(arr.dtype.newbyteorder(override_order))
def _push_field(self, field: ReaderField, skip_sum: bool = False) -> int:
if field.name in self.fields:

94
gguf-py/gguf/gguf_writer.py
View file

@ -26,6 +26,7 @@ from .constants import (
RopeScalingType,
PoolingType,
TokenType,
ExpertGatingFuncType,
)
from .quants import quant_shape_from_byte_shape
@ -568,6 +569,9 @@ class GGUFWriter:
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_description(self, source_id: int, description: str) -> None:
self.add_string(Keys.General.BASE_MODEL_DESCRIPTION.format(id=source_id), description)
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)
@ -580,15 +584,42 @@ class GGUFWriter:
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_dataset_count(self, source_count: int) -> None:
self.add_uint32(Keys.General.DATASET_COUNT, source_count)
def add_dataset_name(self, source_id: int, name: str) -> None:
self.add_string(Keys.General.DATASET_NAME.format(id=source_id), name)
def add_dataset_author(self, source_id: int, author: str) -> None:
self.add_string(Keys.General.DATASET_AUTHOR.format(id=source_id), author)
def add_dataset_version(self, source_id: int, version: str) -> None:
self.add_string(Keys.General.DATASET_VERSION.format(id=source_id), version)
def add_dataset_organization(self, source_id: int, organization: str) -> None:
self.add_string(Keys.General.DATASET_ORGANIZATION.format(id=source_id), organization)
def add_dataset_description(self, source_id: int, description: str) -> None:
self.add_string(Keys.General.DATASET_DESCRIPTION.format(id=source_id), description)
def add_dataset_url(self, source_id: int, url: str) -> None:
self.add_string(Keys.General.DATASET_URL.format(id=source_id), url)
def add_dataset_doi(self, source_id: int, doi: str) -> None:
self.add_string(Keys.General.DATASET_DOI.format(id=source_id), doi)
def add_dataset_uuid(self, source_id: int, uuid: str) -> None:
self.add_string(Keys.General.DATASET_UUID.format(id=source_id), uuid)
def add_dataset_repo_url(self, source_id: int, repo_url: str) -> None:
self.add_string(Keys.General.DATASET_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)
@ -601,6 +632,21 @@ class GGUFWriter:
def add_embedding_length(self, length: int) -> None:
self.add_uint32(Keys.LLM.EMBEDDING_LENGTH.format(arch=self.arch), length)
def add_features_length(self, length: int) -> None:
self.add_uint32(Keys.LLM.FEATURES_LENGTH.format(arch=self.arch), length)
def add_posnet_embedding_length(self, length: int) -> None:
self.add_uint32(Keys.PosNet.EMBEDDING_LENGTH.format(arch=self.arch), length)
def add_posnet_block_count(self, length: int) -> None:
self.add_uint32(Keys.PosNet.BLOCK_COUNT.format(arch=self.arch), length)
def add_convnext_embedding_length(self, length: int) -> None:
self.add_uint32(Keys.ConvNext.EMBEDDING_LENGTH.format(arch=self.arch), length)
def add_convnext_block_count(self, length: int) -> None:
self.add_uint32(Keys.ConvNext.BLOCK_COUNT.format(arch=self.arch), length)
def add_block_count(self, length: int) -> None:
self.add_uint32(Keys.LLM.BLOCK_COUNT.format(arch=self.arch), length)
@ -670,6 +716,15 @@ class GGUFWriter:
def add_expert_weights_scale(self, value: float) -> None:
self.add_float32(Keys.LLM.EXPERT_WEIGHTS_SCALE.format(arch=self.arch), value)
def add_expert_weights_norm(self, value: bool) -> None:
self.add_bool(Keys.LLM.EXPERT_WEIGHTS_NORM.format(arch=self.arch), value)
def add_expert_gating_func(self, value: ExpertGatingFuncType) -> None:
self.add_uint32(Keys.LLM.EXPERT_GATING_FUNC.format(arch=self.arch), value.value)
def add_swin_norm(self, value: bool) -> None:
self.add_bool(Keys.LLM.SWIN_NORM.format(arch=self.arch), value)
def add_rescale_every_n_layers(self, count: int) -> None:
self.add_uint32(Keys.LLM.RESCALE_EVERY_N_LAYERS.format(arch=self.arch), count)
@ -679,15 +734,30 @@ class GGUFWriter:
def add_time_decay_extra_dim(self, dim: int) -> None:
self.add_uint32(Keys.LLM.TIME_DECAY_EXTRA_DIM.format(arch=self.arch), dim)
def add_residual_scale(self, value: float) -> None:
self.add_float32(Keys.LLM.RESIDUAL_SCALE.format(arch=self.arch), value)
def add_embedding_scale(self, value: float) -> None:
self.add_float32(Keys.LLM.EMBEDDING_SCALE.format(arch=self.arch), value)
def add_wkv_head_size(self, size: int) -> None:
self.add_uint32(Keys.WKV.HEAD_SIZE.format(arch=self.arch), size)
def add_token_shift_count(self, count: int) -> None:
self.add_uint32(Keys.LLM.TOKEN_SHIFT_COUNT.format(arch=self.arch), count)
def add_layer_norm_eps(self, value: float) -> None:
self.add_float32(Keys.Attention.LAYERNORM_EPS.format(arch=self.arch), value)
def add_layer_norm_rms_eps(self, value: float) -> None:
self.add_float32(Keys.Attention.LAYERNORM_RMS_EPS.format(arch=self.arch), value)
def add_group_norm_eps(self, value: float) -> None:
self.add_float32(Keys.Attention.GROUPNORM_EPS.format(arch=self.arch), value)
def add_group_norm_groups(self, value: int) -> None:
self.add_uint32(Keys.Attention.GROUPNORM_GROUPS.format(arch=self.arch), value)
def add_causal_attention(self, value: bool) -> None:
self.add_bool(Keys.Attention.CAUSAL.format(arch=self.arch), value)
@ -703,12 +773,18 @@ class GGUFWriter:
def add_sliding_window(self, value: int) -> None:
self.add_uint32(Keys.Attention.SLIDING_WINDOW.format(arch=self.arch), value)
def add_attention_scale(self, value: float) -> None:
self.add_float32(Keys.Attention.SCALE.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)
def add_rope_dimension_count(self, count: int) -> None:
self.add_uint32(Keys.Rope.DIMENSION_COUNT.format(arch=self.arch), count)
def add_rope_dimension_sections(self, dims: Sequence[int]) -> None:
self.add_array(Keys.Rope.DIMENSION_SECTIONS.format(arch=self.arch), dims)
def add_rope_freq_base(self, value: float) -> None:
self.add_float32(Keys.Rope.FREQ_BASE.format(arch=self.arch), value)
@ -781,9 +857,6 @@ class GGUFWriter:
def add_pad_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.PAD_ID, id)
def add_cls_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.CLS_ID, id)
def add_mask_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.MASK_ID, id)
@ -831,15 +904,6 @@ class GGUFWriter:
self.add_string(Keys.Tokenizer.CHAT_TEMPLATE, value)
def add_prefix_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.PREFIX_ID, id)
def add_suffix_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.SUFFIX_ID, id)
def add_middle_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.MIDDLE_ID, id)
def add_eot_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.EOT_ID, id)

150
gguf-py/gguf/metadata.py
View file

@ -41,7 +41,7 @@ class Metadata:
base_models: Optional[list[dict]] = None
tags: Optional[list[str]] = None
languages: Optional[list[str]] = None
datasets: Optional[list[str]] = None
datasets: Optional[list[dict]] = 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:
@ -91,9 +91,11 @@ class Metadata:
# Base Models is received here as an array of models
metadata.base_models = metadata_override.get("general.base_models", metadata.base_models)
# Datasets is received here as an array of datasets
metadata.datasets = metadata_override.get("general.datasets", metadata.datasets)
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:
@ -346,12 +348,12 @@ class Metadata:
use_model_card_metadata("author", "model_creator")
use_model_card_metadata("basename", "model_type")
if "base_model" in model_card:
if "base_model" in model_card or "base_models" in model_card or "base_model_sources" 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)
base_model_value = model_card.get("base_model", model_card.get("base_models", model_card.get("base_model_sources", None)))
if base_model_value is not None:
if isinstance(base_model_value, str):
@ -364,18 +366,106 @@ class Metadata:
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}"
if isinstance(model_id, str):
if model_id.startswith("http://") or model_id.startswith("https://") or model_id.startswith("ssh://"):
base_model["repo_url"] = model_id
# Check if Hugging Face ID is present in URL
if "huggingface.co" in model_id:
match = re.match(r"https?://huggingface.co/([^/]+/[^/]+)$", model_id)
if match:
model_id_component = match.group(1)
model_full_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(model_id_component, total_params)
# Populate model dictionary with extracted components
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
else:
# Likely a Hugging Face ID
model_full_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(model_id, total_params)
# Populate model dictionary with extracted components
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}"
elif isinstance(model_id, dict):
base_model = model_id
else:
logger.error(f"base model entry '{str(model_id)}' not in a known format")
metadata.base_models.append(base_model)
if "datasets" in model_card or "dataset" in model_card or "dataset_sources" in model_card:
# This represents the datasets that this was trained from
metadata_datasets = []
dataset_value = model_card.get("datasets", model_card.get("dataset", model_card.get("dataset_sources", None)))
if dataset_value is not None:
if isinstance(dataset_value, str):
metadata_datasets.append(dataset_value)
elif isinstance(dataset_value, list):
metadata_datasets.extend(dataset_value)
if metadata.datasets is None:
metadata.datasets = []
for dataset_id in metadata_datasets:
# NOTE: model size of base model is assumed to be similar to the size of the current model
dataset = {}
if isinstance(dataset_id, str):
if dataset_id.startswith(("http://", "https://", "ssh://")):
dataset["repo_url"] = dataset_id
# Check if Hugging Face ID is present in URL
if "huggingface.co" in dataset_id:
match = re.match(r"https?://huggingface.co/([^/]+/[^/]+)$", dataset_id)
if match:
dataset_id_component = match.group(1)
dataset_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(dataset_id_component, total_params)
# Populate dataset dictionary with extracted components
if dataset_name_component is not None:
dataset["name"] = Metadata.id_to_title(dataset_name_component)
if org_component is not None:
dataset["organization"] = Metadata.id_to_title(org_component)
if version is not None:
dataset["version"] = version
else:
# Likely a Hugging Face ID
dataset_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(dataset_id, total_params)
# Populate dataset dictionary with extracted components
if dataset_name_component is not None:
dataset["name"] = Metadata.id_to_title(dataset_name_component)
if org_component is not None:
dataset["organization"] = Metadata.id_to_title(org_component)
if version is not None:
dataset["version"] = version
if org_component is not None and dataset_name_component is not None:
dataset["repo_url"] = f"https://huggingface.co/{org_component}/{dataset_name_component}"
elif isinstance(dataset_id, dict):
dataset = dataset_id
else:
logger.error(f"dataset entry '{str(dataset_id)}' not in a known format")
metadata.datasets.append(dataset)
use_model_card_metadata("license", "license")
use_model_card_metadata("license_name", "license_name")
use_model_card_metadata("license_link", "license_link")
@ -386,9 +476,6 @@ class Metadata:
use_array_model_card_metadata("languages", "languages")
use_array_model_card_metadata("languages", "language")
use_array_model_card_metadata("datasets", "datasets")
use_array_model_card_metadata("datasets", "dataset")
# Hugging Face Parameter Heuristics
####################################
@ -458,7 +545,10 @@ class Metadata:
gguf_writer.add_size_label(self.size_label)
if self.license is not None:
gguf_writer.add_license(self.license)
if isinstance(self.license, list):
gguf_writer.add_license(",".join(self.license))
else:
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:
@ -493,6 +583,8 @@ class Metadata:
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 "description" in base_model_entry:
gguf_writer.add_base_model_description(key, base_model_entry["description"])
if "url" in base_model_entry:
gguf_writer.add_base_model_url(key, base_model_entry["url"])
if "doi" in base_model_entry:
@ -502,9 +594,29 @@ class Metadata:
if "repo_url" in base_model_entry:
gguf_writer.add_base_model_repo_url(key, base_model_entry["repo_url"])
if self.datasets is not None:
gguf_writer.add_dataset_count(len(self.datasets))
for key, dataset_entry in enumerate(self.datasets):
if "name" in dataset_entry:
gguf_writer.add_dataset_name(key, dataset_entry["name"])
if "author" in dataset_entry:
gguf_writer.add_dataset_author(key, dataset_entry["author"])
if "version" in dataset_entry:
gguf_writer.add_dataset_version(key, dataset_entry["version"])
if "organization" in dataset_entry:
gguf_writer.add_dataset_organization(key, dataset_entry["organization"])
if "description" in dataset_entry:
gguf_writer.add_dataset_description(key, dataset_entry["description"])
if "url" in dataset_entry:
gguf_writer.add_dataset_url(key, dataset_entry["url"])
if "doi" in dataset_entry:
gguf_writer.add_dataset_doi(key, dataset_entry["doi"])
if "uuid" in dataset_entry:
gguf_writer.add_dataset_uuid(key, dataset_entry["uuid"])
if "repo_url" in dataset_entry:
gguf_writer.add_dataset_repo_url(key, dataset_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)

0
gguf-py/scripts/__init__.py → gguf-py/gguf/scripts/__init__.py
View file

4
gguf-py/scripts/gguf_convert_endian.py → gguf-py/gguf/scripts/gguf_convert_endian.py
View file

@ -11,8 +11,8 @@ from pathlib import Path
import numpy as np
# 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))
if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent.parent / 'gguf-py').exists():
sys.path.insert(0, str(Path(__file__).parent.parent.parent))
import gguf

4
gguf-py/scripts/gguf_dump.py → gguf-py/gguf/scripts/gguf_dump.py
View file

@ -12,8 +12,8 @@ from typing import Any
import numpy as np
# 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))
if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent.parent / 'gguf-py').exists():
sys.path.insert(0, str(Path(__file__).parent.parent.parent))
from gguf import GGUFReader, GGUFValueType, ReaderTensor # noqa: E402

4
gguf-py/scripts/gguf_hash.py → gguf-py/gguf/scripts/gguf_hash.py
View file

@ -13,8 +13,8 @@ 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))
if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent.parent / 'gguf-py').exists():
sys.path.insert(0, str(Path(__file__).parent.parent.parent))
from gguf import GGUFReader # noqa: E402

4
gguf-py/scripts/gguf_new_metadata.py → gguf-py/gguf/scripts/gguf_new_metadata.py
View file

@ -13,8 +13,8 @@ from tqdm import tqdm
from typing import Any, Sequence, NamedTuple
# 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))
if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent.parent / 'gguf-py').exists():
sys.path.insert(0, str(Path(__file__).parent.parent.parent))
import gguf

4
gguf-py/scripts/gguf_set_metadata.py → gguf-py/gguf/scripts/gguf_set_metadata.py
View file

@ -6,8 +6,8 @@ import sys
from pathlib import Path
# 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))
if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent.parent / 'gguf-py').exists():
sys.path.insert(0, str(Path(__file__).parent.parent.parent))
from gguf import GGUFReader # noqa: E402

190
gguf-py/gguf/tensor_mapping.py
View file

@ -13,7 +13,7 @@ class TensorNameMap:
"transformer.wte", # gpt2 gpt-j mpt refact qwen dbrx jais exaone
"transformer.word_embeddings", # falcon
"word_embeddings", # bloom
"model.embed_tokens", # llama-hf nemotron
"model.embed_tokens", # llama-hf nemotron olmoe olmo2 rwkv6qwen2
"tok_embeddings", # llama-pth
"embeddings.word_embeddings", # bert nomic-bert
"language_model.embedding.word_embeddings", # persimmon
@ -42,6 +42,7 @@ class TensorNameMap:
"emb_ln", # nomic-bert
"transformer.norm", # openelm
"rwkv.blocks.0.pre_ln", # rwkv
"backbone.norm", # wavtokenizer
),
# Position embeddings
@ -54,19 +55,20 @@ class TensorNameMap:
# Output
MODEL_TENSOR.OUTPUT: (
"embed_out", # gptneox
"lm_head", # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone
"lm_head", # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone olmoe olmo2 phimoe
"output", # llama-pth bloom internlm2
"word_embeddings_for_head", # persimmon
"lm_head.linear", # phi2
"output_layer", # chatglm
"head", # rwkv
"head.out", # wavtokenizer
),
# Output norm
MODEL_TENSOR.OUTPUT_NORM: (
"gpt_neox.final_layer_norm", # gptneox
"transformer.ln_f", # gpt2 gpt-j falcon jais exaone
"model.norm", # llama-hf baichuan internlm2
"model.norm", # llama-hf baichuan internlm2 olmoe olmo2 phimoe
"norm", # llama-pth
"transformer.norm_f", # mpt dbrx
"ln_f", # refact bloom qwen gpt2
@ -80,6 +82,7 @@ class TensorNameMap:
"transformer.norm", # openelm
"model.norm", # nemotron
"rwkv.ln_out", # rwkv
"backbone.final_layer_norm", # wavtokenizer
),
# Rope frequencies
@ -87,6 +90,13 @@ class TensorNameMap:
"rope.freqs", # llama-pth
"rotary_pos_emb.inv_freq", # chatglm
),
MODEL_TENSOR.ROPE_FACTORS_LONG: (),
MODEL_TENSOR.ROPE_FACTORS_SHORT: (),
MODEL_TENSOR.CONV1D: (
"backbone.embed", # roberta
),
}
block_mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = {
@ -98,7 +108,7 @@ class TensorNameMap:
"transformer.h.{bid}.input_layernorm", # falcon7b
"h.{bid}.input_layernorm", # bloom
"transformer.h.{bid}.ln_mlp", # falcon40b
"model.layers.{bid}.input_layernorm", # llama-hf nemotron
"model.layers.{bid}.input_layernorm", # llama-hf nemotron olmoe phimoe
"layers.{bid}.attention_norm", # llama-pth
"language_model.encoder.layers.{bid}.input_layernorm", # persimmon
"model.layers.{bid}.ln1", # yi
@ -142,7 +152,8 @@ class TensorNameMap:
# Attention query
MODEL_TENSOR.ATTN_Q: (
"model.layers.{bid}.self_attn.q_proj", # llama-hf nemotron
"model.layers.{bid}.self_attn.q_proj", # llama-hf nemotron olmoe olmo2 phimoe
"model.layers.{bid}.self_attn.q_proj_no_perm", # llama-custom
"layers.{bid}.attention.wq", # llama-pth
"encoder.layer.{bid}.attention.self.query", # bert
"transformer.h.{bid}.attn.q_proj", # gpt-j
@ -154,7 +165,8 @@ class TensorNameMap:
# Attention key
MODEL_TENSOR.ATTN_K: (
"model.layers.{bid}.self_attn.k_proj", # llama-hf nemotron
"model.layers.{bid}.self_attn.k_proj", # llama-hf nemotron olmoe olmo2 phimoe
"model.layers.{bid}.self_attn.k_proj_no_perm", # llama-custom
"layers.{bid}.attention.wk", # llama-pth
"encoder.layer.{bid}.attention.self.key", # bert
"transformer.h.{bid}.attn.k_proj", # gpt-j
@ -167,7 +179,7 @@ class TensorNameMap:
# Attention value
MODEL_TENSOR.ATTN_V: (
"model.layers.{bid}.self_attn.v_proj", # llama-hf nemotron
"model.layers.{bid}.self_attn.v_proj", # llama-hf nemotron olmoe olmo2 phimoe
"layers.{bid}.attention.wv", # llama-pth
"encoder.layer.{bid}.attention.self.value", # bert
"transformer.h.{bid}.attn.v_proj", # gpt-j
@ -185,7 +197,8 @@ class TensorNameMap:
"transformer.blocks.{bid}.attn.out_proj", # mpt
"transformer.h.{bid}.self_attention.dense", # falcon
"h.{bid}.self_attention.dense", # bloom
"model.layers.{bid}.self_attn.o_proj", # llama-hf nemotron
"model.layers.{bid}.self_attn.o_proj", # llama-hf nemotron olmoe olmo2 phimoe
"model.layers.{bid}.self_attn.linear_attn", # deci
"layers.{bid}.attention.wo", # llama-pth
"encoder.layer.{bid}.attention.output.dense", # bert
"transformer.h.{bid}.attn.out_proj", # gpt-j
@ -212,7 +225,7 @@ class TensorNameMap:
),
MODEL_TENSOR.ATTN_POST_NORM: (
"model.layers.{bid}.post_attention_layernorm", # gemma2
"model.layers.{bid}.post_attention_layernorm", # gemma2 olmo2
),
# Rotary embeddings
@ -229,7 +242,7 @@ class TensorNameMap:
"transformer.h.{bid}.ln_2", # gpt2 refact qwen jais exaone
"h.{bid}.post_attention_layernorm", # bloom
"transformer.blocks.{bid}.norm_2", # mpt
"model.layers.{bid}.post_attention_layernorm", # llama-hf nemotron
"model.layers.{bid}.post_attention_layernorm", # llama-hf nemotron olmoe phimoe
"layers.{bid}.ffn_norm", # llama-pth
"language_model.encoder.layers.{bid}.post_attention_layernorm", # persimmon
"model.layers.{bid}.ln2", # yi
@ -247,21 +260,26 @@ class TensorNameMap:
# Post feed-forward norm
MODEL_TENSOR.FFN_POST_NORM: (
"model.layers.{bid}.post_feedforward_layernorm", # gemma2
"model.layers.{bid}.post_feedforward_layernorm", # gemma2 olmo2
),
MODEL_TENSOR.FFN_GATE_INP: (
"layers.{bid}.feed_forward.gate", # mixtral
"model.layers.{bid}.block_sparse_moe.gate", # mixtral
"model.layers.{bid}.mlp.gate", # qwen2moe
"transformer.decoder_layer.{bid}.router", # Grok
"transformer.blocks.{bid}.ffn.router.layer", # dbrx
"layers.{bid}.feed_forward.gate", # mixtral
"model.layers.{bid}.block_sparse_moe.gate", # mixtral phimoe
"model.layers.{bid}.mlp.gate", # qwen2moe olmoe
"transformer.decoder_layer.{bid}.router", # Grok
"transformer.blocks.{bid}.ffn.router.layer", # dbrx
"model.layers.{bid}.block_sparse_moe.router.layer", # granitemoe
),
MODEL_TENSOR.FFN_GATE_INP_SHEXP: (
"model.layers.{bid}.mlp.shared_expert_gate", # qwen2moe
),
MODEL_TENSOR.FFN_EXP_PROBS_B: (
"model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3
),
# Feed-forward up
MODEL_TENSOR.FFN_UP: (
"gpt_neox.layers.{bid}.mlp.dense_h_to_4h", # gptneox
@ -269,7 +287,7 @@ class TensorNameMap:
"transformer.blocks.{bid}.ffn.up_proj", # mpt
"transformer.h.{bid}.mlp.dense_h_to_4h", # falcon
"h.{bid}.mlp.dense_h_to_4h", # bloom
"model.layers.{bid}.mlp.up_proj", # llama-hf refact nemotron
"model.layers.{bid}.mlp.up_proj", # llama-hf refact nemotron olmo2
"layers.{bid}.feed_forward.w3", # llama-pth
"encoder.layer.{bid}.intermediate.dense", # bert
"transformer.h.{bid}.mlp.fc_in", # gpt-j
@ -292,15 +310,16 @@ class TensorNameMap:
),
MODEL_TENSOR.FFN_UP_EXP: (
"layers.{bid}.feed_forward.experts.w3", # mixtral (merged)
"transformer.decoder_layer.{bid}.moe.linear_v", # Grok (merged)
"transformer.blocks.{bid}.ffn.experts.mlp.v1", # dbrx
"model.layers.{bid}.mlp.experts.up_proj", # qwen2moe (merged)
"layers.{bid}.feed_forward.experts.w3", # mixtral (merged)
"transformer.decoder_layer.{bid}.moe.linear_v", # Grok (merged)
"transformer.blocks.{bid}.ffn.experts.mlp.v1", # dbrx
"model.layers.{bid}.mlp.experts.up_proj", # qwen2moe olmoe (merged)
"model.layers.{bid}.block_sparse_moe.experts.w3", # phimoe (merged)
),
MODEL_TENSOR.FFN_UP_SHEXP: (
"model.layers.{bid}.mlp.shared_expert.up_proj", # qwen2moe
"model.layers.{bid}.mlp.shared_experts.up_proj", # deepseek2
"model.layers.{bid}.mlp.shared_experts.up_proj", # deepseek deepseek2
),
# AWQ-activation gate
@ -310,7 +329,7 @@ class TensorNameMap:
# Feed-forward gate
MODEL_TENSOR.FFN_GATE: (
"model.layers.{bid}.mlp.gate_proj", # llama-hf refact
"model.layers.{bid}.mlp.gate_proj", # llama-hf refact olmo2
"layers.{bid}.feed_forward.w1", # llama-pth
"transformer.h.{bid}.mlp.w2", # qwen
"transformer.h.{bid}.mlp.c_fc2", # jais
@ -324,15 +343,16 @@ class TensorNameMap:
),
MODEL_TENSOR.FFN_GATE_EXP: (
"layers.{bid}.feed_forward.experts.w1", # mixtral (merged)
"transformer.decoder_layer.{bid}.moe.linear", # Grok (merged)
"transformer.blocks.{bid}.ffn.experts.mlp.w1", # dbrx
"model.layers.{bid}.mlp.experts.gate_proj", # qwen2moe (merged)
"layers.{bid}.feed_forward.experts.w1", # mixtral (merged)
"transformer.decoder_layer.{bid}.moe.linear", # Grok (merged)
"transformer.blocks.{bid}.ffn.experts.mlp.w1", # dbrx
"model.layers.{bid}.mlp.experts.gate_proj", # qwen2moe olmoe (merged)
"model.layers.{bid}.block_sparse_moe.experts.w1", # phimoe (merged)
),
MODEL_TENSOR.FFN_GATE_SHEXP: (
"model.layers.{bid}.mlp.shared_expert.gate_proj", # qwen2moe
"model.layers.{bid}.mlp.shared_experts.gate_proj", # deepseek2
"model.layers.{bid}.mlp.shared_experts.gate_proj", # deepseek deepseek2
),
# Feed-forward down
@ -342,7 +362,7 @@ class TensorNameMap:
"transformer.blocks.{bid}.ffn.down_proj", # mpt
"transformer.h.{bid}.mlp.dense_4h_to_h", # falcon
"h.{bid}.mlp.dense_4h_to_h", # bloom
"model.layers.{bid}.mlp.down_proj", # llama-hf nemotron
"model.layers.{bid}.mlp.down_proj", # llama-hf nemotron olmo2
"layers.{bid}.feed_forward.w2", # llama-pth
"encoder.layer.{bid}.output.dense", # bert
"transformer.h.{bid}.mlp.fc_out", # gpt-j
@ -364,21 +384,23 @@ class TensorNameMap:
),
MODEL_TENSOR.FFN_DOWN_EXP: (
"layers.{bid}.feed_forward.experts.w2", # mixtral (merged)
"transformer.decoder_layer.{bid}.moe.linear_1", # Grok (merged)
"transformer.blocks.{bid}.ffn.experts.mlp.w2", # dbrx
"model.layers.{bid}.mlp.experts.down_proj", # qwen2moe (merged)
"layers.{bid}.feed_forward.experts.w2", # mixtral (merged)
"transformer.decoder_layer.{bid}.moe.linear_1", # Grok (merged)
"transformer.blocks.{bid}.ffn.experts.mlp.w2", # dbrx
"model.layers.{bid}.mlp.experts.down_proj", # qwen2moe olmoe (merged)
"model.layers.{bid}.block_sparse_moe.output_linear", # granitemoe
"model.layers.{bid}.block_sparse_moe.experts.w2", # phimoe (merged)
),
MODEL_TENSOR.FFN_DOWN_SHEXP: (
"model.layers.{bid}.mlp.shared_expert.down_proj", # qwen2moe
"model.layers.{bid}.mlp.shared_experts.down_proj", # deepseek2
"model.layers.{bid}.mlp.shared_experts.down_proj", # deepseek deepseek2
),
MODEL_TENSOR.ATTN_Q_NORM: (
"language_model.encoder.layers.{bid}.self_attention.q_layernorm",
"model.layers.{bid}.self_attn.q_layernorm", # persimmon
"model.layers.{bid}.self_attn.q_norm", # cohere
"model.layers.{bid}.self_attn.q_norm", # cohere olmoe chameleon olmo2
"transformer.blocks.{bid}.attn.q_ln", # sea-lion
"encoder.layer.{bid}.attention.self.layer_norm_q", # jina-bert-v2
"transformer.layers.{bid}.attn.q_norm", # openelm
@ -387,7 +409,7 @@ class TensorNameMap:
MODEL_TENSOR.ATTN_K_NORM: (
"language_model.encoder.layers.{bid}.self_attention.k_layernorm",
"model.layers.{bid}.self_attn.k_layernorm", # persimmon
"model.layers.{bid}.self_attn.k_norm", # cohere
"model.layers.{bid}.self_attn.k_norm", # cohere olmoe chameleon olmo2
"transformer.blocks.{bid}.attn.k_ln", # sea-lion
"encoder.layer.{bid}.attention.self.layer_norm_k", # jina-bert-v2
"transformer.layers.{bid}.attn.k_norm", # openelm
@ -442,34 +464,42 @@ class TensorNameMap:
MODEL_TENSOR.TIME_MIX_W1: (
"rwkv.blocks.{bid}.attention.time_maa_w1", # rwkv v6
"model.layers.{bid}.self_attn.time_maa_w1", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_W2: (
"rwkv.blocks.{bid}.attention.time_maa_w2", # rwkv v6
"model.layers.{bid}.self_attn.time_maa_w2", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_LERP_X: (
"rwkv.blocks.{bid}.attention.time_maa_x", # rwkv v6
"model.layers.{bid}.self_attn.time_maa_x", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_LERP_K: (
"rwkv.blocks.{bid}.attention.time_maa_k", # rwkv v6
"model.layers.{bid}.self_attn.time_maa_k", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_LERP_V: (
"rwkv.blocks.{bid}.attention.time_maa_v", # rwkv v6
"model.layers.{bid}.self_attn.time_maa_v", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_LERP_R: (
"rwkv.blocks.{bid}.attention.time_maa_r", # rwkv v6
"model.layers.{bid}.self_attn.time_maa_r", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_LERP_G: (
"rwkv.blocks.{bid}.attention.time_maa_g", # rwkv v6
"model.layers.{bid}.self_attn.time_maa_g", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_LERP_W: (
"rwkv.blocks.{bid}.attention.time_maa_w", # rwkv v6
"model.layers.{bid}.self_attn.time_maa_w", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_FIRST: (
@ -478,30 +508,37 @@ class TensorNameMap:
MODEL_TENSOR.TIME_MIX_DECAY: (
"rwkv.blocks.{bid}.attention.time_decay", # rwkv v6
"model.layers.{bid}.self_attn.time_decay", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_DECAY_W1: (
"rwkv.blocks.{bid}.attention.time_decay_w1", # rwkv v6
"model.layers.{bid}.self_attn.time_decay_w1", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_DECAY_W2: (
"rwkv.blocks.{bid}.attention.time_decay_w2", # rwkv v6
"model.layers.{bid}.self_attn.time_decay_w2", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_KEY: (
"rwkv.blocks.{bid}.attention.key", # rwkv
"rwkv.blocks.{bid}.attention.key", # rwkv
"model.layers.{bid}.self_attn.k_proj", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_VALUE: (
"rwkv.blocks.{bid}.attention.value", # rwkv
"rwkv.blocks.{bid}.attention.value", # rwkv
"model.layers.{bid}.self_attn.v_proj", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_RECEPTANCE: (
"rwkv.blocks.{bid}.attention.receptance", # rwkv
"model.layers.{bid}.self_attn.q_proj", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_GATE: (
"rwkv.blocks.{bid}.attention.gate", # rwkv
"rwkv.blocks.{bid}.attention.gate", # rwkv
"model.layers.{bid}.self_attn.gate", # rwkv6qwen2
),
MODEL_TENSOR.TIME_MIX_LN: (
@ -509,7 +546,8 @@ class TensorNameMap:
),
MODEL_TENSOR.TIME_MIX_OUTPUT: (
"rwkv.blocks.{bid}.attention.output", # rwkv
"rwkv.blocks.{bid}.attention.output", # rwkv
"model.layers.{bid}.self_attn.o_proj", # rwkv6qwen2
),
MODEL_TENSOR.CHANNEL_MIX_LERP_K: (
@ -674,9 +712,81 @@ class TensorNameMap:
"encoder.block.{bid}.layer.1.DenseReluDense.wo", # t5
),
############################################################################
# TODO: these do not belong to block_mappings_cfg - move them to mappings_cfg
MODEL_TENSOR.ENC_OUTPUT_NORM: (
"encoder.final_layer_norm", # t5
),
MODEL_TENSOR.CLS: (
"classifier", # jina
"classifier.dense", # roberta
),
MODEL_TENSOR.CLS_OUT: (
"classifier.out_proj", # roberta
),
#############################################################################
MODEL_TENSOR.CONVNEXT_DW: (
"backbone.convnext.{bid}.dwconv", # wavtokenizer
),
MODEL_TENSOR.CONVNEXT_NORM: (
"backbone.convnext.{bid}.norm", # wavtokenizer
),
MODEL_TENSOR.CONVNEXT_PW1: (
"backbone.convnext.{bid}.pwconv1", # wavtokenizer
),
MODEL_TENSOR.CONVNEXT_PW2: (
"backbone.convnext.{bid}.pwconv2", # wavtokenizer
),
MODEL_TENSOR.CONVNEXT_GAMMA: (
"backbone.convnext.{bid}.gamma", # wavtokenizer
),
MODEL_TENSOR.POSNET_CONV1: (
"backbone.posnet.{bid}.conv1", # wavtokenizer
),
MODEL_TENSOR.POSNET_CONV2: (
"backbone.posnet.{bid}.conv2", # wavtokenizer
),
MODEL_TENSOR.POSNET_NORM: (
"backbone.posnet.{bid}.norm", # wavtokenizer
),
MODEL_TENSOR.POSNET_NORM1: (
"backbone.posnet.{bid}.norm1", # wavtokenizer
),
MODEL_TENSOR.POSNET_NORM2: (
"backbone.posnet.{bid}.norm2", # wavtokenizer
),
MODEL_TENSOR.POSNET_ATTN_NORM: (
"backbone.posnet.{bid}.norm", # wavtokenizer
),
MODEL_TENSOR.POSNET_ATTN_Q: (
"backbone.posnet.{bid}.q", # wavtokenizer
),
MODEL_TENSOR.POSNET_ATTN_K: (
"backbone.posnet.{bid}.k", # wavtokenizer
),
MODEL_TENSOR.POSNET_ATTN_V: (
"backbone.posnet.{bid}.v", # wavtokenizer
),
MODEL_TENSOR.POSNET_ATTN_OUT: (
"backbone.posnet.{bid}.proj_out", # wavtokenizer
),
}
# architecture-specific block mappings

26
gguf-py/gguf/vocab.py
View file

@ -122,8 +122,30 @@ class SpecialVocab:
tokenizer = json.load(f)
if self.load_merges:
merges = tokenizer.get('model', {}).get('merges')
if isinstance(merges, list) and merges and isinstance(merges[0], str):
self.merges = merges
if isinstance(merges, list) and merges:
if isinstance(merges[0], str):
self.merges = merges
elif isinstance(merges[0], list) and len(merges[0]) == 2 and isinstance(merges[0][0], str):
# New format since transformers 4.45 to support spaces in merges
# ref: https://github.com/ggerganov/llama.cpp/issues/9692
# TODO: internally store as the new format instead of converting to old
if any(' ' in s for pair in merges for s in pair):
logger.warning(f'Spaces in merges detected, encoding as {chr(ord(" ") + 256)!r}')
self.merges = [
' '.join(
[
# ensure the spaces are properly encoded
''.join(
chr(ord(c) + 256) if c == ' ' else c
for c in part
)
for part in pair
]
)
for pair in merges
]
else:
raise ValueError("Unknown tokenizer merges format")
added_tokens = tokenizer.get('added_tokens', {})
else:
added_tokens = {}

11
gguf-py/pyproject.toml
View file

@ -1,12 +1,11 @@
[tool.poetry]
name = "gguf"
version = "0.10.0"
version = "0.15.0"
description = "Read and write ML models in GGUF for GGML"
authors = ["GGML <ggml@ggml.ai>"]
packages = [
{include = "gguf"},
{include = "gguf/py.typed"},
{include = "scripts"},
]
readme = "README.md"
homepage = "https://ggml.ai"
@ -33,7 +32,7 @@ requires = ["poetry-core>=1.0.0"]
build-backend = "poetry.core.masonry.api"
[tool.poetry.scripts]
gguf-convert-endian = "scripts:gguf_convert_endian_entrypoint"
gguf-dump = "scripts:gguf_dump_entrypoint"
gguf-set-metadata = "scripts:gguf_set_metadata_entrypoint"
gguf-new-metadata = "scripts:gguf_new_metadata_entrypoint"
gguf-convert-endian = "gguf.scripts:gguf_convert_endian_entrypoint"
gguf-dump = "gguf.scripts:gguf_dump_entrypoint"
gguf-set-metadata = "gguf.scripts:gguf_set_metadata_entrypoint"
gguf-new-metadata = "gguf.scripts:gguf_new_metadata_entrypoint"

37
gguf-py/tests/test_metadata.py
View file

@ -182,8 +182,43 @@ class TestMetadataMethod(unittest.TestCase):
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']
expect.datasets=[{'name': 'OpenHermes 2.5', 'organization': 'Teknium', 'version': '2.5', 'repo_url': 'https://huggingface.co/teknium/OpenHermes-2.5'}]
self.assertEqual(got, expect)
# Base Model spec is inferred from model id
model_card = {'base_models': 'teknium/OpenHermes-2.5'}
expect = gguf.Metadata(base_models=[{'name': 'OpenHermes 2.5', 'organization': 'Teknium', 'version': '2.5', 'repo_url': 'https://huggingface.co/teknium/OpenHermes-2.5'}])
got = gguf.Metadata.apply_metadata_heuristic(gguf.Metadata(), model_card, None, None)
self.assertEqual(got, expect)
# Base Model spec is only url
model_card = {'base_models': ['https://huggingface.co/teknium/OpenHermes-2.5']}
expect = gguf.Metadata(base_models=[{'name': 'OpenHermes 2.5', 'organization': 'Teknium', 'version': '2.5', 'repo_url': 'https://huggingface.co/teknium/OpenHermes-2.5'}])
got = gguf.Metadata.apply_metadata_heuristic(gguf.Metadata(), model_card, None, None)
self.assertEqual(got, expect)
# Base Model spec is given directly
model_card = {'base_models': [{'name': 'OpenHermes 2.5', 'organization': 'Teknium', 'version': '2.5', 'repo_url': 'https://huggingface.co/teknium/OpenHermes-2.5'}]}
expect = gguf.Metadata(base_models=[{'name': 'OpenHermes 2.5', 'organization': 'Teknium', 'version': '2.5', 'repo_url': 'https://huggingface.co/teknium/OpenHermes-2.5'}])
got = gguf.Metadata.apply_metadata_heuristic(gguf.Metadata(), model_card, None, None)
self.assertEqual(got, expect)
# Dataset spec is inferred from model id
model_card = {'datasets': 'teknium/OpenHermes-2.5'}
expect = gguf.Metadata(datasets=[{'name': 'OpenHermes 2.5', 'organization': 'Teknium', 'version': '2.5', 'repo_url': 'https://huggingface.co/teknium/OpenHermes-2.5'}])
got = gguf.Metadata.apply_metadata_heuristic(gguf.Metadata(), model_card, None, None)
self.assertEqual(got, expect)
# Dataset spec is only url
model_card = {'datasets': ['https://huggingface.co/teknium/OpenHermes-2.5']}
expect = gguf.Metadata(datasets=[{'name': 'OpenHermes 2.5', 'organization': 'Teknium', 'version': '2.5', 'repo_url': 'https://huggingface.co/teknium/OpenHermes-2.5'}])
got = gguf.Metadata.apply_metadata_heuristic(gguf.Metadata(), model_card, None, None)
self.assertEqual(got, expect)
# Dataset spec is given directly
model_card = {'datasets': [{'name': 'OpenHermes 2.5', 'organization': 'Teknium', 'version': '2.5', 'repo_url': 'https://huggingface.co/teknium/OpenHermes-2.5'}]}
expect = gguf.Metadata(datasets=[{'name': 'OpenHermes 2.5', 'organization': 'Teknium', 'version': '2.5', 'repo_url': 'https://huggingface.co/teknium/OpenHermes-2.5'}])
got = gguf.Metadata.apply_metadata_heuristic(gguf.Metadata(), model_card, None, None)
self.assertEqual(got, expect)
def test_apply_metadata_heuristic_from_hf_parameters(self):

2
gguf-py/tests/test_quants.py
View file

@ -136,7 +136,7 @@ def compare_tensors(t1: np.ndarray, t2: np.ndarray, qtype: GGMLQuantizationType)
logger.debug(f"Sample bad block ({diff_bits[bad_block_id]} differing bits):\n{t1[bad_block_id]}\nReference:\n{t2[bad_block_id]}")
sum_diff_bits = np.sum(diff_bits)
logger.debug(f"{sum_diff_bits} bits differ ({100 * sum_diff_bits/(x.size * 8):.6f}%)")
logger.debug(f"{sum_diff_bits} bits differ ({100 * sum_diff_bits / (x.size * 8):.6f}%)")
return False