ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-05-01 11:51:53 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge mainline - Aug 12 2024 (#17)

Browse Source 
* Merge mainline

* Fix after merge

* Remove CI check

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2024-08-12 15:14:32 +02:00
committed by GitHub
parent f5d1af61d7
commit 8f43e55103
164 changed files with 9041 additions and 3404 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
gguf-py/gguf/constants.py
View File

@@ -161,6 +161,7 @@ class Keys:
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"
@@ -216,6 +217,7 @@ class MODEL_ARCH(IntEnum):
CHATGLM = auto()
BITNET = auto()
T5 = auto()
T5ENCODER = auto()
JAIS = auto()
@@ -343,6 +345,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.CHATGLM: "chatglm",
MODEL_ARCH.BITNET: "bitnet",
MODEL_ARCH.T5: "t5",
MODEL_ARCH.T5ENCODER: "t5encoder",
MODEL_ARCH.JAIS: "jais",
}
@@ -1035,6 +1038,21 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.ENC_FFN_UP,
MODEL_TENSOR.ENC_OUTPUT_NORM,
],
MODEL_ARCH.T5ENCODER: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT,
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,
@@ -1145,6 +1163,19 @@ 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
IQ1_BN = 34,
IQ2_BN = 35,
Q8_K64 = 36,
IQ2_K = 37,
IQ3_K = 38,
IQ4_K = 39,
IQ5_K = 40,
IQ6_K = 41,
IQ2_TN = 42,
# TODO: add GGMLFileType from ggml_ftype in ggml.h
@@ -1157,7 +1188,7 @@ class LlamaFileType(IntEnum):
MOSTLY_F16 = 1 # except 1d tensors
MOSTLY_Q4_0 = 2 # except 1d tensors
MOSTLY_Q4_1 = 3 # except 1d tensors
MOSTLY_Q4_1_SOME_F16 = 4 # tok_embeddings.weight and output.weight are F16
# MOSTLY_Q4_1_SOME_F16 = 4 # tok_embeddings.weight and output.weight are F16
# MOSTLY_Q4_2 = 5 # support has been removed
# MOSTLY_Q4_3 = 6 # support has been removed
MOSTLY_Q8_0 = 7 # except 1d tensors
@@ -1186,6 +1217,18 @@ 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_IQ1_BN = 36, # except 1d tensors
MOSTLY_IQ2_BN = 37, # except 1d tensors
MOSTLY_IQ2_K = 38, # except 1d tensors
MOSTLY_IQ3_K = 39, # except 1d tensors
MOSTLY_IQ4_K = 40, # except 1d tensors
MOSTLY_IQ5_K = 41, # except 1d tensors
MOSTLY_IQ6_K = 42, # except 1d tensors
MOSTLY_IQ2_TN = 43, # except 1d tensors
GUESSED = 1024 # not specified in the model file
@@ -1259,6 +1302,9 @@ 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),
}
@@ -1327,3 +1373,4 @@ KEY_TOKENIZER_PRIFIX_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

14
gguf-py/gguf/gguf_writer.py
View File

@@ -312,6 +312,8 @@ class GGUFWriter:
self.add_key_value(key, val, GGUFValueType.STRING)
def add_array(self, key: str, val: Sequence[Any]) -> None:
if len(val) == 0:
return
self.add_key_value(key, val, GGUFValueType.ARRAY)
@staticmethod
@@ -826,6 +828,9 @@ class GGUFWriter:
def add_eot_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.EOT_ID, id)
def add_eom_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.EOM_ID, id)
def _pack(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> bytes:
pack_prefix = ''
if not skip_pack_prefix:
@@ -845,7 +850,14 @@ class GGUFWriter:
encoded_val = val.encode("utf-8") if isinstance(val, str) else val
kv_data += self._pack("Q", len(encoded_val))
kv_data += encoded_val
elif vtype == GGUFValueType.ARRAY and isinstance(val, Sequence) and val:
elif vtype == GGUFValueType.ARRAY:
if not isinstance(val, Sequence):
raise ValueError("Invalid GGUF metadata array, expecting sequence")
if len(val) == 0:
raise ValueError("Invalid GGUF metadata array. Empty array")
if isinstance(val, bytes):
ltype = GGUFValueType.UINT8
else:

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

@@ -191,6 +191,8 @@ class LazyBase(ABC, metaclass=LazyMeta):
class LazyNumpyTensor(LazyBase):
_tensor_type = np.ndarray
shape: tuple[int, ...] # Makes the type checker happy in quants.py
@classmethod
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,

131
gguf-py/gguf/metadata.py
View File

@@ -174,7 +174,7 @@ class Metadata:
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] == '.':
if org_component is not None and len(org_component) > 0 and org_component[0] == '.':
org_component = None
name_parts: list[str] = model_full_name_component.split('-')
@@ -284,20 +284,67 @@ class Metadata:
########################
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")
def use_model_card_metadata(metadata_key: str, model_card_key: str):
if model_card_key in model_card and getattr(metadata, metadata_key, None) is None:
setattr(metadata, metadata_key, model_card.get(model_card_key))
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")
def use_array_model_card_metadata(metadata_key: str, model_card_key: str):
# Note: Will append rather than replace if already exist
tags_value = model_card.get(model_card_key, None)
if tags_value is None:
return
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")
current_value = getattr(metadata, metadata_key, None)
if current_value is None:
current_value = []
if isinstance(tags_value, str):
current_value.append(tags_value)
elif isinstance(tags_value, list):
current_value.extend(tags_value)
setattr(metadata, metadata_key, current_value)
# LLAMA.cpp's direct internal convention
# (Definitely not part of hugging face formal/informal standard)
#########################################
use_model_card_metadata("name", "name")
use_model_card_metadata("author", "author")
use_model_card_metadata("version", "version")
use_model_card_metadata("organization", "organization")
use_model_card_metadata("description", "description")
use_model_card_metadata("finetune", "finetune")
use_model_card_metadata("basename", "basename")
use_model_card_metadata("size_label", "size_label")
use_model_card_metadata("source_url", "url")
use_model_card_metadata("source_doi", "doi")
use_model_card_metadata("source_uuid", "uuid")
use_model_card_metadata("source_repo_url", "repo_url")
# LLAMA.cpp's huggingface style convention
# (Definitely not part of hugging face formal/informal standard... but with model_ appended to match their style)
###########################################
use_model_card_metadata("name", "model_name")
use_model_card_metadata("author", "model_author")
use_model_card_metadata("version", "model_version")
use_model_card_metadata("organization", "model_organization")
use_model_card_metadata("description", "model_description")
use_model_card_metadata("finetune", "model_finetune")
use_model_card_metadata("basename", "model_basename")
use_model_card_metadata("size_label", "model_size_label")
use_model_card_metadata("source_url", "model_url")
use_model_card_metadata("source_doi", "model_doi")
use_model_card_metadata("source_uuid", "model_uuid")
use_model_card_metadata("source_repo_url", "model_repo_url")
# Hugging Face Direct Convention
#################################
# 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'
use_model_card_metadata("name", "model_name")
use_model_card_metadata("author", "model_creator")
use_model_card_metadata("basename", "model_type")
if "base_model" in model_card:
# This represents the parent models that this is based on
@@ -329,58 +376,18 @@ class Metadata:
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")
use_model_card_metadata("license", "license")
use_model_card_metadata("license_name", "license_name")
use_model_card_metadata("license_link", "license_link")
if "license_name" in model_card and metadata.license_name is None:
metadata.license_name = model_card.get("license_name")
use_array_model_card_metadata("tags", "tags")
use_array_model_card_metadata("tags", "pipeline_tag")
if "license_link" in model_card and metadata.license_link is None:
metadata.license_link = model_card.get("license_link")
use_array_model_card_metadata("languages", "languages")
use_array_model_card_metadata("languages", "language")
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)
use_array_model_card_metadata("datasets", "datasets")
use_array_model_card_metadata("datasets", "dataset")
# Hugging Face Parameter Heuristics
####################################

1217
gguf-py/gguf/quants.py
View File

File diff suppressed because it is too large Load Diff