mirror/OpenLLM
1
0
Fork 0
mirror of https://github.com/bentoml/OpenLLM.git synced 2026-03-10 11:10:27 -04:00
Code Issues Packages Projects Releases Wiki Activity

perf: unify LLM interface (#518)

Browse Source 
Signed-off-by: Aaron <29749331+aarnphm@users.noreply.github.com>
Signed-off-by: paperspace <29749331+aarnphm@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
This commit is contained in:
Aaron Pham
2023-11-06 20:39:43 -05:00
committed by GitHub
parent f2639879af
commit e2029c934b
136 changed files with 9646 additions and 11244 deletions
Download Patch File Download Diff File Expand all files Collapse all files

57
openllm-python/src/openllm/_quantisation.py
View File

@@ -8,6 +8,8 @@ import transformers
from openllm_core._typing_compat import LiteralQuantise
from openllm_core._typing_compat import overload
from openllm_core.exceptions import MissingDependencyError
from openllm_core.utils import is_autoawq_available
from openllm_core.utils import is_autogptq_available
from openllm_core.utils import is_bitsandbytes_available
from openllm_core.utils import is_optimum_supports_gptq
@@ -20,25 +22,36 @@ if t.TYPE_CHECKING:
logger = logging.getLogger(__name__)
@overload
def infer_quantisation_config(cls: type[LLM[t.Any, t.Any]], quantise: t.Literal['int8', 'int4'], **attrs: t.Any) -> tuple[transformers.BitsAndBytesConfig, DictStrAny]:
def infer_quantisation_config(self: LLM[t.Any, t.Any], quantise: t.Literal['int8', 'int4'], **attrs: t.Any) -> tuple[transformers.BitsAndBytesConfig, DictStrAny]:
...
@overload
def infer_quantisation_config(cls: type[LLM[t.Any, t.Any]], quantise: t.Literal['gptq'], **attrs: t.Any) -> tuple[transformers.GPTQConfig, DictStrAny]:
def infer_quantisation_config(self: LLM[t.Any, t.Any], quantise: t.Literal['gptq'], **attrs: t.Any) -> tuple[transformers.GPTQConfig, DictStrAny]:
...
def infer_quantisation_config(cls: type[LLM[t.Any, t.Any]], quantise: LiteralQuantise, **attrs: t.Any) -> tuple[transformers.BitsAndBytesConfig | transformers.GPTQConfig, DictStrAny]:
@overload
def infer_quantisation_config(self: LLM[t.Any, t.Any], quantise: t.Literal['awq'], **attrs: t.Any) -> tuple[transformers.AwqConfig, DictStrAny]:
...
def infer_quantisation_config(self: LLM[t.Any, t.Any], quantise: LiteralQuantise,
**attrs: t.Any) -> tuple[transformers.BitsAndBytesConfig | transformers.GPTQConfig | transformers.AwqConfig, DictStrAny]:
# 8 bit configuration
int8_threshold = attrs.pop('llm_int8_threshhold', 6.0)
int8_enable_fp32_cpu_offload = attrs.pop('llm_int8_enable_fp32_cpu_offload', False)
int8_skip_modules: list[str] | None = attrs.pop('llm_int8_skip_modules', None)
int8_has_fp16_weight = attrs.pop('llm_int8_has_fp16_weight', False)
# shared arguments for gptq and awq
bits = attrs.pop('bits', 4)
group_size = attrs.pop('group_size', 128)
def create_awq_config() -> transformers.AwqConfig:
zero_point = attrs.pop('zero_point', True)
return transformers.AwqConfig(bits=bits, group_size=group_size, zero_point=zero_point)
def create_gptq_config() -> transformers.GPTQConfig:
gptq_bits = attrs.pop('bits', 4)
gptq_tokenizer = attrs.pop('tokenizer', None)
gptq_tokenizer = attrs.pop('tokenizer', self.model_id)
gptq_dataset = attrs.pop('dataset', 'c4')
gptq_group_size = attrs.pop('group_size', 128)
gptq_damp_percent = attrs.pop('damp_percent', 0.1)
gptq_desc_act = attrs.pop('desc_act', False)
gptq_sym = attrs.pop('sym', True)
@@ -50,10 +63,10 @@ def infer_quantisation_config(cls: type[LLM[t.Any, t.Any]], quantise: LiteralQua
gptq_batch_size = attrs.pop('batch_size', 1)
gptq_pad_token_id = attrs.pop('pad_token_id', None)
gptq_disable_exllama = attrs.pop('disable_exllama', False)
return transformers.GPTQConfig(bits=gptq_bits,
return transformers.GPTQConfig(bits=bits,
tokenizer=gptq_tokenizer,
dataset=gptq_dataset,
group_size=gptq_group_size,
group_size=group_size,
damp_percent=gptq_damp_percent,
desc_act=gptq_desc_act,
sym=gptq_sym,
@@ -67,25 +80,22 @@ def infer_quantisation_config(cls: type[LLM[t.Any, t.Any]], quantise: LiteralQua
disable_exllama=gptq_disable_exllama)
def create_int8_config(int8_skip_modules: list[str] | None) -> transformers.BitsAndBytesConfig:
if int8_skip_modules is None: int8_skip_modules = []
if 'lm_head' not in int8_skip_modules and cls.config_class.__openllm_model_type__ == 'causal_lm':
logger.debug("Skipping 'lm_head' for quantization for %s", cls.__name__)
int8_skip_modules.append('lm_head')
# if int8_skip_modules is None: int8_skip_modules = []
# if 'lm_head' not in int8_skip_modules and self.config_class.__openllm_model_type__ == 'causal_lm':
# logger.debug("Skipping 'lm_head' for quantization for %s", self.__name__)
# int8_skip_modules.append('lm_head')
return transformers.BitsAndBytesConfig(load_in_8bit=True,
llm_int8_enable_fp32_cpu_offload=int8_enable_fp32_cpu_offload,
llm_int8_threshhold=int8_threshold,
llm_int8_skip_modules=int8_skip_modules,
llm_int8_has_fp16_weight=int8_has_fp16_weight,
)
llm_int8_has_fp16_weight=int8_has_fp16_weight)
# 4 bit configuration
int4_compute_dtype = attrs.pop('bnb_4bit_compute_dtype', torch.bfloat16)
int4_quant_type = attrs.pop('bnb_4bit_quant_type', 'nf4')
int4_use_double_quant = attrs.pop('bnb_4bit_use_double_quant', True)
# NOTE: Quantization setup
# quantize is a openllm.LLM feature, where we can quantize the model
# with bitsandbytes or quantization aware training.
# NOTE: Quantization setup quantize is a openllm.LLM feature, where we can quantize the model with bitsandbytes or quantization aware training.
if not is_bitsandbytes_available():
raise RuntimeError("Quantization requires bitsandbytes to be installed. Make sure to install OpenLLM with 'pip install \"openllm[fine-tune]\"'")
if quantise == 'int8': quantisation_config = create_int8_config(int8_skip_modules)
@@ -96,12 +106,15 @@ def infer_quantisation_config(cls: type[LLM[t.Any, t.Any]], quantise: LiteralQua
bnb_4bit_use_double_quant=int4_use_double_quant)
elif quantise == 'gptq':
if not is_autogptq_available() or not is_optimum_supports_gptq():
logger.warning(
"'quantize=\"gptq\"' requires 'auto-gptq' and 'optimum>=0.12' to be installed (not available with local environment). Make sure to have 'auto-gptq' available locally: 'pip install \"openllm[gptq]\"'. OpenLLM will fallback to int8 with bitsandbytes."
)
quantisation_config = create_int8_config(int8_skip_modules)
raise MissingDependencyError(
"'quantize=\"gptq\"' requires 'auto-gptq' and 'optimum>=0.12' to be installed (missing or failed to import). Make sure to do 'pip install \"openllm[gptq]\"'")
else:
quantisation_config = create_gptq_config()
elif quantise == 'awq':
if not is_autoawq_available():
raise MissingDependencyError("quantize='awq' requires 'auto-awq' to be installed (missing or failed to import). Make sure to do 'pip install \"openllm[awq]\"'.")
else:
quantisation_config = create_awq_config()
else:
raise ValueError(f"'quantize' must be one of ['int8', 'int4', 'gptq'], got {quantise} instead.")
raise ValueError(f"'quantize' must be one of ['int8', 'int4', 'gptq', 'awq'], got {quantise} instead.")
return quantisation_config, attrs