Implement AWQ quantization support for LLaMA (#1032)

Co-authored-by: Robert Irvine <robert@seamlessml.com>
Co-authored-by: root <rirv938@gmail.com>
Co-authored-by: Casper <casperbh.96@gmail.com>
Co-authored-by: julian-q <julianhquevedo@gmail.com>
This commit is contained in:
Woosuk Kwon
2023-09-16 00:03:37 -07:00
committed by GitHub
co-authored by Robert Irvine root Casper julian-q
parent b9fe4616f9
commit e3e79e9e8a
19 changed files with 1178 additions and 208 deletions
+28 -2
View File
@@ -8,7 +8,8 @@ from transformers import PretrainedConfig
from vllm.config import ModelConfig
from vllm.model_executor.models import * # pylint: disable=wildcard-import
from vllm.model_executor.weight_utils import initialize_dummy_weights
from vllm.model_executor.weight_utils import (get_quant_config,
initialize_dummy_weights)
# TODO(woosuk): Lazy-load the model classes.
_MODEL_REGISTRY = {
@@ -30,6 +31,11 @@ _MODEL_REGISTRY = {
"RWForCausalLM": FalconForCausalLM,
}
# FIXME(woosuk): Remove this once all models support quantization.
_MODEL_CLASSES_SUPPORT_QUANTIZATION = [
LlamaForCausalLM,
]
@contextlib.contextmanager
def _set_default_torch_dtype(dtype: torch.dtype):
@@ -52,10 +58,30 @@ def _get_model_architecture(config: PretrainedConfig) -> Type[nn.Module]:
def get_model(model_config: ModelConfig) -> nn.Module:
model_class = _get_model_architecture(model_config.hf_config)
# Get the quantization config.
quant_config = None
if model_config.quantization is not None:
if model_class not in _MODEL_CLASSES_SUPPORT_QUANTIZATION:
raise ValueError(
f"Quantization is not supported for {model_class}.")
quant_config = get_quant_config(model_config.quantization,
model_config.model,
model_config.download_dir)
supported_dtypes = quant_config.get_supported_act_dtypes()
if model_config.dtype not in supported_dtypes:
raise ValueError(
f"{model_config.dtype} is not supported for quantization "
f"method {model_config.quantization}. Supported dtypes: "
f"{supported_dtypes}")
with _set_default_torch_dtype(model_config.dtype):
# Create a model instance.
# The weights will be initialized as empty tensors.
model = model_class(model_config.hf_config)
if model_class in _MODEL_CLASSES_SUPPORT_QUANTIZATION:
model = model_class(model_config.hf_config, quant_config)
else:
model = model_class(model_config.hf_config)
if model_config.load_format == "dummy":
model = model.cuda()
# NOTE(woosuk): For accurate performance evaluation, we assign