TP/quantization/weight loading refactor part 2 - Refactor quantized linear logic and extend quantization support to all models (#1622)

Refactor the tensor parallelism, quantization, and weight-loading codes.

Summary of the new features enabled by this PR:
- **All models** are able to be quantized with AWQ and SqueezeLLM, and [soon GPTQ](https://github.com/vllm-project/vllm/pull/1580).
- Model loading code became much simpler.
- Support model parallelism for all MQA/GQA models when the number of key/value heads is smaller than the tensor parallel size.
This commit is contained in:
Zhuohan Li
2023-11-15 22:50:41 -08:00
committed by GitHub
parent 660a7fcfa4
commit 7076fa1c9f
36 changed files with 2159 additions and 2508 deletions
+4 -16
View File
@@ -37,13 +37,6 @@ _MODEL_REGISTRY = {
"YiForCausalLM": YiForCausalLM,
}
# FIXME(woosuk): Remove this once all models support quantization.
_MODEL_CLASSES_SUPPORT_QUANTIZATION = [
LlamaForCausalLM,
MistralForCausalLM,
YiForCausalLM,
]
@contextlib.contextmanager
def _set_default_torch_dtype(dtype: torch.dtype):
@@ -67,12 +60,9 @@ 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
# Get the (maybe quantized) linear method.
linear_method = 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)
@@ -90,14 +80,12 @@ def get_model(model_config: ModelConfig) -> nn.Module:
f"{model_config.dtype} is not supported for quantization "
f"method {model_config.quantization}. Supported dtypes: "
f"{supported_dtypes}")
linear_method = quant_config.get_linear_method()
with _set_default_torch_dtype(model_config.dtype):
# Create a model instance.
# The weights will be initialized as empty tensors.
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)
model = model_class(model_config.hf_config, linear_method)
if model_config.load_format == "dummy":
model = model.cuda()
# NOTE(woosuk): For accurate performance evaluation, we assign