Support FP8 Quantization and Inference Run on Intel Gaudi (HPU) using INC (Intel Neural Compressor) (#12010)
Signed-off-by: Nir David <ndavid@habana.ai> Signed-off-by: Uri Livne <ulivne@habana.ai> Co-authored-by: Uri Livne <ulivne@habana.ai>
This commit is contained in:
@@ -36,6 +36,7 @@ QuantizationMethods = Literal[
|
||||
"torchao",
|
||||
"auto-round",
|
||||
"rtn",
|
||||
"inc",
|
||||
]
|
||||
QUANTIZATION_METHODS: list[str] = list(get_args(QuantizationMethods))
|
||||
|
||||
@@ -104,6 +105,7 @@ def get_quantization_config(quantization: str) -> type[QuantizationConfig]:
|
||||
from .gptq_marlin import GPTQMarlinConfig
|
||||
from .gptq_marlin_24 import GPTQMarlin24Config
|
||||
from .hqq_marlin import HQQMarlinConfig
|
||||
from .inc import INCConfig
|
||||
from .ipex_quant import IPEXConfig
|
||||
from .marlin import MarlinConfig
|
||||
from .modelopt import ModelOptFp8Config, ModelOptNvFp4Config
|
||||
@@ -144,7 +146,8 @@ def get_quantization_config(quantization: str) -> type[QuantizationConfig]:
|
||||
"moe_wna16": MoeWNA16Config,
|
||||
"torchao": TorchAOConfig,
|
||||
"auto-round": AutoRoundConfig,
|
||||
"rtn": RTNConfig
|
||||
"rtn": RTNConfig,
|
||||
"inc": INCConfig,
|
||||
}
|
||||
# Update the `method_to_config` with customized quantization methods.
|
||||
method_to_config.update(_CUSTOMIZED_METHOD_TO_QUANT_CONFIG)
|
||||
@@ -157,4 +160,4 @@ __all__ = [
|
||||
"QuantizationMethods",
|
||||
"get_quantization_config",
|
||||
"QUANTIZATION_METHODS",
|
||||
]
|
||||
]
|
||||
|
||||
@@ -0,0 +1,61 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
#
|
||||
# Intel Gaudi supports quantization of various modules and functions,
|
||||
# including, but not limited to `Linear`, `KVCache`, `Matmul` and `Softmax`.
|
||||
# During model loading,
|
||||
# INC will patch layers with quantization/dequantization operators.
|
||||
# Meanwhile, INC will convert original weight to target datatype
|
||||
# and loading to target device.
|
||||
# static scaling should be provided through Quant_CONFIG:
|
||||
# `QUANT_CONFIG` is an environment variable,
|
||||
# that points to the measurement or quantization JSON config file.
|
||||
# The measurement configuration file is used during the calibration procedure,
|
||||
# to collect measurements for a given model.
|
||||
# The quantization configuration is used during inference.
|
||||
# For more information, please refer to:
|
||||
# https://docs.habana.ai/en/v1.21.1/PyTorch/vLLM_Inference/vLLM_FP8_Inference.html
|
||||
|
||||
from typing import Any, Optional
|
||||
|
||||
import torch
|
||||
|
||||
from vllm.model_executor.layers.fused_moe.layer import (
|
||||
FusedMoE, UnquantizedFusedMoEMethod)
|
||||
from vllm.model_executor.layers.linear import (LinearBase,
|
||||
UnquantizedLinearMethod)
|
||||
from vllm.model_executor.layers.quantization import QuantizationMethods
|
||||
from vllm.model_executor.layers.quantization.base_config import (
|
||||
QuantizationConfig, QuantizeMethodBase)
|
||||
|
||||
|
||||
class INCConfig(QuantizationConfig):
|
||||
"""Config class for FP8 using Intel Neural Compressor."""
|
||||
|
||||
@classmethod
|
||||
def get_name(cls) -> QuantizationMethods:
|
||||
return "inc"
|
||||
|
||||
@classmethod
|
||||
def get_supported_act_dtypes(cls) -> list[torch.dtype]:
|
||||
return [torch.bfloat16]
|
||||
|
||||
@classmethod
|
||||
def from_config(cls, config: dict[str, Any]) -> "INCConfig":
|
||||
raise AssertionError
|
||||
|
||||
def get_quant_method(self, layer: torch.nn.Module,
|
||||
prefix: str) -> Optional["QuantizeMethodBase"]:
|
||||
if isinstance(layer, LinearBase):
|
||||
return UnquantizedLinearMethod()
|
||||
elif isinstance(layer, FusedMoE):
|
||||
return UnquantizedFusedMoEMethod(layer.moe_config)
|
||||
return None
|
||||
|
||||
@classmethod
|
||||
def get_min_capability(cls) -> int:
|
||||
raise AssertionError
|
||||
|
||||
@staticmethod
|
||||
def get_config_filenames() -> list[str]:
|
||||
return []
|
||||
@@ -6,9 +6,12 @@ import torch
|
||||
import torch.nn as nn
|
||||
|
||||
from vllm.config import LoadConfig, ModelConfig, VllmConfig
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.model_loader.utils import (
|
||||
initialize_model, process_weights_after_loading, set_default_torch_dtype)
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
class BaseModelLoader(ABC):
|
||||
"""Base class for model loaders."""
|
||||
@@ -32,11 +35,16 @@ class BaseModelLoader(ABC):
|
||||
model_config: ModelConfig) -> nn.Module:
|
||||
"""Load a model with the given configurations."""
|
||||
device_config = vllm_config.device_config
|
||||
target_device = torch.device(device_config.device)
|
||||
load_config = vllm_config.load_config
|
||||
load_device = device_config.device if load_config.device is None else \
|
||||
load_config.device
|
||||
target_device = torch.device(load_device)
|
||||
with set_default_torch_dtype(model_config.dtype):
|
||||
with target_device:
|
||||
model = initialize_model(vllm_config=vllm_config,
|
||||
model_config=model_config)
|
||||
|
||||
logger.debug("Loading weights on %s ...", load_device)
|
||||
# Quantization does not happen in `load_weights` but after it
|
||||
self.load_weights(model, model_config)
|
||||
process_weights_after_loading(model, model_config, target_device)
|
||||
|
||||
@@ -152,8 +152,8 @@ def get_quant_config(model_config: ModelConfig,
|
||||
quant_cls = get_quantization_config(model_config.quantization)
|
||||
|
||||
# GGUF doesn't have config file
|
||||
if model_config.quantization == "gguf":
|
||||
return quant_cls.from_config({})
|
||||
if model_config.quantization in ("gguf", "inc"):
|
||||
return quant_cls()
|
||||
|
||||
# Read the quantization config from the HF model config, if available.
|
||||
hf_quant_config = getattr(model_config.hf_config, "quantization_config",
|
||||
|
||||
Reference in New Issue
Block a user