SYLARandHongsheng Liu
d1e230c95b
Signed-off-by: lishunyang12 <lishunyang12@163.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
31 lines
1.1 KiB
Python
31 lines
1.1 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
"""Unified quantization framework for vLLM-OMNI.
|
|
|
|
Delegates to vLLM's quantization registry (35+ methods, all platforms).
|
|
Adds per-component quantization for multi-stage models.
|
|
|
|
from vllm_omni.quantization import build_quant_config
|
|
|
|
config = build_quant_config("fp8")
|
|
config = build_quant_config({"transformer": {"method": "fp8"}, "vae": None})
|
|
"""
|
|
|
|
from .component_config import ComponentQuantizationConfig, resolve_component_quant_config
|
|
from .factory import SUPPORTED_QUANTIZATION_METHODS, build_quant_config, register_quantization_override
|
|
from .inc_config import OmniINCConfig
|
|
|
|
# Heavy configs are NOT imported here to avoid pulling in
|
|
# optional dependencies (pynvml, torch_npu) at module load time.
|
|
# Import them directly when needed:
|
|
# from vllm_omni.quantization.mxfp8_config import DiffusionMXFP8Config
|
|
|
|
__all__ = [
|
|
"build_quant_config",
|
|
"ComponentQuantizationConfig",
|
|
"resolve_component_quant_config",
|
|
"OmniINCConfig",
|
|
"SUPPORTED_QUANTIZATION_METHODS",
|
|
"register_quantization_override",
|
|
]
|