[Online quantization] Extend re-quantization API to MOE (showcase NVFP4 modelopt/CT/quark to any online MOE method) (#60276)

Signed-off-by: Felix Marty <Felix.Marty@amd.com>
Signed-off-by: Tan Pin Siang <tanpinsiang@gmail.com>
Co-authored-by: vllmellm <vllm.ellm@embeddedllm.com>
Co-authored-by: Tan Pin Siang <tanpinsiang@gmail.com>
This commit is contained in:
fxmarty-amd
2026-10-14 15:49:37 +08:00
committed by GitHub
co-authored by vllmellm Tan Pin Siang
parent e1726d9142
commit c9bfa574ce
16 changed files with 1353 additions and 148 deletions
@@ -0,0 +1,45 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import torch
from vllm.model_executor.layers.quantization.base_config import QuantizeMethodBase
class OnlineRequantizationMixin:
"""Shared lifecycle for converting serialized weights before quantization."""
def __init__(self, *args, **kwargs) -> None:
super().__init__(*args, **kwargs)
self.requantization_source: QuantizeMethodBase | None = None
self.requantization_source_parameters: dict[str, torch.nn.Parameter] = {}
def set_requantization_source(self, source_method: QuantizeMethodBase) -> None:
"""Configure serialized-weight conversion before online quantization."""
self.requantization_source = source_method
self.uses_meta_device = False
def create_requantization_source_weights(
self, layer: torch.nn.Module, *weight_args, **extra_weight_attrs
) -> bool:
"""Delegate weight creation to the source and track its parameters."""
if self.requantization_source is None:
return False
existing_parameter_names = set(layer._parameters)
self.requantization_source.create_weights(
layer, *weight_args, **extra_weight_attrs
)
self.requantization_source_parameters = {
name: parameter
for name, parameter in layer._parameters.items()
if name not in existing_parameter_names and parameter is not None
}
return True
def release_requantization_source_weights(self, layer: torch.nn.Module) -> None:
"""Release serialized parameters after successful requantization."""
for name, source_parameter in self.requantization_source_parameters.items():
if layer._parameters.get(name) is source_parameter:
delattr(layer, name)
self.requantization_source_parameters.clear()