[Quantization] Layer re-quantization for linear layers through online quantization API (MXFP8 -> FP8 PTPC showcase) (#55684)

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>
Co-authored-by: Misha Goin <mgoin64@gmail.com>
This commit is contained in:
fxmarty-amd
2026-10-06 11:38:18 -05:00
committed by GitHub
co-authored by vllmellm Tan Pin Siang Misha Goin
parent 29f955cce0
commit f8d92cae41
12 changed files with 422 additions and 45 deletions
@@ -2,6 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import inspect
from collections.abc import Callable
from contextlib import nullcontext
from functools import wraps
from weakref import WeakKeyDictionary, WeakSet
@@ -321,7 +322,21 @@ def _reload_attention_scales(layer: torch.nn.Module, info: LayerReloadingInfo) -
_get_weight_loader(param)(*args.args, **args.kwargs)
if quant_method is not None:
quant_method.process_weights_after_loading(layer)
from vllm.model_executor.layers.quantization.online.fp8 import OnlineLinearBase
from vllm.model_executor.layers.quantization.online.moe_base import (
OnlineMoEMethodBase,
)
from vllm.model_executor.model_loader.utils import (
online_quantization_timing_context,
)
timing_context = (
online_quantization_timing_context(layer, quant_method)
if isinstance(quant_method, (OnlineLinearBase, OnlineMoEMethodBase))
else nullcontext()
)
with timing_context:
quant_method.process_weights_after_loading(layer)
_copy_and_restore_kernel_tensors(layer, info)
@@ -356,7 +371,21 @@ def _layerwise_process(layer: torch.nn.Module, info: LayerReloadingInfo):
# Process weights (quantization, repacking, etc.)
quant_method = getattr(layer, "quant_method", None)
if isinstance(quant_method, QuantizeMethodBase):
quant_method.process_weights_after_loading(layer)
from vllm.model_executor.layers.quantization.online.fp8 import OnlineLinearBase
from vllm.model_executor.layers.quantization.online.moe_base import (
OnlineMoEMethodBase,
)
from vllm.model_executor.model_loader.utils import (
online_quantization_timing_context,
)
timing_context = (
online_quantization_timing_context(layer, quant_method)
if isinstance(quant_method, (OnlineLinearBase, OnlineMoEMethodBase))
else nullcontext()
)
with timing_context:
quant_method.process_weights_after_loading(layer)
# Re-reconcile parameter TP state: process_weights_after_loading may
# have re-created Parameters (stamped with the global rank), which would
# otherwise break replicated (disable_tp) weights on a subsequent reload.