[Quantization] Support online quantization with partially pre-quantized checkpoints (#51392)

Signed-off-by: Felix Marty <Felix.Marty@amd.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Codex <codex@openai.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
fxmarty-amd
2026-09-04 11:46:06 -07:00
committed by GitHub
co-authored by Codex mgoin
parent a11dfcff89
commit 5690b02c03
25 changed files with 846 additions and 175 deletions
@@ -12,6 +12,7 @@ if TYPE_CHECKING:
)
from vllm.model_executor.layers.fused_moe import RoutedExperts
from vllm.model_executor.layers.fused_moe.config import FusedMoEConfig
from vllm.model_executor.layers.fused_moe.oracle.int8 import (
convert_to_int8_moe_kernel_format,
make_int8_moe_kernel,
@@ -38,9 +39,9 @@ class Int8OnlineMoEMethod(OnlineMoEMethodBase):
def __init__(
self,
*,
layer: torch.nn.Module,
moe: FusedMoEConfig,
):
super().__init__(layer.moe_config)
super().__init__(moe)
self.int8_backend, self.experts_cls = select_int8_moe_backend(
config=self.moe,
weight_key=kInt8StaticChannelSym,