[CPU] Support cpu compressed-tensor w8a8 int8 moe (#42920)

Signed-off-by: yuwenzho <yuwen.zhou@intel.com>
Signed-off-by: Yuwen Zhou <yuwen.zhou@intel.com>
This commit is contained in:
Yuwen Zhou
2026-06-29 03:04:05 +00:00
committed by GitHub
parent a2abce646f
commit 58d6a6e60a
6 changed files with 410 additions and 9 deletions
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import torch
from compressed_tensors.quantization import (
QuantizationArgs,
@@ -20,6 +19,7 @@ from vllm.model_executor.layers.fused_moe.config import (
FusedMoEQuantConfig,
)
from vllm.model_executor.layers.fused_moe.oracle.int8 import (
convert_to_int8_moe_kernel_format,
make_int8_moe_kernel,
make_int8_moe_quant_config,
select_int8_moe_backend,
@@ -31,7 +31,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
kInt8DynamicTokenSym,
kInt8StaticChannelSym,
)
from vllm.model_executor.utils import set_weight_attrs
from vllm.model_executor.utils import replace_parameter, set_weight_attrs
logger = init_logger(__name__)
@@ -142,6 +142,14 @@ class CompressedTensorsW8A8Int8MoEMethod(CompressedTensorsMoEMethod):
layer.w2_input_scale = None
def process_weights_after_loading(self, layer: RoutedExperts) -> None:
w13, w2 = convert_to_int8_moe_kernel_format(
int8_backend=self.int8_backend,
w13=layer.w13_weight,
w2=layer.w2_weight,
)
replace_parameter(layer, "w13_weight", w13)
replace_parameter(layer, "w2_weight", w2)
self.moe_quant_config = self.get_fused_moe_quant_config(layer)
assert self.experts_cls is not None
self.moe_kernel = make_int8_moe_kernel(
@@ -193,3 +201,27 @@ class CompressedTensorsW8A8Int8MoEMethod(CompressedTensorsMoEMethod):
shared_experts=shared_experts,
shared_experts_input=shared_experts_input,
)
def apply_monolithic(
self,
layer: RoutedExperts,
x: torch.Tensor,
router_logits: torch.Tensor,
input_ids: torch.Tensor | None = None,
) -> torch.Tensor:
assert self.is_monolithic
assert self.moe_kernel is not None
return self.moe_kernel.apply_monolithic(
x,
layer.w13_weight,
layer.w2_weight,
router_logits,
activation=layer.activation,
global_num_experts=layer.global_num_experts,
expert_map=layer.expert_map,
apply_router_weight_on_input=layer.apply_router_weight_on_input,
num_expert_group=layer.num_expert_group,
topk_group=layer.topk_group,
e_score_correction_bias=layer.e_score_correction_bias,
routed_scaling_factor=layer.routed_scaling_factor,
)