[Kernels] Clean up FusedMoeMethodBase and modular kernel setup. Remove extra arguments from modular kernel methods. (#22035)

Signed-off-by: Bill Nell <bnell@redhat.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
This commit is contained in:
bnellnm
2025-08-15 14:46:00 -04:00
committed by GitHub
co-authored by Michael Goin
parent 48b01fd4d4
commit 8ad7285ea2
54 changed files with 2010 additions and 1293 deletions
@@ -7,7 +7,8 @@ import torch
from vllm import _custom_ops as ops
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe import (FusedMoE, FusedMoEMethodBase,
from vllm.model_executor.layers.fused_moe import (FusedMoE, FusedMoEConfig,
FusedMoEMethodBase,
FusedMoeWeightScaleSupported)
from vllm.model_executor.layers.quantization.utils.mxfp4_utils import (
OCP_MX_BLOCK_SIZE)
@@ -25,6 +26,9 @@ __all__ = [
class QuarkMoEMethod(FusedMoEMethodBase):
def __init__(self, moe: FusedMoEConfig):
super().__init__(moe)
@staticmethod
def get_moe_method(
quant_config: "QuarkConfig", # type: ignore # noqa E501 # noqa F821
@@ -42,17 +46,24 @@ class QuarkMoEMethod(FusedMoEMethodBase):
input_config = layer_quant_config.get("input_tensors")
if quant_config._is_fp8_w8a8(weight_config, input_config):
return QuarkW8A8Fp8MoEMethod(weight_config, input_config)
return QuarkW8A8Fp8MoEMethod(weight_config, input_config,
module.moe_config)
elif quant_config._is_mx_fp4(weight_config, input_config):
return QuarkW4A4MXFp4MoEMethod(weight_config, input_config)
return QuarkW4A4MXFp4MoEMethod(weight_config, input_config,
module.moe_config)
else:
raise RuntimeError("Unsupported FusedMoe scheme")
class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
def __init__(self, weight_config: dict[str, Any], input_config: dict[str,
Any]):
def __init__(
self,
weight_config: dict[str, Any],
input_config: dict[str, Any],
moe: FusedMoEConfig,
):
super().__init__(moe)
self.weight_quant = weight_config
self.input_quant = input_config
@@ -215,6 +226,8 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
logical_to_physical_map: Optional[torch.Tensor] = None,
logical_replica_count: Optional[torch.Tensor] = None,
) -> torch.Tensor:
assert self.fused_experts is None
if enable_eplb:
raise NotImplementedError(
"EPLB not supported for `QuarkW8A8Fp8MoEMethod` yet.")
@@ -231,7 +244,8 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
num_expert_group=num_expert_group,
custom_routing_function=custom_routing_function,
scoring_func=scoring_func,
e_score_correction_bias=e_score_correction_bias)
e_score_correction_bias=e_score_correction_bias,
indices_type=self.topk_indices_dtype)
return fused_experts(
x,
@@ -253,8 +267,13 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
class QuarkW4A4MXFp4MoEMethod(QuarkMoEMethod):
def __init__(self, weight_config: dict[str, Any], input_config: dict[str,
Any]):
def __init__(
self,
weight_config: dict[str, Any],
input_config: dict[str, Any],
moe: FusedMoEConfig,
):
super().__init__(moe)
self.weight_quant = weight_config
self.input_quant = input_config
@@ -369,6 +388,7 @@ class QuarkW4A4MXFp4MoEMethod(QuarkMoEMethod):
logical_to_physical_map: Optional[torch.Tensor] = None,
logical_replica_count: Optional[torch.Tensor] = None,
) -> torch.Tensor:
assert self.fused_experts is None
if enable_eplb:
raise NotImplementedError(
@@ -386,7 +406,8 @@ class QuarkW4A4MXFp4MoEMethod(QuarkMoEMethod):
num_expert_group=num_expert_group,
custom_routing_function=custom_routing_function,
scoring_func=scoring_func,
e_score_correction_bias=e_score_correction_bias)
e_score_correction_bias=e_score_correction_bias,
indices_type=self.topk_indices_dtype)
out = fused_experts(
x,