[Refactor] Move NVFP4 GEMM management into NvFp4LinearKernel (#39129)

Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2026-04-09 15:05:36 -04:00
committed by GitHub
parent 0d310ffbeb
commit 2800706f06
11 changed files with 697 additions and 377 deletions
@@ -10,7 +10,10 @@ from torch.nn.parameter import Parameter
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm.config import get_current_vllm_config
from vllm.logger import init_logger
from vllm.model_executor.kernels.linear import init_fp8_linear_kernel
from vllm.model_executor.kernels.linear import (
init_fp8_linear_kernel,
init_nvfp4_linear_kernel,
)
from vllm.model_executor.layers.attention import Attention, MLAAttention
from vllm.model_executor.layers.fused_moe.activation import MoEActivation
from vllm.model_executor.layers.fused_moe.config import (
@@ -70,12 +73,6 @@ from vllm.model_executor.layers.quantization.utils.mxfp8_utils import (
Mxfp8LinearOp,
mxfp8_e4m3_quantize,
)
from vllm.model_executor.layers.quantization.utils.nvfp4_utils import (
NvFp4LinearBackend,
apply_nvfp4_linear,
convert_to_nvfp4_linear_kernel_format,
select_nvfp4_linear_backend,
)
from vllm.model_executor.layers.quantization.utils.quant_utils import (
GroupShape,
create_fp8_quant_key,
@@ -1090,11 +1087,7 @@ class ModelOptNvFp4LinearMethod(LinearMethodBase):
def __init__(self, quant_config: ModelOptNvFp4Config) -> None:
self.quant_config = quant_config
self.marlin_input_dtype = None
self.backend = select_nvfp4_linear_backend()
self.swizzle = None
if self.backend == NvFp4LinearBackend.EMULATION:
self.swizzle = False
self.kernel = init_nvfp4_linear_kernel()
def create_weights(
self,
@@ -1201,7 +1194,7 @@ class ModelOptNvFp4LinearMethod(LinearMethodBase):
)
# Convert layer to NVFP4 linear kernel format
convert_to_nvfp4_linear_kernel_format(self.backend, layer)
self.kernel.process_weights_after_loading(layer)
def apply(
self,
@@ -1209,13 +1202,7 @@ class ModelOptNvFp4LinearMethod(LinearMethodBase):
x: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
return apply_nvfp4_linear(
backend=self.backend,
layer=layer,
x=x,
bias=bias,
swizzle=self.swizzle,
)
return self.kernel.apply_weights(layer=layer, x=x, bias=bias)
class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):