[gpt-oss] flashinfer mxfp4 (#22339)

Signed-off-by: simon-mo <xmo@berkeley.edu>
Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
Co-authored-by: simon-mo <xmo@berkeley.edu>
This commit is contained in:
Yongye Zhu
2025-08-06 12:37:27 -07:00
committed by GitHub
co-authored by simon-mo
parent 31f5dc5b2a
commit 31f09c615f
5 changed files with 453 additions and 3 deletions
@@ -1,5 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Callable, Optional
import torch
from vllm.utils import direct_register_custom_op
@@ -7,6 +9,26 @@ from vllm.utils import direct_register_custom_op
OCP_MX_BLOCK_SIZE = 32
def _can_support_mxfp4(use_grouped_topk: bool = False,
topk_group: Optional[int] = None,
num_expert_group: Optional[int] = None,
expert_map: Optional[torch.Tensor] = None,
custom_routing_function: Optional[Callable] = None,
e_score_correction_bias: Optional[torch.Tensor] = None,
apply_router_weight_on_input: bool = False,
scoring_func: str = "softmax",
activation: str = "silu",
expert_load_view: Optional[torch.Tensor] = None,
logical_to_physical_map: Optional[torch.Tensor] = None,
logical_replica_count: Optional[torch.Tensor] = None):
return not (use_grouped_topk or topk_group or num_expert_group
or expert_map or custom_routing_function
or e_score_correction_bias or apply_router_weight_on_input
or scoring_func != "softmax" or activation != "silu"
or expert_load_view or logical_to_physical_map
or logical_replica_count)
def _dequant_mxfp4(x: torch.Tensor, scale: torch.Tensor,
float_dtype: torch.dtype) -> torch.Tensor:
try: