DP/EP Support for gpt-oss with deepep-ht comm kernel on SM100 (#23608)

This commit is contained in:
Yongye Zhu
2025-08-27 17:33:21 -04:00
committed by GitHub
parent 853c371fc3
commit 082cc07ef8
11 changed files with 365 additions and 12 deletions
@@ -66,11 +66,10 @@ def _can_support_mxfp4(use_grouped_topk: bool = False,
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 != "swigluoai"
or expert_load_view or logical_to_physical_map
or logical_replica_count)
or custom_routing_function or e_score_correction_bias
or apply_router_weight_on_input or scoring_func != "softmax"
or activation != "swigluoai" or expert_load_view
or logical_to_physical_map or logical_replica_count)
def _dequant_mxfp4(x: torch.Tensor, scale: torch.Tensor,
@@ -0,0 +1,20 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import torch
from vllm.logger import init_logger
logger = init_logger(__name__)
def mxfp8_quantize(x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
try:
from flashinfer import mxfp8_quantize
except ImportError as err:
raise ImportError("The package `flashinfer` is required to do "
"MX-FP8 quantization. Please install it with" \
"`pip install flashinfer`") from err
return mxfp8_quantize(x, is_sf_swizzled_layout=False)