[ROCm][CI] Loosen block-FP8 fused MoE test tolerance for large-K shapes (#48847)

Signed-off-by: Stefan Koncarevic <Stefan.Koncarevic@amd.com>
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
stefankoncarevic
2026-08-07 01:19:26 -07:00
committed by GitHub
co-authored by Andreas Karatzas mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 58fcaa0baa
commit 0de0362ea1
2 changed files with 56 additions and 11 deletions
@@ -5,7 +5,7 @@ import torch
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm._aiter_ops import rocm_aiter_ops
from vllm.forward_context import get_forward_context
from vllm.forward_context import get_forward_context, is_forward_context_available
from vllm.model_executor.layers.fused_moe.activation import MoEActivation
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEConfig,
@@ -27,6 +27,9 @@ __all__ = [
]
# TODO(akaratza): Remove
# _get_padding_mask and patch_gating_output after
# https://github.com/ROCm/aiter/pull/4530 is released and consumed by vLLM.
def _get_padding_mask() -> torch.Tensor | None:
"""
Retrieves a boolean mask with non-padding (0) and padding (1) tokens.
@@ -36,7 +39,13 @@ def _get_padding_mask() -> torch.Tensor | None:
slot_mapping[num_tokens_unpadded:num_tokens_padded].fill_(-1)
in gpu_model_runner.py.
Direct kernel callers do not have model-runner padding metadata, so no
mask is needed when there is no forward context.
"""
if not is_forward_context_available():
return None
forward_context = get_forward_context()
# model runner v2.