[CI][ROCm] Reuse AITER's prebuilt MoE module in the modular-kernel harness (#60633)

Signed-off-by: Stefan Koncarevic <stefan.koncarevic@amd.com>
This commit is contained in:
stefankoncarevic
2026-10-08 11:18:05 -05:00
committed by GitHub
parent 9f3d70e081
commit f4dde3132c
@@ -1,6 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from dataclasses import dataclass
from dataclasses import dataclass, replace
from types import SimpleNamespace
from typing import Any
@@ -702,6 +702,21 @@ def make_modular_kernel(
return modular_kernel
def _shuffle_weights_for_aiter(rank_weights: WeightTensors) -> WeightTensors:
"""Pre-shuffle weights so AITER selects its prebuilt `preshuffle_on` module.
Production shuffles in `process_weights_after_loading`; this harness builds its
tensors directly, so without this every call asks for a `preshuffle_off` module
that is not prebuilt, and JIT-compiles a kernel the image already ships.
"""
from vllm._aiter_ops import rocm_aiter_ops
w1, w2 = rocm_aiter_ops.shuffle_weights(rank_weights.w1, rank_weights.w2)
w1.is_shuffled = True
w2.is_shuffled = True
return replace(rank_weights, w1=w1, w2=w2)
def _maybe_convert_weights_for_experts(
config: Config,
rank_weights: WeightTensors,
@@ -715,6 +730,15 @@ def _maybe_convert_weights_for_experts(
fe_type = config.fused_experts_type
fe_name = getattr(fe_type, "__name__", "")
# AITER's prebuilt modules carry a gfx950 instance table even on gfx942 and reject
# intermediate sizes that are not a multiple of 256, but only for the fp8 per-tensor
# and per-token schemes. Those keep JIT-compiling until
# https://github.com/ROCm/aiter/issues/5766 is fixed.
if fe_name == "AiterExperts" and (
config.quant_dtype is None or config.quant_block_shape is not None
):
return _shuffle_weights_for_aiter(rank_weights)
backend: Fp8MoeBackend | None = None
if fe_name == "TrtLlmFp8ExpertsModular":
backend = Fp8MoeBackend.FLASHINFER_TRTLLM