Skip to content

fix(rocm): Enable non-gated MoE (is_act_and_mul=False) support on ROCm #32244

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
tjtanaa merged 1 commit into from
Jan 16, 2026
Merged

fix(rocm): Enable non-gated MoE (is_act_and_mul=False) support on ROCm #32244

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 8 additions & 4 deletions vllm/model_executor/layers/fused_moe/layer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -448,9 +448,13 @@ def __init__(
)

# ROCm aiter shared experts fusion
self.rocm_aiter_fmoe_enabled = rocm_aiter_ops.is_fused_moe_enabled()
# AITER only supports gated activations (silu/gelu), so disable it
# for non-gated MoE (is_act_and_mul=False)
self.rocm_aiter_fmoe_enabled = (
rocm_aiter_ops.is_fused_moe_enabled() and is_act_and_mul
)
self.aiter_fmoe_shared_expert_enabled = (
rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
rocm_aiter_ops.is_fusion_moe_shared_experts_enabled() and is_act_and_mul
)

self.num_fused_shared_experts = (
Expand Down Expand Up @@ -619,9 +623,9 @@ def _get_quant_method() -> FusedMoEMethodBase:
# for heuristic purposes, so it must be initialized first.
self.quant_method: FusedMoEMethodBase = _get_quant_method()

if not self.moe_config.is_act_and_mul and not current_platform.is_cuda():
if not self.moe_config.is_act_and_mul and not current_platform.is_cuda_alike():
raise NotImplementedError(
"is_act_and_mul=False is supported only for CUDA for now"
"is_act_and_mul=False is supported only for CUDA and ROCm for now"
)

if self.enable_eplb and not self.quant_method.supports_eplb:
Expand Down
7 changes: 7 additions & 0 deletions vllm/model_executor/layers/fused_moe/unquantized_fused_moe_method.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,6 +8,7 @@

import vllm.envs as envs
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm._aiter_ops import rocm_aiter_ops
from vllm.logger import init_logger
from vllm.model_executor.custom_op import CustomOp
from vllm.model_executor.layers.fused_moe.config import (
Expand Down Expand Up @@ -56,6 +57,12 @@ def __init__(self, moe: FusedMoEConfig):
use_ep=self.moe.moe_parallel_config.use_ep,
use_dp=self.moe.moe_parallel_config.dp_size > 1,
)

# AITER only supports gated activations (silu/gelu), so disable it
# for non-gated MoE (is_act_and_mul=False)
self.rocm_aiter_moe_enabled = (
rocm_aiter_ops.is_fused_moe_enabled() and moe.is_act_and_mul
)
self.kernel: mk.FusedMoEModularKernel | None = None

@property
Expand Down