A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 12, 22:15 UTC
[ROCm][BugFix] Fix shared expert loading error when disable `VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS` (#28633)
Signed-off-by: ganyi <ygan@amd.com>
Pleaplusone committed
Nov 20, 2025 at 15:50 UTC
7218f83992c7d61fc3845ea24407a1f3b909713e
1 file changed
+5
-2
vllm/model_executor/models/deepseek_v2.py
+5
-2
@@ -287,7 +287,10 @@ class DeepseekV2MoE(nn.Module):
287
)
288
289
self.is_rocm_aiter_moe_enabled = rocm_aiter_ops.is_fused_moe_enabled()
290
- if config.n_shared_experts is None or self.is_rocm_aiter_moe_enabled:
290
+ self.is_fusion_moe_shared_experts_enabled = (
291
+ rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
292
+ )
293
+ if config.n_shared_experts is None or self.is_fusion_moe_shared_experts_enabled:
294
self.shared_experts = None
295
else:
296
intermediate_size = config.moe_intermediate_size * config.n_shared_experts
@@ -327,7 +330,7 @@ class DeepseekV2MoE(nn.Module):
330
num_redundant_experts=self.n_redundant_experts,
331
is_sequence_parallel=self.is_sequence_parallel,
332
n_shared_experts=config.n_shared_experts
330
- if rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
333
+ if self.is_fusion_moe_shared_experts_enabled
334
else None,
335
)
336