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 16, 22:00 UTC
[Kernel] Raise an exception in MoE kernel if the batch size is larger then 65k (#5939)
Cody Yu committed
Jun 29, 2024 at 06:04 UTC
f7dac83d95ae38973b425a8bb2d3a3df9fe9a9c2
1 file changed
+5
vllm/model_executor/layers/fused_moe/fused_moe.py
+5
@@ -423,6 +423,11 @@ def fused_experts(hidden_states: torch.Tensor,
423
M, _ = hidden_states.shape
424
E, N, _ = w1.shape
425
426
+ if M > 65536:
427
+ # https://github.com/vllm-project/vllm/issues/5938
428
+ raise ValueError("MoE kernel does not support more than 65536 tokens, "
429
+ f"but got {M}")
430
+
431
if override_config:
432
config = override_config
433
else: