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 14, 05:30 UTC
[Bug] Fix Assertion error DeepEP/csrc/kernels/intranode.cu:928: 'false and Unsupported type' (#26532)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Oct 13, 2025 at 18:26 UTC
7200a21cd14d48d548be4ce4af1870d3caf1a994
2 files changed
+5
-1
vllm/model_executor/layers/fused_moe/deepep_ht_prepare_finalize.py
+4
@@ -336,7 +336,11 @@ class DeepEPHTPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
336
apply_router_weight_on_input=apply_router_weight_on_input,
337
)
338
dbo_yield_and_switch_from_compute_to_comm()
339
+ assert fused_expert_output.dtype == torch.bfloat16, (
340
+ f"Expected fused_expert_output bfloat16, got {fused_expert_output.dtype}"
341
+ )
342
combined_x, _, event = self.buffer.combine(
343
+ # HT combine only supports BF16
344
x=fused_expert_output,
345
handle=handle,
346
topk_weights=None,
vllm/model_executor/layers/fused_moe/modular_kernel.py
+1
-1
@@ -984,7 +984,7 @@ class FusedMoEModularKernel(torch.nn.Module):
984
assert num_chunks == 0
985
workspace13 = None
986
workspace2 = None
987
- fused_out = torch.empty_like(a1q)
987
+ fused_out = torch.empty_like(a1q, dtype=in_dtype)
988
else:
989
assert num_chunks > 0
990
workspace13, workspace2, fused_out = self._allocate_buffers(