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 4, 07:30 UTC
[NVFP4][fix] Fix `layer.weight` -> `w13` typo in NVFP4 MOE emulation kernel preparation (#41630)
Signed-off-by: Felix Marty <Felix.Marty@amd.com>
fxmarty-amd committed
May 4, 2026 at 22:13 UTC
9c07342fdc556efd4b70a6c4183cac699bfcd1cd
1 file changed
+1
-1
vllm/model_executor/layers/fused_moe/oracle/nvfp4.py
+1
-1
@@ -381,7 +381,7 @@ def convert_to_nvfp4_moe_kernel_format(
381
elif nvfp4_backend == NvFp4MoeBackend.EMULATION:
382
# Move the E2M1 lookup table to the device now, because
383
# `.to(device)` is not allowed during CUDA graph capture.
384
- kE2M1ToFloat_handle.val = kE2M1ToFloat_handle.val.to(layer.weight.device)
384
+ kE2M1ToFloat_handle.val = kE2M1ToFloat_handle.val.to(w13.device)
385
386
if a13_scale is None or a2_scale is None:
387
raise ValueError(