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 13, 12:00 UTC
Fix Fused MoE LoRA Triton kernel bug (#28450)
Signed-off-by: chaojun-zhang <chaojun.zhang@intel.com>
Chaojun Zhang committed
Nov 11, 2025 at 20:46 UTC
7dbe6d81d6f17abe93389d97d417e4886467546f
1 file changed
+2
-1
vllm/lora/ops/triton_ops/fused_moe_lora_op.py
+2
-1
@@ -26,7 +26,7 @@ def _get_ptr(lora_weights: list[torch.Tensor], device: torch.device):
26
tensor_ptrs = []
27
for lora_weight in lora_weights:
28
tensor_ptrs.append(lora_weight.data_ptr())
29
- ptr_tensor = torch.tensor(tensor_ptrs, device=device)
29
+ ptr_tensor = torch.tensor(tensor_ptrs, device=device, dtype=torch.uint64)
30
31
_LORA_PTR_DICT[key] = ptr_tensor
32
return _LORA_PTR_DICT.get(key)
@@ -85,6 +85,7 @@ def _fused_moe_lora_kernel(
85
GROUP_SIZE_M: tl.constexpr,
86
SPLIT_K: tl.constexpr,
87
USE_GDC: tl.constexpr,
88
+ launch_pdl: tl.constexpr,
89
IS_PRIMARY: tl.constexpr,
90
):
91
pid = tl.program_id(axis=0)