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
[Misc] Slight improve deepgemm print (#24085)
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Jee Jee Li committed
Sep 5, 2025 at 00:06 UTC
94866d7c9387e4e71944080fd30497bcb59db399
1 file changed
+3
-2
vllm/model_executor/layers/fused_moe/deep_gemm_moe.py
+3
-2
@@ -57,13 +57,14 @@ def _valid_deep_gemm(hidden_states: torch.Tensor, w1: torch.Tensor,
57
if not _valid_deep_gemm_shape(M, N, K):
58
logger.debug_once(
59
"DeepGemm disabled due to unaligned problem size. "
60
- "M: %s, N: %s, K: %s. M should >= align size "
61
- "and N and K must be multiples of %s."
60
+ "M: %s, N: %s, K: %s. M should >= %s "
61
+ "and N and K must be multiples of %s. "
62
"This is not an error and we will fall back to triton.",
63
M,
64
N,
65
K,
66
align,
67
+ align,
68
)
69
return False
70
elif N <= 512: