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 15, 01:00 UTC
[Log] Debugging Log with more Information (#20770)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Jul 17, 2025 at 20:19 UTC
8a8fc946398c34a3b23786c9cb7bf217e223b268
2 files changed
+37
-13
vllm/model_executor/layers/fused_moe/cutlass_moe.py
+17
-9
@@ -571,34 +571,42 @@ def _valid_cutlass_block_scaled_grouped_gemm(
571
572
_, K, N = w2.size()
573
if not _valid_cutlass_block_scaled_grouped_gemm_shape(N, K):
574
- logger.debug(
575
- "CutlassBlockScaledGroupedGemm disabled: unalinged problem size.")
574
+ logger.debug_once(
575
+ "CutlassBlockScaledGroupedGemm disabled: unaligned problem size. "
576
+ "N: %s, K: %s",
577
+ N,
578
+ K,
579
+ )
580
return False
581
582
if (w1.dtype != torch.float8_e4m3fn or w2.dtype != torch.float8_e4m3fn):
579
- logger.debug(
580
- "CutlassBlockScaledGroupedGemm disabled: invalid weight dtype(s).")
583
+ logger.debug_once(
584
+ "CutlassBlockScaledGroupedGemm disabled: invalid weight dtype(s). "
585
+ "w1.dtype: %s, w2.dtype: %s",
586
+ w1.dtype,
587
+ w2.dtype,
588
+ )
589
return False
590
591
if expert_map is not None:
584
- logger.debug(
592
+ logger.debug_once(
593
"CutlassBlockScaledGroupedGemm disabled: expert_parallel is"
594
" not supported.")
595
return False
596
597
if activation != "silu":
590
- logger.debug(
598
+ logger.debug_once(
599
"CutlassBlockScaledGroupedGemm disabled: only activation silu is"
600
" supported.")
601
return False
602
603
if apply_router_weight_on_input:
596
- logger.debug("CutlassBlockScaledGroupedGemm disabled:"
597
- " apply_router_weight_on_input is not supported.")
604
+ logger.debug_once("CutlassBlockScaledGroupedGemm disabled:"
605
+ " apply_router_weight_on_input is not supported.")
606
return False
607
608
if inplace:
601
- logger.debug(
609
+ logger.debug_once(
610
"CutlassBlockScaledGroupedGemm disabled: inplace is not supported."
611
)
612
return False
vllm/model_executor/layers/fused_moe/deep_gemm_moe.py
+20
-4
@@ -50,17 +50,33 @@ def _valid_deep_gemm(hidden_states: torch.Tensor, w1: torch.Tensor,
50
M = hidden_states.size(0)
51
_, K, N = w2.size()
52
if not _valid_deep_gemm_shape(M, N, K):
53
- logger.debug("DeepGemm disabled: unaligned problem size.")
53
+ logger.debug_once(
54
+ "DeepGemm disabled: unaligned problem size. M: %s, N: %s, K: %s",
55
+ M,
56
+ N,
57
+ K,
58
+ )
59
return False
60
61
if (w1.dtype != torch.float8_e4m3fn or w2.dtype != torch.float8_e4m3fn):
57
- logger.debug("DeepGemm disabled: invalid weight dtype(s).")
62
+ logger.debug_once(
63
+ "DeepGemm disabled: invalid weight dtype(s). "
64
+ "w1.dtype: %s, w2.dtype: %s",
65
+ w1.dtype,
66
+ w2.dtype,
67
+ )
68
return False
69
70
if (not hidden_states.is_contiguous() or not w1.is_contiguous()
71
or not w2.is_contiguous()):
62
- logger.debug(
63
- "DeepGemm disabled: weights or activations not contiguous.")
72
+ logger.debug_once(
73
+ "DeepGemm disabled: weights or activations not contiguous. "
74
+ "hidden_states.is_contiguous(): %s, w1.is_contiguous(): %s, "
75
+ "w2.is_contiguous(): %s",
76
+ hidden_states.is_contiguous(),
77
+ w1.is_contiguous(),
78
+ w2.is_contiguous(),
79
+ )
80
return False
81
82
return True