frozenleaves/vllm
1
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2025-10-20 23:03:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Log] Debugging Log with more Information (#20770)

Browse Source 
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye
2025-07-17 20:19:46 -04:00
committed by GitHub
parent 4de7146351
commit 8a8fc94639
2 changed files with 37 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
vllm/model_executor/layers/fused_moe/cutlass_moe.py
View File

@ -571,34 +571,42 @@ def _valid_cutlass_block_scaled_grouped_gemm(
_, K, N = w2.size()
if not _valid_cutlass_block_scaled_grouped_gemm_shape(N, K):
logger.debug(
"CutlassBlockScaledGroupedGemm disabled: unalinged problem size.")
logger.debug_once(
"CutlassBlockScaledGroupedGemm disabled: unaligned problem size. "
"N: %s, K: %s",
N,
K,
)
return False
if (w1.dtype != torch.float8_e4m3fn or w2.dtype != torch.float8_e4m3fn):
logger.debug(
"CutlassBlockScaledGroupedGemm disabled: invalid weight dtype(s).")
logger.debug_once(
"CutlassBlockScaledGroupedGemm disabled: invalid weight dtype(s). "
"w1.dtype: %s, w2.dtype: %s",
w1.dtype,
w2.dtype,
)
return False
if expert_map is not None:
logger.debug(
logger.debug_once(
"CutlassBlockScaledGroupedGemm disabled: expert_parallel is"
" not supported.")
return False
if activation != "silu":
logger.debug(
logger.debug_once(
"CutlassBlockScaledGroupedGemm disabled: only activation silu is"
" supported.")
return False
if apply_router_weight_on_input:
logger.debug("CutlassBlockScaledGroupedGemm disabled:"
" apply_router_weight_on_input is not supported.")
logger.debug_once("CutlassBlockScaledGroupedGemm disabled:"
" apply_router_weight_on_input is not supported.")
return False
if inplace:
logger.debug(
logger.debug_once(
"CutlassBlockScaledGroupedGemm disabled: inplace is not supported."
)
return False

24
vllm/model_executor/layers/fused_moe/deep_gemm_moe.py
View File

@ -50,17 +50,33 @@ def _valid_deep_gemm(hidden_states: torch.Tensor, w1: torch.Tensor,
M = hidden_states.size(0)
_, K, N = w2.size()
if not _valid_deep_gemm_shape(M, N, K):
logger.debug("DeepGemm disabled: unaligned problem size.")
logger.debug_once(
"DeepGemm disabled: unaligned problem size. M: %s, N: %s, K: %s",
M,
N,
K,
)
return False
if (w1.dtype != torch.float8_e4m3fn or w2.dtype != torch.float8_e4m3fn):
logger.debug("DeepGemm disabled: invalid weight dtype(s).")
logger.debug_once(
"DeepGemm disabled: invalid weight dtype(s). "
"w1.dtype: %s, w2.dtype: %s",
w1.dtype,
w2.dtype,
)
return False
if (not hidden_states.is_contiguous() or not w1.is_contiguous()
or not w2.is_contiguous()):
logger.debug(
"DeepGemm disabled: weights or activations not contiguous.")
logger.debug_once(
"DeepGemm disabled: weights or activations not contiguous. "
"hidden_states.is_contiguous(): %s, w1.is_contiguous(): %s, "
"w2.is_contiguous(): %s",
hidden_states.is_contiguous(),
w1.is_contiguous(),
w2.is_contiguous(),
)
return False
return True