A high-throughput and memory-efficient inference and serving engine for LLMs
[Log] Update deepgemm log (#45857)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Jun 17, 2026 at 11:34 UTC
9c7c74bf1023774cf0e247bf77a038989ee272db
3 files changed
+11
-6
tests/kernels/moe/modular_kernel_tools/common.py
+4
-1
@@ -328,7 +328,10 @@ class Config:
328
if self.needs_deep_ep_v2() and not has_deep_ep_v2():
329
return False, "Needs DeepEP v2, but DeepEP v2 not available."
330
if self.needs_deep_gemm() and not has_deep_gemm():
331
- return False, "Needs DeepGEMM, but DeepGEMM not available."
331
+ return (
332
+ False,
333
+ "Needs DeepGEMM, but the current vLLM environment does not provide it.",
334
+ )
335
if self.needs_aiter() and not has_aiter(): # noqa: SIM103
336
return False, "Needs Aiter, but Aiter not available."
337
if self.needs_mori() and not has_mori(): # noqa: SIM103
vllm/model_executor/layers/sparse_attn_indexer.py
+2
-1
@@ -442,7 +442,8 @@ class SparseAttnIndexer(CustomOp):
442
self.use_fp4_cache = use_fp4_cache
443
if current_platform.is_cuda() and not has_deep_gemm():
444
raise RuntimeError(
445
- "Sparse Attention Indexer CUDA op requires DeepGEMM to be installed."
445
+ "Sparse Attention Indexer CUDA op requires DeepGEMM support in "
446
+ "the current vLLM environment."
447
)
448
449
def forward_native(
vllm/utils/deep_gemm.py
+5
-4
@@ -120,8 +120,9 @@ def is_deep_gemm_e8m0_used() -> bool:
120
def _missing(*_: Any, **__: Any) -> NoReturn:
121
"""Placeholder for unavailable DeepGEMM backend."""
122
raise RuntimeError(
123
- "DeepGEMM backend is not available or outdated. Please install or "
124
- "update the `deep_gemm` to a newer version to enable FP8 kernels."
123
+ "DeepGEMM backend is unavailable in the current vLLM environment, "
124
+ "or the available DeepGEMM package does not provide the required APIs "
125
+ "for these kernels."
126
)
127
128
@@ -156,7 +157,7 @@ def _import_deep_gemm():
157
logger.debug_once("Imported deep_gemm module from site-packages")
158
return module
159
except ImportError:
159
- logger.debug_once(
160
+ logger.info_once(
161
"deep_gemm not found in site-packages, "
162
"trying vendored vllm.third_party.deep_gemm"
163
)
@@ -167,7 +168,7 @@ def _import_deep_gemm():
168
logger.debug_once("Imported deep_gemm module from vllm.third_party.deep_gemm")
169
return module
170
except ImportError:
170
- logger.debug_once("Vendored deep_gemm not found either")
171
+ logger.info_once("Vendored deep_gemm not found either")
172
except Exception as e:
173
# The vendored module may raise RuntimeError during _C.init()
174
# if JIT include files are missing (e.g. incomplete wheel).