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 10, 06:30 UTC
[Misc] improve warning/assert messages (#32226)
Signed-off-by: cjackal <44624812+cjackal@users.noreply.github.com>
cjackal committed
Jan 13, 2026 at 12:11 UTC
15b33ff06473d3efddefde4990363f1aecff1d48
6 files changed
+19
-19
vllm/compilation/compiler_interface.py
+2
-2
@@ -546,7 +546,7 @@ class InductorAdaptor(CompilerInterface):
546
hash_str, example_inputs, True, False
547
)
548
assert inductor_compiled_graph is not None, (
549
- "Inductor cache lookup failed. Please remove"
549
+ "Inductor cache lookup failed. Please remove "
550
f"the cache directory and try again." # noqa
551
)
552
elif torch.__version__ >= "2.6":
@@ -557,7 +557,7 @@ class InductorAdaptor(CompilerInterface):
557
hash_str, example_inputs, True, None, constants
558
)
559
assert inductor_compiled_graph is not None, (
560
- "Inductor cache lookup failed. Please remove"
560
+ "Inductor cache lookup failed. Please remove "
561
f"the cache directory and try again." # noqa
562
)
563
vllm/config/compilation.py
+4
-4
@@ -949,8 +949,8 @@ class CompilationConfig:
949
)
950
if self.cudagraph_mode == CUDAGraphMode.PIECEWISE:
951
logger.warning_once(
952
- "Piecewise compilation with empty splitting_ops do not"
953
- "contains piecewise cudagraph. Setting cudagraph_"
952
+ "Piecewise compilation with empty splitting_ops does not "
953
+ "contain piecewise cudagraph. Setting cudagraph_"
954
"mode to NONE. Hint: If you are using attention "
955
"backends that support cudagraph, consider manually "
956
"setting cudagraph_mode to FULL or FULL_DECODE_ONLY "
@@ -959,8 +959,8 @@ class CompilationConfig:
959
self.cudagraph_mode = CUDAGraphMode.NONE
960
elif self.cudagraph_mode == CUDAGraphMode.FULL_AND_PIECEWISE:
961
logger.warning_once(
962
- "Piecewise compilation with empty splitting_ops do "
963
- "not contains piecewise cudagraph. Setting "
962
+ "Piecewise compilation with empty splitting_ops does "
963
+ "not contain piecewise cudagraph. Setting "
964
"cudagraph_mode to FULL."
965
)
966
self.cudagraph_mode = CUDAGraphMode.FULL
vllm/config/model.py
+1
-1
@@ -1494,7 +1494,7 @@ class ModelConfig:
1494
1495
if self.runner_type != "pooling" and head_dtype != self.dtype:
1496
logger.warning_once(
1497
- "`head_dtype` currently only supports pooling models."
1497
+ "`head_dtype` currently only supports pooling models, "
1498
"fallback to model dtype [%s].",
1499
self.dtype,
1500
)
vllm/config/vllm.py
+10
-10
@@ -672,9 +672,9 @@ class VllmConfig:
672
and self.compilation_config.mode != CompilationMode.VLLM_COMPILE
673
):
674
logger.warning(
675
- "Inductor compilation was disabled by user settings,"
676
- "Optimizations settings that are only active during"
677
- "Inductor compilation will be ignored."
675
+ "Inductor compilation was disabled by user settings, "
676
+ "optimizations settings that are only active during "
677
+ "inductor compilation will be ignored."
678
)
679
680
def has_blocked_weights():
@@ -790,7 +790,7 @@ class VllmConfig:
790
791
logger.warning_once(
792
"--kv-sharing-fast-prefill requires changes on model side for "
793
- "correctness and to realize prefill savings. "
793
+ "correctness and to realize prefill savings."
794
)
795
# TODO: Move after https://github.com/vllm-project/vllm/pull/26847 lands
796
self._set_compile_ranges()
@@ -813,7 +813,7 @@ class VllmConfig:
813
and not self.cache_config.enable_prefix_caching
814
):
815
logger.warning(
816
- "KV cache events are on, but prefix caching is not enabled."
816
+ "KV cache events are on, but prefix caching is not enabled. "
817
"Use --enable-prefix-caching to enable."
818
)
819
if (
@@ -822,9 +822,9 @@ class VllmConfig:
822
and not self.kv_events_config.enable_kv_cache_events
823
):
824
logger.warning(
825
- "KV cache events are disabled,"
826
- "but the scheduler is configured to publish them."
827
- "Modify KVEventsConfig.enable_kv_cache_events"
825
+ "KV cache events are disabled, "
826
+ "but the scheduler is configured to publish them. "
827
+ "Modify KVEventsConfig.enable_kv_cache_events "
828
"to True to enable."
829
)
830
current_platform.check_and_update_config(self)
@@ -893,7 +893,7 @@ class VllmConfig:
893
else "pipeline parallelism"
894
)
895
logger.warning_once(
896
- "Sequence parallelism not supported with"
896
+ "Sequence parallelism not supported with "
897
"native rms_norm when using %s, "
898
"this will likely lead to an error.",
899
regime,
@@ -910,7 +910,7 @@ class VllmConfig:
910
logger.warning_once(
911
"No piecewise cudagraph for executing cascade attention."
912
" Will fall back to eager execution if a batch runs "
913
- "into cascade attentions"
913
+ "into cascade attentions."
914
)
915
916
if self.compilation_config.cudagraph_mode.requires_piecewise_compilation():
vllm/lora/ops/triton_ops/utils.py
+1
-1
@@ -170,7 +170,7 @@ def load_lora_op_config(op_type: str, add_inputs: bool | None) -> dict | None:
170
171
config_path = Path(f"{user_defined_config_folder}/{config_fname}")
172
if not config_path.exists():
173
- logger.warning_once(f"No LoRA kernel configs founded in {config_path}")
173
+ logger.warning_once(f"No LoRA kernel configs found in {config_path}")
174
return None
175
176
# Load json
vllm/v1/attention/backends/fa_utils.py
+1
-1
@@ -67,7 +67,7 @@ def get_flash_attn_version(requires_alibi: bool = False) -> int | None:
67
# 3. fallback for unsupported combinations
68
if device_capability.major == 10 and fa_version == 3:
69
logger.warning_once(
70
- "Cannot use FA version 3 on Blackwell platform "
70
+ "Cannot use FA version 3 on Blackwell platform, "
71
"defaulting to FA version 2."
72
)
73
fa_version = 2