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
[Doc]: fix typos in Python scripts (#23828)
Signed-off-by: Didier Durand <durand.didier@gmail.com>
Didier Durand committed
Aug 28, 2025 at 14:37 UTC
d3da2eea546b33b9444519f99c26721f7344117f
10 files changed
+12
-12
vllm/compilation/backends.py
+2
-2
@@ -271,7 +271,7 @@ def split_graph(graph: fx.GraphModule,
271
outputs.append(
272
SplitItem(name, graph_id, (graph_id in split_op_graphs), module))
273
274
- # sort by intetger graph_id, rather than string name
274
+ # sort by integer graph_id, rather than string name
275
outputs.sort(key=lambda x: x.graph_id)
276
277
return split_gm, outputs
@@ -424,7 +424,7 @@ class VllmBackend:
424
425
# if the model is initialized with a non-empty prefix,
426
# then usually it's enough to use that prefix,
427
- # e.g. launguage_model, vision_model, etc.
427
+ # e.g. language_model, vision_model, etc.
428
# when multiple parts are initialized as independent
429
# models, we need to use the model_tag to distinguish
430
# them, e.g. backbone (default), eagle_head, etc.
vllm/config/cache.py
+1
-1
@@ -115,7 +115,7 @@ class CacheConfig:
115
116
In some KV sharing setups, e.g. YOCO (https://arxiv.org/abs/2405.05254),
117
some layers can skip tokens corresponding to prefill. This flag enables
118
- attention metadata for eligible layers to be overriden with metadata
118
+ attention metadata for eligible layers to be overridden with metadata
119
necessary for implementing this optimization in some models (e.g. Gemma3n)
120
"""
121
vllm/engine/arg_utils.py
+1
-1
@@ -1053,7 +1053,7 @@ class EngineArgs:
1053
self.trust_remote_code, self.revision,
1054
self.code_revision, self.config_format)
1055
1056
- # if loading a SpeculatorsConfig, load the specualtive_config
1056
+ # if loading a SpeculatorsConfig, load the speculative_config
1057
# details from the config directly
1058
# no user input required / expected
1059
if isinstance(hf_config, SpeculatorsConfig):
vllm/entrypoints/chat_utils.py
+1
-1
@@ -640,7 +640,7 @@ class BaseMultiModalContentParser(ABC):
640
def __init__(self) -> None:
641
super().__init__()
642
643
- # stores model placehodlers list with corresponding
643
+ # stores model placeholders list with corresponding
644
# general MM placeholder:
645
# {
646
# "<##IMAGE##>": ["<image>", "<image>", "<image>"],
vllm/entrypoints/openai/api_server.py
+1
-1
@@ -1096,7 +1096,7 @@ if envs.VLLM_SERVER_DEV_MODE:
1096
raise HTTPException(status_code=HTTPStatus.BAD_REQUEST.value,
1097
detail="Missing 'method' in request body")
1098
# For security reason, only serialized string args/kwargs are passed.
1099
- # User-defined `method` is responsible for deseralization if needed.
1099
+ # User-defined `method` is responsible for deserialization if needed.
1100
args: list[str] = body.get("args", [])
1101
kwargs: dict[str, str] = body.get("kwargs", {})
1102
timeout: Optional[float] = body.get("timeout")
vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py
+1
-1
@@ -579,7 +579,7 @@ class CompressedTensorsConfig(QuantizationConfig):
579
format = scheme_dict.get("format")
580
581
# Find the sparsity scheme of the layer
582
- # assume that fused layers inerhit first component's sparsity scheme
582
+ # assume that fused layers inherit first component's sparsity scheme
583
sparsity_targets = (self.sparsity_scheme_map.keys() -
584
set(self.sparsity_ignore_list))
585
sparsity_scheme: Optional[SparsityCompressionConfig] = None
vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
+1
-1
@@ -71,7 +71,7 @@ class CompressedTensorsMoEMethod(FusedMoEMethodBase):
71
) -> "CompressedTensorsMoEMethod":
72
# TODO: @dsikka: refactor this to use schemes as other kernels
73
# are supported + check if the layer is being ignored.
74
- # Check if a using "Linear" to select scheems
74
+ # Check if a using "Linear" to select schemes
75
if "Linear" in quant_config.target_scheme_map:
76
matched_target = "Linear"
77
else:
vllm/v1/cudagraph_dispatcher.py
+2
-2
@@ -11,7 +11,7 @@ logger = init_logger(__name__)
11
12
class CudagraphDispatcher:
13
"""
14
- Runtime cudagraph dispatcher to dispach keys for multiple set of cudagraphs.
14
+ Runtime cudagraph dispatcher to dispatch keys for multiple set of cudagraphs.
15
16
The dispatcher stores two sets of dispatch keys, one for PIECEWISE and one
17
for FULL cudagraph runtime mode. The keys are initialized depending on
@@ -21,7 +21,7 @@ class CudagraphDispatcher:
21
22
At runtime, the dispatch method generates the runtime cudagraph mode (FULL,
23
PIECEWISE, or NONE for no cudagraph) and the valid key (batch descriptor)
24
- based on the input key. After dispatching (commuicate via forward context),
24
+ based on the input key. After dispatching (communicate via forward context),
25
the cudagraph wrappers will trust the dispatch key to do either capturing
26
or replaying (if mode matched), or pass through to the underlying runnable
27
without cudagraph (if mode no match or mode is NONE).
vllm/v1/worker/block_table.py
+1
-1
@@ -110,7 +110,7 @@ class BlockTable:
110
self.block_table_cpu.fill_(0)
111
112
def get_device_tensor(self) -> torch.Tensor:
113
- """Ruturns the device tensor of the block table."""
113
+ """Returns the device tensor of the block table."""
114
return self.block_table
115
116
def get_cpu_tensor(self) -> torch.Tensor:
vllm/v1/worker/cpu_model_runner.py
+1
-1
@@ -43,7 +43,7 @@ class CPUModelRunner(GPUModelRunner):
43
Args:
44
scheduler_output: The scheduler output.
45
"""
46
- # Attention free models have zero kv_cache_goups, however models
46
+ # Attention free models have zero kv_cache_groups, however models
47
# like Mamba are also attention free but use the kv_cache for
48
# keeping its internal state. This is why we check the number
49
# of kv_cache groups instead of solely checking