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 comments (#24417)
Signed-off-by: Didier Durand <durand.didier@gmail.com>
Didier Durand committed
Sep 8, 2025 at 09:22 UTC
f4962a6d55a340ebb569d377c842deff7611d8f7
12 files changed
+12
-12
examples/offline_inference/chat_with_tools.py
+1
-1
@@ -143,5 +143,5 @@ outputs = llm.chat(messages, sampling_params, tools=tools)
143
144
print(outputs[0].outputs[0].text.strip())
145
# yields
146
-# 'The weather in Dallas, TX is 85 degrees fahrenheit. '
146
+# 'The weather in Dallas, TX is 85 degrees Fahrenheit. '
147
# 'It is partly cloudly, with highs in the 90's.'
vllm/attention/backends/mla/common.py
+1
-1
@@ -1052,7 +1052,7 @@ class MLACommonImpl(MLAAttentionImpl[T], Generic[T]):
1052
return layer.weight
1053
1054
# we currently do not have quantized bmm's which are needed for
1055
- # `W_UV` and `W_UK_T`, we we just store fp16/bf16 copies and perform
1055
+ # `W_UV` and `W_UK_T`, we just store fp16/bf16 copies and perform
1056
# the bmm's in 16-bit, the extra memory overhead of this is fairly low
1057
kv_b_proj_weight = get_and_maybe_dequant_weights(self.kv_b_proj).T
1058
assert kv_b_proj_weight.shape == (
vllm/config/__init__.py
+1
-1
@@ -1169,7 +1169,7 @@ class ModelConfig:
1169
]
1170
# Any custom overrides will be in quantization_methods so we place
1171
# them at the start of the list so custom overrides have preference
1172
- # over the built in ones.
1172
+ # over the built-in ones.
1173
quantization_methods = quantization_methods + overrides
1174
1175
# Detect which checkpoint is it
vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+1
-1
@@ -770,7 +770,7 @@ class NixlConnectorWorker:
770
# with joint KV for each block. This minimizes the overhead in
771
# registerMem allowing faster descs queries. In order to be able to
772
# split on kv_heads dim as required by heterogeneous TP, one must
773
- # be able to index K/V separately. Hence the we double the number
773
+ # be able to index K/V separately. Hence we double the number
774
# of 'virtual' regions here and halve `block_len` below.
775
self.num_regions *= 2
776
vllm/engine/arg_utils.py
+1
-1
@@ -1159,7 +1159,7 @@ class EngineArgs:
1159
# Note(hc): In the current implementation of decode context
1160
# parallel(DCP), tp_size needs to be divisible by dcp_size,
1161
# because the world size does not change by dcp, it simply
1162
- # reuse the GPUs of TP group, and split one TP group into
1162
+ # reuses the GPUs of TP group, and split one TP group into
1163
# tp_size//dcp_size DCP groups.
1164
assert self.tensor_parallel_size % self.decode_context_parallel_size \
1165
== 0, (
vllm/engine/multiprocessing/client.py
+1
-1
@@ -235,7 +235,7 @@ class MQLLMEngineClient(EngineClient):
235
# therefore we have to inform that the current
236
# processed requests failed as well. Send back a dead
237
# engine error give this feedback and also give a
238
- # 'hint' to the server to shutdown next.
238
+ # 'hint' to the server to shut down next.
239
exception = self.dead_error
240
241
if request_id is None:
vllm/entrypoints/openai/cli_args.py
+1
-1
@@ -204,7 +204,7 @@ schema. Example: `[{"type": "text", "text": "Hello world!"}]`"""
204
frontend_kwargs["lora_modules"]["type"] = optional_type(str)
205
frontend_kwargs["lora_modules"]["action"] = LoRAParserAction
206
207
- # Special case: Middleware needs append action
207
+ # Special case: Middleware needs to append action
208
frontend_kwargs["middleware"]["action"] = "append"
209
frontend_kwargs["middleware"]["type"] = str
210
if "nargs" in frontend_kwargs["middleware"]:
vllm/entrypoints/openai/tool_parsers/llama4_pythonic_tool_parser.py
+1
-1
@@ -176,7 +176,7 @@ class Llama4PythonicToolParser(ToolParser):
176
index] += delta.function.arguments
177
178
# HACK: serving_chat.py inspects the internal state of tool parsers
179
- # when determining it's final streaming delta, automatically
179
+ # when determining its final streaming delta, automatically
180
# adding autocompleted JSON.
181
# These two lines avoid that nonsense while ensuring finish_reason
182
# is set to tool_calls when at least one tool is called.
vllm/entrypoints/openai/tool_parsers/mistral_tool_parser.py
+1
-1
@@ -143,7 +143,7 @@ class MistralToolParser(ToolParser):
143
except json.JSONDecodeError:
144
# use a regex to find the part corresponding to the tool call.
145
# NOTE: This use case should not happen if the model is trained
146
- # correctly. It's a easy possible fix so it's included, but
146
+ # correctly. It's an easy possible fix so it's included, but
147
# can be brittle for very complex / highly nested tool calls
148
raw_tool_call = self.tool_call_regex.findall(tool_content)[0]
149
function_call_arr = json.loads(raw_tool_call)
vllm/model_executor/layers/fused_moe/modular_kernel.py
+1
-1
@@ -302,7 +302,7 @@ class FusedMoEPrepareAndFinalize(ABC):
302
def max_num_tokens_per_rank(self) -> Optional[int]:
303
"""
304
Some PrepareFinalize All2All implementations are batched. Meaning,
305
- they can processes only as set of tokens at a time. This
305
+ they can process only as set of tokens at a time. This
306
function returns the batch size i.e the maximum number of tokens
307
the implementation can process at a time.
308
Return None if there are no such restrictions.
vllm/model_executor/layers/quantization/utils/marlin_utils.py
+1
-1
@@ -201,7 +201,7 @@ def marlin_make_workspace(output_size_per_partition: int,
201
def marlin_make_workspace_new(device: torch.device,
202
max_blocks_per_sm: int = 1) -> torch.Tensor:
203
# In the new marlin kernel, we use the num of threadblocks as workspace
204
- # size. The num of threadblocks is is sms_count * max_blocks_per_sm.
204
+ # size. The num of threadblocks is sms_count * max_blocks_per_sm.
205
sms = torch.cuda.get_device_properties(device).multi_processor_count
206
return torch.zeros(sms * max_blocks_per_sm,
207
dtype=torch.int,
vllm/v1/worker/block_table.py
+1
-1
@@ -98,7 +98,7 @@ class BlockTable:
98
# here because M (max_model_len) is not necessarily divisible by
99
# block_size.
100
if self.dcp_world_size > 1:
101
- # Note(hc): The DCP implement store kvcache with a interleave
101
+ # Note(hc): The DCP implement store kvcache with an interleave
102
# style, the kvcache for the token whose token_idx is i is
103
# always stored on the GPU whose dcp_rank equals i % cp_world_size:
104