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]: fixing typos to improve docs (#24480)
Signed-off-by: Didier Durand <durand.didier@gmail.com>
Didier Durand committed
Sep 9, 2025 at 08:06 UTC
46876dff3283cdf7d764bf6133ea3aadcd507b7d
9 files changed
+12
-12
docs/features/tool_calling.md
+1
-1
@@ -169,7 +169,7 @@ All Llama 3.1, 3.2 and 4 models should be supported.
169
170
The tool calling that is supported is the [JSON-based tool calling](https://llama.meta.com/docs/model-cards-and-prompt-formats/llama3_1/#json-based-tool-calling). For [pythonic tool calling](https://github.com/meta-llama/llama-models/blob/main/models/llama3_2/text_prompt_format.md#zero-shot-function-calling) introduced by the Llama-3.2 models, see the `pythonic` tool parser below. As for Llama 4 models, it is recommended to use the `llama4_pythonic` tool parser.
171
172
-Other tool calling formats like the built in python tool calling or custom tool calling are not supported.
172
+Other tool calling formats like the built-in python tool calling or custom tool calling are not supported.
173
174
Known issues:
175
docs/getting_started/installation/gpu/rocm.inc.md
+1
-1
@@ -119,7 +119,7 @@ Currently, there are no pre-built ROCm wheels.
119
This may take 5-10 minutes. Currently, `pip install .` does not work for ROCm installation.
120
121
!!! tip
122
- - Triton flash attention is used by default. For benchmarking purposes, it is recommended to run a warm up step before collecting perf numbers.
122
+ - Triton flash attention is used by default. For benchmarking purposes, it is recommended to run a warm-up step before collecting perf numbers.
123
- Triton flash attention does not currently support sliding window attention. If using half precision, please use CK flash-attention for sliding window support.
124
- To use CK flash-attention or PyTorch naive attention, please use this flag `export VLLM_USE_TRITON_FLASH_ATTN=0` to turn off triton flash attention.
125
- The ROCm version of PyTorch, ideally, should match the ROCm driver version.
examples/tool_chat_template_phi4_mini.jinja
+2
-2
@@ -9,7 +9,7 @@
9
<|system|>
10
{{ system_message }}
11
{%- if tools %}
12
-In addition to plain text responses, you can chose to call one or more of the provided functions.
12
+In addition to plain text responses, you can choose to call one or more of the provided functions.
13
14
Use the following rule to decide when to call a function:
15
* if the response can be generated from your internal knowledge (e.g., as in the case of queries like "What is the capital of Poland?"), do so
@@ -19,7 +19,7 @@ If you decide to call functions:
19
* prefix function calls with functools marker (no closing marker required)
20
* all function calls should be generated in a single JSON list formatted as functools[{"name": [function name], "arguments": [function arguments as JSON]}, ...]
21
* follow the provided JSON schema. Do not hallucinate arguments or values. Do to blindly copy values from the provided samples
22
- * respect the argument type formatting. E.g., if the type if number and format is float, write value 7 as 7.0
22
+ * respect the argument type formatting. E.g., if the type is number and format is float, write value 7 as 7.0
23
* make sure you pick the right functions that match the user intent
24
25
tests/engine/test_executor.py
+1
-1
@@ -25,7 +25,7 @@ class CustomUniExecutor(UniProcExecutor):
25
timeout: Optional[float] = None,
26
args: tuple = (),
27
kwargs: Optional[dict] = None) -> list[Any]:
28
- # Drop marker to show that this was ran
28
+ # Drop marker to show that this was run
29
with open(".marker", "w"):
30
...
31
return super().collective_rpc(method, timeout, args, kwargs)
tests/entrypoints/offline_mode/test_offline_mode.py
+2
-2
@@ -79,7 +79,7 @@ def test_offline_mode(monkeypatch: pytest.MonkeyPatch):
79
)
80
81
# Need to re-import huggingface_hub
82
- # and friends to setup offline mode
82
+ # and friends to set up offline mode
83
_re_import_modules()
84
# Cached model files should be used in offline mode
85
for model_config in MODEL_CONFIGS:
@@ -136,7 +136,7 @@ def test_model_from_huggingface_offline(monkeypatch: pytest.MonkeyPatch):
136
disable_connect,
137
)
138
# Need to re-import huggingface_hub
139
- # and friends to setup offline mode
139
+ # and friends to set up offline mode
140
_re_import_modules()
141
engine_args = EngineArgs(model="facebook/opt-125m")
142
LLM(**dataclasses.asdict(engine_args))
tests/kernels/utils.py
+1
-1
@@ -1247,7 +1247,7 @@ def baseline_scaled_mm(a: torch.Tensor,
1247
# then we would expand a to:
1248
# a = [[1, 1, 2, 2],
1249
# [3, 3, 4, 4]]
1250
- # NOTE this function this function does not explicitly broadcast dimensions
1250
+ # NOTE this function does not explicitly broadcast dimensions
1251
# with an extent of 1, since this can be done implicitly by pytorch
1252
def group_broadcast(t, shape):
1253
for i, s in enumerate(shape):
tests/models/language/generation/test_hybrid.py
+2
-2
@@ -301,7 +301,7 @@ def test_fail_upon_inc_requests_and_finished_requests_lt_available_blocks(
301
finished_requests_ids is larger than the maximum mamba block capacity.
302
303
This could generally happen due to the fact that hybrid does support
304
- statelessness mechanism where it can cleanup new incoming requests in
304
+ statelessness mechanism where it can clean up new incoming requests in
305
a single step.
306
"""
307
try:
@@ -322,7 +322,7 @@ def test_state_cleanup(
322
This test is for verifying that the Hybrid state is cleaned up between
323
steps.
324
325
- If its not cleaned, an error would be expected.
325
+ If it's not cleaned, an error would be expected.
326
"""
327
try:
328
with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
tests/tpu/test_quantization_accuracy.py
+1
-1
@@ -28,7 +28,7 @@ ACCURACY_CONFIGS = [
28
expected_value=0.76), # no bias
29
# NOTE(rob): We cannot re-initialize vLLM in the same process for TPU,
30
# so only one of these tests can run in a single call to pytest. As
31
- # a follow up, move this into the LM-EVAL section of the CI.
31
+ # a follow-up, move this into the LM-EVAL section of the CI.
32
# GSM8KAccuracyTestConfig(
33
# model_name="neuralmagic/Qwen2-7B-Instruct-quantized.w8a8",
34
# expected_value=0.66), # bias in QKV layers
vllm/distributed/parallel_state.py
+1
-1
@@ -1117,7 +1117,7 @@ def initialize_model_parallel(
1117
"decode context model parallel group is already initialized")
1118
# Note(hc): In the current implementation of decode context parallel,
1119
# dcp_size must not exceed tp_size, because the world size does not
1120
- # change by DCP, it simply reuse the GPUs of TP group, and split one
1120
+ # change by DCP, it simply reuses the GPUs of TP group, and split one
1121
# TP group into tp_size//dcp_size DCP groups.
1122
group_ranks = all_ranks.reshape(
1123
-1, decode_context_model_parallel_size).unbind(0)