@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / fad73be1a5

Fork 0

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 (#24077)

Signed-off-by: Didier Durand <durand.didier@gmail.com>

Didier Durand committed Sep 2, 2025 at 11:38 UTC fad73be1a54391efc5b974c86a077b421b58dace

14 files changed +19 -19

tests/async_engine/test_api_server.py

+1 -1

@@ -98,7 +98,7 @@ def test_api_server(api_server, distributed_executor_backend: str):

98 pool.join()

100 # check cancellation stats

101 - # give it some times to update the stats

101 + # give it some time to update the stats

102 time.sleep(1)

103

104 num_aborted_requests = requests.get(

tests/core/block/e2e/test_correctness.py

+2 -2

@@ -439,10 +439,10 @@ def test_auto_prefix_caching_with_preemption(baseline_llm_generator,

439 @pytest.mark.parametrize("seed", [1])

440 def test_auto_prefix_caching_after_eviction_start(baseline_llm_generator,

441 test_llm_generator):

442 - """Verify block manager v2 with auto prefix caching could works normal

442 + """Verify block manager v2 with auto prefix caching could work normally

443 even when eviction started.

444 With APC enabled, all blocks are held by native block at the beginning.

445 - Then blocks are managed by evictor instead. If cache hit at the evitor's

445 + Then blocks are managed by evictor instead. If cache hit at the evictor's

446 block, then it could be reused, or we need to recompute its kv cache.

447 """

448 output_len = 10

tests/engine/test_arg_utils.py

+1 -1

@@ -167,7 +167,7 @@ def test_get_kwargs():

167 # dict should have json tip in help

168 json_tip = "Should either be a valid JSON string or JSON keys"

169 assert json_tip in kwargs["json_tip"]["help"]

170 - # nested config should should construct the nested config

170 + # nested config should construct the nested config

171 assert kwargs["nested_config"]["type"]('{"field": 2}') == NestedConfig(2)

172

173

tests/kernels/moe/test_deepep_deepgemm_moe.py

+1 -1

@@ -282,7 +282,7 @@ def triton_impl(a: torch.Tensor, topk_ids: torch.Tensor,

282 a1_scale=a1_scale,

283 block_shape=block_shape,

284 # Make sure this is set to False so we

285 - # dont end up comparing the same implementation.

285 + # don't end up comparing the same implementation.

286 allow_deep_gemm=False)

287

288

tests/lora/test_add_lora.py

+2 -2

@@ -59,10 +59,10 @@ async def requests_processing_time(llm,

59 @pytest.mark.asyncio

60 async def test_add_lora(chatglm3_lora_files):

61 """

62 - The add_lora function is used to pre-load some LoRA adapters into the

62 + The add_lora function is used to preload some LoRA adapters into the

63 engine in anticipation of future requests using these adapters. To test

64 this functionality, we use the async engine to process some requests - We

65 - do it twice, once with add_lora() pre-loading and once without.

65 + do it twice, once with add_lora() preloading and once without.

67 We measure the request processing time in both cases and expect the time

68 to be lesser in the case with add_lora() calls.

tests/lora/test_lora_allowed_token_ids.py

+2 -2

@@ -18,7 +18,7 @@ def test_allowed_token_ids_with_lora_vocab(llama_2_7b_base_huggingface_id,

18 adapters that define additional tokens.

19 """

21 - # Setup a base model compatible with the sql_lora_files adapter and

21 + # Set up a base model compatible with the sql_lora_files adapter and

22 # a known number of tokens in the base model.

23 model_config = ModelConfig(

24 model=llama_2_7b_base_huggingface_id,

@@ -84,7 +84,7 @@ def test_allowed_token_ids_with_lora_adapter_no_vocab(

84 adapters that do not define additional tokens.

85 """

87 - # Setup a base model compatible with the qwen25vl_lora_files adapter and

87 + # Set up a base model compatible with the qwen25vl_lora_files adapter and

88 # a known number of tokens in the base model.

89 model_config = ModelConfig(

90 model=qwen25vl_base_huggingface_id,

tests/models/language/generation/test_common.py

+1 -1

@@ -13,7 +13,7 @@ from ...registry import HF_EXAMPLE_MODELS

13 from ...utils import check_logprobs_close

15 # These have unsupported head_dim for FA. We do not

16 -# not have a clean way to fall back, so we fail with

16 +# have a clean way to fall back, so we fail with

17 # a clear msg when it happens.

18 # https://github.com/vllm-project/vllm/issues/14524

19 REQUIRES_V0 = ["microsoft/phi-2", "stabilityai/stablelm-3b-4e1t"]

tests/models/language/generation/test_mistral.py

+2 -2

@@ -20,7 +20,7 @@ MISTRAL_FORMAT_MODELS = [

20 "mistralai/Mistral-7B-Instruct-v0.3",

21 # uses the v3-Tekken tokenizer

22 "mistralai/Ministral-8B-Instruct-2410",

23 - # Mistral-Nemo is to big for CI, but passes locally

23 + # Mistral-Nemo is too big for CI, but passes locally

24 # "mistralai/Mistral-Nemo-Instruct-2407"

25 ]

@@ -273,7 +273,7 @@ def test_mistral_function_calling(vllm_runner, model: str, dtype: str) -> None:

273

274

275 def test_mistral_function_call_nested_json():

276 - """Ensure that the function-name regex captures the entire outer-most

276 + """Ensure that the function-name regex captures the entire outermost

277 JSON block, including nested braces."""

278

279 # Create a minimal stub tokenizer that provides the few attributes the

tests/models/multimodal/generation/test_qwen2_vl.py

+2 -2

@@ -154,7 +154,7 @@ def batch_make_image_embeddings(

154 embed_counter += cur_batch_embed_len

155 image_counter += cur_batch_image_count

156

157 - # ensure we don't lost any images or embeddings

157 + # ensure we don't lose any images or embeddings

158 assert embed_counter == image_embeds.size(0)

159 assert image_counter == image_grid_thw.size(0)

160 assert len(image_batches) == len(result)

@@ -238,7 +238,7 @@ def batch_make_video_embeddings(

238 embed_counter += cur_batch_embed_len

239 video_counter += cur_batch_video_count

240

241 - # ensure we don't lost any videos or embeddings

241 + # ensure we don't lose any videos or embeddings

242 assert embed_counter == video_embeds.size(0)

243 assert video_counter == video_grid_thw.size(0)

244 assert len(video_batches) == len(result)

tests/v1/core/test_kv_cache_utils.py

+1 -1

@@ -247,7 +247,7 @@ def test_free_kv_cache_block_queue_append_n():

247

248 def test_free_kv_cache_block_queue_popleft_n():

249 blocks = [KVCacheBlock(block_id=i) for i in range(6)]

250 - # Create a empty FreeKVCacheBlockQueue with these blocks

250 + # Create an empty FreeKVCacheBlockQueue with these blocks

251 queue = FreeKVCacheBlockQueue(

252 [blocks[1], blocks[3], blocks[5], blocks[4], blocks[0], blocks[2]])

253 assert queue.num_free_blocks == 6

tests/v1/executor/test_executor.py

+1 -1

@@ -27,7 +27,7 @@ class CustomMultiprocExecutor(MultiprocExecutor):

27 kwargs: Optional[dict] = None,

28 non_block: bool = False,

29 unique_reply_rank: Optional[int] = None) -> list[Any]:

30 - # Drop marker to show that this was ran

30 + # Drop marker to show that this was run

31 with open(".marker", "w"):

32 ...

33 return super().collective_rpc(method, timeout, args, kwargs)

tests/v1/spec_decode/test_eagle.py

+1 -1

@@ -183,7 +183,7 @@ def test_load_model(mock_get_model, mock_get_layers, mock_get_pp_group, method,

183 mock_pp_group.world_size = pp_size

184 mock_get_pp_group.return_value = mock_pp_group

185

186 - # Setup the target model mock with a custom class so that

186 + # Set up the target model mock with a custom class so that

187 # isinstance() checks match the expected type.

188 class _TargetModelStub(LlamaForCausalLM):

189 model: mock.MagicMock

tests/v1/test_kv_sharing.py

+1 -1

@@ -30,7 +30,7 @@ def test_initialize_kv_cache_for_kv_sharing_different_attn_groups():

30 }

32 # Layers 0 and 1 both belong in KV cache group 0

33 - # However, if they have have different attention backends, they will be

33 + # However, if they have different attention backends, they will be

34 # placed in different attention groups for KV cache group 0

35 kv_cache_groups = [

36 KVCacheGroupSpec(["model.layers.0", "model.layers.1"],

tests/v1/worker/test_gpu_model_runner.py

+1 -1

@@ -702,7 +702,7 @@ def test_hybrid_attention_mamba_tensor_shapes(monkeypatch):

702 KVCacheTensors for the attention and mamba layers

703 (via _reshape_kv_cache_tensors function). This test verifies

704 that the views are compatible: writing a mamba block

705 - will not corrupt an attention block and vice-versa

705 + will not corrupt an attention block and vice versa

706 '''

707

708 current_platform.seed_everything(42)