6a895197fa · tduncan/Nvidia-Optimization-Stack-VLLM

.buildkite/scripts/upload-nightly-wheels.sh

+1 -1

@@ -72,7 +72,7 @@ obj_json="objects.json"

72 aws s3api list-objects-v2 --bucket "$BUCKET" --prefix "$SUBPATH/" --delimiter / --output json > "$obj_json"

73 mkdir -p "$INDICES_OUTPUT_DIR"

74

75 -# call script to generate indicies for all existing wheels

75 +# call script to generate indices for all existing wheels

76 # this indices have relative paths that could work as long as it is next to the wheel directory in s3

77 # i.e., the wheels are always in s3://vllm-wheels/<commit>/

78 # and indices can be placed in /<commit>/, or /nightly/, or /<version>/

.buildkite/test-amd.yaml

+2 -2

@@ -467,7 +467,7 @@ steps:

467 - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api

468 - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine

469

470 -# TODO: Add the "V1 Test attetion (MI300)" test group

470 +# TODO: Add the "V1 Test attention (MI300)" test group

471

472 - label: V1 Test attention (H100) # 10min

473 mirror_hardwares: [amdexperimental, amdproduction]

@@ -2174,7 +2174,7 @@ steps:

2174 - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api

2175 - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine

2176

2177 -# TODO: Add the "V1 Test attetion (MI300)" test group

2177 +# TODO: Add the "V1 Test attention (MI300)" test group

2178

2179 - label: V1 Test attention (H100) # 10min

2180 mirror_hardwares: [amdexperimental]

.pre-commit-config.yaml

+1 -1

@@ -13,7 +13,7 @@ repos:

13 args: [--output-format, github, --fix]

14 - id: ruff-format

15 - repo: https://github.com/crate-ci/typos

16 - rev: v1.38.1

16 + rev: v1.43.5

17 hooks:

18 - id: typos

19 args: [--force-exclude]

benchmarks/attention_benchmarks/common.py

+1 -1

@@ -30,7 +30,7 @@ def batch_spec_sort_key(spec: str) -> tuple[int, int, int]:

30 max_kv_len = max(r.kv_len for r in requests) if requests else 0

31 return (batch_size, max_q_len, max_kv_len)

32 except Exception:

33 - # Fallback for unparseable specs

33 + # Fallback for unparsable specs

34 return (0, 0, 0)

35

36

benchmarks/kernels/benchmark_2d_silu_mul_fp8_quant.py

+1 -1

@@ -202,7 +202,7 @@ def test_correctness(T: int, N: int):

202 # reference output

203 ref_out_q, ref_out_s = output_from_impl(ImplType.REFERENCE)

204

205 - # test ouptut

205 + # test output

206 out_q, out_s = output_from_impl(

207 ImplType.SILU_MUL_PER_TOKEN_GROUP_QUANT_FP8_COLMAJOR

208 )

csrc/cpu/cpu_attn_amx.hpp

+1 -1

@@ -420,7 +420,7 @@ class AttentionImpl<ISA::AMX, scalar_t, head_dim> {

420 const int64_t block_size, const int64_t block_size_stride) {

421 // For AMX 2D tiles, size of each line is 64 bytes

422 constexpr int64_t amx_tile_row_size = AMX_TILE_ROW_BYTES;

423 - // For AMX B martix, N always is 16

423 + // For AMX B matrix, N always is 16

424 constexpr int64_t amx_b_tile_n_size = AMX_TILE_ROW_BYTES / 4;

425 constexpr int64_t amx_b_tile_k_size = amx_tile_row_size / sizeof(scalar_t);

426 // For now suppose block_size is divisible by amx_tile_column_num

csrc/cpu/torch_bindings.cpp

+1 -1

@@ -4,7 +4,7 @@

4

5 #include <torch/library.h>

6

7 -// Note: overwrite the external defination for sharing same name between

7 +// Note: overwrite the external definition for sharing same name between

8 // libraries use different ISAs.

9 #define TORCH_EXTENSION_NAME _C

10

csrc/moe/moe_align_sum_kernels.cu

+2 -2

@@ -35,11 +35,11 @@ __global__ void batched_moe_align_block_size_kernel(

35 int32_t const block_ids_size = sorted_ids_size / block_size;

36 int32_t const SENTINEL =

37 num_batches * max_tokens_per_batch; // To denote invalid entries.

38 - // Intialize sorted_ids

38 + // Initialize sorted_ids

39 for (size_t i = threadIdx.x; i < sorted_ids_size; i += stride) {

40 sorted_ids[i] = SENTINEL;

41 }

42 - // Intialize expert_ids with -1

42 + // Initialize expert_ids with -1

43 for (size_t i = threadIdx.x; i < block_ids_size; i += stride) {

44 block_ids[i] = -1;

45 }

csrc/quantization/activation_kernels.cu

+1 -1

@@ -542,7 +542,7 @@ __global__ void silu_mul_fp8_quant_deep_gemm_kernel(

542 if (!lane_id) {

543 // Store scales.

544 if constexpr (std::is_same<scale_t, uint8_t>::value) {

545 - // Packed UE8MO format. Remove Mantissa.

545 + // Packed UE8M0 format. Remove Mantissa.

546 *y_s_ptr = reinterpret_cast<int16_t&>(y_s) >> 7;

547

548 bool const jump_pack = (current_group_id + 1) % 4 == 0;

csrc/rocm/skinny_gemms.cu

+1 -1

@@ -1476,7 +1476,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)

1476 #endif

1477

1478 // B[] staging is cooperative across GrpsShrB, so sync here before reading

1479 - // back. This wait is currently inserted by compiler, but not gauranteed.

1479 + // back. This wait is currently inserted by compiler, but not guaranteed.

1480 asm volatile("s_waitcnt 0");

1481 __syncthreads();

1482

docs/design/cuda_graphs.md

+1 -1

@@ -98,7 +98,7 @@ The goal of this structure is to uniquely identify a (padded) batch with minimal

98

99 ### `CudagraphDispatcher`

100

101 -The [CudagraphDispatcher][vllm.v1.cudagraph_dispatcher.CudagraphDispatcher] takes responsibility for maintaining two sets of valid dispatching keys, one set for `FULL` runtime mode and one set for `PIECEWISE` runtime mode, and dispatches the correct runtime mode and the dispatching keys before executing the model's forwards. It will take in the initial key (a rough batch_descriptor for the padded input) and return the selected runtime mode and the final batch_descriptor, then tell the CUDAGraphWarpper instances that decision through forward contexts. Notice that `CudagraphDispatcher` is the only source of truth for available CUDA Graph keys and `CUDAGraphWrapper` instances can blindly trust the forward context on what CUDA Graphs to dispatch to. This lets us simplify the wrapper code and centralize the logic in the dispatcher.

101 +The [CudagraphDispatcher][vllm.v1.cudagraph_dispatcher.CudagraphDispatcher] takes responsibility for maintaining two sets of valid dispatching keys, one set for `FULL` runtime mode and one set for `PIECEWISE` runtime mode, and dispatches the correct runtime mode and the dispatching keys before executing the model's forwards. It will take in the initial key (a rough batch_descriptor for the padded input) and return the selected runtime mode and the final batch_descriptor, then tell the CUDAGraphWrapper instances that decision through forward contexts. Notice that `CudagraphDispatcher` is the only source of truth for available CUDA Graph keys and `CUDAGraphWrapper` instances can blindly trust the forward context on what CUDA Graphs to dispatch to. This lets us simplify the wrapper code and centralize the logic in the dispatcher.

102

103 The dispatching keys are initialized through the dispatcher's `initialize_cudagraph_keys` method, which is called by the gpu_model_runner after all possible attention backends are initialized. This is where we can get much fancier in the future and “prepare” all kinds of CUDA Graphs combinations. For now, we just append available keys based on the valid combos of `decode_mode`/`mixed_mode` of `cudagraph_mode` and `cudagraph_capture_sizes` in the compilation config.

104

docs/design/fused_moe_modular_kernel.md

+1 -1

@@ -47,7 +47,7 @@ The TopK Weight Application and Reduction components happen right after the Unpe

47 Please find the implementations of TopKWeightAndReduce [here](../../vllm/model_executor/layers/fused_moe/topk_weight_and_reduce.py).

48

49 `FusedMoEPrepareAndFinalizeModular::finalize()` method accepts a `TopKWeightAndReduce` argument that is invoked inside the method.

50 -The `FusedMoEModularKernel` acts as a bridge between the `FusedMoEExpertsModular` and `FusedMoEPerpareAndFinalize` implementations to determine where the TopK Weight Application and Reduction happens.

50 +The `FusedMoEModularKernel` acts as a bridge between the `FusedMoEExpertsModular` and `FusedMoEPrepareAndFinalize` implementations to determine where the TopK Weight Application and Reduction happens.

51

52 * `FusedMoEExpertsModular::finalize_weight_and_reduce_impl` method returns `TopKWeightAndReduceNoOp` if the `FusedMoEExpertsModular` implementation does the weight application and reduction itself.

53 * `FusedMoEExpertsModular::finalize_weight_and_reduce_impl` method returns `TopKWeightAndReduceContiguous` / `TopKWeightAndReduceNaiveBatched` / `TopKWeightAndReduceDelegate` if the `FusedMoEExpertsModular` implementation needs the `FusedMoEPrepareAndFinalizeModular::finalize()` to do the weight application and reduction.

docs/design/logits_processors.md

+1 -1

@@ -352,7 +352,7 @@ The `BatchUpdate` abstraction models the persistent batch as a list of requests,

352 (s, d, UNIDIRECTIONAL or SWAP)

353 ```

354

355 - * If the Move specifies `UNIDRECTIONAL`:

355 + * If the Move specifies `UNIDIRECTIONAL`:

356

357 * The request at index `s` is moved to index `d`; index `s` becomes an empty slot

358

docs/design/plugin_system.md

+1 -1

@@ -141,7 +141,7 @@ Every plugin has three parts:

141 - triton ops

142 Custom way doesn't work for triton ops now.

143

144 -7. (optional) Implement other plugable modules, such as lora, graph backend, quantization, mamba attention backend, etc.

144 +7. (optional) Implement other pluggable modules, such as lora, graph backend, quantization, mamba attention backend, etc.

145

146 ## Compatibility Guarantee

147

docs/models/pooling_models.md

+2 -2

@@ -641,7 +641,7 @@ Then you obtain the sparse embeddings like this:

641 curl -s http://localhost:8000/pooling -H "Content-Type: application/json" -d '{

642 "model": "BAAI/bge-m3",

643 "task": "token_classify",

644 - "input": ["What is BGE M3?", "Defination of BM25"]

644 + "input": ["What is BGE M3?", "Definition of BM25"]

645 }'

646 ```

647

@@ -657,7 +657,7 @@ You can obtain the colbert embeddings like this:

657 curl -s http://localhost:8000/pooling -H "Content-Type: application/json" -d '{

658 "model": "BAAI/bge-m3",

659 "task": "token_embed",

660 - "input": ["What is BGE M3?", "Defination of BM25"]

660 + "input": ["What is BGE M3?", "Definition of BM25"]

661 }'

662 ```

663

examples/online_serving/dashboards/grafana/query_statistics.json

+1 -1

@@ -349,7 +349,7 @@

349 "defaults": {

350 "color": { "mode": "thresholds" },

351 "mappings": [

352 - { "options": { "Calcultion": { "index": 0, "text": "Last (not null)" } }, "type": "value" }

352 + { "options": { "Calculation": { "index": 0, "text": "Last (not null)" } }, "type": "value" }

353 ],

354 "thresholds": {

355 "mode": "absolute",

pyproject.toml

+25 -164

@@ -124,193 +124,54 @@ python = "./.venv"

124

125 [tool.typos.files]

126 # these files may be written in non english words

127 -extend-exclude = ["tests/models/fixtures/*", "tests/prompts/*",

128 - "benchmarks/sonnet.txt", "tests/lora/data/*", "build/*",

129 - "vllm/third_party/*", "vllm/entrypoints/serve/instrumentator/static/*",

130 - "docs/governance/process.md"]

131 -ignore-hidden = true

132 -ignore-files = true

133 -ignore-dot = true

134 -ignore-vcs = true

135 -ignore-global = true

136 -ignore-parent = true

127 +extend-exclude = ["tests/models/fixtures/*", "tests/prompts/*", "tests/tokenizers_/*",

128 + "benchmarks/sonnet.txt", "tests/lora/data/*", "examples/pooling/token_embed/*", "build/*",

129 + "vllm/third_party/*", "vllm/entrypoints/serve/instrumentator/static/*", "tests/entrypoints/openai/test_transcription_validation.py",

130 + "docs/governance/process.md", "tests/v1/engine/test_fast_incdec_prefix_err.py", ".git/*"]

131 +ignore-hidden = false

132

133 [tool.typos.default]

139 -binary = false

140 -check-filename = false

141 -check-file = true

142 -unicode = true

143 -ignore-hex = true

144 -identifier-leading-digits = false

145 -locale = "en"

146 -extend-ignore-identifiers-re = ["NVML_*", ".*Unc.*", ".*_thw",

147 - ".*UE8M0.*", ".*[UE4M3|ue4m3].*", ".*eles.*",

148 - ".*[Tt]h[rR].*"]

149 -extend-ignore-words-re = []

150 -extend-ignore-re = []

134 +extend-ignore-identifiers-re = [".*[Uu][Ee][0-9][Mm][0-9].*"]

135

136 [tool.typos.default.extend-identifiers]

137 bbc5b7ede = "bbc5b7ede"

154 -womens_doubles = "womens_doubles"

155 -v_2nd = "v_2nd"

156 -# splitted_input = "splitted_input"

138 NOOPs = "NOOPs"

158 -typ = "typ"

139 nin_shortcut = "nin_shortcut"

160 -UperNetDecoder = "UperNetDecoder"

161 -subtile = "subtile"

140 cudaDevAttrMaxSharedMemoryPerBlockOptin = "cudaDevAttrMaxSharedMemoryPerBlockOptin"

163 -SFOuput = "SFOuput"

164 -# huggingface transformers repo uses these words

141 +

142 depthwise_seperable_out_channel = "depthwise_seperable_out_channel"

166 -DepthWiseSeperableConv1d = "DepthWiseSeperableConv1d"

167 -depthwise_seperable_CNN = "depthwise_seperable_CNN"

143 +pard_token = "pard_token"

144 +ptd_token_id = "ptd_token_id"

145 +ser_de = "ser_de"

146 +shared_memory_per_block_optin = "shared_memory_per_block_optin"

147 +FoPE = "FoPE"

148 +k_ot = "k_ot"

149 +view_seperator = "view_seperator"

150 +inverse_std_variences = "inverse_std_variences"

151

152 [tool.typos.default.extend-words]

153 iy = "iy"

171 -tendencias = "tendencias"

154 indx = "indx"

155 # intel cpu features

156 tme = "tme"

157 dout = "dout"

158 Pn = "Pn"

159 arange = "arange"

160 +thw = "thw"

161 +subtile = "subtile"

162 +HSA = "HSA"

163 +setp = "setp"

164 +CPY = "CPY"

165 +thr = "thr"

166 +Thr = "Thr"

167 PARD = "PARD"

168 pard = "pard"

169 AKS = "AKS"

181 -

182 -[tool.typos.type.py]

183 -extend-glob = []

184 -extend-ignore-identifiers-re = []

185 -extend-ignore-words-re = []

186 -extend-ignore-re = []

187 -

188 -[tool.typos.type.py.extend-identifiers]

189 -arange = "arange"

190 -NDArray = "NDArray"

191 -EOFError = "EOFError"

192 -fo = "fo"

170 ba = "ba"

194 -

195 -[tool.typos.type.py.extend-words]

196 -ba = "ba"

197 -nd = "nd"

198 -

199 -[tool.typos.type.cpp]

200 -extend-glob = ["*.cu"]

201 -extend-ignore-identifiers-re = []

202 -extend-ignore-words-re = []

203 -extend-ignore-re = []

204 -

205 -[tool.typos.type.cpp.extend-identifiers]

206 -countr_one = "countr_one"

207 -k_ot = "k_ot"

208 -ot = "ot"

209 -

210 -[tool.typos.type.cpp.extend-words]

211 -

212 -[tool.typos.type.rust]

213 -extend-glob = []

214 -extend-ignore-identifiers-re = []

215 -extend-ignore-words-re = []

216 -extend-ignore-re = []

217 -

218 -[tool.typos.type.rust.extend-identifiers]

219 -flate2 = "flate2"

220 -

221 -[tool.typos.type.rust.extend-words]

222 -ser = "ser"

223 -

224 -[tool.typos.type.lock]

225 -extend-glob = []

226 -check-file = false

227 -extend-ignore-identifiers-re = []

228 -extend-ignore-words-re = []

229 -extend-ignore-re = []

230 -

231 -[tool.typos.type.lock.extend-identifiers]

232 -

233 -[tool.typos.type.lock.extend-words]

234 -

235 -[tool.typos.type.jl]

236 -extend-glob = []

237 -extend-ignore-identifiers-re = []

238 -extend-ignore-words-re = []

239 -extend-ignore-re = []

240 -

241 -[tool.typos.type.jl.extend-identifiers]

242 -

243 -[tool.typos.type.jl.extend-words]

244 -modul = "modul"

245 -egals = "egals"

246 -usig = "usig"

247 -egal = "egal"

248 -

249 -[tool.typos.type.go]

250 -extend-glob = []

251 -extend-ignore-identifiers-re = []

252 -extend-ignore-words-re = []

253 -extend-ignore-re = []

254 -

255 -[tool.typos.type.go.extend-identifiers]

256 -flate = "flate"

257 -

258 -[tool.typos.type.go.extend-words]

259 -

260 -[tool.typos.type.css]

261 -extend-glob = []

262 -extend-ignore-identifiers-re = []

263 -extend-ignore-words-re = []

264 -extend-ignore-re = []

265 -

266 -[tool.typos.type.css.extend-identifiers]

171 +fo = "fo"

172 nd = "nd"

268 -

269 -[tool.typos.type.css.extend-words]

270 -

271 -[tool.typos.type.man]

272 -extend-glob = []

273 -extend-ignore-identifiers-re = []

274 -extend-ignore-words-re = []

275 -extend-ignore-re = []

276 -

277 -[tool.typos.type.man.extend-identifiers]

278 -Nd = "Nd"

279 -

280 -[tool.typos.type.man.extend-words]

281 -

282 -[tool.typos.type.cert]

283 -extend-glob = []

284 -check-file = false

285 -extend-ignore-identifiers-re = []

286 -extend-ignore-words-re = []

287 -extend-ignore-re = []

288 -

289 -[tool.typos.type.cert.extend-identifiers]

290 -

291 -[tool.typos.type.cert.extend-words]

292 -

293 -[tool.typos.type.sh]

294 -extend-glob = []

295 -extend-ignore-identifiers-re = []

296 -extend-ignore-words-re = []

297 -extend-ignore-re = []

298 -

299 -[tool.typos.type.sh.extend-identifiers]

300 -ot = "ot"

301 -

302 -[tool.typos.type.sh.extend-words]

303 -

304 -[tool.typos.type.vimscript]

305 -extend-glob = []

306 -extend-ignore-identifiers-re = []

307 -extend-ignore-words-re = []

308 -extend-ignore-re = []

309 -

310 -[tool.typos.type.vimscript.extend-identifiers]

311 -windo = "windo"

312 -

313 -[tool.typos.type.vimscript.extend-words]

173 +eles = "eles"

174 +datas = "datas"

175

176 [tool.uv]

177 no-build-isolation-package = ["torch"]

tests/compile/test_decorator.py

+1 -1

@@ -234,7 +234,7 @@ def test_conditional_compile_enable_if(use_inductor_graph_partition, monkeypatch

234 expected_num_backend_compilations = 4

235

236 # A has support_torch_compile but enable_if fn returns False

237 - # enalbe_if will be True for B, so we expect mod1 and mod2

237 + # enable_if will be True for B, so we expect mod1 and mod2

238 # to be compiled

239 with compilation_counter.expect(

240 num_graphs_seen=2,

tests/compile/test_wrapper.py

+1 -1

@@ -95,7 +95,7 @@ def test_torch_compile_wrapper(use_bytecode_hook, monkeypatch):

95 f"Expected {expected1}, got {result1}"

96 )

97

98 - # Second call should triger another compilation

98 + # Second call should trigger another compilation

99 x2 = torch.tensor([1, 2, 3])

100 result2 = wrapper(x2)

101 expected2 = torch.tensor([100, 200, 300])

tests/kernels/attention/test_attention.py

+1 -1

@@ -444,7 +444,7 @@ def ref_multi_query_kv_attention(

444

445

446 @pytest.mark.parametrize("attention_cls", [Attention, MMEncoderAttention])

447 -def test_num_heads_not_divisble_by_num_kv_heads(attention_cls: type) -> None:

447 +def test_num_heads_not_divisible_by_num_kv_heads(attention_cls: type) -> None:

448 head_size = 64

449 scale = float(1.0 / (head_size**0.5))

450 num_heads = 16

tests/kernels/moe/test_modular_kernel_combinations.py

+3 -3

@@ -162,7 +162,7 @@ Ns = [1024]

162 TOPKs = [4, 1]

163 Es = [32]

164 DTYPEs = [torch.bfloat16]

165 -FUSED_MOE_CHUNK_SIZEs = [None, 16]

165 +FUSED_MOE_CHUNK_SIZES = [None, 16]

166

167

168 def is_nyi_config(config: Config) -> bool:

@@ -192,7 +192,7 @@ def generate_valid_test_cases(

192 DTYPEs,

193 MK_QUANT_CONFIGS,

194 product(prepare_finalize_types, MK_FUSED_EXPERT_TYPES),

195 - FUSED_MOE_CHUNK_SIZEs,

195 + FUSED_MOE_CHUNK_SIZES,

196 ):

197 total = total + 1

198

@@ -266,7 +266,7 @@ def test_modular_kernel_combinations_multigpu(

266 if cuda_device_count_stateless() < world_size:

267 pytest.skip(

268 f"Not enough GPUs available to run, got "

269 - f"{cuda_device_count_stateless()} exepected "

269 + f"{cuda_device_count_stateless()} expected "

270 f"{world_size}."

271 )

272

tests/models/language/generation/test_mistral.py

+2 -2

@@ -87,7 +87,7 @@ MSGS = [

87 {

88 "role": "user",

89 "content": "Could you please rewrite the below article? \n\n My English needs "

90 - "improvving, maybe I make errors.",

90 + "improving, maybe I make errors.",

91 },

92 {

93 "role": "assistant",

@@ -98,7 +98,7 @@ MSGS = [

98 "type": "function",

99 "function": {

100 "name": "rewrite",

101 - "arguments": '{"text":"My English needs improvving, maybe '

101 + "arguments": '{"text":"My English needs improving, maybe '

102 'I make errors."}',

103 },

104 }

tests/models/language/pooling/test_bge_m3.py

+1 -1

@@ -14,7 +14,7 @@ MAX_MODEL_LEN = 512

14

15

16 # Example from https://huggingface.co/BAAI/bge-m3

17 -sentences_1 = ["What is BGE M3?", "Defination of BM25"]

17 +sentences_1 = ["What is BGE M3?", "Definition of BM25"]

18 sentences_2 = [

19 "BGE M3 is an embedding model supporting dense retrieval, "

20 "lexical matching and multi-vector interaction.",

tests/models/multimodal/generation/vlm_utils/model_utils.py

+2 -2

@@ -719,7 +719,7 @@ def isaac_patch_hf_runner(hf_model: HfRunner) -> HfRunner:

719 # Convert to tuple or None

720 all_hidden_states = tuple(hidden_states_list) if output_hidden_states else None

721

722 - # Include hiden_states for compatibility with hidden_states_to_seq_logprobs()

722 + # Include hidden_states for compatibility with hidden_states_to_seq_logprobs()

723 return BaseModelOutputWithPast(

724 last_hidden_state=hidden_states,

725 past_key_values=past_key_values,

@@ -1226,7 +1226,7 @@ def voxtral_patch_hf_runner(hf_model: "HfRunner") -> "HfRunner":

1226 dicts (accepting ``url``, ``path``, or ``base64`` audio) rather than

1227 the standard ``processor(text=, audio=, sampling_rate=)`` interface.

1228 2. HfRunner.get_inputs cannot handle multi-audio per prompt because it

1229 - mis-unpacks ``[(arr1, sr1), (arr2, sr2)]`` via a ``len == 2`` check.

1229 + incorrectly unpacks ``[(arr1, sr1), (arr2, sr2)]`` via a ``len == 2`` check.

1230

1231 We override ``get_inputs`` to build conversation dicts and call

1232 ``apply_chat_template`` directly, bypassing both issues. We also wrap

tests/quantization/test_blackwell_moe.py

+1 -1

@@ -25,7 +25,7 @@ def set_test_environment():

25 os.environ["FLASHINFER_NVCC_THREADS"] = "16"

26

27

28 -# Overide the backbone layers to 4 for faster startup

28 +# Override the backbone layers to 4 for faster startup

29 HF_OVERRIDE_TEXT = {

30 "num_layers": 4,

31 "num_hidden_layers": 4,

tests/renderers/test_hf.py

+2 -2

@@ -206,8 +206,8 @@ def test_resolve_chat_template_kwargs(sample_json_schema, model, expected_kwargs

206

207 chat_template_kwargs = {

208 # both unused

209 - "unsed_kwargs_1": 123,

210 - "unsed_kwargs_2": "abc",

209 + "unused_kwargs_1": 123,

210 + "unused_kwargs_2": "abc",

211 # should not appear

212 "chat_template": "{% Hello world! %}",

213 "tokenize": True,

tests/test_config.py

+4 -4

@@ -853,7 +853,7 @@ def test_vllm_config_defaults_are_none():

853

854

855 @pytest.mark.parametrize(

856 - ("model_id", "compiliation_config", "optimization_level"),

856 + ("model_id", "compilation_config", "optimization_level"),

857 [

858 (

859 None,

@@ -895,7 +895,7 @@ def test_vllm_config_defaults_are_none():

895 ("RedHatAI/DeepSeek-V2.5-1210-FP8", CompilationConfig(), OptimizationLevel.O3),

896 ],

897 )

898 -def test_vllm_config_defaults(model_id, compiliation_config, optimization_level):

898 +def test_vllm_config_defaults(model_id, compilation_config, optimization_level):

899 """Test that optimization-level defaults are correctly applied."""

900

901 model_config = None

@@ -903,12 +903,12 @@ def test_vllm_config_defaults(model_id, compiliation_config, optimization_level)

903 model_config = ModelConfig(model_id)

904 vllm_config = VllmConfig(

905 model_config=model_config,

906 - compilation_config=compiliation_config,

906 + compilation_config=compilation_config,

907 optimization_level=optimization_level,

908 )

909 else:

910 vllm_config = VllmConfig(

911 - compilation_config=compiliation_config,

911 + compilation_config=compilation_config,

912 optimization_level=optimization_level,

913 )

914 # Use the global optimization level defaults

tests/tool_parsers/test_seed_oss_tool_parser.py

+2 -2

@@ -106,7 +106,7 @@ def test_extract_tool_calls_no_tools(seed_oss_tool_parser):

106 @pytest.mark.parametrize(

107 ids=[

108 "tool_call_0_thinking_budget",

109 - "tool_call_512_thinkg_budget",

109 + "tool_call_512_thinking_budget",

110 "tool_call_unlimited_thinking_budget",

111 ],

112 argnames=["model_output", "expected_tool_calls", "expected_content"],

@@ -308,7 +308,7 @@ def stream_delta_message_generator(

308 @pytest.mark.parametrize(

309 ids=[

310 "tool_call_0_thinking_budget",

311 - "tool_call_512_thinkg_budget",

311 + "tool_call_512_thinking_budget",

312 "tool_call_unlimited_thinking_budget",

313 ],

314 argnames=["model_output", "expected_tool_calls", "expected_content"],

tests/transformers_utils/test_repo_utils.py

+4 -4

@@ -34,10 +34,10 @@ def test_list_filtered_repo_files(

34 subfolder.mkdir()

35 (path_tmp_dir / "json_file.json").touch()

36 (path_tmp_dir / "correct_2.txt").touch()

37 - (path_tmp_dir / "uncorrect.txt").touch()

38 - (path_tmp_dir / "uncorrect.jpeg").touch()

37 + (path_tmp_dir / "incorrect.txt").touch()

38 + (path_tmp_dir / "incorrect.jpeg").touch()

39 (subfolder / "correct.txt").touch()

40 - (subfolder / "uncorrect_sub.txt").touch()

40 + (subfolder / "incorrect_sub.txt").touch()

41

42 def _glob_path() -> list[str]:

43 return [

@@ -86,7 +86,7 @@ def test_one_filtered_repo_files(allow_patterns: list[str], expected_bool: bool)

86 path_tmp_dir = Path(tmp_dir)

87 subfolder = path_tmp_dir / "subfolder"

88 subfolder.mkdir()

89 - (path_tmp_dir / "uncorrect.jpeg").touch()

89 + (path_tmp_dir / "incorrect.jpeg").touch()

90 (subfolder / "correct.txt").touch()

91

92 def _glob_path() -> list[str]:

tests/v1/core/test_kv_cache_utils.py

+1 -1

@@ -308,7 +308,7 @@ def test_free_kv_cache_block_queue_append_n():

308

309 # Create an empty FreeKVCacheBlockQueue

310 invalid_queue = FreeKVCacheBlockQueue([])

311 - # set prev_free_block to None and this will cause assertation in append_n

311 + # set prev_free_block to None and this will cause assertion in append_n

312 invalid_queue.fake_free_list_tail.prev_free_block = None

313 with pytest.raises(AssertionError):

314 # Append 1 block

tests/v1/core/test_prefix_caching.py

+4 -4

@@ -2304,22 +2304,22 @@ def test_block_lookup_cache_single_block_per_key():

2304 assert cache.get_one_block(key0) is block0

2305 assert cache.get_one_block(key1) is block1

2306 assert cache.get_one_block(key2) is None

2307 - # No block poped due to block_id mismatch

2307 + # No block popped due to block_id mismatch

2308 assert cache.pop(key0, 100) is None

2309 assert cache.get_one_block(key0) is block0

2310 assert cache.get_one_block(key1) is block1

2311 assert cache.get_one_block(key2) is None

2312 - # block poped with (key0, block ID 0)

2312 + # block popped with (key0, block ID 0)

2313 assert cache.pop(key0, 0) is block0

2314 assert cache.get_one_block(key0) is None

2315 assert cache.get_one_block(key1) is block1

2316 assert cache.get_one_block(key2) is None

2317 - # No block poped due to block_id mismatch

2317 + # No block popped due to block_id mismatch

2318 assert cache.pop(key0, 1) is None

2319 assert cache.get_one_block(key0) is None

2320 assert cache.get_one_block(key1) is block1

2321 assert cache.get_one_block(key2) is None

2322 - # block poped with (key1, block ID 1)

2322 + # block popped with (key1, block ID 1)

2323 assert cache.pop(key1, 1) is block1

2324 assert cache.get_one_block(key0) is None

2325 assert cache.get_one_block(key1) is None

tests/v1/core/test_priority_scheduler_random.py

+2 -2

@@ -140,7 +140,7 @@ def _mock_draft_token_ids(

140 return DraftTokenIds(req_ids=request_ids, draft_token_ids=sampled_token_ids)

141

142

143 -def _chech_valid_scheduler_output(

143 +def _check_valid_scheduler_output(

144 scheduler_output: SchedulerOutput,

145 seen_request_ids: set[str],

146 seen_mm_hashes: set[str],

@@ -242,7 +242,7 @@ def test_priority_scheduling_blast(

242 )

243 scheduler.add_request(req)

244 scheduler_output = scheduler.schedule()

245 - _chech_valid_scheduler_output(

245 + _check_valid_scheduler_output(

246 scheduler_output, seen_request_ids, seen_mm_hashes

247 )

248 model_output = _mock_execute_model(

tests/v1/core/test_scheduler.py

+9 -9

@@ -1116,7 +1116,7 @@ def _step_until_done(

1116

1117

1118 def _step_until_kv_transfer_finished(scheduler: Scheduler, req_ids: list[str]):

1119 - """Cycle requests through a KV transfer cyle."""

1119 + """Cycle requests through a KV transfer cycle."""

1120

1121 # Requests should first transition to WAITING_FOR_REMOTE_KVS

1122 output = scheduler.schedule()

@@ -2714,7 +2714,7 @@ def _assert_right_encoder_inputs(

2714 if expected_total_reqs == 0:

2715 return

2716

2717 - # Number of expected enocder inputs should match number of requests

2717 + # Number of expected encoder inputs should match number of requests

2718 if expected_encoder_inputs:

2719 assert check_exist and requests is not None # only support expect input exist

2720 assert len(requests) == len(expected_encoder_inputs)

@@ -2964,7 +2964,7 @@ def test_ec_connector_with_partial_cache_hit_multi_round(use_kv_connector):

2964 )

2965 scheduler.update_from_output(output, model_output)

2966

2967 - # request1 is finished after outputing 1 token

2967 + # request1 is finished after outputting 1 token

2968 # Finish request

2969 scheduler.finish_requests(request1.request_id, RequestStatus.FINISHED_LENGTH_CAPPED)

2970

@@ -3060,14 +3060,14 @@ def test_ec_connector_schedule_multiple_requests(cache_exist, use_kv_connector):

3060 for request in requests:

3061 scheduler.add_request(request)

3062

3063 - # Set up to test different encoder cache exsistence scenario after preemption

3063 + # Set up to test different encoder cache existence scenario after preemption

3064 # Order of getting encoder cache should be: local cache -> connector-> compute

3065 scheduler.ec_connector.update_state_after_alloc = Mock(

3066 wraps=scheduler.ec_connector.update_state_after_alloc

3067 )

3068

3069 if cache_exist == "local":

3070 - # Allocate cache to cache manager manually to mimick

3070 + # Allocate cache to cache manager manually to mimic

3071 for req in requests:

3072 scheduler.encoder_cache_manager.allocate(req, 0)

3073 else:

@@ -3384,13 +3384,13 @@ def test_priority_scheduling_ec_connector_preemption_and_resumption(

3384 pooler_output=[],

3385 )

3386 # Finish the requests to make room for the preempted requests to resume

3387 - # req_high is finished after outputing 2 tokens

3387 + # req_high is finished after outputting 2 tokens

3388 scheduler.update_from_output(output, model_output)

3389 scheduler.finish_requests(

3390 request_high.request_id, RequestStatus.FINISHED_LENGTH_CAPPED

3391 )

3392

3393 - # Set up to test different encoder cache exsistence scenario after preemption

3393 + # Set up to test different encoder cache existence scenario after preemption

3394 # Order of getting encoder cache should be: local cache -> connector-> compute

3395 # By default, the cache should still exist in local in this test case

3396 if cache_exist != "local":

@@ -3483,7 +3483,7 @@ def test_ec_connector_allocate_encoder_tokens_with_external_load(use_kv_connecto

3483 ec_role="ec_consumer",

3484 )

3485

3486 - # Limit the number of availiable slots of EncoderCacheManager

3486 + # Limit the number of available slots of EncoderCacheManager

3487 scheduler.encoder_cache_manager = EncoderCacheManager(cache_size=32)

3488

3489 # Create MM request1

@@ -3574,7 +3574,7 @@ def test_ec_connector_allocate_encoder_tokens_with_external_load(use_kv_connecto

3574 )

3575 scheduler.update_from_output(output, model_output)

3576

3577 - # request1 is finished after outputing 1 token

3577 + # request1 is finished after outputting 1 token

3578 # Finish request

3579 scheduler.finish_requests(request1.request_id, RequestStatus.FINISHED_LENGTH_CAPPED)

3580 assert scheduler.get_num_unfinished_requests() == 1

tests/v1/e2e/test_mamba_prefix_cache.py

+2 -2

@@ -76,11 +76,11 @@ def get_fake_sample_fn() -> SamplerOutput:

76 ),

77 logprobs_tensors=None,

78 )

79 - accpeted_tokens = prompt_token_ids[

79 + accepted_tokens = prompt_token_ids[

80 first_token_id_index : first_token_id_index

81 + min(num_accepted_tokens, logits.shape[0])

82 ]

83 - sampled_token_ids = accpeted_tokens

83 + sampled_token_ids = accepted_tokens

84 return SamplerOutput(

85 sampled_token_ids=torch.tensor(

86 [sampled_token_ids], device="cuda", dtype=torch.int32

tests/v1/entrypoints/llm/test_struct_output_generate.py

+1 -1

@@ -911,7 +911,7 @@ def test_structured_output_with_structural_tag(backend: str):

911 ),

912 )

913

914 - prompt = "Hello and repete hello 10 times, do not say anything else. Only say hello hello hello, now start"

914 + prompt = "Hello and repeat hello 10 times, do not say anything else. Only say hello hello hello, now start"

915 outputs = llm.generate(prompt, sampling_params=sampling_params, use_tqdm=True)

916 assert outputs is not None

917 for output in outputs:

tests/v1/kv_connector/unit/test_kv_connector_lifecycle.py renamed

tests/v1/kv_connector/unit/test_moriio_connector.py

+5 -5

@@ -99,7 +99,7 @@ def _setup_kv_transfer_request(request, remote_host="127.0.0.1", fake_port=4789)

99 return request

100

101

102 -class FakeMorIIOWrapper:

102 +class FakeMoRIIOWrapper:

103 # A fake MoRIIOWrapper for testing purposes

104 def __init__(self, *args, **kwargs):

105 pass

@@ -168,7 +168,7 @@ class FakeMorIIOWrapper:

168 pass

169

170

171 -class FakeMorIIOConnectorWorker(MoRIIOConnectorWorker):

171 +class FakeMoRIIOConnectorWorker(MoRIIOConnectorWorker):

172 # Define a fake remote engine id for testing

173 REMOTE_ENGINE_ID = "remote_engine"

174

@@ -373,7 +373,7 @@ def test_read_mode_loads_remote_block_ids(moriio_read_mode):

373 # Set remote block ids to be fetched.

374 request.kv_transfer_params["remote_block_ids"] = block_list

375

376 - # Remote Prefill, triggers MorIIOConnectorMetadata.

376 + # Remote Prefill, triggers MoRIIOConnectorMetadata.

377

378 scheduler_output = scheduler.schedule()

379 kv_connector_metadata = scheduler_output.kv_connector_metadata

@@ -451,7 +451,7 @@ def test_register_kv_caches(mock_parallel_groups):

451

452 with set_current_vllm_config(vllm_config):

453 connector = MoRIIOConnector(vllm_config, KVConnectorRole.WORKER)

454 - connector.connector_worker = FakeMorIIOConnectorWorker(

454 + connector.connector_worker = FakeMoRIIOConnectorWorker(

455 vllm_config, connector.engine_id, hand_shake_latency=0

456 )

457

@@ -528,7 +528,7 @@ def test_moriio_handshake_returns_metadata(mock_parallel_groups):

528 with (

529 patch(

530 "vllm.distributed.kv_transfer.kv_connector.v1.moriio.moriio_engine.MoRIIOWrapper",

531 - FakeMorIIOWrapper,

531 + FakeMoRIIOWrapper,

532 ),

533 ):

534 handshake_port = _find_free_port()

tests/v1/kv_connector/unit/test_nixl_connector.py

+4 -4

@@ -460,9 +460,9 @@ class FakeNixlConnectorWorker(NixlConnectorWorker):

460

461 # When remote tp_size > local tp_size, handshake with multiple

462 # remote ranks.

463 - num_hanshakes = 1 if tp_ratio > 0 else -tp_ratio

463 + num_handshakes = 1 if tp_ratio > 0 else -tp_ratio

464 remote_agents: dict[int, str] = {}

465 - for remote_tp_rank in range(num_hanshakes):

465 + for remote_tp_rank in range(num_handshakes):

466 remote_agent_name = self.add_remote_agent(

467 NixlAgentMetadata(

468 engine_id=self.REMOTE_ENGINE_ID,

@@ -688,7 +688,7 @@ class TestNixlHandshake:

688 )

689 check_handshake(2)

690

691 - # NOTE flexiblity: a second remote with higher number of ranks is

691 + # NOTE flexibility: a second remote with higher number of ranks is

692 # discovered. This is not a scenario we actively support right now, but

693 # the connector allows it.

694 worker.REMOTE_ENGINE_ID = "remote_engine_2"

@@ -1766,7 +1766,7 @@ def test_aborted_request_removed_from_worker_in_batch(default_vllm_config, dist_

1766 req = create_request(request_id=1, do_remote_decode=True, max_tokens=1)

1767 scheduler.add_request(req)

1768

1769 - # First scheduling pass - examinate build_connector_meta output

1769 + # First scheduling pass - examine build_connector_meta output

1770 sched_out = scheduler.schedule()

1771 kv_meta = sched_out.kv_connector_metadata

1772 assert kv_meta is not None

tests/v1/sample/test_logprobs.py

+1 -1

@@ -36,7 +36,7 @@ SAMPLE_PROMPT = BatchLogprobsComposition.SAMPLE_PROMPT

36 # non-associative and sensitive to batch geometry. The ref LLM (no spec

37 # decode, default scheduling) and the spec-decode LLM (chunked prefill,

38 # different effective batch sizes) follow different reduction orders,

39 -# producing numerically divergent logprobs that get mis-attributed to

39 +# producing numerically divergent logprobs that get misattributed to

40 # spec-decode incorrectness.

41 #

42 # Force LLM instances into an identical, deterministic execution

tests/v1/sample/test_rejection_sampler.py

+3 -3

@@ -726,7 +726,7 @@ def test_frequency_penalties(rejection_sampler):

726 spec_tokens = [[1, 1, 1], [], [1, 1, 1]]

727 output_tokens = [[1, 1, 1, 1], [7], [1, 1, 1, 1]] # 1, 7 and 1 are the bonus tokens

728

729 - num_requsts = len(spec_tokens)

729 + num_requests = len(spec_tokens)

730 logits = create_logits_tensor(output_tokens, token_idx_to_override=15)

731 metadata = create_sampling_metadata(

732 all_greedy=True,

@@ -734,8 +734,8 @@ def test_frequency_penalties(rejection_sampler):

734 spec_token_ids=spec_tokens,

735 prompt_token_ids=torch.tensor([[5, 6, 7], [6, 7, 8], [7, 8, 9]], device=DEVICE),

736 frequency_penalties=[1.5, 1.5, 0.7],

737 - presence_penalties=[0.0] * num_requsts,

738 - repetition_penalties=[1.0] * num_requsts,

737 + presence_penalties=[0.0] * num_requests,

738 + repetition_penalties=[1.0] * num_requests,

739 )

740 bonus_token_tensor = torch.tensor(

741 [output_tokens[i][-1] for i in range(len(output_tokens))], device=logits.device

vllm/_custom_ops.py

+2 -2

@@ -3106,7 +3106,7 @@ def cpu_attn_get_scheduler_metadata(

3106 isa: str,

3107 enable_kv_split: bool,

3108 ) -> torch.Tensor:

3109 - sheduler_metadata = torch.ops._C.get_scheduler_metadata(

3109 + scheduler_metadata = torch.ops._C.get_scheduler_metadata(

3110 num_reqs,

3111 num_heads,

3112 num_kv_heads,

@@ -3119,7 +3119,7 @@ def cpu_attn_get_scheduler_metadata(

3119 isa,

3120 enable_kv_split,

3121 )

3122 - return sheduler_metadata

3122 + return scheduler_metadata

3123

3124

3125 def cpu_attn_reshape_and_cache(

vllm/config/compilation.py

+1 -1

@@ -872,7 +872,7 @@ class CompilationConfig:

872 )

873

874 # Currently only eager and inductor backend are supported.

875 - # for piecewise compilation. Custom backends are not suppported for

875 + # for piecewise compilation. Custom backends are not supported for

876 # piecewise compilation. Update when more backends are supported.

877 if self.mode == CompilationMode.VLLM_COMPILE and self.backend not in [

878 "",

vllm/config/observability.py

+1 -1

@@ -59,7 +59,7 @@ class ObservabilityConfig:

59

60 enable_layerwise_nvtx_tracing: bool = False

61 """Enable layerwise NVTX tracing. This traces the execution of each layer or

62 - module in the model and attach informations such as input/output shapes to

62 + module in the model and attach information such as input/output shapes to

63 nvtx range markers. Noted that this doesn't work with CUDA graphs enabled."""

64

65 enable_mfu_metrics: bool = False

vllm/config/vllm.py

+1 -1

@@ -592,7 +592,7 @@ class VllmConfig:

592

593 If the user configuration does not specify a value for a default field

594 and if the default field is still None after all user selections are

595 - applied, then default values will be applied to the field. User speciied

595 + applied, then default values will be applied to the field. User specified

596 fields will not be overridden by the default.

597

598 Args:

vllm/distributed/eplb/policy/default.py

+1 -1

@@ -44,7 +44,7 @@ class DefaultEplbPolicy(AbstractEplbPolicy):

44 rank_in_pack = np.zeros_like(pack_index, dtype=np.int64)

45 return pack_index, rank_in_pack

46

47 - # Sort and get indices in decending order

47 + # Sort and get indices in descending order

48 indices = np.argsort(-weight, axis=-1)

49

50 pack_index = np.full((num_layers, num_groups), -1, dtype=np.int64)

vllm/distributed/kv_transfer/kv_connector/v1/base.py

+1 -1

@@ -129,7 +129,7 @@ class KVConnectorRole(enum.Enum):

129 class KVConnectorHandshakeMetadata(ABC): # noqa: B024

130 """

131 Metadata used for out of band connector handshake between

132 - P/D workers. This needs to serializeable.

132 + P/D workers. This needs to serializable.

133 """

134

135 pass

vllm/distributed/kv_transfer/kv_connector/v1/lmcache_integration/vllm_v1_adapter.py

+2 -2

@@ -398,7 +398,7 @@ class ReqMeta:

398 )

399

400

401 -def need_gpu_interm_buffer(lmcache_config: LMCacheEngineConfig):

401 +def need_gpu_interim_buffer(lmcache_config: LMCacheEngineConfig):

402 return not lmcache_config.enable_pd

403

404

@@ -497,7 +497,7 @@ def _init_lmcache_engine(

497 use_mla,

498 )

499

500 - use_gpu = need_gpu_interm_buffer(lmcache_config)

500 + use_gpu = need_gpu_interim_buffer(lmcache_config)

501 vllm_gpu_connector: (

502 VLLMBufferLayerwiseGPUConnector

503 | VLLMPagedMemGPUConnectorV2

vllm/distributed/kv_transfer/kv_connector/v1/mooncake/mooncake_connector.py

+7 -7

@@ -481,7 +481,7 @@ class MooncakeConnectorWorker:

481 )

482

483 self._remote_agents: dict[EngineId, dict[int, dict[int, str]]] = {}

484 - self._pending_bootstrap_querys: dict[str, asyncio.Event] = {}

484 + self._pending_bootstrap_queries: dict[str, asyncio.Event] = {}

485 self.side_channel_port: int = 0 # we will bind it in register_kv_caches()

486 self.engine_id: EngineId = engine_id

487 self.tp_rank = get_tensor_model_parallel_rank()

@@ -1077,7 +1077,7 @@ class MooncakeConnectorWorker:

1077 response = self._xfer_resp_decoder.decode(ret_msg)

1078 if response.status == MooncakeXferResponseStatus.ERROR:

1079 logger.error(

1080 - "Error happens during tranfering kvcache for %s: %s",

1080 + "Error happens during transferring kvcache for %s: %s",

1081 req_ids,

1082 response.err_msg,

1083 )

@@ -1140,8 +1140,8 @@ class MooncakeConnectorWorker:

1140 )

1141

1142 # Always notify others regardless of connection success or failure.

1143 - self._pending_bootstrap_querys[remote_bootstrap_addr].set()

1144 - del self._pending_bootstrap_querys[remote_bootstrap_addr]

1143 + self._pending_bootstrap_queries[remote_bootstrap_addr].set()

1144 + del self._pending_bootstrap_queries[remote_bootstrap_addr]

1145

1146 def receive_kv(

1147 self,

@@ -1171,11 +1171,11 @@ class MooncakeConnectorWorker:

1171 pull_metas: dict[ReqId, PullReqMeta],

1172 ):

1173 remote_bootstrap_addr = next(iter(pull_metas.values())).remote_bootstrap_addr

1174 - if remote_bootstrap_addr not in self._pending_bootstrap_querys:

1175 - self._pending_bootstrap_querys[remote_bootstrap_addr] = asyncio.Event()

1174 + if remote_bootstrap_addr not in self._pending_bootstrap_queries:

1175 + self._pending_bootstrap_queries[remote_bootstrap_addr] = asyncio.Event()

1176 await self._connect_to_prefiller_bootstrap(remote_bootstrap_addr)

1177 else:

1178 - await self._pending_bootstrap_querys[remote_bootstrap_addr].wait()

1178 + await self._pending_bootstrap_queries[remote_bootstrap_addr].wait()

1179

1180 if remote_engine_id not in self._remote_agents:

1181 logger.error(

vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py

+1 -1

@@ -720,7 +720,7 @@ class OffloadPromMetrics(KVConnectorPromMetrics):

720 per_engine_labelvalues: dict[int, list[object]],

721 ):

722 super().__init__(vllm_config, metric_types, labelnames, per_engine_labelvalues)

723 - # (engine_idx, transfer_tupe) -> (metric with bounded labels)

723 + # (engine_idx, transfer_type) -> (metric with bounded labels)

724 self.histogram_transfer_size: dict[tuple[int, str], PromMetricT] = {}

725 self.counter_kv_bytes: dict[tuple[int, str], PromMetricT] = {}

726 self.counter_kv_transfer_time: dict[tuple[int, str], PromMetricT] = {}

vllm/entrypoints/openai/responses/serving.py

+3 -3

@@ -1647,9 +1647,9 @@ class OpenAIServingResponses(OpenAIServing):

1647 # TODO: in streaming, we noticed this bug:

1648 # https://github.com/vllm-project/vllm/issues/25697

1649 await self._initialize_tool_sessions(request, context, exit_stack)

1650 - processer = self._process_harmony_streaming_events

1650 + processor = self._process_harmony_streaming_events

1651 else:

1652 - processer = self._process_simple_streaming_events

1652 + processor = self._process_simple_streaming_events

1653 # TODO Hanchen make sampling params to include the structural tag

1654

1655 initial_response = ResponsesResponse.from_request(

@@ -1677,7 +1677,7 @@ class OpenAIServingResponses(OpenAIServing):

1677 )

1678

1679 try:

1680 - async for event_data in processer(

1680 + async for event_data in processor(

1681 request,

1682 sampling_params,

1683 result_generator,

vllm/envs.py

+1 -1

@@ -1520,7 +1520,7 @@ environment_variables: dict[str, Callable[[], Any]] = {

1520 os.getenv("VLLM_DEEPEP_BUFFER_SIZE_MB", "1024")

1521 ),

1522 # Force DeepEP to use intranode kernel for inter-node communication in

1523 - # high throughput mode. This is useful archive higher prefill throuhgput

1523 + # high throughput mode. This is useful archive higher prefill throughput

1524 # on system supports multi-node nvlink (e.g GB200).

1525 "VLLM_DEEPEP_HIGH_THROUGHPUT_FORCE_INTRA_NODE": lambda: bool(

1526 int(os.getenv("VLLM_DEEPEP_HIGH_THROUGHPUT_FORCE_INTRA_NODE", "0"))

vllm/forward_context.py

+1 -1

@@ -175,7 +175,7 @@ class DPMetadata:

175 # Get the cumulative tokens across sequence parallel ranks.

176 # In this case the input to the MoEs will be distributed w.r.t both

177 # DP and TP rank.

178 - # When sp_size==1, this is just the cummulative num tokens across DP.

178 + # When sp_size==1, this is just the cumulative num tokens across DP.

179 def cu_tokens_across_sp(self, sp_size: int) -> torch.Tensor:

180 num_tokens_across_sp_cpu = (

181 self.num_tokens_across_dp_cpu - 1 + sp_size

vllm/lora/layers/row_parallel_linear.py

+2 -2

@@ -57,10 +57,10 @@ class RowParallelLinearWithLoRA(BaseLinearLayerWithLoRA):

57 input_parallel = input_

58 else:

59 # TODO: simplify code below

60 - splitted_input = split_tensor_along_last_dim(

60 + split_input = split_tensor_along_last_dim(

61 input_, num_partitions=self.tp_size

62 )

63 - input_parallel = splitted_input[self.tp_rank].contiguous()

63 + input_parallel = split_input[self.tp_rank].contiguous()

64

65 # Matrix multiply.

66 bias_ = (

vllm/lora/lora_model.py

+3 -3

@@ -11,7 +11,7 @@ from vllm.lora.lora_weights import LoRALayerWeights

11 from vllm.lora.peft_helper import PEFTHelper

12 from vllm.lora.utils import (

13 get_lora_id,

14 - is_base_embeddding_weights,

14 + is_base_embedding_weights,

15 parse_fine_tuned_lora_name,

16 )

17 from vllm.model_executor.model_loader.tensorizer import TensorizerConfig

@@ -86,7 +86,7 @@ class LoRAModel:

86 pin_memory = str(device) == "cpu" and is_pin_memory_available()

87 loras: dict[str, LoRALayerWeights] = {}

88 for tensor_name, tensor in tensors.items():

89 - if is_base_embeddding_weights(tensor_name):

89 + if is_base_embedding_weights(tensor_name):

90 continue

91 # Skip modules based on model-defined prefixes (e.g., MTP layers)

92 if skip_prefixes and cls._should_skip_module(tensor_name, skip_prefixes):

@@ -162,7 +162,7 @@ class LoRAModel:

162

163 def check_unexpected_modules(modules: dict):

164 for lora_module in modules.keys(): # noqa

165 - if is_base_embeddding_weights(lora_module):

165 + if is_base_embedding_weights(lora_module):

166 continue

167 # Handle PEFT file format where experts.base_layer is the

168 # gate_up_proj and experts is the down_proj

vllm/lora/utils.py

+1 -1

@@ -193,7 +193,7 @@ def parse_fine_tuned_lora_name(

193 raise ValueError(f"{name} is unsupported LoRA weight")

194

195

196 -def is_base_embeddding_weights(name: str) -> bool:

196 +def is_base_embedding_weights(name: str) -> bool:

197 # hardcoded subfixes for input & output embedding weights

198 embedding_suffixes = (

199 ".embed_tokens.base_layer.weight",

vllm/model_executor/kernels/linear/mixed_precision/cpu.py

+1 -1

@@ -82,7 +82,7 @@ class CPUWNA16LinearKernel(MPLinearKernel):

82 weight = weight.permute(0, 2, 1).reshape(input_size, output_size).contiguous()

83 weight = pack_quantized_values_into_int32(weight, self.config.weight_type, 1)

84 # make 16 output channel as a block and transpose to the make

85 - # the block contigous

85 + # the block contiguous

86 weight = (

87 weight.view(input_size, -1, 16 // pack_factor)

88 .permute(1, 0, 2)

vllm/model_executor/layers/attention/mla_attention.py

+1 -1

@@ -2540,7 +2540,7 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):

2540 )

2541 # workspace

2542 # |------- N tokens --------|--------- N*dcp_size tokens ----------|

2543 - # |<- use for loca_gather ->|<--------- use for allgather -------->|

2543 + # |<- use for local_gather ->|<--------- use for allgather -------->|

2544 allgather_offset = workspace.shape[0] // (dcp_world_size + 1)

2545 assert allgather_offset * (dcp_world_size + 1) == workspace.shape[0]

2546 assert toks <= allgather_offset

vllm/model_executor/layers/fused_moe/flashinfer_cutlass_moe.py

+1 -1

@@ -394,5 +394,5 @@ class FlashInferExperts(mk.FusedMoEExpertsModular):

394

395 def moe_sum(self, input: torch.Tensor, output: torch.Tensor) -> None:

396 # No support for LoRA in flashinfer_cutlass_fused_moe.

397 - # See TODOs in flashinfer functions runMoe and runMoeMinLantency.

397 + # See TODOs in flashinfer functions runMoe and runMoeMinLatency.

398 raise NotImplementedError("LoRA is not supported for flashinfer_cutlass_moe")

vllm/model_executor/layers/fused_moe/fused_marlin_moe.py

+1 -1

@@ -409,7 +409,7 @@ def batched_fused_marlin_moe(

409 Note that the moe_align_block_size function indicates,

410 - What rows of the A matrix (hidden_states) to access during the

411 matmul, via sorted_ids output.

412 - - What expert_id to use for each block matmul, via expert_ids ouptut.

412 + - What expert_id to use for each block matmul, via expert_ids output.

413

414 In the batched version, the tokens are already grouped/batched by experts

415 they subscribe to. Due to this, we can represent the batched hidden_states

vllm/model_executor/layers/fused_moe/modular_kernel.py

+3 -3

@@ -606,7 +606,7 @@ class FusedMoEExperts(ABC):

606 """

607 Whether the kernel supports deployment in particular parallel config.

608

609 - Can be overriden if a kernel does not support EP, SP or some other

609 + Can be overridden if a kernel does not support EP, SP or some other

610 configuration.

611 """

612 raise NotImplementedError

@@ -620,7 +620,7 @@ class FusedMoEExperts(ABC):

620 """

621 Whether the kernel supports a routing method (e.g. GroupedTopK).

622

623 - Can be overriden by monolithic kernels that execute the router

623 + Can be overridden by monolithic kernels that execute the router

624 in addition to the experts if certain routers are not supported.

625 """

626 return True

@@ -633,7 +633,7 @@ class FusedMoEExperts(ABC):

633 """

634 Whether a kernel supports a particular dtype for router logits input.

635

636 - Can be overriden by monolithic kernels that execute the router

636 + Can be overridden by monolithic kernels that execute the router

637 in addition to the experts if certain dtypes are not supported.

638 """

639 return True

vllm/model_executor/layers/linear.py

+2 -2

@@ -1502,10 +1502,10 @@ class RowParallelLinear(LinearBase):

1502 if self.input_is_parallel:

1503 input_parallel = input_

1504 else:

1505 - splitted_input = split_tensor_along_last_dim(

1505 + split_input = split_tensor_along_last_dim(

1506 input_, num_partitions=self.tp_size

1507 )

1508 - input_parallel = splitted_input[self.tp_rank].contiguous()

1508 + input_parallel = split_input[self.tp_rank].contiguous()

1509

1510 # Matrix multiply.

1511 assert self.quant_method is not None

vllm/model_executor/layers/mla.py

+1 -1

@@ -35,7 +35,7 @@ class MultiHeadLatentAttentionWrapper(PluggableLayer):

35 """Pluggable MLA layer which allows OOT backends to add

36 custom implementations of the outer MLA layer (including rope & o_proj).

37 Note that currently oot platforms can still use CustomOp.register_oot to

38 - replace MLA layer entirly, although we use PluggableLayer to register

38 + replace MLA layer entirely, although we use PluggableLayer to register

39 this layer now.

40

41 This class takes positions and hidden_states as input.

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py

+1 -1

@@ -191,7 +191,7 @@ class CompressedTensorsConfig(QuantizationConfig):

191 """

192 Helper function to update target_scheme_map

193 since linear layers get fused into FusedMoE

194 - targetting 'Linear' needs to also match

194 + targeting 'Linear' needs to also match

195 FusedMoE modules.

196 """

197 if (

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py

+1 -1

@@ -2445,7 +2445,7 @@ class CompressedTensorsW4A8Fp8MoEMethod(CompressedTensorsMoEMethod):

2445 w2_scale=layer.w2_weight_scale, # group scale

2446 g1_alphas=layer.w13_weight_chan_scale,

2447 g2_alphas=layer.w2_weight_chan_scale,

2448 - per_act_token_quant=True, # always use dynamc per-token

2448 + per_act_token_quant=True, # always use dynamic per-token

2449 per_out_ch_quant=True, # always use per-channel

2450 )

2451

vllm/model_executor/layers/quantization/cpu_wna16.py

+1 -1

@@ -261,7 +261,7 @@ class CPUAWQLinearMethod(LinearMethodBase):

261

262 zeros = pack_cols(zeros, bits, group_num, output_size).contiguous()

263 # make 16 output channel as a block and transpose to

264 - # the make the block contigous

264 + # the make the block contiguous

265 weight = pack_cols(weight, bits, input_size, output_size)

266 weight = (

267 weight.view(input_size, -1, 16 // pack_factor)

vllm/model_executor/layers/quantization/torchao.py

+1 -1

@@ -199,7 +199,7 @@ class TorchAOConfig(QuantizationConfig):

199

200 @classmethod

201 def from_config_dict_json(cls, config_dict_json: str) -> "TorchAOConfig":

202 - """Iniitalize class from a config_dict json string, got from

202 + """Initialize class from a config_dict json string, got from

203 torchao_config_object = some AOBaseConfig object

204 json.dumps(config_to_dict(torchao_config_object))

205 """

vllm/model_executor/layers/quantization/utils/fp8_utils.py

+1 -1

@@ -255,7 +255,7 @@ def _flashinfer_fp8_blockscale_gemm_impl(

255

256 This batch-size-dependent selection is essential for maintaining model accuracy.

257 Benchmarks on GSM8K show a significant accuracy gap (88% vs 95%) for DeepSeek-V3.1

258 - when using FlashInfer's DeepGEMM on M>=32. The M < 32 strategy fixes the accurracy

258 + when using FlashInfer's DeepGEMM on M>=32. The M < 32 strategy fixes the accuracy

259 drop.

260

261 Args:

vllm/model_executor/layers/quantization/utils/machete_utils.py

+2 -2

@@ -39,7 +39,7 @@ def query_machete_supported_group_sizes(act_type: torch.dtype) -> list[int]:

39

40

41 def check_machete_supports_shape(

42 - in_features: int, out_featrues: int

42 + in_features: int, out_features: int

43 ) -> tuple[bool, str | None]:

44 if in_features % MACHETE_PREPACKED_BLOCK_SHAPE[0] != 0:

45 return (

@@ -47,7 +47,7 @@ def check_machete_supports_shape(

47 "Input features size must be divisible by "

48 f"{MACHETE_PREPACKED_BLOCK_SHAPE[0]}",

49 )

50 - if out_featrues % MACHETE_PREPACKED_BLOCK_SHAPE[1] != 0:

50 + if out_features % MACHETE_PREPACKED_BLOCK_SHAPE[1] != 0:

51 return (

52 False,

53 "Output features size must be divisible by "

vllm/model_executor/layers/rotary_embedding/common.py

+2 -2

@@ -237,7 +237,7 @@ class ApplyRotaryEmb(CustomOp):

237 Arguments of apply_rotary_emb() in vllm_flash_attn:

238 x: [batch_size, seq_len, nheads, headdim]

239 cos, sin: [seqlen_rotary, rotary_dim / 2]

240 - interleaved: defalut as False (Neox-style).

240 + interleaved: default as False (Neox-style).

241 ...

242 """

243 interleaved = not self.is_neox_style

@@ -259,7 +259,7 @@ class ApplyRotaryEmb(CustomOp):

259 Arguments of apply_rotary() in flash_attn:

260 x: [batch_size, seq_len, nheads, headdim]

261 cos, sin: [seqlen_rotary, rotary_dim / 2]

262 - interleaved: defalut as False (Neox-style).

262 + interleaved: default as False (Neox-style).

263 ...

264 """

265 interleaved = not self.is_neox_style

vllm/model_executor/models/ernie45_vl_moe.py

+6 -6

@@ -342,7 +342,7 @@ class Ernie4_5_VLMoeMoE(nn.Module):

342 visual_token_mask = visual_token_mask.repeat(1, self.hidden_size).bool()

343 text_token_mask = ~visual_token_mask

344 final_experts_hidden_states = torch.zeros_like(hidden_states)

345 - final_shared_ouput = (

345 + final_shared_output = (

346 torch.zeros_like(hidden_states) if self.has_shared_experts else None

347 )

348

@@ -356,26 +356,26 @@ class Ernie4_5_VLMoeMoE(nn.Module):

356 text_router_logits, _ = self.text_experts_gate(

357 text_hidden_states.to(dtype=torch.float32)

358 )

359 - text_shared_ouput, text_experts_output = self.text_experts(

359 + text_shared_output, text_experts_output = self.text_experts(

360 hidden_states=text_hidden_states, router_logits=text_router_logits

361 )

362 final_experts_hidden_states[text_token_mask] = text_experts_output.flatten()

363 if self.has_shared_experts:

364 - final_shared_ouput[text_token_mask] = text_shared_ouput.flatten()

364 + final_shared_output[text_token_mask] = text_shared_output.flatten()

365

366 vision_router_logits, _ = self.vision_experts_gate(

367 vision_hidden_states.to(dtype=torch.float32)

368 )

369 - vision_shared_ouput, vision_experts_output = self.vision_experts(

369 + vision_shared_output, vision_experts_output = self.vision_experts(

370 hidden_states=vision_hidden_states, router_logits=vision_router_logits

371 )

372 final_experts_hidden_states[visual_token_mask] = (

373 vision_experts_output.flatten()

374 )

375 if self.has_shared_experts:

376 - final_shared_ouput[visual_token_mask] = vision_shared_ouput.flatten()

376 + final_shared_output[visual_token_mask] = vision_shared_output.flatten()

377

378 - final_hidden_states = (final_shared_ouput, final_experts_hidden_states)

378 + final_hidden_states = (final_shared_output, final_experts_hidden_states)

379 else:

380 # only text modal input

381 text_router_logits, _ = self.text_experts_gate(

vllm/model_executor/models/fireredasr2.py

+1 -1

@@ -107,7 +107,7 @@ class Conv2dSubsampling(nn.Module):

107 )

108

109 self.subsampling = 4

110 - left_context = right_context = 3 # both exclude currect frame

110 + left_context = right_context = 3 # both exclude current frame

111 self.context = left_context + 1 + right_context # 7

112

113 def forward(

vllm/model_executor/models/funasr.py

+9 -9

@@ -115,7 +115,7 @@ class EncoderLayerSANM(nn.Module):

115 hidden_states: torch.Tensor,

116 mask: torch.Tensor | None = None,

117 cache=None,

118 - mask_shfit_chunk=None,

118 + mask_shift_chunk=None,

119 mask_att_chunk_encoder=None,

120 ):

121 residual = hidden_states

@@ -125,14 +125,14 @@ class EncoderLayerSANM(nn.Module):

125 hidden_states = residual + self.self_attn(

126 hidden_states,

127 mask,

128 - mask_shfit_chunk=mask_shfit_chunk,

128 + mask_shift_chunk=mask_shift_chunk,

129 mask_att_chunk_encoder=mask_att_chunk_encoder,

130 )

131 else:

132 hidden_states = self.self_attn(

133 hidden_states,

134 mask,

135 - mask_shfit_chunk=mask_shfit_chunk,

135 + mask_shift_chunk=mask_shift_chunk,

136 mask_att_chunk_encoder=mask_att_chunk_encoder,

137 )

138

@@ -140,7 +140,7 @@ class EncoderLayerSANM(nn.Module):

140 hidden_states = self.norm2(hidden_states)

141 hidden_states = residual + self.feed_forward(hidden_states)

142

143 - return hidden_states, mask, cache, mask_shfit_chunk, mask_att_chunk_encoder

143 + return hidden_states, mask, cache, mask_shift_chunk, mask_att_chunk_encoder

144

145

146 class MultiHeadedAttentionSANM(nn.Module):

@@ -183,13 +183,13 @@ class MultiHeadedAttentionSANM(nn.Module):

183 self,

184 inputs: torch.Tensor,

185 mask: torch.Tensor,

186 - mask_shfit_chunk: torch.Tensor = None,

186 + mask_shift_chunk: torch.Tensor = None,

187 ):

188 b, t, d = inputs.size()

189 if mask is not None:

190 mask = torch.reshape(mask, (b, -1, 1))

191 - if mask_shfit_chunk is not None:

192 - mask = mask * mask_shfit_chunk

191 + if mask_shift_chunk is not None:

192 + mask = mask * mask_shift_chunk

193 inputs = inputs * mask

194

195 x = inputs.transpose(1, 2)

@@ -243,11 +243,11 @@ class MultiHeadedAttentionSANM(nn.Module):

243 self,

244 hidden_states: torch.Tensor,

245 mask: torch.Tensor,

246 - mask_shfit_chunk: torch.Tensor = None,

246 + mask_shift_chunk: torch.Tensor = None,

247 mask_att_chunk_encoder: torch.Tensor = None,

248 ):

249 q_h, k_h, v_h, v = self.forward_qkv(hidden_states)

250 - fsmn_memory = self.forward_fsmn(v, mask, mask_shfit_chunk)

250 + fsmn_memory = self.forward_fsmn(v, mask, mask_shift_chunk)

251 q_h = q_h * self.d_k ** (-0.5)

252 scores = torch.matmul(q_h, k_h.transpose(-2, -1))

253 att_outs = self.forward_attention(v_h, scores, mask, mask_att_chunk_encoder)

vllm/model_executor/models/isaac.py

+1 -1

@@ -646,7 +646,7 @@ class IsaacImageProcessor:

646 return_tensors: str | TensorType | None,

647 **kwargs: Unpack[IsaacImageProcessorKwargs],

648 ) -> BatchFeature:

649 - """Preprocess images into format compatibile with vLLM input processing."""

649 + """Preprocess images into format compatible with vLLM input processing."""

650

651 all_pixel_values: list[torch.Tensor] = []

652 all_image_grids: list[torch.Tensor] = []

vllm/model_executor/models/keye.py

+1 -1

@@ -299,7 +299,7 @@ class KeyeVisionEmbeddings(nn.Module):

299 )

300 (

301 batch_size,

302 - squence_len,

302 + sequence_len,

303 channel,

304 height,

305 width,

vllm/model_executor/models/longcat_flash.py

+8 -8

@@ -238,7 +238,7 @@ class LongcatRouter(nn.Module):

238 self,

239 config: FlashConfig,

240 zero_expert_num: int,

241 - rounter_params_dtype: torch.dtype,

241 + router_params_dtype: torch.dtype,

242 prefix: str = "",

243 ):

244 super().__init__()

@@ -252,12 +252,12 @@ class LongcatRouter(nn.Module):

252 config.hidden_size,

253 self.n_routed_experts,

254 bias=config.router_bias,

255 - params_dtype=rounter_params_dtype,

255 + params_dtype=router_params_dtype,

256 quant_config=None,

257 prefix=f"{prefix}.classifier",

258 )

259 self.e_score_correction_bias = nn.Parameter(

260 - torch.zeros((self.n_routed_experts), dtype=rounter_params_dtype)

260 + torch.zeros((self.n_routed_experts), dtype=router_params_dtype)

261 )

262

263 def forward(self, hidden_states):

@@ -281,14 +281,14 @@ class LongcatMoe(nn.Module):

281 super().__init__()

282 self.hidden_size = hidden_size

283 # Gate always runs at half / full precision for now.

284 - self.rounter_params_dtype = params_dtype

284 + self.router_params_dtype = params_dtype

285 if config.router_dtype == "float32":

286 - self.rounter_params_dtype = torch.float32

286 + self.router_params_dtype = torch.float32

287

288 self.router = LongcatRouter(

289 config=config,

290 zero_expert_num=config.zero_expert_num,

291 - rounter_params_dtype=self.rounter_params_dtype,

291 + router_params_dtype=self.router_params_dtype,

292 prefix=f"{prefix}.gate",

293 )

294

@@ -309,7 +309,7 @@ class LongcatMoe(nn.Module):

309 prefix=f"{prefix}.experts",

310 enable_eplb=enable_eplb,

311 routed_scaling_factor=config.routed_scaling_factor,

312 - router_logits_dtype=self.rounter_params_dtype,

312 + router_logits_dtype=self.router_params_dtype,

313 )

314

315 def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:

@@ -329,7 +329,7 @@ class LongcatMoe(nn.Module):

329 hidden_states_padded = hidden_states

330

331 router_logits_full = self.router(

332 - hidden_states_padded.to(self.rounter_params_dtype)

332 + hidden_states_padded.to(self.router_params_dtype)

333 )

334

335 # ZeroExpertFusedMoE handles routing memoization and zero expert computation

vllm/model_executor/models/molmo2.py

+4 -4

@@ -1321,14 +1321,14 @@ def get_image_size(image: ImageInput) -> ImageSize:

1321 raise ValueError(f"Unknown image type: {type(image)}")

1322

1323

1324 -def exif_tranpose(

1324 +def exif_transpose(

1325 images: ImageInput | None,

1326 ) -> ImageInput | None:

1327 if images is None:

1328 return None

1329 if images is not None and isinstance(images, (list, tuple)):

1330 images = [

1331 - exif_tranpose(img) if isinstance(img, Image) else img for img in images

1331 + exif_transpose(img) if isinstance(img, Image) else img for img in images

1332 ]

1333 elif images is not None and isinstance(images, Image):

1334 images = ImageOps.exif_transpose(images)

@@ -1667,7 +1667,7 @@ class Molmo2ProcessorWrapper:

1667 **kwargs: object,

1668 ) -> BatchFeature:

1669 inputs = [text]

1670 - images = exif_tranpose(images)

1670 + images = exif_transpose(images)

1671 if getattr(self.processor, "image_processor", None) is not None:

1672 inputs.append(images)

1673 if getattr(self.processor, "video_processor", None) is not None:

@@ -2352,7 +2352,7 @@ class Molmo2MultiModalProcessor(BaseMultiModalProcessor[Molmo2ProcessingInfo]):

2352 def get_image_replacement_molmo2(item_idx: int) -> list[int]:

2353 images = mm_items.get_items("image", ImageProcessorItems)

2354 image = images.get(item_idx)

2355 - image = exif_tranpose(image)

2355 + image = exif_transpose(image)

2356

2357 resize_nrows, resize_cols = processor.get_base_grid_size(is_video=False)

2358 if use_single_crop_col_tokens is not None:

vllm/model_executor/models/nemotron_h.py

+2 -2

@@ -349,7 +349,7 @@ class NemotronHMoEDecoderLayer(nn.Module):

349 super().__init__()

350 self.config = config

351

352 - # Get per-layer config for heterogeneous models if exsist

352 + # Get per-layer config for heterogeneous models if exists

353 get_layer_config = getattr(config, "get_nemotron_h_config_for_layer", None)

354 layer_config = get_layer_config(layer_idx) if get_layer_config else config

355

@@ -517,7 +517,7 @@ class NemotronHAttentionDecoderLayer(nn.Module):

517 ) -> None:

518 super().__init__()

519

520 - # Get per-layer config for heterogeneous models if exsist

520 + # Get per-layer config for heterogeneous models if exists

521 get_layer_config = getattr(config, "get_nemotron_h_config_for_layer", None)

522 layer_config = get_layer_config(layer_idx) if get_layer_config else config

523

vllm/model_executor/models/paddleocr_vl.py

+1 -1

@@ -486,7 +486,7 @@ class SiglipVisionEmbeddings(nn.Module):

486 )

487 (

488 batch_size,

489 - squence_len,

489 + sequence_len,

490 channel,

491 height,

492 width,

vllm/model_executor/models/phi4mm_audio.py

+6 -6

@@ -689,19 +689,19 @@ class ConformerEncoder(TransformerEncoderBase):

689 default False.

690 ext_pw_out_channel: int, optional

691 the number of channel for CNN

692 - before depthwise_seperable_CNN.

692 + before depthwise_separable_CNN.

693 If 0 then use linear. default 0.

694 ext_pw_kernel_size: int, optional

695 - kernel size of N before depthwise_seperable_CNN.

695 + kernel size of N before depthwise_separable_CNN.

696 only work for ext_pw_out_channel > 0.

697 default 1

698 depthwise_seperable_out_channel: int, optional

699 the number of channel for

700 - depthwise_seperable_CNN.

700 + depthwise_separable_CNN.

701 default 256.

702 depthwise_multiplier: int, optional

703 the number of multiplier for

704 - depthwise_seperable_CNN.

704 + depthwise_separable_CNN.

705 default 1.

706 chunk_se: int, optional

707 0 for offline SE.

@@ -711,7 +711,7 @@ class ConformerEncoder(TransformerEncoderBase):

711 by only the current chunk.

712 default 0.

713 kernel_size: int, optional

714 - the number of kernels for depthwise_seperable_CNN.

714 + the number of kernels for depthwise_separable_CNN.

715 default 3.

716 activation: str, optional

717 FeedForward block activation.

@@ -721,7 +721,7 @@ class ConformerEncoder(TransformerEncoderBase):

721 activation function used in ConvModule part

722 of the conformer, default "relu".

723 conv_glu_type: str, optional

724 - activation used use glu in depthwise_seperable_CNN,

724 + activation used use glu in depthwise_separable_CNN,

725 default "sigmoid"

726 bias_in_glu: bool, optional

727 if set to True, use additive bias in the weight module

vllm/model_executor/models/phi4mm_utils.py

+3 -3

@@ -217,8 +217,8 @@ class GLUPointWiseConv(nn.Module):

217 return x

218

219

220 -class DepthWiseSeperableConv1d(nn.Module):

221 - """DepthWiseSeperableConv1d module used in Convnet module

220 +class DepthWiseSeparableConv1d(nn.Module):

221 + """DepthWiseSeparableConv1d module used in ConvNet module

222 for the conformer, for more details see:

223 https://arxiv.org/pdf/2005.08100v1.pdf

224

@@ -390,7 +390,7 @@ class ConvModule(nn.Module):

390 else:

391 padding = (kernel_size - 1) // 2

392

393 - self.dw_sep_conv_1d = DepthWiseSeperableConv1d(

393 + self.dw_sep_conv_1d = DepthWiseSeparableConv1d(

394 input_dim,

395 depthwise_seperable_out_channel,

396 kernel_size,

vllm/model_executor/models/qwen2_vl.py

+1 -1

@@ -916,7 +916,7 @@ class Qwen2VLProcessingInfo(BaseProcessingInfo):

916 self, max_pixels: int | None = None

917 ) -> ImageSize:

918 # NOTE: Simply processing a huge size with _get_vision_info might not give a

919 - # size that maximizes the number of featrues, i.e., the number of (merged)

919 + # size that maximizes the number of features, i.e., the number of (merged)

920 # patches. This is because the number of patches limits the allowed aspect

921 # ratios. For example, suppose the maximum number of patches is 1280. A square

922 # image cannot be broken down into 1280 patches, so feeding a giant square image

vllm/model_executor/models/step3_vl.py

+2 -2

@@ -459,14 +459,14 @@ class Step3VLProcessor:

459 image_inputs = {}

460 text_inputs = self.tokenizer(text)

461 else:

462 - splitted_images_data = self._split_images(images)

462 + split_images_data = self._split_images(images)

463 pixel_values_lst = []

464 patch_pixel_values_lst = []

465 patch_newline_mask_lst = []

466 image_repl_str_lst = []

467 image_repl_ids_lst = []

468 num_patches = []

469 - for raw_img, img_patches, patch_newline_mask in splitted_images_data:

469 + for raw_img, img_patches, patch_newline_mask in split_images_data:

470 pixel_values_lst.extend(self._convert_images_to_pixel_values([raw_img]))

471

472 if len(img_patches) > 0:

vllm/model_executor/models/step3p5.py

+1 -1

@@ -353,7 +353,7 @@ class FusedMoEBlock(nn.Module):

353 if swiglu_limit not in (None, 0):

354 swiglu_limit = float(swiglu_limit)

355 assert swiglu_limit == 7.0, (

356 - "Swiglu limit in fused moe block only suport 7.0 now."

356 + "Swiglu limit in fused moe block only support 7.0 now."

357 )

358 activation = "swiglustep"

359 logger.debug(

vllm/reasoning/ernie45_reasoning_parser.py

+3 -3

@@ -18,7 +18,7 @@ logger = init_logger(__name__)

18 class Ernie45ReasoningParser(BaseThinkingReasoningParser):

19 """

20 Reasoning parser for Ernie45 thinking model.

21 - The Ernie45 thinking model ouput format is

21 + The Ernie45 thinking model output format is

22 abc\n</think>\n\n<response>\ndef\n</response>\n

23 or abc\n</think>\ndef

24 """

@@ -73,7 +73,7 @@ class Ernie45ReasoningParser(BaseThinkingReasoningParser):

73 Extract reasoning content from a delta message.

74 Handles streaming output where previous + delta = current.

75 Uses token IDs for faster processing.

76 - The Ernie45 thinking model ouput format is

76 + The Ernie45 thinking model output format is

77 abc\n</think>\n\n<response>\ndef\n</response>\n

78 or abc\n</think>\ndef

79 - 'abc' goes to reasoning

@@ -148,7 +148,7 @@ class Ernie45ReasoningParser(BaseThinkingReasoningParser):

148 ) -> tuple[str | None, str | None]:

149 """

150 Extract reasoning content from the model output.

151 - The Ernie45 thinking model ouput format is

151 + The Ernie45 thinking model output format is

152 abc\n</think>\n\n\n<response>\ndef\n</response>\n

153 or abc\n</think>\ndef

154 - 'abc' goes to reasoning

vllm/renderers/hf.py

+1 -1

@@ -564,7 +564,7 @@ def replace_vision_chunk_video_placeholder(

564 mm_data: "MultiModalDataDict",

565 video_placeholder: str | None,

566 ) -> str | list[int]:

567 - # get video placehoder, replace it with runtime video-chunk prompts

567 + # get video placeholder, replace it with runtime video-chunk prompts

568 if video_placeholder and isinstance(prompt_raw, str):

569 video_prompts = build_video_prompts_from_mm_data(mm_data)

570

vllm/renderers/inputs/preprocess.py

+1 -1

@@ -1,5 +1,5 @@

1 """

2 -Schemas and utilites for preprocessing inputs.

2 +Schemas and utilities for preprocessing inputs.

3 """

4

5 # SPDX-License-Identifier: Apache-2.0

vllm/renderers/inputs/tokenize.py

+1 -1

@@ -1,5 +1,5 @@

1 """

2 -Schemas and utilites for tokenization inputs.

2 +Schemas and utilities for tokenization inputs.

3 """

4

5 # SPDX-License-Identifier: Apache-2.0

vllm/tokenizers/mistral.py

+2 -2

@@ -169,7 +169,7 @@ def _prepare_apply_chat_template_tools_and_messages(

169 tool.pop(tool_key)

170 logger.warning_once(

171 f"'{tool_key}' is not supported by mistral-common for tools. "

172 - "It has been poped from the tool definition."

172 + "It has been popped from the tool definition."

173 )

174 if tool["type"] == "function":

175 function_keys = list(tool["function"].keys())

@@ -178,7 +178,7 @@ def _prepare_apply_chat_template_tools_and_messages(

178 tool["function"].pop(function_key)

179 logger.warning_once(

180 f"'{function_key}' is not supported by mistral-common "

181 - "for function tools. It has been poped from the "

181 + "for function tools. It has been popped from the "

182 "function definition."

183 )

184 else:

vllm/transformers_utils/processors/ovis2_5.py

+1 -1

@@ -402,7 +402,7 @@ class Ovis2_5Processor(ProcessorMixin):

402 images = [images]

403 elif video is not None:

404 is_video = True

405 - # type of vidoe in dummy_mm_data is np.ndarray

405 + # type of video in dummy_mm_data is np.ndarray

406 if isinstance(video, np.ndarray):

407 images = []

408 for i in range(video.shape[0]):

vllm/v1/attention/backends/cpu_attn.py

+2 -2

@@ -174,7 +174,7 @@ class CPUAttentionMetadataBuilder(AttentionMetadataBuilder[CPUAttentionMetadata]

174 query_start_loc = query_start_loc[: num_decodes + 1]

175 block_table_tensor = block_table_tensor[:num_decodes]

176

177 - sheduler_metadata = ops.cpu_attn_get_scheduler_metadata(

177 + scheduler_metadata = ops.cpu_attn_get_scheduler_metadata(

178 num_reqs=num_reqs,

179 num_heads=self.num_heads,

180 num_kv_heads=self.num_kv_heads,

@@ -197,7 +197,7 @@ class CPUAttentionMetadataBuilder(AttentionMetadataBuilder[CPUAttentionMetadata]

197 seq_lens=seq_lens,

198 block_table=block_table_tensor,

199 slot_mapping=slot_mapping,

200 - scheduler_metadata=sheduler_metadata,

200 + scheduler_metadata=scheduler_metadata,

201 causal=causal,

202 use_sdpa_prefill=self.use_sdpa_prefill,

203 num_decode_tokens=num_decode_tokens,

vllm/v1/attention/backends/mamba_attn.py

+1 -1

@@ -383,7 +383,7 @@ class BaseMambaAttentionMetadataBuilder(AttentionMetadataBuilder[M], abc.ABC):

383

384 # Return a tensor of shape (#requests, #max blocks)

385 state_indices_tensor = common_attn_metadata.block_table_tensor

386 - # Additional cache-related varaiables:

386 + # Additional cache-related variables:

387 mamba_block_size = self.kv_cache_spec.block_size

388 (

389 block_idx_last_computed_token,

vllm/v1/attention/backends/mla/flashmla_sparse.py

+3 -3

@@ -49,14 +49,14 @@ if TYPE_CHECKING:

49

50 logger = init_logger(__name__)

51

52 -# For FP8 sparse attention we have two impelementations:

52 +# For FP8 sparse attention we have two implementations:

53 # 1. Mixed batch mode: use the FP8 decode kernel for both prefill and decode this is

54 # done by treating all tokens as single batch.

55 # 2. Separate prefill and decode mode: use the BF16 prefill kernel for prefill

56 # (upconverting the FP8 cache to BF16 then calling the prefill kernel) and using

57 # the FP8 decode kernel for decode.

58 # Currently we use #1 when the number of heads per rank is low (i.e. TP) since the BF16

59 -# prefill kernel requires padding the numer of heads to 128 while the decode does not

59 +# prefill kernel requires padding the number of heads to 128 while the decode does not

60 # so when the per ranke head count is below MIN_HEADS_FOR_BF16_PREFILL we use the mixed

61 # batch mode (#2).

62 MIN_HEADS_FOR_BF16_PREFILL = 32

@@ -126,7 +126,7 @@ class FlashMLASparseBackend(AttentionBackend):

126 cache_dtype_str: str = "auto",

127 ) -> tuple[int, ...]:

128 if cache_dtype_str == "fp8_ds_mla":

129 - # custom storage fromat is 656 bytes

129 + # custom storage format is 656 bytes

130 # see FlashMLA readme.md for details

131 return (num_blocks, block_size, 656)

132 else:

vllm/v1/attention/backends/rocm_aiter_fa.py

+3 -3

@@ -370,7 +370,7 @@ class AiterFlashAttentionMetadata:

370 slot_mapping: torch.Tensor

371 block_table: torch.Tensor

372

373 - # prefill and deocde split

373 + # prefill and decode split

374 num_decodes: int

375 num_decode_tokens: int

376 num_prefills: int

@@ -1099,7 +1099,7 @@ class AiterFlashAttentionImpl(AttentionImpl):

1099 extend_tokens_slice = slice(

1100 num_decode_tokens, num_decode_tokens + num_extend_tokens

1101 )

1102 - extend_querys = query[extend_tokens_slice]

1102 + extend_queries = query[extend_tokens_slice]

1103 extend_keys = key[extend_tokens_slice]

1104 extend_values = value[extend_tokens_slice]

1105 extend_outputs = output[extend_tokens_slice]

@@ -1110,7 +1110,7 @@ class AiterFlashAttentionImpl(AttentionImpl):

1110 v_scale = attn_metadata.v_scale

1111 self.extend_forward(

1112 attn_metadata=attn_metadata,

1113 - query=extend_querys,

1113 + query=extend_queries,

1114 key=extend_keys,

1115 value=extend_values,

1116 key_cache=key_cache,

vllm/v1/core/single_type_kv_cache_manager.py

+1 -1

@@ -863,7 +863,7 @@ class MambaManager(SingleTypeKVCacheManager):

863 ):

864 # Mamba can't rely on blocks generated by other requests in the current step

865 # To put it in the next step, we return num_gpu_blocks + 1 so

866 - # that kv_cache_manager will think there is no enough blocks to allocte now

866 + # that kv_cache_manager will think there is no enough blocks to allocate now

867 # and don't schedule it in the current step.

868 return self.block_pool.num_gpu_blocks + 1

869 if self.mamba_cache_mode != "align":

vllm/v1/engine/core.py

+2 -2

@@ -1724,11 +1724,11 @@ class DPEngineCoreProc(EngineCoreProc):

1724 """

1725 Send notifications to EngineCoreClient, which can then forward

1726 the notifications to other engine core processes. It is used for:

1727 - 1) In scale up: new core engines to notify exisiting core engines

1727 + 1) In scale up: new core engines to notify existing core engines

1728 that they are ready;

1729 2) In scale down: removing core engines to notify EngineCoreClient

1730 so EngineCoreClient can release their ray placement groups;

1731 - 3) Both scale up/down: to notify EngineCoreClient that exisiting

1731 + 3) Both scale up/down: to notify EngineCoreClient that existing

1732 core engines have already switched to the new parallel setup.

1733 """

1734 if vllm_config is None:

vllm/v1/engine/input_processor.py

+1 -1

@@ -194,7 +194,7 @@ class InputProcessor:

194 @staticmethod

195 def assign_request_id(request: EngineCoreRequest):

196 """Replace the externally supplied request ID with an internal request ID

197 - that adds 8 random characters in order to ensure uniquness.

197 + that adds 8 random characters in order to ensure uniqueness.

198 """

199 if request.external_req_id is not None:

200 raise ValueError(

vllm/v1/kv_offload/worker/cpu_gpu.py

+1 -1

@@ -197,7 +197,7 @@ class SingleDirectionOffloadingHandler(OffloadingHandler):

197 transfer = self._transfers.popleft()

198 transfer_time = (

199 transfer.start_event.elapsed_time(transfer.end_event) * 1e-3

200 - ) # elapsed_time is in miliseconds

200 + ) # elapsed_time is in milliseconds

201 result = TransferResult(

202 job_id=transfer.job_id,

203 success=True,

vllm/v1/worker/gpu_model_runner.py

+11 -11

@@ -905,7 +905,7 @@ class GPUModelRunner(

905 Args:

906 scheduler_output: The scheduler output.

907 """

908 - # Attention free models have zero kv_cache_goups, however models

908 + # Attention free models have zero kv_cache_groups, however models

909 # like Mamba are also attention free but use the kv_cache for

910 # keeping its internal state. This is why we check the number

911 # of kv_cache groups instead of solely checking

@@ -1065,7 +1065,7 @@ class GPUModelRunner(

1065 # of the request. for example:

1066 # fist step: num_computed_tokens = 0, spec_tokens = [],

1067 # prev_num_draft_len = 0.

1068 - # second step: num_computed_tokens = 100(prompt lenth),

1068 + # second step: num_computed_tokens = 100(prompt length),

1069 # spec_tokens = [a,b], prev_num_draft_len = 0.

1070 # third step: num_computed_tokens = 100 + 2, spec_tokens = [c,d],

1071 # prev_num_draft_len = 2.

@@ -1412,30 +1412,30 @@ class GPUModelRunner(

1412 prev_draft_token_indices.extend(range(start, start + draft_len))

1413 indices_match &= prev_index == flattened_index

1414 max_flattened_index = max(max_flattened_index, flattened_index)

1415 - num_commmon_tokens = len(sample_flattened_indices)

1415 + num_common_tokens = len(sample_flattened_indices)

1416 total_without_spec = total_num_scheduled_tokens - total_num_spec_tokens

1417 - if num_commmon_tokens < total_without_spec:

1417 + if num_common_tokens < total_without_spec:

1418 # If not all requests are decodes from the last iteration,

1419 # We need to copy the input_ids_cpu to the GPU first.

1420 self.input_ids.copy_to_gpu(total_num_scheduled_tokens)

1421 if self.enable_prompt_embeds:

1422 self.inputs_embeds.copy_to_gpu(total_num_scheduled_tokens)

1423 self.is_token_ids.copy_to_gpu(total_num_scheduled_tokens)

1424 - if num_commmon_tokens == 0:

1424 + if num_common_tokens == 0:

1425 # No requests in common with the previous iteration

1426 # So input_ids.cpu will have all the input ids.

1427 return

1428 - if indices_match and max_flattened_index == (num_commmon_tokens - 1):

1428 + if indices_match and max_flattened_index == (num_common_tokens - 1):

1429 # Common-case optimization: the batch is unchanged

1430 # and no reordering happened.

1431 # The indices are both the same permutation of 0..N-1 so

1432 # we can copy directly using a single slice.

1433 - self.input_ids.gpu[:num_commmon_tokens].copy_(

1434 - self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],

1433 + self.input_ids.gpu[:num_common_tokens].copy_(

1434 + self.input_batch.prev_sampled_token_ids[:num_common_tokens, 0],

1435 non_blocking=True,

1436 )

1437 if self.enable_prompt_embeds:

1438 - self.is_token_ids.gpu[:num_commmon_tokens] = True

1438 + self.is_token_ids.gpu[:num_common_tokens] = True

1439 return

1440 # Upload the index tensors asynchronously so the scatter can be non-blocking.

1441 sampled_tokens_index_tensor = torch.tensor(

@@ -4383,7 +4383,7 @@ class GPUModelRunner(

4383 self.model.compile(fullgraph=True, backend=backend)

4384 return

4385 # for other compilation modes, cudagraph behavior is controlled by

4386 - # CudagraphWraper and CudagraphDispatcher of vllm.

4386 + # CudagraphWrapper and CudagraphDispatcher of vllm.

4387

4388 # wrap the model with full cudagraph wrapper if needed.

4389 cudagraph_mode = self.compilation_config.cudagraph_mode

@@ -4444,7 +4444,7 @@ class GPUModelRunner(

4444 :param weights_path: path to load weights from if weights_iterator is not

4445 provided. Use path of original model if neither is provided.

4446 :param is_checkpoint_format: set to False if weights have already been processed

4447 - into kernel format (repacking, renaming, ect.)

4447 + into kernel format (repacking, renaming, etc.)

4448 """

4449 # TODO(@kylesayrs): generalize to all runners and loaders

4450 # argument validation

[Bugfix][CI] fix typos (#34934)