51010a1807 · tduncan/Nvidia-Optimization-Stack-VLLM

vllm/attention/backends/flashinfer.py

+1 -1

@@ -438,7 +438,7 @@ class FlashInferMetadata(AttentionMetadata):

438 not in supported_head_sizes:

439 raise ValueError(

440 f"Only {supported_head_sizes} are supported for head_dim,",

441 - f"received {self.head_dim}.")

441 + f" received {self.head_dim}.")

442

443 def begin_forward(self):

444 if self.num_prefill_tokens > 0:

vllm/attention/backends/mla/common.py

+1 -1

@@ -533,7 +533,7 @@ class MLACommonMetadata(AttentionMetadata):

533 not in supported_head_sizes:

534 raise ValueError(

535 f"Only {supported_head_sizes} are supported for head_dim,",

536 - f"received {self.head_dim}.")

536 + f" received {self.head_dim}.")

537

538 @property

539 def prefill_metadata(self) -> Optional["MLACommonMetadata"]:

vllm/attention/backends/rocm_flash_attn.py

+2 -2

@@ -497,7 +497,7 @@ class ROCmFlashAttentionImpl(AttentionImpl):

497 if logits_soft_cap is not None:

498 raise ValueError(

499 "ROCm Triton FlashAttention does not support attention"

500 - "logits soft capping."

500 + " logits soft capping."

501 " please try using the ROCm CK "

502 "FA backend instead by setting the env var "

503 "`VLLM_USE_TRITON_FLASH_ATTN=0`")

@@ -528,7 +528,7 @@ class ROCmFlashAttentionImpl(AttentionImpl):

528 if self.use_naive_attn:

529 if logits_soft_cap is not None:

530 raise ValueError(

531 - "ROCm Naive FlashAttention does not support"

531 + "ROCm Naive FlashAttention does not support "

532 "attention logits soft capping.")

533

534 self.attn_func = _sdpa_attention

vllm/config.py

+6 -6

@@ -924,8 +924,8 @@ class ModelConfig:

924 layers_block_type_value = getattr(self.hf_config,

925 "layers_block_type", None)

926 if layers_block_type_value is None:

927 - raise ValueError("The model is an hybrid without a"

928 - "layers_block_type in the hf_config,"

927 + raise ValueError("The model is an hybrid without a "

928 + "layers_block_type in the hf_config, "

929 "cannot determine the num of "

930 f"{block_type.value} layers")

931

@@ -2516,7 +2516,7 @@ def _get_and_verify_dtype(

2516

2517 if current_platform.is_hpu() and config_dtype == torch.float16:

2518 logger.info(

2519 - "For HPU, we cast models to bfloat16 instead of"

2519 + "For HPU, we cast models to bfloat16 instead of "

2520 "using float16 by default. Please specify `dtype` if you "

2521 "want to use float16.")

2522 torch_dtype = torch.bfloat16

@@ -2732,7 +2732,7 @@ class DecodingConfig:

2732 backend=self.guided_decoding_backend).backend_name

2733 if backend not in valid_guided_backends:

2734 raise ValueError(f"Invalid guided_decoding_backend '{backend},"

2735 - f"must be one of {valid_guided_backends}")

2735 + f" must be one of {valid_guided_backends}")

2736

2737

2738 @dataclass

@@ -3008,7 +3008,7 @@ class CompilationConfig(BaseModel):

3008 def model_post_init(self, __context: Any) -> None:

3009 if not self.enable_reshape and self.enable_fusion:

3010 logger.warning_once(

3011 - "Fusion enabled but reshape elimination disabled."

3011 + "Fusion enabled but reshape elimination disabled. "

3012 "RMSNorm + quant (fp8) fusion might not work")

3013

3014 pass_config: PassConfig = Field(default_factory=PassConfig)

@@ -3563,7 +3563,7 @@ def set_current_vllm_config(vllm_config: VllmConfig, check_compile=False):

3563 logger.warning(

3564 "`torch.compile` is turned on, but the model %s"

3565 " does not support it. Please open an issue on GitHub"

3566 - "if you want it to be supported.",

3566 + " if you want it to be supported.",

3567 vllm_config.model_config.model)

3568 _current_vllm_config = old_vllm_config

3569

vllm/distributed/device_communicators/pynccl_wrapper.py

+2 -2

@@ -227,10 +227,10 @@ class NCCLLibrary:

227 self.lib = NCCLLibrary.path_to_library_cache[so_file]

228 except Exception as e:

229 logger.error(

230 - "Failed to load NCCL library from %s ."

230 + "Failed to load NCCL library from %s. "

231 "It is expected if you are not running on NVIDIA/AMD GPUs."

232 "Otherwise, the nccl library might not exist, be corrupted "

233 - "or it does not support the current platform %s."

233 + "or it does not support the current platform %s. "

234 "If you already have the library, please set the "

235 "environment variable VLLM_NCCL_SO_PATH"

236 " to point to the correct nccl library path.", so_file,

vllm/distributed/kv_transfer/kv_pipe/mooncake_pipe.py

+1 -1

@@ -137,7 +137,7 @@ class MooncakeTransferEngine:

137 if metadata_backend not in supported_backend:

138 raise ValueError(

139 "Mooncake Configuration error. `metadata_backend`"

140 - f"should be one of {supported_backend}.")

140 + f" should be one of {supported_backend}.")

141

142 self.engine.initializeExt(local_hostname, metadata_server,

143 protocol, device_name, metadata_backend)

vllm/entrypoints/chat_utils.py

+1 -1

@@ -823,7 +823,7 @@ def _parse_chat_message_content_part(

823 # content is empty, log a warning and skip

824 if part_type in VALID_MESSAGE_CONTENT_MM_PART_TYPES and not content:

825 logger.warning(

826 - "Skipping multimodal part (type: '%s')"

826 + "Skipping multimodal part (type: '%s') "

827 "with empty / unparsable content.", part_type)

828 return None

829

vllm/entrypoints/llm.py

+1 -1

@@ -1342,7 +1342,7 @@ class LLM:

1342 return params

1343

1344 if params.guided_decoding is not None:

1345 - raise ValueError("Cannot set both guided_options_request and"

1345 + raise ValueError("Cannot set both guided_options_request and "

1346 "params.guided_decoding.")

1347

1348 params.guided_decoding = GuidedDecodingParams(

vllm/entrypoints/openai/api_server.py

+1 -1

@@ -575,7 +575,7 @@ async def do_rerank(request: RerankRequest, raw_request: Request):

575 async def do_rerank_v1(request: RerankRequest, raw_request: Request):

576 logger.warning_once(

577 "To indicate that the rerank API is not part of the standard OpenAI"

578 - " API, we have located it at `/rerank`. Please update your client"

578 + " API, we have located it at `/rerank`. Please update your client "

579 "accordingly. (Note: Conforms to JinaAI rerank API)")

580

581 return await do_rerank(request, raw_request)

vllm/executor/ray_distributed_executor.py

+1 -1

@@ -513,7 +513,7 @@ class RayDistributedExecutor(DistributedExecutorBase):

513 if cupy_spec is None and envs.VLLM_USE_RAY_COMPILED_DAG_NCCL_CHANNEL:

514 raise ValueError(

515 "cupy is not installed but required since "

516 - "VLLM_USE_RAY_COMPILED_DAG_NCCL_CHANNEL is set."

516 + "VLLM_USE_RAY_COMPILED_DAG_NCCL_CHANNEL is set. "

517 "Run `pip install ray[adag]` and check cupy installation.")

518

519 def _compiled_ray_dag(self, enable_asyncio: bool):

vllm/executor/ray_utils.py

+1 -1

@@ -317,7 +317,7 @@ def initialize_ray_cluster(

317 if parallel_config.world_size > device_bundles:

318 raise ValueError(

319 f"The number of required {device_str}s exceeds the total "

320 - f"number of available {device_str}s in the placement group."

320 + f"number of available {device_str}s in the placement group. "

321 f"Required number of devices: {parallel_config.world_size}. "

322 f"Total number of devices: {device_bundles}.")

323 else:

vllm/lora/models.py

+1 -1

@@ -437,7 +437,7 @@ class LoRAModelManager(AdapterModelManager):

437 def pin_adapter(self, lora_id: int) -> bool:

438 """Pin a LoRAModel in the manager cache."""

439 raise NotImplementedError(

440 - "Pinning is not supported in LoRAModelManager."

440 + "Pinning is not supported in LoRAModelManager. "

441 "Use LRUCacheLoRAModelManager for pinning") # type: ignore

442

443 def _set_adapter_mapping(self, mapping: LoRAMapping) -> None:

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

+1 -1

@@ -71,7 +71,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):

71 if not (self.weight_quant.strategy == QuantizationStrategy.TENSOR

72 and self.input_quant.strategy == QuantizationStrategy.TENSOR):

73 raise ValueError(

74 - "For FP8 Fused MoE layers, only per-tensor scales"

74 + "For FP8 Fused MoE layers, only per-tensor scales "

75 "for weights and activations are supported. Found "

76 f"{self.weight_quant}, {self.input_quant}")

77

vllm/model_executor/layers/quantization/gptq.py

+1 -1

@@ -74,7 +74,7 @@ class GPTQConfig(QuantizationConfig):

74 def __repr__(self) -> str:

75 return (f"GPTQConfig(weight_bits={self.weight_bits}, "

76 f"group_size={self.group_size}, "

77 - f"desc_act={self.desc_act}),"

77 + f"desc_act={self.desc_act}), "

78 f"lm_head_quantized={self.lm_head_quantized}), "

79 f"dynamic={self.dynamic}")

80

vllm/model_executor/layers/quantization/modelopt.py

+1 -1

@@ -56,7 +56,7 @@ class ModelOptFp8Config(QuantizationConfig):

56 quant_method = quant_config["quant_algo"]

57 is_checkpoint_fp8_serialized = ("FP8" in quant_method)

58 if not is_checkpoint_fp8_serialized:

59 - raise ValueError("ModelOpt currently only supports static FP8"

59 + raise ValueError("ModelOpt currently only supports static FP8 "

60 "quantization in vLLM. Please check the "

61 "`hf_quant_config.json` file for your model's "

62 "quant configuration.")

vllm/model_executor/layers/quantization/neuron_quant.py

+2 -2

@@ -25,8 +25,8 @@ class NeuronQuantConfig(QuantizationConfig):

25 if self.quant_dtype not in SUPPORTED_QUANT_DTYPE_LIST:

26 raise ValueError(

27 f"Neuron quantization datatype {self.quant_dtype} is not valid,"

28 - f"the quantization datatype should match one of the below types"

29 - f"{SUPPORTED_QUANT_DTYPE_LIST}")

28 + f" the quantization datatype should match one of the below "

29 + f"types {SUPPORTED_QUANT_DTYPE_LIST}")

30 self.dequant_dtype = dequant_dtype

31 self.quantize_method = quantize_method

32

vllm/model_executor/layers/quantization/quark/quark_moe.py

+1 -1

@@ -55,7 +55,7 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):

55 if not (weight_qscheme == "per_tensor"

56 and input_qscheme == "per_tensor"):

57 raise ValueError(

58 - "For FP8 Fused MoE layers, only per-tensor scales"

58 + "For FP8 Fused MoE layers, only per-tensor scales "

59 "for weights and activations are supported. Found "

60 f"{weight_qscheme}, {input_qscheme}") # noqa E501

61

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

+1 -1

@@ -118,7 +118,7 @@ def verify_marlin_supports_shape(output_size_per_partition: int,

118 and input_size_per_partition % group_size != 0):

119 raise ValueError(

120 f"Weight input_size_per_partition = {input_size_per_partition}"

121 - f" is not divisible by group_size = {group_size}."

121 + f" is not divisible by group_size = {group_size}. "

122 "Consider reducing tensor_parallel_size or running "

123 "with --quantization gptq.")

124

vllm/model_executor/model_loader/loader.py

+1 -1

@@ -1088,7 +1088,7 @@ class BitsAndBytesModelLoader(BaseModelLoader):

1088 self.model_type = type(model).__name__

1089

1090 logger.info("Loading weights with BitsAndBytes quantization. "

1091 - " May take a while ...")

1091 + "May take a while ...")

1092

1093 quant_config = getattr(model_config.hf_config, "quantization_config",

1094 None)

vllm/model_executor/models/deepseek_vl2.py

+1 -1

@@ -562,7 +562,7 @@ class DeepseekVLV2ForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):

562 # 3D tensor

563 return list(torch.unbind(image_data, dim=0))

564 raise ValueError(

565 - "We expect batched 2D tensors;"

565 + "We expect batched 2D tensors; "

566 "this can be either a list of 2D tensors or a single 3D tensor."

567 )

568

vllm/model_executor/models/fuyu.py

+1 -1

@@ -290,7 +290,7 @@ class FuyuForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):

290 expected_expr = str(expected_dims)

291 raise ValueError(

292 "The expected shape of pixel values per image per batch "

293 - f" per patch is {expected_expr}. "

293 + f"per patch is {expected_expr}. "

294 f"You supplied {tuple(d.shape)}.")

295

296 for d in data:

vllm/model_executor/models/gritlm.py

+4 -4

@@ -90,8 +90,8 @@ class GritLMPooler(nn.Module):

90

91 # Return no instruction in case of missing BOS token.

92 if prompt_token_ids[0] != self.token_ids["<s>"]:

93 - logger.warning("BOS token not found in prompt,"

94 - "thus using empty string for instruction."

93 + logger.warning("BOS token not found in prompt, "

94 + "thus using empty string for instruction. "

95 "GritLM requires BOS token in prompt.")

96 return instruction_len

97

@@ -111,8 +111,8 @@ class GritLMPooler(nn.Module):

111 if found_embed_pattern_idx != -1:

112 instruction_len = found_embed_pattern_idx + len(embed_pattern_ids)

113 else:

114 - logger.warning("Query instruction not found in prompt,"

115 - "thus using BOS token as instruction instead."

114 + logger.warning("Query instruction not found in prompt, "

115 + "thus using BOS token as instruction instead. "

116 "GritLM requires query instruction in prompt.")

117 instruction_len = 1

118

vllm/model_executor/models/minicpmv.py

+1 -1

@@ -673,7 +673,7 @@ class MiniCPMVMultiModalProcessor(BaseMultiModalProcessor[_I]):

673 for modality, count in counts.items():

674 if modality not in inputs or not inputs[modality]:

675 raise ValueError(f"None input data of {modality}."

676 - "But prompt requires.")

676 + " But prompt requires.")

677 counter_key = self.get_modality_num_counter(modality)

678 if len(inputs[modality][counter_key]) != count:

679 raise ValueError(f"The prompt requires {count} "

vllm/model_executor/models/phi3v.py

+1 -1

@@ -639,7 +639,7 @@ class Phi3VForCausalLM(nn.Module, SupportsMultiModal, SupportsPP,

639 # 3D tensor

640 return list(torch.unbind(image_data, dim=0))

641 raise ValueError(

642 - "We expect batched 2D tensors;"

642 + "We expect batched 2D tensors; "

643 "this can be either a list of 2D tensors or a single 3D tensor."

644 )

645

vllm/model_executor/models/prithvi_geospatial_mae.py

+2 -2

@@ -153,8 +153,8 @@ class PrithviGeoSpatialMAE(nn.Module, IsAttentionFree, SupportsMultiModal):

153 vllm_config.model_config.hf_config.to_dict()["pretrained_cfg"])

154 if self.model is None:

155 raise ValueError(

156 - "Unsupported task."

157 - "Only SemanticSegmentationTask is supported for now"

156 + "Unsupported task. "

157 + "Only SemanticSegmentationTask is supported for now "

158 "by PrithviGeospatialMAE.")

159

160 def _parse_and_validate_multimodal_data(

vllm/multimodal/profiling.py

+1 -1

@@ -160,7 +160,7 @@ class MultiModalProfiler(Generic[_I]):

160

161 if mm_counts.keys() != mm_max_tokens_per_item.keys():

162 raise AssertionError(

163 - "The keys returned by `get_supported_mm_limits`"

163 + "The keys returned by `get_supported_mm_limits` "

164 f"({set(mm_counts.keys())}) should be the same as those "

165 "returned by `get_mm_max_tokens_per_item` "

166 f"({set(mm_max_tokens_per_item.keys())})")

vllm/platforms/cuda.py

+1 -1

@@ -190,7 +190,7 @@ class CudaPlatformBase(Platform):

190 "Cannot use FlashAttention-2 backend for FP8 KV cache.")

191 logger.warning(

192 "Please use FlashInfer backend with FP8 KV Cache for "

193 - "better performance by setting environment variable "

193 + "better performance by setting environment variable "

194 "VLLM_ATTENTION_BACKEND=FLASHINFER")

195 target_backend = _Backend.XFORMERS

196 elif block_size % 16 != 0:

vllm/platforms/openvino.py

+1 -1

@@ -97,7 +97,7 @@ class OpenVinoPlatform(Platform):

97

98 if envs.VLLM_OPENVINO_CPU_KV_CACHE_PRECISION == "u8":

99 if not OpenVinoPlatform.is_openvino_cpu():

100 - logger.info("VLLM_OPENVINO_CPU_KV_CACHE_PRECISION is"

100 + logger.info("VLLM_OPENVINO_CPU_KV_CACHE_PRECISION is "

101 "ignored for GPU, f16 data type will be used.")

102 cache_config.cache_dtype = ov.Type.f16

103 else:

vllm/platforms/xpu.py

+1 -1

@@ -73,7 +73,7 @@ class XPUPlatform(Platform):

73 logger.warning(

74 "bfloat16 is only supported on Intel Data Center GPU, "

75 "Intel Arc GPU is not supported yet. Your device is %s,"

76 - "which is not supported. will fallback to float16",

76 + " which is not supported. will fallback to float16",

77 cls.get_device_name())

78 model_config.dtype = torch.float16

79 if not model_config.enforce_eager:

vllm/prompt_adapter/models.py

+1 -1

@@ -226,7 +226,7 @@ class PromptAdapterModelManager(AdapterModelManager):

226 def pin_adapter(self, prompt_adapter_id: int) -> bool:

227 """Pin a PromptAdapterModel in the manager cache."""

228 raise NotImplementedError(

229 - "Pinning is not supported in PromptAdapterModelManager."

229 + "Pinning is not supported in PromptAdapterModelManager. "

230 "Use LRUCachePromptAdapterModelManager for pinning"

231 ) # type: ignore

232

vllm/spec_decode/draft_model_runner.py

+1 -1

@@ -16,7 +16,7 @@ try:

16 ROCmFlashAttentionMetadata as FlashAttentionMetadata)

17 except (ModuleNotFoundError, ImportError) as err:

18 raise RuntimeError(

19 - "Draft model speculative decoding currently only supports"

19 + "Draft model speculative decoding currently only supports "

20 "CUDA and ROCm flash attention backend.") from err

21

22 from vllm.logger import init_logger

vllm/transformers_utils/configs/jais.py

+4 -4

@@ -212,26 +212,26 @@ class JAISConfig(PretrainedConfig):

212 if (not isinstance(self.alibi_scaling, dict)

213 or len(self.alibi_scaling) != 2):

214 raise ValueError(

215 - "`alibi_scaling` must be a dictionary with two fields,"

215 + "`alibi_scaling` must be a dictionary with two fields, "

216 "`type` and `factor` or `type` and `train_seq_len`, "

217 f"got {self.alibi_scaling}")

218 alibi_scaling_type = self.alibi_scaling.get("type", None)

219 alibi_scaling_factor = self.alibi_scaling.get("factor", None)

220 alibi_dynamic_scaling = self.alibi_scaling.get("train_seq_len", None)

221 if alibi_scaling_type is None or alibi_scaling_type != "linear":

222 - raise ValueError(f"`alibi_scaling`'s type field must be 'linear',"

222 + raise ValueError(f"`alibi_scaling`'s type field must be 'linear', "

223 f"got {alibi_scaling_type}")

224 if (alibi_scaling_factor is not None

225 and not isinstance(alibi_scaling_factor, float)

226 or (alibi_scaling_factor is not None

227 and alibi_scaling_factor <= 1.0)):

228 raise ValueError(

229 - f"`alibi_scaling`'s factor field must be a float > 1.0,"

229 + f"`alibi_scaling`'s factor field must be a float > 1.0, "

230 f"got {alibi_scaling_factor}")

231 if (alibi_dynamic_scaling is not None

232 and not isinstance(alibi_dynamic_scaling, int)

233 or (alibi_dynamic_scaling is not None

234 and alibi_dynamic_scaling <= 1)):

235 raise ValueError(

236 - f"`alibi_scaling`'s `train_seq_len` field must be an"

236 + f"`alibi_scaling`'s `train_seq_len` field must be an "

237 f"integer > 1, got {alibi_dynamic_scaling}")

vllm/utils.py

+3 -3

@@ -447,7 +447,7 @@ def get_ip() -> str:

447 logger.warning(

448 "The environment variable HOST_IP is deprecated and ignored, as"

449 " it is often used by Docker and other software to"

450 - "interact with the container's network stack. Please "

450 + " interact with the container's network stack. Please "

451 "use VLLM_HOST_IP instead to set the IP address for vLLM processes"

452 " to communicate with each other.")

453 if host_ip:

@@ -2091,8 +2091,8 @@ def set_ulimit(target_soft_limit=65535):

2091 (target_soft_limit, current_hard))

2092 except ValueError as e:

2093 logger.warning(

2094 - "Found ulimit of %s and failed to automatically increase"

2095 - "with error %s. This can cause fd limit errors like"

2094 + "Found ulimit of %s and failed to automatically increase "

2095 + "with error %s. This can cause fd limit errors like "

2096 "`OSError: [Errno 24] Too many open files`. Consider "

2097 "increasing with ulimit -n", current_soft, e)

2098

vllm/v1/worker/gpu_worker.py

+1 -1

@@ -277,5 +277,5 @@ def _check_if_gpu_supports_dtype(torch_dtype: torch.dtype):

277 raise ValueError(

278 "Bfloat16 is only supported on GPUs with compute capability "

279 f"of at least 8.0. Your {gpu_name} GPU {compute_str}. "

280 - "You can use float16 instead by explicitly setting the"

280 + "You can use float16 instead by explicitly setting the "

281 "`dtype` flag in CLI, for example: --dtype=half.")

vllm/worker/openvino_worker.py

+1 -1

@@ -545,7 +545,7 @@ class OpenVINOWorker(LoraNotSupportedWorkerBase):

545 "value. This may cause low performance due to "

546 "occupying the majority of available system "

547 "memory. Please consider decreasing "

548 - "gpu_memory_utilization or explicitly setting"

548 + "gpu_memory_utilization or explicitly setting "

549 "`VLLM_OPENVINO_KVCACHE_SPACE` (GB) environment "

550 "variable.", memory_utilization)

551

vllm/worker/worker.py

+2 -2

@@ -525,7 +525,7 @@ def _check_if_gpu_supports_dtype(torch_dtype: torch.dtype):

525 raise ValueError(

526 "Bfloat16 is only supported on GPUs with compute capability "

527 f"of at least 8.0. Your {gpu_name} GPU {compute_str}. "

528 - "You can use float16 instead by explicitly setting the"

528 + "You can use float16 instead by explicitly setting the "

529 "`dtype` flag in CLI, for example: --dtype=half.")

530

531

@@ -533,7 +533,7 @@ def raise_if_cache_size_invalid(num_gpu_blocks, block_size, is_attention_free,

533 max_model_len) -> None:

534 if is_attention_free and num_gpu_blocks != 0:

535 raise ValueError("No memory should be allocated for the cache blocks "

536 - f"for an attention-free model, but {num_gpu_blocks}"

536 + f"for an attention-free model, but {num_gpu_blocks} "

537 "blocks are allocated.")

538 if not is_attention_free and num_gpu_blocks <= 0:

539 raise ValueError("No available memory for the cache blocks. "

[Misc] set single whitespace between log sentences (#13771)