5157338ed9 · tduncan/Nvidia-Optimization-Stack-VLLM

benchmarks/kernels/benchmark_lora.py

+1 -1

@@ -89,7 +89,7 @@ def make_prompt_lora_mapping(num_prompts: int, num_active_loras: int,

89 sort_by_lora_id: bool,

90 device: str) -> torch.Tensor:

91 """

92 - All prompts are mapped to a Lora ID in range [0, num_active_loras).

92 + All prompts are mapped to a LoRA ID in range [0, num_active_loras).

93 where 0 refers to first lora, 1 refers to second lora and so on.

94 """

95 assert num_active_loras > 0

docs/source/features/lora.md

+1 -1

@@ -170,7 +170,7 @@ Now, you can specify a base_model_name alongside the name and path using JSON fo

170

171 To provide the backward compatibility support, you can still use the old key-value format (name=path), but the `base_model_name` will remain unspecified in that case.

172

173 -## Lora model lineage in model card

173 +## LoRA model lineage in model card

174

175 The new format of `--lora-modules` is mainly to support the display of parent model information in the model card. Here's an explanation of how your current response supports this:

176

tests/core/test_scheduler.py

+1 -1

@@ -491,7 +491,7 @@ def test_prefill_schedule_max_lora():

491 lora_path="abc"))

492 scheduler.add_seq_group(seq_group)

493 # Add two more requests to verify lora is prioritized.

494 - # 0: Lora, 1: Lora, 2: regular, 3: regular

494 + # 0: LoRA, 1: LoRA, 2: regular, 3: regular

495 # In the first iteration, index 0, 2 is scheduled.

496 # If a request is not scheduled because it hits max lora, it is

497 # prioritized. Verify that.

tests/entrypoints/openai/test_cli_args.py

+1 -1

@@ -26,7 +26,7 @@ def serve_parser():

26 return make_arg_parser(parser)

27

28

29 -### Tests for Lora module parsing

29 +### Tests for LoRA module parsing

30 def test_valid_key_value_format(serve_parser):

31 # Test old format: name=path

32 args = serve_parser.parse_args([

tests/entrypoints/openai/test_serving_models.py

+10 -10

@@ -8,8 +8,8 @@ import pytest

8 from vllm.config import ModelConfig

9 from vllm.engine.protocol import EngineClient

10 from vllm.entrypoints.openai.protocol import (ErrorResponse,

11 - LoadLoraAdapterRequest,

12 - UnloadLoraAdapterRequest)

11 + LoadLoRAAdapterRequest,

12 + UnloadLoRAAdapterRequest)

13 from vllm.entrypoints.openai.serving_models import (BaseModelPath,

14 OpenAIServingModels)

15 from vllm.lora.request import LoRARequest

@@ -51,7 +51,7 @@ async def test_serving_model_name():

51 @pytest.mark.asyncio

52 async def test_load_lora_adapter_success():

53 serving_models = await _async_serving_models_init()

54 - request = LoadLoraAdapterRequest(lora_name="adapter",

54 + request = LoadLoRAAdapterRequest(lora_name="adapter",

55 lora_path="/path/to/adapter2")

56 response = await serving_models.load_lora_adapter(request)

57 assert response == LORA_LOADING_SUCCESS_MESSAGE.format(lora_name='adapter')

@@ -62,7 +62,7 @@ async def test_load_lora_adapter_success():

62 @pytest.mark.asyncio

63 async def test_load_lora_adapter_missing_fields():

64 serving_models = await _async_serving_models_init()

65 - request = LoadLoraAdapterRequest(lora_name="", lora_path="")

65 + request = LoadLoRAAdapterRequest(lora_name="", lora_path="")

66 response = await serving_models.load_lora_adapter(request)

67 assert isinstance(response, ErrorResponse)

68 assert response.type == "InvalidUserInput"

@@ -72,14 +72,14 @@ async def test_load_lora_adapter_missing_fields():

72 @pytest.mark.asyncio

73 async def test_load_lora_adapter_duplicate():

74 serving_models = await _async_serving_models_init()

75 - request = LoadLoraAdapterRequest(lora_name="adapter1",

75 + request = LoadLoRAAdapterRequest(lora_name="adapter1",

76 lora_path="/path/to/adapter1")

77 response = await serving_models.load_lora_adapter(request)

78 assert response == LORA_LOADING_SUCCESS_MESSAGE.format(

79 lora_name='adapter1')

80 assert len(serving_models.lora_requests) == 1

81

82 - request = LoadLoraAdapterRequest(lora_name="adapter1",

82 + request = LoadLoRAAdapterRequest(lora_name="adapter1",

83 lora_path="/path/to/adapter1")

84 response = await serving_models.load_lora_adapter(request)

85 assert isinstance(response, ErrorResponse)

@@ -91,12 +91,12 @@ async def test_load_lora_adapter_duplicate():

91 @pytest.mark.asyncio

92 async def test_unload_lora_adapter_success():

93 serving_models = await _async_serving_models_init()

94 - request = LoadLoraAdapterRequest(lora_name="adapter1",

94 + request = LoadLoRAAdapterRequest(lora_name="adapter1",

95 lora_path="/path/to/adapter1")

96 response = await serving_models.load_lora_adapter(request)

97 assert len(serving_models.lora_requests) == 1

98

99 - request = UnloadLoraAdapterRequest(lora_name="adapter1")

99 + request = UnloadLoRAAdapterRequest(lora_name="adapter1")

100 response = await serving_models.unload_lora_adapter(request)

101 assert response == LORA_UNLOADING_SUCCESS_MESSAGE.format(

102 lora_name='adapter1')

@@ -106,7 +106,7 @@ async def test_unload_lora_adapter_success():

106 @pytest.mark.asyncio

107 async def test_unload_lora_adapter_missing_fields():

108 serving_models = await _async_serving_models_init()

109 - request = UnloadLoraAdapterRequest(lora_name="", lora_int_id=None)

109 + request = UnloadLoRAAdapterRequest(lora_name="", lora_int_id=None)

110 response = await serving_models.unload_lora_adapter(request)

111 assert isinstance(response, ErrorResponse)

112 assert response.type == "InvalidUserInput"

@@ -116,7 +116,7 @@ async def test_unload_lora_adapter_missing_fields():

116 @pytest.mark.asyncio

117 async def test_unload_lora_adapter_not_found():

118 serving_models = await _async_serving_models_init()

119 - request = UnloadLoraAdapterRequest(lora_name="nonexistent_adapter")

119 + request = UnloadLoRAAdapterRequest(lora_name="nonexistent_adapter")

120 response = await serving_models.unload_lora_adapter(request)

121 assert isinstance(response, ErrorResponse)

122 assert response.type == "NotFoundError"

tests/lora/test_layers.py

+9 -9

@@ -14,16 +14,16 @@ from vllm.config import LoRAConfig

14 from vllm.lora.fully_sharded_layers import (

15 ColumnParallelLinearWithShardedLoRA,

16 MergedColumnParallelLinearWithShardedLoRA,

17 - MergedQKVParallelLinearWithShardedLora, QKVParallelLinearWithShardedLora,

17 + MergedQKVParallelLinearWithShardedLoRA, QKVParallelLinearWithShardedLoRA,

18 RowParallelLinearWithShardedLoRA)

19 # yapf conflicts with isort for this block

20 # yapf: disable

21 from vllm.lora.layers import (BaseLayerWithLoRA, ColumnParallelLinearWithLoRA,

22 - LinearScalingRotaryEmbeddingWithLora,

22 + LinearScalingRotaryEmbeddingWithLoRA,

23 LogitsProcessorWithLoRA, LoRAMapping,

24 MergedColumnParallelLinearWithLoRA,

25 - MergedQKVParallelLinearWithLora,

26 - QKVParallelLinearWithLora,

25 + MergedQKVParallelLinearWithLoRA,

26 + QKVParallelLinearWithLoRA,

27 ReplicatedLinearWithLoRA,

28 RowParallelLinearWithLoRA,

29 VocabParallelEmbeddingWithLoRA)

@@ -866,9 +866,9 @@ def test_column_parallel_packed(dist_init, num_loras, repeats, fully_shard,

866 bias=False,

867 params_dtype=torch.float16)

868 linear.weight.data = torch.rand_like(linear.weight.data)

869 - lora_linear = (MergedQKVParallelLinearWithLora(linear)

869 + lora_linear = (MergedQKVParallelLinearWithLoRA(linear)

870 if not fully_shard else

871 - MergedQKVParallelLinearWithShardedLora(linear))

871 + MergedQKVParallelLinearWithShardedLoRA(linear))

872 else:

873 linear = QKVParallelLinear(4096,

874 64,

@@ -876,9 +876,9 @@ def test_column_parallel_packed(dist_init, num_loras, repeats, fully_shard,

876 bias=False,

877 params_dtype=torch.float16)

878 linear.weight.data = torch.rand_like(linear.weight.data)

879 - lora_linear = QKVParallelLinearWithLora(

879 + lora_linear = QKVParallelLinearWithLoRA(

880 linear

881 - ) if not fully_shard else QKVParallelLinearWithShardedLora(linear)

881 + ) if not fully_shard else QKVParallelLinearWithShardedLoRA(linear)

882

883 @dataclass

884 class FakeConfig:

@@ -1024,7 +1024,7 @@ def test_rotary_embedding_long_context(dist_init, num_loras, device,

1024 base,

1025 is_neox_style,

1026 )

1027 - lora_rope = LinearScalingRotaryEmbeddingWithLora(rope)

1027 + lora_rope = LinearScalingRotaryEmbeddingWithLoRA(rope)

1028 lora_rope.set_mapping(punica_wrapper)

1029 lora_rope.create_lora_weights(max_loras, lora_config)

1030 linear_rope = get_rope(head_size, rotary_dim, max_position, base,

tests/lora/test_long_context.py

+2 -2

@@ -8,7 +8,7 @@ import pytest

8

9 import vllm

10 from vllm import SamplingParams

11 -from vllm.lora.layers import LinearScalingRotaryEmbeddingWithLora

11 +from vllm.lora.layers import LinearScalingRotaryEmbeddingWithLoRA

12 from vllm.lora.request import LoRARequest

13 from vllm.model_executor.layers.rotary_embedding import (

14 LinearScalingRotaryEmbedding)

@@ -151,7 +151,7 @@ def test_rotary_emb_replaced(dist_init):

151 if "rotary_emb" in module_name:

152 if "base_layer" not in module_name:

153 rotary_emb_count += 1

154 - assert isinstance(module, LinearScalingRotaryEmbeddingWithLora)

154 + assert isinstance(module, LinearScalingRotaryEmbeddingWithLoRA)

155 else:

156 assert isinstance(module, LinearScalingRotaryEmbedding)

157 # Llama 2 has 32 layers.

vllm/engine/llm_engine.py

+1 -1

@@ -1629,7 +1629,7 @@ class LLMEngine:

1629 max_tokens_requests: List[int] = []

1630 finished_reason_requests: List[str] = []

1631

1632 - # Lora requests

1632 + # LoRA requests

1633 running_lora_adapters = dict(

1634 collectionsCounter([

1635 running_request.lora_request.lora_name

vllm/entrypoints/openai/api_server.py

+5 -5

@@ -53,7 +53,7 @@ from vllm.entrypoints.openai.protocol import (ChatCompletionRequest,

53 EmbeddingResponse,

54 EmbeddingResponseData,

55 ErrorResponse,

56 - LoadLoraAdapterRequest,

56 + LoadLoRAAdapterRequest,

57 PoolingChatRequest,

58 PoolingCompletionRequest,

59 PoolingRequest, PoolingResponse,

@@ -63,7 +63,7 @@ from vllm.entrypoints.openai.protocol import (ChatCompletionRequest,

63 TokenizeResponse,

64 TranscriptionRequest,

65 TranscriptionResponse,

66 - UnloadLoraAdapterRequest)

66 + UnloadLoRAAdapterRequest)

67 from vllm.entrypoints.openai.reasoning_parsers import ReasoningParserManager

68 # yapf: enable

69 from vllm.entrypoints.openai.serving_chat import OpenAIServingChat

@@ -690,12 +690,12 @@ if envs.VLLM_TORCH_PROFILER_DIR:

690

691 if envs.VLLM_ALLOW_RUNTIME_LORA_UPDATING:

692 logger.warning(

693 - "Lora dynamic loading & unloading is enabled in the API server. "

693 + "LoRA dynamic loading & unloading is enabled in the API server. "

694 "This should ONLY be used for local development!")

695

696 @router.post("/v1/load_lora_adapter",

697 dependencies=[Depends(validate_json_request)])

698 - async def load_lora_adapter(request: LoadLoraAdapterRequest,

698 + async def load_lora_adapter(request: LoadLoRAAdapterRequest,

699 raw_request: Request):

700 handler = models(raw_request)

701 response = await handler.load_lora_adapter(request)

@@ -707,7 +707,7 @@ if envs.VLLM_ALLOW_RUNTIME_LORA_UPDATING:

707

708 @router.post("/v1/unload_lora_adapter",

709 dependencies=[Depends(validate_json_request)])

710 - async def unload_lora_adapter(request: UnloadLoraAdapterRequest,

710 + async def unload_lora_adapter(request: UnloadLoRAAdapterRequest,

711 raw_request: Request):

712 handler = models(raw_request)

713 response = await handler.unload_lora_adapter(request)

vllm/entrypoints/openai/protocol.py

+2 -2

@@ -1431,12 +1431,12 @@ class DetokenizeResponse(OpenAIBaseModel):

1431 prompt: str

1432

1433

1434 -class LoadLoraAdapterRequest(BaseModel):

1434 +class LoadLoRAAdapterRequest(BaseModel):

1435 lora_name: str

1436 lora_path: str

1437

1438

1439 -class UnloadLoraAdapterRequest(BaseModel):

1439 +class UnloadLoRAAdapterRequest(BaseModel):

1440 lora_name: str

1441 lora_int_id: Optional[int] = Field(default=None)

1442

vllm/entrypoints/openai/serving_models.py

+7 -7

@@ -9,10 +9,10 @@ from typing import List, Optional, Union

9 from vllm.config import ModelConfig

10 from vllm.engine.protocol import EngineClient

11 from vllm.entrypoints.openai.protocol import (ErrorResponse,

12 - LoadLoraAdapterRequest,

12 + LoadLoRAAdapterRequest,

13 ModelCard, ModelList,

14 ModelPermission,

15 - UnloadLoraAdapterRequest)

15 + UnloadLoRAAdapterRequest)

16 from vllm.logger import init_logger

17 from vllm.lora.request import LoRARequest

18 from vllm.prompt_adapter.request import PromptAdapterRequest

@@ -88,7 +88,7 @@ class OpenAIServingModels:

88 if self.static_lora_modules is None:

89 return

90 for lora in self.static_lora_modules:

91 - load_request = LoadLoraAdapterRequest(lora_path=lora.path,

91 + load_request = LoadLoRAAdapterRequest(lora_path=lora.path,

92 lora_name=lora.name)

93 load_result = await self.load_lora_adapter(

94 request=load_request, base_model_name=lora.base_model_name)

@@ -140,7 +140,7 @@ class OpenAIServingModels:

140

141 async def load_lora_adapter(

142 self,

143 - request: LoadLoraAdapterRequest,

143 + request: LoadLoRAAdapterRequest,

144 base_model_name: Optional[str] = None

145 ) -> Union[ErrorResponse, str]:

146 error_check_ret = await self._check_load_lora_adapter_request(request)

@@ -177,7 +177,7 @@ class OpenAIServingModels:

177

178 async def unload_lora_adapter(

179 self,

180 - request: UnloadLoraAdapterRequest) -> Union[ErrorResponse, str]:

180 + request: UnloadLoRAAdapterRequest) -> Union[ErrorResponse, str]:

181 error_check_ret = await self._check_unload_lora_adapter_request(request

182 )

183 if error_check_ret is not None:

@@ -192,7 +192,7 @@ class OpenAIServingModels:

192 return f"Success: LoRA adapter '{lora_name}' removed successfully."

193

194 async def _check_load_lora_adapter_request(

195 - self, request: LoadLoraAdapterRequest) -> Optional[ErrorResponse]:

195 + self, request: LoadLoRAAdapterRequest) -> Optional[ErrorResponse]:

196 # Check if both 'lora_name' and 'lora_path' are provided

197 if not request.lora_name or not request.lora_path:

198 return create_error_response(

@@ -214,7 +214,7 @@ class OpenAIServingModels:

214

215 async def _check_unload_lora_adapter_request(

216 self,

217 - request: UnloadLoraAdapterRequest) -> Optional[ErrorResponse]:

217 + request: UnloadLoRAAdapterRequest) -> Optional[ErrorResponse]:

218 # Check if either 'lora_name' or 'lora_int_id' is provided

219 if not request.lora_name and not request.lora_int_id:

220 return create_error_response(

vllm/lora/fully_sharded_layers.py

+6 -6

@@ -13,8 +13,8 @@ from vllm.distributed.communication_op import (

13 from vllm.distributed.parallel_state import get_tensor_model_parallel_rank

14 from vllm.lora.layers import (ColumnParallelLinearWithLoRA,

15 MergedColumnParallelLinearWithLoRA,

16 - MergedQKVParallelLinearWithLora,

17 - QKVParallelLinearWithLora,

16 + MergedQKVParallelLinearWithLoRA,

17 + QKVParallelLinearWithLoRA,

18 RowParallelLinearWithLoRA)

19

20 if TYPE_CHECKING:

@@ -167,9 +167,9 @@ class MergedColumnParallelLinearWithShardedLoRA(

167 )

168

169

170 -class QKVParallelLinearWithShardedLora(QKVParallelLinearWithLora):

170 +class QKVParallelLinearWithShardedLoRA(QKVParallelLinearWithLoRA):

171 """

172 - Differs from QKVParallelLinearWithLora by slicing the

172 + Differs from QKVParallelLinearWithLoRA by slicing the

173 LoRA A's also.

174

175 Based on S-LoRA, slicing happens along the rank dim.

@@ -202,9 +202,9 @@ class QKVParallelLinearWithShardedLora(QKVParallelLinearWithLora):

202 )

203

204

205 -class MergedQKVParallelLinearWithShardedLora(MergedQKVParallelLinearWithLora):

205 +class MergedQKVParallelLinearWithShardedLoRA(MergedQKVParallelLinearWithLoRA):

206 """

207 - Differs from MergedQKVParallelLinearWithLora by slicing the

207 + Differs from MergedQKVParallelLinearWithLoRA by slicing the

208 LoRA A's also.

209

210 Based on S-LoRA, slicing happens along the rank dim.

vllm/lora/layers.py

+4 -4

@@ -363,7 +363,7 @@ class BaseLinearLayerWithLoRA(BaseLayerWithLoRA):

363 embeddings_tensor: Optional[torch.Tensor],

364 lora_bias: Optional[torch.Tensor] = None,

365 ):

366 - # Except for QKVParallelLinearWithLora and

366 + # Except for QKVParallelLinearWithLoRA and

367 # MergedColumnParallelLinearWithLoRA, all other linear LoRA layers

368 # store weights in a tuple of size 1. These two layers will

369 # override this function.

@@ -686,7 +686,7 @@ class MergedColumnParallelLinearWithLoRA(ColumnParallelLinearWithLoRA):

686 and len(packed_modules_list) == 2)

687

688

689 -class QKVParallelLinearWithLora(ColumnParallelLinearWithLoRA):

689 +class QKVParallelLinearWithLoRA(ColumnParallelLinearWithLoRA):

690 """

691 ColumnParallelLinear layer that is specifically designed for

692 qkv_proj. Certain models, such as chatglm3 and baichuan-7b,

@@ -754,7 +754,7 @@ class QKVParallelLinearWithLora(ColumnParallelLinearWithLoRA):

754 packed_modules_list) == 1

755

756

757 -class MergedQKVParallelLinearWithLora(MergedColumnParallelLinearWithLoRA):

757 +class MergedQKVParallelLinearWithLoRA(MergedColumnParallelLinearWithLoRA):

758 """MergedColumnParallelLinear layer that is composed of 3 sublayers (slices)

759 packed together in qkv proj fashion

760 (q_proj + k_proj + v_proj -> qkv_proj).

@@ -1120,7 +1120,7 @@ class LogitsProcessorWithLoRA(BaseLayerWithLoRA):

1120 return False

1121

1122

1123 -class LinearScalingRotaryEmbeddingWithLora(BaseLayerWithLoRA):

1123 +class LinearScalingRotaryEmbeddingWithLoRA(BaseLayerWithLoRA):

1124 """Implements RoPE-scaled embeddings with linear scaling for

1125 multiple LoRA adapters with a specialized kernel.

1126

vllm/lora/models.py

+5 -5

@@ -20,7 +20,7 @@ from vllm.adapter_commons.utils import (add_adapter, deactivate_adapter,

20 from vllm.config import LoRAConfig

21 from vllm.logger import init_logger

22 from vllm.lora.layers import (BaseLayerWithLoRA,

23 - LinearScalingRotaryEmbeddingWithLora,

23 + LinearScalingRotaryEmbeddingWithLoRA,

24 LoRAMapping)

25 from vllm.lora.lora import LoRALayerWeights, PackedLoRALayerWeights

26 from vllm.lora.peft_helper import PEFTHelper

@@ -201,7 +201,7 @@ class LoRAModel(AdapterModel):

201 expected_lora_modules: Name of modules that are expected to be

202 replaced by lora.

203 peft_helper: Loaded lora configuration information.

204 - lora_model_id: Lora model id. If not given, automatically set by

204 + lora_model_id: LoRA model id. If not given, automatically set by

205 a global counter.

206 device: Device where the lora model is loaded.

207 dtype: dtype of the lora model weights.

@@ -480,9 +480,9 @@ class LoRAModelManager(AdapterModelManager):

480 from_layer(module, self.lora_slots, self.lora_config,

481 packed_moduled_lst, self.model.config))

482

483 - # LinearScalingRotaryEmbeddingWithLora is used to handle

483 + # LinearScalingRotaryEmbeddingWithLoRA is used to handle

484 # long context lora. Register relevant metadata.

485 - if isinstance(new_module, LinearScalingRotaryEmbeddingWithLora):

485 + if isinstance(new_module, LinearScalingRotaryEmbeddingWithLoRA):

486 self.long_lora_context = LongContextLoRAContext(

487 new_module.scaling_factors, new_module.rotary_dim)

488 self.scaling_factor_to_offset = \

@@ -527,7 +527,7 @@ class LoRAModelManager(AdapterModelManager):

527 bias_enabled = self.lora_config.bias_enabled

528 if (not self._match_target_modules(module_name)

529 or not isinstance(module, BaseLayerWithLoRA)

530 - or isinstance(module, LinearScalingRotaryEmbeddingWithLora)

530 + or isinstance(module, LinearScalingRotaryEmbeddingWithLoRA)

531 or self._filter_unsupported_mm_module(module_name)):

532 continue

533 parts = module_name.split(".")

vllm/lora/peft_helper.py

+1 -1

@@ -42,7 +42,7 @@ class PEFTHelper:

42

43 def _validate_features(self) -> List[str]:

44 """

45 - Check if there are any unsupported Lora features.

45 + Check if there are any unsupported LoRA features.

46 """

47 error_msg = []

48 if self.modules_to_save:

vllm/lora/punica_wrapper/punica_base.py

+1 -1

@@ -314,7 +314,7 @@ class PunicaWrapperBase(PunicaWrapperABC):

314 def long_lora_indices(self) -> torch.Tensor:

315 """

316 This property provides access to the indices used for long context

317 - lora, specifically for LinearScalingRotaryEmbeddingWithLora.

317 + lora, specifically for LinearScalingRotaryEmbeddingWithLoRA.

318 """

319 long_lora_len = self.indices_len[4]

320 return self._long_lora_indices[:long_lora_len]

vllm/lora/utils.py

+9 -9

@@ -15,17 +15,17 @@ from vllm.logger import init_logger

15 from vllm.lora.fully_sharded_layers import (

16 ColumnParallelLinearWithShardedLoRA,

17 MergedColumnParallelLinearWithShardedLoRA,

18 - MergedQKVParallelLinearWithShardedLora, QKVParallelLinearWithShardedLora,

18 + MergedQKVParallelLinearWithShardedLoRA, QKVParallelLinearWithShardedLoRA,

19 RowParallelLinearWithShardedLoRA)

20 # being imported for _all_lora_classes below

21 # yapf conflicts with isort for this block

22 # yapf: disable

23 from vllm.lora.layers import (BaseLayerWithLoRA, ColumnParallelLinearWithLoRA,

24 - LinearScalingRotaryEmbeddingWithLora,

24 + LinearScalingRotaryEmbeddingWithLoRA,

25 LogitsProcessorWithLoRA,

26 MergedColumnParallelLinearWithLoRA,

27 - MergedQKVParallelLinearWithLora,

28 - QKVParallelLinearWithLora,

27 + MergedQKVParallelLinearWithLoRA,

28 + QKVParallelLinearWithLoRA,

29 ReplicatedLinearWithLoRA,

30 RowParallelLinearWithLoRA,

31 VocabParallelEmbeddingWithLoRA)

@@ -41,17 +41,17 @@ _all_lora_classes: Set[Type[BaseLayerWithLoRA]] = {

41 VocabParallelEmbeddingWithLoRA,

42 ColumnParallelLinearWithLoRA,

43 MergedColumnParallelLinearWithLoRA,

44 - QKVParallelLinearWithLora,

45 - MergedQKVParallelLinearWithLora,

44 + QKVParallelLinearWithLoRA,

45 + MergedQKVParallelLinearWithLoRA,

46 RowParallelLinearWithLoRA,

47 ReplicatedLinearWithLoRA,

48 LogitsProcessorWithLoRA,

49 ColumnParallelLinearWithShardedLoRA,

50 - QKVParallelLinearWithShardedLora,

50 + QKVParallelLinearWithShardedLoRA,

51 MergedColumnParallelLinearWithShardedLoRA,

52 - MergedQKVParallelLinearWithShardedLora,

52 + MergedQKVParallelLinearWithShardedLoRA,

53 RowParallelLinearWithShardedLoRA,

54 - LinearScalingRotaryEmbeddingWithLora,

54 + LinearScalingRotaryEmbeddingWithLoRA,

55 }

56

57

vllm/spec_decode/proposer_worker_base.py

+2 -2

@@ -6,10 +6,10 @@ from typing import List, Optional, Set, Tuple

6 from vllm.model_executor.layers.sampler import SamplerOutput

7 from vllm.sequence import ExecuteModelRequest

8 from vllm.spec_decode.interfaces import SpeculativeProposer

9 -from vllm.worker.worker_base import LoraNotSupportedWorkerBase

9 +from vllm.worker.worker_base import LoRANotSupportedWorkerBase

10

11

12 -class ProposerWorkerBase(LoraNotSupportedWorkerBase, SpeculativeProposer):

12 +class ProposerWorkerBase(LoRANotSupportedWorkerBase, SpeculativeProposer):

13 """Interface for proposer workers"""

14

15 @abstractmethod

vllm/spec_decode/spec_decode_worker.py

+2 -2

@@ -47,7 +47,7 @@ from vllm.spec_decode.util import (Timer, create_logprobs_output,

47 get_sampled_token_logprobs, nvtx_range,

48 split_batch_by_proposal_len)

49 from vllm.utils import resolve_obj_by_qualname

50 -from vllm.worker.worker_base import LoraNotSupportedWorkerBase, WorkerBase

50 +from vllm.worker.worker_base import LoRANotSupportedWorkerBase, WorkerBase

51

52 logger = init_logger(__name__)

53

@@ -118,7 +118,7 @@ def create_spec_worker(*args, **kwargs) -> "SpecDecodeWorker":

118

119 # Reminder: Please update docs/source/features/compatibility_matrix.md

120 # If the feature combo become valid

121 -class SpecDecodeWorker(LoraNotSupportedWorkerBase):

121 +class SpecDecodeWorker(LoRANotSupportedWorkerBase):

122 """Worker which implements speculative decoding.

123

124 Speculative decoding reduces decoding per-token latency by using a proposal

vllm/transformers_utils/configs/arctic.py

+1 -1

@@ -21,7 +21,7 @@ ARCTIC_PRETRAINED_CONFIG_ARCHIVE_MAP = {

21

22

23 @dataclass

24 -class ArcticLoraConfig:

24 +class ArcticLoRAConfig:

25 lora_r: int = 64

26 lora_alpha: float = 16

27 shard_base_weights: bool = False

vllm/worker/neuron_worker.py

+2 -2

@@ -13,11 +13,11 @@ from vllm.model_executor.layers.sampler import SamplerOutput

13 from vllm.sequence import ExecuteModelRequest

14 from vllm.worker.neuron_model_runner import NeuronModelRunner

15 from vllm.worker.worker_base import (LocalOrDistributedWorkerBase,

16 - LoraNotSupportedWorkerBase, WorkerBase,

16 + LoRANotSupportedWorkerBase, WorkerBase,

17 WorkerInput)

18

19

20 -class NeuronWorker(LoraNotSupportedWorkerBase, LocalOrDistributedWorkerBase):

20 +class NeuronWorker(LoRANotSupportedWorkerBase, LocalOrDistributedWorkerBase):

21 """A worker class that executes the model on a group of neuron cores.

22 """

23

vllm/worker/openvino_worker.py

+2 -2

@@ -24,7 +24,7 @@ from vllm.sampling_params import SamplingParams

24 from vllm.sequence import ExecuteModelRequest, SequenceGroupMetadata

25 from vllm.utils import bind_kv_cache

26 from vllm.worker.openvino_model_runner import OpenVINOModelRunner

27 -from vllm.worker.worker_base import LoraNotSupportedWorkerBase, WorkerBase

27 +from vllm.worker.worker_base import LoRANotSupportedWorkerBase, WorkerBase

28

29 logger = init_logger(__name__)

30

@@ -203,7 +203,7 @@ class OpenVINOCacheEngine:

203 return dtype_size * total

204

205

206 -class OpenVINOWorker(LoraNotSupportedWorkerBase):

206 +class OpenVINOWorker(LoRANotSupportedWorkerBase):

207 """A worker class that executes the model on OpenVINO backend.

208

209 Each worker is associated with a single OpenVINO device. The worker is

vllm/worker/tpu_worker.py

+2 -2

@@ -17,13 +17,13 @@ from vllm.sequence import ExecuteModelRequest

17 from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, bind_kv_cache, get_dtype_size

18 from vllm.worker.tpu_model_runner import ExecutionMode, TPUModelRunner

19 from vllm.worker.worker_base import (LocalOrDistributedWorkerBase,

20 - LoraNotSupportedWorkerBase, WorkerBase,

20 + LoRANotSupportedWorkerBase, WorkerBase,

21 WorkerInput)

22

23 logger = init_logger(__name__)

24

25

26 -class TPUWorker(LoraNotSupportedWorkerBase, LocalOrDistributedWorkerBase):

26 +class TPUWorker(LoRANotSupportedWorkerBase, LocalOrDistributedWorkerBase):

27

28 def __init__(

29 self,

vllm/worker/worker_base.py

+1 -1

@@ -189,7 +189,7 @@ class DelegateWorkerBase(WorkerBase):

189 return getattr(self.worker, attr)

190

191

192 -class LoraNotSupportedWorkerBase(WorkerBase):

192 +class LoRANotSupportedWorkerBase(WorkerBase):

193 """Partial implementation of WorkerBase that raises exceptions when LoRA

194 methods are invoked.

195 """

vllm/worker/xpu_worker.py

+2 -2

@@ -18,13 +18,13 @@ from vllm.model_executor import set_random_seed

18 from vllm.platforms import current_platform

19 from vllm.worker.cache_engine import CacheEngine

20 from vllm.worker.worker import Worker

21 -from vllm.worker.worker_base import LoraNotSupportedWorkerBase, WorkerBase

21 +from vllm.worker.worker_base import LoRANotSupportedWorkerBase, WorkerBase

22 from vllm.worker.xpu_model_runner import XPUModelRunner

23

24 logger = init_logger(__name__)

25

26

27 -class XPUWorker(LoraNotSupportedWorkerBase, Worker):

27 +class XPUWorker(LoRANotSupportedWorkerBase, Worker):

28 """A worker class that executes (a partition of) the model on a GPU.

29

30 Each worker is associated with a single XPU device. The worker is

[Misc] Improve LoRA spelling (#13831)