A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 15:45 UTC
Revert "[Core] Rename `PromptInputs` to `PromptType`, and `inputs` to `prompt`" (#8750)
Simon Mo committed
Sep 23, 2024 at 22:45 UTC
3185fb0ccae73816018d0936c03171b7cf1ba2f8
18 files changed
+162
-157
benchmarks/benchmark_latency.py
+4
-4
@@ -11,7 +11,7 @@ from tqdm import tqdm
11
12
from vllm import LLM, SamplingParams
13
from vllm.engine.arg_utils import DEVICE_OPTIONS, EngineArgs
14
-from vllm.inputs import PromptType
14
+from vllm.inputs import PromptInputs
15
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
16
from vllm.utils import FlexibleArgumentParser
17
@@ -61,7 +61,7 @@ def main(args: argparse.Namespace):
61
dummy_prompt_token_ids = np.random.randint(10000,
62
size=(args.batch_size,
63
args.input_len))
64
- dummy_prompts: List[PromptType] = [{
64
+ dummy_inputs: List[PromptInputs] = [{
65
"prompt_token_ids": batch
66
} for batch in dummy_prompt_token_ids.tolist()]
67
@@ -74,13 +74,13 @@ def main(args: argparse.Namespace):
74
],
75
on_trace_ready=torch.profiler.tensorboard_trace_handler(
76
str(profile_dir))) as p:
77
- llm.generate(dummy_prompts,
77
+ llm.generate(dummy_inputs,
78
sampling_params=sampling_params,
79
use_tqdm=False)
80
print(p.key_averages())
81
else:
82
start_time = time.perf_counter()
83
- llm.generate(dummy_prompts,
83
+ llm.generate(dummy_inputs,
84
sampling_params=sampling_params,
85
use_tqdm=False)
86
end_time = time.perf_counter()
docs/source/dev/multimodal/multimodal_index.rst
+1
-1
@@ -8,7 +8,7 @@ Multi-Modality
8
vLLM provides experimental support for multi-modal models through the :mod:`vllm.multimodal` package.
9
10
Multi-modal inputs can be passed alongside text and token prompts to :ref:`supported models <supported_vlms>`
11
-via the ``multi_modal_data`` field in :class:`vllm.inputs.PromptType`.
11
+via the ``multi_modal_data`` field in :class:`vllm.inputs.PromptInputs`.
12
13
Currently, vLLM only has built-in support for image data. You can extend vLLM to process additional modalities
14
by following :ref:`this guide <adding_multimodal_plugin>`.
docs/source/dev/offline_inference/llm_inputs.rst
+1
-1
@@ -1,7 +1,7 @@
1
LLM Inputs
2
==========
3
4
-.. autodata:: vllm.inputs.PromptType
4
+.. autodata:: vllm.inputs.PromptInputs
5
6
.. autoclass:: vllm.inputs.TextPrompt
7
:show-inheritance:
docs/source/models/vlm.rst
+1
-1
@@ -27,7 +27,7 @@ The :class:`~vllm.LLM` class can be instantiated in much the same way as languag
27
We have removed all vision language related CLI args in the ``0.5.1`` release. **This is a breaking change**, so please update your code to follow
28
the above snippet. Specifically, ``image_feature_size`` can no longer be specified as we now calculate that internally for each model.
29
30
-To pass an image to the model, note the following in :class:`vllm.inputs.PromptType`:
30
+To pass an image to the model, note the following in :class:`vllm.inputs.PromptInputs`:
31
32
* ``prompt``: The prompt should follow the format that is documented on HuggingFace.
33
* ``multi_modal_data``: This is a dictionary that follows the schema defined in :class:`vllm.multimodal.MultiModalDataDict`.
tests/mq_llm_engine/test_error_handling.py
+6
-6
@@ -61,7 +61,7 @@ async def test_evil_forward(tmp_socket):
61
62
# Throws an error in first forward pass.
63
with pytest.raises(RAISED_ERROR):
64
- async for _ in client.generate(prompt="Hello my name is",
64
+ async for _ in client.generate(inputs="Hello my name is",
65
sampling_params=SamplingParams(),
66
request_id=uuid.uuid4()):
67
pass
@@ -69,7 +69,7 @@ async def test_evil_forward(tmp_socket):
69
70
# Engine is errored, should get ENGINE_DEAD_ERROR.
71
with pytest.raises(MQEngineDeadError):
72
- async for _ in client.generate(prompt="Hello my name is",
72
+ async for _ in client.generate(inputs="Hello my name is",
73
sampling_params=SamplingParams(),
74
request_id=uuid.uuid4()):
75
pass
@@ -118,7 +118,7 @@ async def test_failed_health_check(tmp_socket):
118
119
# Generate call should throw ENGINE_DEAD_ERROR
120
with pytest.raises(MQEngineDeadError):
121
- async for _ in client.generate(prompt="Hello my name is",
121
+ async for _ in client.generate(inputs="Hello my name is",
122
sampling_params=SamplingParams(),
123
request_id=uuid.uuid4()):
124
pass
@@ -165,7 +165,7 @@ async def test_failed_abort(tmp_socket):
165
# with reference to the original KeyError("foo")
166
with pytest.raises(MQEngineDeadError) as execinfo:
167
async for _ in client.generate(
168
- prompt="Hello my name is",
168
+ inputs="Hello my name is",
169
sampling_params=SamplingParams(max_tokens=2000),
170
request_id=uuid.uuid4()):
171
pass
@@ -190,7 +190,7 @@ async def test_bad_request(tmp_socket):
190
191
# Invalid request should fail, but not crash the server.
192
with pytest.raises(ValueError):
193
- async for _ in client.generate(prompt="Hello my name is",
193
+ async for _ in client.generate(inputs="Hello my name is",
194
sampling_params=SamplingParams(),
195
request_id="abcd-1",
196
lora_request=LoRARequest(
@@ -199,7 +199,7 @@ async def test_bad_request(tmp_socket):
199
pass
200
201
# This request should be okay.
202
- async for _ in client.generate(prompt="Hello my name is",
202
+ async for _ in client.generate(inputs="Hello my name is",
203
sampling_params=SamplingParams(),
204
request_id="abcd-2"):
205
pass
tests/mq_llm_engine/utils.py
+1
-1
@@ -20,7 +20,7 @@ async def generate(
20
count = 0
21
async for out in client.generate(
22
request_id=request_id,
23
- prompt="Hello my name is Robert and",
23
+ inputs="Hello my name is Robert and",
24
sampling_params=SamplingParams(max_tokens=num_tokens,
25
temperature=0)):
26
vllm/__init__.py
+2
-2
@@ -5,7 +5,7 @@ from vllm.engine.async_llm_engine import AsyncLLMEngine
5
from vllm.engine.llm_engine import LLMEngine
6
from vllm.entrypoints.llm import LLM
7
from vllm.executor.ray_utils import initialize_ray_cluster
8
-from vllm.inputs import PromptType, TextPrompt, TokensPrompt
8
+from vllm.inputs import PromptInputs, TextPrompt, TokensPrompt
9
from vllm.model_executor.models import ModelRegistry
10
from vllm.outputs import (CompletionOutput, EmbeddingOutput,
11
EmbeddingRequestOutput, RequestOutput)
@@ -19,7 +19,7 @@ __all__ = [
19
"__version_tuple__",
20
"LLM",
21
"ModelRegistry",
22
- "PromptType",
22
+ "PromptInputs",
23
"TextPrompt",
24
"TokensPrompt",
25
"SamplingParams",
vllm/engine/async_llm_engine.py
+13
-11
@@ -17,7 +17,7 @@ from vllm.engine.metrics_types import StatLoggerBase
17
from vllm.executor.executor_base import ExecutorAsyncBase
18
from vllm.executor.gpu_executor import GPUExecutorAsync
19
from vllm.executor.ray_utils import initialize_ray_cluster
20
-from vllm.inputs import PromptType
20
+from vllm.inputs import PromptInputs
21
from vllm.logger import init_logger
22
from vllm.lora.request import LoRARequest
23
from vllm.model_executor.layers.sampler import SamplerOutput
@@ -405,7 +405,7 @@ class _AsyncLLMEngine(LLMEngine):
405
async def add_request_async(
406
self,
407
request_id: str,
408
- prompt: PromptType,
408
+ inputs: PromptInputs,
409
params: Union[SamplingParams, PoolingParams],
410
arrival_time: Optional[float] = None,
411
lora_request: Optional[LoRARequest] = None,
@@ -420,7 +420,7 @@ class _AsyncLLMEngine(LLMEngine):
420
arrival_time = time.time()
421
422
preprocessed_inputs = await self.input_preprocessor.preprocess_async(
423
- prompt,
423
+ inputs,
424
request_id=request_id,
425
lora_request=lora_request,
426
prompt_adapter_request=prompt_adapter_request,
@@ -777,7 +777,7 @@ class AsyncLLMEngine:
777
async def add_request(
778
self,
779
request_id: str,
780
- prompt: PromptType,
780
+ inputs: PromptInputs,
781
params: Union[SamplingParams, PoolingParams],
782
arrival_time: Optional[float] = None,
783
lora_request: Optional[LoRARequest] = None,
@@ -797,7 +797,7 @@ class AsyncLLMEngine:
797
stream = self._request_tracker.add_request(
798
request_id,
799
verbose=self.log_requests,
800
- prompt=prompt,
800
+ inputs=inputs,
801
params=params,
802
arrival_time=arrival_time or time.time(),
803
lora_request=lora_request,
@@ -808,7 +808,7 @@ class AsyncLLMEngine:
808
809
async def generate(
810
self,
811
- prompt: PromptType,
811
+ inputs: PromptInputs,
812
sampling_params: SamplingParams,
813
request_id: str,
814
lora_request: Optional[LoRARequest] = None,
@@ -822,7 +822,8 @@ class AsyncLLMEngine:
822
from the LLMEngine to the caller.
823
824
Args:
825
- prompt: The prompt to the LLM. See :class:`~vllm.inputs.PromptType`
825
+ inputs: The inputs to the LLM. See
826
+ :class:`~vllm.inputs.PromptInputs`
827
for more details about the format of each input.
828
sampling_params: The sampling parameters of the request.
829
request_id: The unique id of the request.
@@ -880,7 +881,7 @@ class AsyncLLMEngine:
881
"""
882
async for output in await self.add_request(
883
request_id,
883
- prompt,
884
+ inputs,
885
sampling_params,
886
lora_request=lora_request,
887
trace_headers=trace_headers,
@@ -890,7 +891,7 @@ class AsyncLLMEngine:
891
892
async def encode(
893
self,
893
- prompt: PromptType,
894
+ inputs: PromptInputs,
895
pooling_params: PoolingParams,
896
request_id: str,
897
lora_request: Optional[LoRARequest] = None,
@@ -903,7 +904,8 @@ class AsyncLLMEngine:
904
from the LLMEngine to the caller.
905
906
Args:
906
- prompt: The prompt to the LLM. See :class:`~vllm.inputs.PromptType`
907
+ inputs: The inputs to the LLM. See
908
+ :class:`~vllm.inputs.PromptInputs`
909
for more details about the format of each input.
910
pooling_params: The pooling parameters of the request.
911
request_id: The unique id of the request.
@@ -957,7 +959,7 @@ class AsyncLLMEngine:
959
"""
960
async for output in await self.add_request(
961
request_id,
960
- prompt,
962
+ inputs,
963
pooling_params,
964
lora_request=lora_request,
965
trace_headers=trace_headers,
vllm/engine/llm_engine.py
+5
-4
@@ -29,7 +29,7 @@ from vllm.executor.executor_base import ExecutorBase
29
from vllm.executor.gpu_executor import GPUExecutor
30
from vllm.executor.ray_utils import initialize_ray_cluster
31
from vllm.inputs import (INPUT_REGISTRY, EncoderDecoderLLMInputs,
32
- InputRegistry, LLMInputs, PromptType)
32
+ InputRegistry, LLMInputs, PromptInputs)
33
from vllm.inputs.preprocess import InputPreprocessor
34
from vllm.logger import init_logger
35
from vllm.lora.request import LoRARequest
@@ -689,7 +689,7 @@ class LLMEngine:
689
def add_request(
690
self,
691
request_id: str,
692
- prompt: PromptType,
692
+ inputs: PromptInputs,
693
params: Union[SamplingParams, PoolingParams],
694
arrival_time: Optional[float] = None,
695
lora_request: Optional[LoRARequest] = None,
@@ -704,7 +704,8 @@ class LLMEngine:
704
705
Args:
706
request_id: The unique ID of the request.
707
- prompt: The prompt to the LLM. See :class:`~vllm.inputs.PromptType`
707
+ inputs: The inputs to the LLM. See
708
+ :class:`~vllm.inputs.PromptInputs`
709
for more details about the format of each input.
710
params: Parameters for sampling or pooling.
711
:class:`~vllm.SamplingParams` for text generation.
@@ -744,7 +745,7 @@ class LLMEngine:
745
arrival_time = time.time()
746
747
preprocessed_inputs = self.input_preprocessor.preprocess(
747
- prompt,
748
+ inputs,
749
request_id=request_id,
750
lora_request=lora_request,
751
prompt_adapter_request=prompt_adapter_request,
vllm/engine/multiprocessing/__init__.py
+2
-2
@@ -3,7 +3,7 @@ from enum import Enum
3
from typing import List, Mapping, Optional, Union
4
5
from vllm import PoolingParams
6
-from vllm.inputs import PromptType
6
+from vllm.inputs import PromptInputs
7
from vllm.lora.request import LoRARequest
8
from vllm.outputs import RequestOutput
9
from vllm.prompt_adapter.request import PromptAdapterRequest
@@ -23,7 +23,7 @@ class MQEngineDeadError(RuntimeError):
23
24
@dataclass
25
class RPCProcessRequest:
26
- prompt: PromptType
26
+ inputs: PromptInputs
27
params: Union[SamplingParams, PoolingParams]
28
request_id: str
29
lora_request: Optional[LoRARequest] = None
vllm/engine/multiprocessing/client.py
+11
-9
@@ -25,7 +25,7 @@ from vllm.engine.multiprocessing import (ENGINE_DEAD_ERROR, IPC_DATA_EXT,
25
RPCStartupResponse)
26
# yapf: enable
27
from vllm.envs import VLLM_RPC_TIMEOUT
28
-from vllm.inputs import PromptType
28
+from vllm.inputs import PromptInputs
29
from vllm.logger import init_logger
30
from vllm.lora.request import LoRARequest
31
from vllm.outputs import EmbeddingRequestOutput, RequestOutput
@@ -375,7 +375,7 @@ class MQLLMEngineClient:
375
376
def generate(
377
self,
378
- prompt: PromptType,
378
+ inputs: PromptInputs,
379
sampling_params: SamplingParams,
380
request_id: str,
381
lora_request: Optional[LoRARequest] = None,
@@ -389,7 +389,8 @@ class MQLLMEngineClient:
389
from the LLMEngine to the caller.
390
391
Args:
392
- prompt: The prompt to the LLM. See :class:`~vllm.inputs.PromptType`
392
+ inputs: The inputs to the LLM. See
393
+ :class:`~vllm.inputs.PromptInputs`
394
for more details about the format of each input.
395
sampling_params: The sampling parameters of the request.
396
request_id: The unique id of the request.
@@ -398,13 +399,13 @@ class MQLLMEngineClient:
399
prompt_adapter_request: Prompt Adapter request to use
400
for generation, if any.
401
"""
401
- return self._process_request(prompt, sampling_params, request_id,
402
+ return self._process_request(inputs, sampling_params, request_id,
403
lora_request, trace_headers,
404
prompt_adapter_request)
405
406
def encode(
407
self,
407
- prompt: PromptType,
408
+ inputs: PromptInputs,
409
pooling_params: PoolingParams,
410
request_id: str,
411
lora_request: Optional[LoRARequest] = None,
@@ -417,7 +418,8 @@ class MQLLMEngineClient:
418
from the LLMEngine to the caller.
419
420
Args:
420
- prompt: The prompt to the LLM. See :class:`~vllm.inputs.PromptType`
421
+ inputs: The inputs to the LLM. See
422
+ :class:`~vllm.inputs.PromptInputs`
423
for more details about the format of each input.
424
pooling_params: The pooling parameters of the request.
425
request_id: The unique id of the request.
@@ -428,12 +430,12 @@ class MQLLMEngineClient:
430
The output `EmbeddingRequestOutput` objects from the LLMEngine
431
for the request.
432
"""
431
- return self._process_request(prompt, pooling_params, request_id,
433
+ return self._process_request(inputs, pooling_params, request_id,
434
lora_request, trace_headers)
435
436
async def _process_request(
437
self,
436
- prompt: PromptType,
438
+ inputs: PromptInputs,
439
params: Union[SamplingParams, PoolingParams],
440
request_id: str,
441
lora_request: Optional[LoRARequest] = None,
@@ -466,7 +468,7 @@ class MQLLMEngineClient:
468
469
request_bytes = pickle.dumps(
470
RPCProcessRequest(
469
- prompt=prompt,
471
+ inputs=inputs,
472
params=params,
473
request_id=request_id,
474
lora_request=lora_request,
vllm/engine/multiprocessing/engine.py
+1
-1
@@ -252,7 +252,7 @@ class MQLLMEngine:
252
try:
253
self.engine.add_request(
254
request_id=request_id,
255
- prompt=request.prompt,
255
+ inputs=request.inputs,
256
params=request.params,
257
lora_request=request.lora_request,
258
trace_headers=request.trace_headers,
vllm/engine/protocol.py
+4
-4
@@ -3,7 +3,7 @@ from typing import (AsyncGenerator, List, Mapping, Optional, Protocol,
3
4
from vllm.config import DecodingConfig, ModelConfig
5
from vllm.core.scheduler import SchedulerOutputs
6
-from vllm.inputs.data import PromptType
6
+from vllm.inputs.data import PromptInputs
7
from vllm.lora.request import LoRARequest
8
from vllm.model_executor.layers.sampler import SamplerOutput
9
from vllm.outputs import EmbeddingRequestOutput, RequestOutput
@@ -35,19 +35,19 @@ class EngineClient(Protocol):
35
36
def generate(
37
self,
38
- prompt: PromptType,
38
+ inputs: PromptInputs,
39
sampling_params: SamplingParams,
40
request_id: str,
41
lora_request: Optional[LoRARequest] = None,
42
trace_headers: Optional[Mapping[str, str]] = None,
43
prompt_adapter_request: Optional[PromptAdapterRequest] = None
44
) -> AsyncGenerator[RequestOutput, None]:
45
- """Generate outputs for a request."""
45
+ """Generates outputs for a request"""
46
...
47
48
def encode(
49
self,
50
- prompt: PromptType,
50
+ inputs: PromptInputs,
51
pooling_params: PoolingParams,
52
request_id: str,
53
lora_request: Optional[LoRARequest] = None,
vllm/entrypoints/llm.py
+38
-42
@@ -12,7 +12,7 @@ from vllm.entrypoints.chat_utils import (ChatCompletionMessageParam,
12
apply_hf_chat_template,
13
apply_mistral_chat_template,
14
parse_chat_messages)
15
-from vllm.inputs import PromptType, TextPrompt, TokensPrompt
15
+from vllm.inputs import PromptInputs, TextPrompt, TokensPrompt
16
from vllm.inputs.parse import parse_and_batch_prompt
17
from vllm.logger import init_logger
18
from vllm.lora.request import LoRARequest
@@ -293,8 +293,8 @@ class LLM:
293
@overload
294
def generate(
295
self,
296
- prompts: Union[PromptType, Sequence[PromptType]],
297
- /,
296
+ inputs: Union[PromptInputs, Sequence[PromptInputs]],
297
+ /, # We may enable `inputs` keyword after removing the old API
298
*,
299
sampling_params: Optional[Union[SamplingParams,
300
Sequence[SamplingParams]]] = None,
@@ -311,7 +311,7 @@ class LLM:
311
)
312
def generate(
313
self,
314
- prompts: Union[Union[PromptType, Sequence[PromptType]],
314
+ prompts: Union[Union[PromptInputs, Sequence[PromptInputs]],
315
Optional[Union[str, List[str]]]] = None,
316
sampling_params: Optional[Union[SamplingParams,
317
Sequence[SamplingParams]]] = None,
@@ -329,9 +329,7 @@ class LLM:
329
into a single list and pass it to this method.
330
331
Args:
332
- prompts: The prompts to the LLM. You may pass a sequence of prompts
333
- for batch inference. See :class:`~vllm.inputs.PromptType`
334
- for more details about the format of each prompts.
332
+ inputs: A list of inputs to generate completions for.
333
sampling_params: The sampling parameters for text generation. If
334
None, we use the default sampling parameters.
335
When it is a single value, it is applied to every prompt.
@@ -357,13 +355,12 @@ class LLM:
355
"models (XForCausalLM, XForConditionalGeneration).")
356
357
if prompt_token_ids is not None:
360
- parsed_prompts = self._convert_v1_inputs(
358
+ inputs = self._convert_v1_inputs(
359
prompts=cast(Optional[Union[str, List[str]]], prompts),
360
prompt_token_ids=prompt_token_ids,
361
)
362
else:
365
- parsed_prompts = cast(Union[PromptType, Sequence[PromptType]],
366
- prompts)
363
+ inputs = cast(Union[PromptInputs, Sequence[PromptInputs]], prompts)
364
365
if isinstance(guided_options_request, dict):
366
if len(guided_options_request) > 1:
@@ -378,7 +375,7 @@ class LLM:
375
sampling_params = SamplingParams()
376
377
self._validate_and_add_requests(
381
- prompts=parsed_prompts,
378
+ inputs=inputs,
379
params=sampling_params,
380
lora_request=lora_request,
381
prompt_adapter_request=prompt_adapter_request,
@@ -533,9 +530,9 @@ class LLM:
530
conversation, mm_data = parse_chat_messages(messages, model_config,
531
tokenizer)
532
536
- prompt_data: Union[str, List[int]]
533
+ prompt: Union[str, List[int]]
534
if isinstance(tokenizer, MistralTokenizer):
538
- prompt_data = apply_mistral_chat_template(
535
+ prompt = apply_mistral_chat_template(
536
tokenizer,
537
messages=messages,
538
chat_template=chat_template,
@@ -543,7 +540,7 @@ class LLM:
540
tools=tools,
541
)
542
else:
546
- prompt_data = apply_hf_chat_template(
543
+ prompt = apply_hf_chat_template(
544
tokenizer,
545
conversation=conversation,
546
chat_template=chat_template,
@@ -551,17 +548,17 @@ class LLM:
548
tools=tools,
549
)
550
554
- prompt: PromptType
555
- if is_list_of(prompt_data, int):
556
- prompt = TokensPrompt(prompt_token_ids=prompt_data)
551
+ inputs: PromptInputs
552
+ if is_list_of(prompt, int):
553
+ inputs = TokensPrompt(prompt_token_ids=prompt)
554
else:
558
- prompt = TextPrompt(prompt=prompt_data)
555
+ inputs = TextPrompt(prompt=prompt)
556
557
if mm_data is not None:
561
- prompt["multi_modal_data"] = mm_data
558
+ inputs["multi_modal_data"] = mm_data
559
560
return self.generate(
564
- prompt,
561
+ inputs,
562
sampling_params=sampling_params,
563
use_tqdm=use_tqdm,
564
lora_request=lora_request,
@@ -631,8 +628,8 @@ class LLM:
628
@overload
629
def encode(
630
self,
634
- prompts: Union[PromptType, Sequence[PromptType]],
635
- /,
631
+ inputs: Union[PromptInputs, Sequence[PromptInputs]],
632
+ /, # We may enable `inputs` keyword after removing the old API
633
*,
634
pooling_params: Optional[Union[PoolingParams,
635
Sequence[PoolingParams]]] = None,
@@ -649,7 +646,7 @@ class LLM:
646
)
647
def encode(
648
self,
652
- prompts: Union[Union[PromptType, Sequence[PromptType]],
649
+ prompts: Union[Union[PromptInputs, Sequence[PromptInputs]],
650
Optional[Union[str, List[str]]]] = None,
651
pooling_params: Optional[Union[PoolingParams,
652
Sequence[PoolingParams]]] = None,
@@ -665,9 +662,9 @@ class LLM:
662
into a single list and pass it to this method.
663
664
Args:
668
- prompts: The prompts to the LLM. You may pass a sequence of prompts
669
- for batch inference. See :class:`~vllm.inputs.PromptType`
670
- for more details about the format of each prompts.
665
+ inputs: The inputs to the LLM. You may pass a sequence of inputs for
666
+ batch inference. See :class:`~vllm.inputs.PromptInputs`
667
+ for more details about the format of each input.
668
pooling_params: The pooling parameters for pooling. If None, we
669
use the default pooling parameters.
670
use_tqdm: Whether to use tqdm to display the progress bar.
@@ -690,20 +687,19 @@ class LLM:
687
)
688
689
if prompt_token_ids is not None:
693
- parsed_prompts = self._convert_v1_inputs(
690
+ inputs = self._convert_v1_inputs(
691
prompts=cast(Optional[Union[str, List[str]]], prompts),
692
prompt_token_ids=prompt_token_ids,
693
)
694
else:
698
- parsed_prompts = cast(Union[PromptType, Sequence[PromptType]],
699
- prompts)
695
+ inputs = cast(Union[PromptInputs, Sequence[PromptInputs]], prompts)
696
697
if pooling_params is None:
698
# Use default pooling params.
699
pooling_params = PoolingParams()
700
701
self._validate_and_add_requests(
706
- prompts=parsed_prompts,
702
+ inputs=inputs,
703
params=pooling_params,
704
lora_request=lora_request,
705
prompt_adapter_request=prompt_adapter_request,
@@ -747,9 +743,9 @@ class LLM:
743
raise ValueError("Either prompts or prompt_token_ids must be "
744
"provided.")
745
750
- parsed_prompts: List[PromptType] = []
746
+ inputs: List[PromptInputs] = []
747
for i in range(num_requests):
752
- item: PromptType
748
+ item: PromptInputs
749
750
if prompts is not None:
751
item = TextPrompt(prompt=prompts[i])
@@ -758,24 +754,24 @@ class LLM:
754
else:
755
raise AssertionError
756
761
- parsed_prompts.append(item)
757
+ inputs.append(item)
758
763
- return parsed_prompts
759
+ return inputs
760
761
def _validate_and_add_requests(
762
self,
767
- prompts: Union[PromptType, Sequence[PromptType]],
763
+ inputs: Union[PromptInputs, Sequence[PromptInputs]],
764
params: Union[SamplingParams, Sequence[SamplingParams], PoolingParams,
765
Sequence[PoolingParams]],
766
lora_request: Optional[Union[Sequence[LoRARequest], LoRARequest]],
767
prompt_adapter_request: Optional[PromptAdapterRequest],
768
guided_options: Optional[GuidedDecodingRequest] = None,
769
) -> None:
774
- if isinstance(prompts, (str, dict)):
770
+ if isinstance(inputs, (str, dict)):
771
# Convert a single prompt to a list.
776
- prompts = [prompts]
772
+ inputs = [inputs]
773
778
- num_requests = len(prompts)
774
+ num_requests = len(inputs)
775
if isinstance(params, list) and len(params) != num_requests:
776
raise ValueError("The lengths of prompts and params "
777
"must be the same.")
@@ -792,9 +788,9 @@ class LLM:
788
sp.output_kind = RequestOutputKind.FINAL_ONLY
789
790
# Add requests to the engine.
795
- for i, prompt in enumerate(prompts):
791
+ for i, request_inputs in enumerate(inputs):
792
self._add_request(
797
- prompt,
793
+ request_inputs,
794
params[i] if isinstance(params, Sequence) else params,
795
lora_request=lora_request[i] if isinstance(
796
lora_request, Sequence) else lora_request,
@@ -803,7 +799,7 @@ class LLM:
799
800
def _add_request(
801
self,
806
- prompt: PromptType,
802
+ inputs: PromptInputs,
803
params: Union[SamplingParams, PoolingParams],
804
lora_request: Optional[LoRARequest] = None,
805
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
@@ -811,7 +807,7 @@ class LLM:
807
request_id = str(next(self.request_counter))
808
self.llm_engine.add_request(
809
request_id,
814
- prompt,
810
+ inputs,
811
params,
812
lora_request=lora_request,
813
prompt_adapter_request=prompt_adapter_request,
vllm/inputs/__init__.py
+3
-3
@@ -1,5 +1,5 @@
1
from .data import (EncoderDecoderLLMInputs, ExplicitEncoderDecoderPrompt,
2
- LLMInputs, PromptType, SingletonPrompt, TextPrompt,
2
+ LLMInputs, PromptInputs, SingletonPromptInputs, TextPrompt,
3
TokensPrompt, build_explicit_enc_dec_prompt,
4
to_enc_dec_tuple_list, zip_enc_dec_prompts)
5
from .registry import InputContext, InputRegistry
@@ -16,8 +16,8 @@ See also:
16
__all__ = [
17
"TextPrompt",
18
"TokensPrompt",
19
- "PromptType",
20
- "SingletonPrompt",
19
+ "PromptInputs",
20
+ "SingletonPromptInputs",
21
"ExplicitEncoderDecoderPrompt",
22
"LLMInputs",
23
"EncoderDecoderLLMInputs",
vllm/inputs/data.py
+15
-11
@@ -33,7 +33,7 @@ class TokensPrompt(TypedDict):
33
"""
34
35
36
-SingletonPrompt = Union[str, TextPrompt, TokensPrompt]
36
+SingletonPromptInputs = Union[str, TextPrompt, TokensPrompt]
37
"""
38
Set of possible schemas for a single LLM input:
39
@@ -46,7 +46,7 @@ which may be utilized for encoder/decoder models when
46
the user desires to express both the encoder & decoder
47
prompts explicitly, i.e. :class:`ExplicitEncoderDecoderPrompt`
48
49
-A prompt of type :class:`SingletonPromptType` may be employed
49
+A prompt of type :class:`SingletonPromptInputs` may be employed
50
as (1) input to a decoder-only model, (2) input to
51
the encoder of an encoder/decoder model, in the scenario
52
where the decoder-prompt is not specified explicitly, or
@@ -55,12 +55,12 @@ more than one prompt, i.e. :class:`ExplicitEncoderDecoderPrompt`
55
"""
56
57
_T1_co = TypeVar("_T1_co",
58
- bound=SingletonPrompt,
59
- default=SingletonPrompt,
58
+ bound=SingletonPromptInputs,
59
+ default=SingletonPromptInputs,
60
covariant=True)
61
_T2_co = TypeVar("_T2_co",
62
- bound=SingletonPrompt,
63
- default=SingletonPrompt,
62
+ bound=SingletonPromptInputs,
63
+ default=SingletonPromptInputs,
64
covariant=True)
65
66
@@ -72,7 +72,7 @@ class ExplicitEncoderDecoderPrompt(TypedDict, Generic[_T1_co, _T2_co]):
72
73
The encoder and decoder prompts, respectively,
74
may formatted according to any of the
75
- :class:`SingletonPromptType` schemas, and are not
75
+ :class:`SingletonPromptInputs` schemas, and are not
76
required to have the same schema.
77
78
Only the encoder prompt may have multi-modal data.
@@ -81,7 +81,7 @@ class ExplicitEncoderDecoderPrompt(TypedDict, Generic[_T1_co, _T2_co]):
81
be used as an input to a decoder-only model,
82
and that the `encoder_prompt` and `decoder_prompt`
83
fields of this data structure themselves must be
84
- :class:`SingletonPromptType` instances.
84
+ :class:`SingletonPromptInputs` instances.
85
"""
86
87
encoder_prompt: _T1_co
@@ -89,7 +89,7 @@ class ExplicitEncoderDecoderPrompt(TypedDict, Generic[_T1_co, _T2_co]):
89
decoder_prompt: Optional[_T2_co]
90
91
92
-PromptType = Union[SingletonPrompt, ExplicitEncoderDecoderPrompt]
92
+PromptInputs = Union[SingletonPromptInputs, ExplicitEncoderDecoderPrompt]
93
"""
94
Set of possible schemas for an LLM input, including
95
both decoder-only and encoder/decoder input types:
@@ -140,8 +140,12 @@ class EncoderDecoderLLMInputs(LLMInputs):
140
"""
141
142
143
-_T1 = TypeVar("_T1", bound=SingletonPrompt, default=SingletonPrompt)
144
-_T2 = TypeVar("_T2", bound=SingletonPrompt, default=SingletonPrompt)
143
+_T1 = TypeVar("_T1",
144
+ bound=SingletonPromptInputs,
145
+ default=SingletonPromptInputs)
146
+_T2 = TypeVar("_T2",
147
+ bound=SingletonPromptInputs,
148
+ default=SingletonPromptInputs)
149
150
151
def build_explicit_enc_dec_prompt(
vllm/inputs/parse.py
+11
-11
@@ -5,7 +5,7 @@ from typing_extensions import TypeIs
5
from vllm.utils import is_list_of
6
7
from .data import (EncoderDecoderLLMInputs, ExplicitEncoderDecoderPrompt,
8
- LLMInputs, PromptType, SingletonPrompt, TextPrompt,
8
+ LLMInputs, PromptInputs, SingletonPromptInputs, TextPrompt,
9
TokensPrompt)
10
11
@@ -81,23 +81,23 @@ class ParsedTokensPrompt(TypedDict):
81
82
83
def parse_singleton_prompt(
84
- prompt: SingletonPrompt,
84
+ inputs: SingletonPromptInputs,
85
) -> Union[ParsedStrPrompt, ParsedTextPrompt, ParsedTokensPrompt]:
86
- if isinstance(prompt, str):
87
- return ParsedStrPrompt(type="str", content=prompt)
88
- elif isinstance(prompt, dict):
89
- if "prompt_token_ids" in prompt:
86
+ if isinstance(inputs, str):
87
+ return ParsedStrPrompt(type="str", content=inputs)
88
+ elif isinstance(inputs, dict):
89
+ if "prompt_token_ids" in inputs:
90
return ParsedTokensPrompt(type="tokens",
91
- content=prompt) # type: ignore
92
- elif "prompt" in prompt:
93
- return ParsedTextPrompt(type="text", content=prompt)
91
+ content=inputs) # type: ignore
92
+ elif "prompt" in inputs:
93
+ return ParsedTextPrompt(type="text", content=inputs)
94
95
raise TypeError("inputs must be a string, TextPrompt, or TokensPrompt")
96
97
98
def is_explicit_encoder_decoder_prompt(
99
- prompt: PromptType) -> TypeIs[ExplicitEncoderDecoderPrompt]:
100
- return isinstance(prompt, dict) and "encoder_prompt" in prompt
99
+ inputs: PromptInputs) -> TypeIs[ExplicitEncoderDecoderPrompt]:
100
+ return isinstance(inputs, dict) and "encoder_prompt" in inputs
101
102
103
def is_valid_encoder_decoder_llm_inputs(
vllm/inputs/preprocess.py
+43
-43
@@ -9,8 +9,8 @@ from vllm.lora.request import LoRARequest
9
from vllm.prompt_adapter.request import PromptAdapterRequest
10
from vllm.transformers_utils.tokenizer_group import BaseTokenizerGroup
11
12
-from .data import (EncoderDecoderLLMInputs, LLMInputs, PromptType,
13
- SingletonPrompt)
12
+from .data import (EncoderDecoderLLMInputs, LLMInputs, PromptInputs,
13
+ SingletonPromptInputs)
14
from .parse import is_explicit_encoder_decoder_prompt, parse_singleton_prompt
15
16
if TYPE_CHECKING:
@@ -206,7 +206,7 @@ class InputPreprocessor:
206
207
def _extract_prompt_components(
208
self,
209
- prompt: SingletonPrompt,
209
+ inputs: SingletonPromptInputs,
210
request_id: str,
211
lora_request: Optional[LoRARequest] = None,
212
) -> PromptComponents:
@@ -216,7 +216,7 @@ class InputPreprocessor:
216
Arguments:
217
218
* request_id
219
- * prompt: single encoder or decoder input prompt
219
+ * inputs: single encoder or decoder input prompt
220
* lora_request: this is only valid for decoder prompts
221
222
Returns:
@@ -226,24 +226,24 @@ class InputPreprocessor:
226
* multi_modal_data
227
'''
228
229
- parsed = parse_singleton_prompt(prompt)
229
+ parsed = parse_singleton_prompt(inputs)
230
231
if parsed["type"] == "str":
232
- prompt_text = parsed["content"]
232
+ prompt = parsed["content"]
233
prompt_token_ids = self._tokenize_prompt(
234
- prompt_text,
234
+ prompt,
235
request_id=request_id,
236
lora_request=lora_request,
237
)
238
multi_modal_data = None
239
elif parsed["type"] == "tokens":
240
- prompt_text = None
240
+ prompt = None
241
prompt_token_ids = parsed["content"]["prompt_token_ids"]
242
multi_modal_data = parsed["content"].get("multi_modal_data")
243
elif parsed["type"] == "text":
244
- prompt_text = parsed["content"]["prompt"]
244
+ prompt = parsed["content"]["prompt"]
245
prompt_token_ids = self._tokenize_prompt(
246
- prompt_text,
246
+ prompt,
247
request_id=request_id,
248
lora_request=lora_request,
249
)
@@ -251,33 +251,33 @@ class InputPreprocessor:
251
else:
252
assert_never(parsed)
253
254
- return prompt_text, prompt_token_ids, multi_modal_data
254
+ return prompt, prompt_token_ids, multi_modal_data
255
256
async def _extract_prompt_components_async(
257
self,
258
- prompt: SingletonPrompt,
258
+ inputs: SingletonPromptInputs,
259
request_id: str,
260
lora_request: Optional[LoRARequest] = None,
261
) -> PromptComponents:
262
"""Async version of :meth:`_extract_prompt_components`."""
263
- parsed = parse_singleton_prompt(prompt)
263
+ parsed = parse_singleton_prompt(inputs)
264
265
if parsed["type"] == "str":
266
- prompt_text = parsed["content"]
266
+ prompt = parsed["content"]
267
prompt_token_ids = await self._tokenize_prompt_async(
268
- prompt_text,
268
+ prompt,
269
request_id=request_id,
270
lora_request=lora_request,
271
)
272
multi_modal_data = None
273
elif parsed["type"] == "tokens":
274
- prompt_text = None
274
+ prompt = None
275
prompt_token_ids = parsed["content"]["prompt_token_ids"]
276
multi_modal_data = parsed["content"].get("multi_modal_data")
277
elif parsed["type"] == "text":
278
- prompt_text = parsed["content"]["prompt"]
278
+ prompt = parsed["content"]["prompt"]
279
prompt_token_ids = await self._tokenize_prompt_async(
280
- prompt_text,
280
+ prompt,
281
request_id=request_id,
282
lora_request=lora_request,
283
)
@@ -285,7 +285,7 @@ class InputPreprocessor:
285
else:
286
assert_never(parsed)
287
288
- return prompt_text, prompt_token_ids, multi_modal_data
288
+ return prompt, prompt_token_ids, multi_modal_data
289
290
def _build_enc_dec_llm_inputs(
291
self,
@@ -311,7 +311,7 @@ class InputPreprocessor:
311
312
def _process_encoder_decoder_prompt(
313
self,
314
- prompt: PromptType,
314
+ inputs: PromptInputs,
315
request_id: str,
316
) -> EncoderDecoderLLMInputs:
317
'''
@@ -339,7 +339,7 @@ class InputPreprocessor:
339
340
Arguments:
341
342
- * prompt: an input prompt
342
+ * inputs: an input prompt
343
* request_id
344
345
Returns:
@@ -350,13 +350,13 @@ class InputPreprocessor:
350
encoder_comps: PromptComponents
351
decoder_comps: DecoderPromptComponents
352
353
- if is_explicit_encoder_decoder_prompt(prompt):
353
+ if is_explicit_encoder_decoder_prompt(inputs):
354
encoder_comps = self._extract_prompt_components(
355
- prompt["encoder_prompt"],
355
+ inputs["encoder_prompt"],
356
request_id=request_id,
357
)
358
359
- if (decoder_input := prompt["decoder_prompt"]) is None:
359
+ if (decoder_input := inputs["decoder_prompt"]) is None:
360
decoder_comps = None, None, None
361
else:
362
decoder_comps = self._extract_prompt_components(
@@ -365,7 +365,7 @@ class InputPreprocessor:
365
)
366
else:
367
encoder_comps = self._extract_prompt_components(
368
- prompt,
368
+ inputs,
369
request_id=request_id,
370
)
371
@@ -375,20 +375,20 @@ class InputPreprocessor:
375
376
async def _process_encoder_decoder_prompt_async(
377
self,
378
- prompt: PromptType,
378
+ inputs: PromptInputs,
379
request_id: str,
380
) -> EncoderDecoderLLMInputs:
381
"""Async version of :meth:`_process_encoder_decoder_prompt`."""
382
encoder_comps: PromptComponents
383
decoder_comps: DecoderPromptComponents
384
385
- if is_explicit_encoder_decoder_prompt(prompt):
385
+ if is_explicit_encoder_decoder_prompt(inputs):
386
encoder_task = self._extract_prompt_components_async(
387
- prompt["encoder_prompt"],
387
+ inputs["encoder_prompt"],
388
request_id=request_id,
389
)
390
391
- if (decoder_input := prompt["decoder_prompt"]) is None:
391
+ if (decoder_input := inputs["decoder_prompt"]) is None:
392
encoder_comps = await encoder_task
393
decoder_comps = None, None, None
394
else:
@@ -401,7 +401,7 @@ class InputPreprocessor:
401
encoder_task, decoder_task)
402
else:
403
encoder_comps = await self._extract_prompt_components_async(
404
- prompt,
404
+ inputs,
405
request_id=request_id,
406
)
407
@@ -425,7 +425,7 @@ class InputPreprocessor:
425
426
def _process_decoder_only_prompt(
427
self,
428
- prompt: SingletonPrompt,
428
+ inputs: SingletonPromptInputs,
429
request_id: str,
430
lora_request: Optional[LoRARequest] = None,
431
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
@@ -436,7 +436,7 @@ class InputPreprocessor:
436
437
Arguments:
438
439
- * prompt: input prompt
439
+ * inputs: input prompt
440
* request_id
441
* lora_request
442
* prompt_adapter_request
@@ -447,7 +447,7 @@ class InputPreprocessor:
447
'''
448
449
prompt_comps = self._extract_prompt_components(
450
- prompt,
450
+ inputs,
451
request_id=request_id,
452
lora_request=lora_request,
453
)
@@ -459,14 +459,14 @@ class InputPreprocessor:
459
460
async def _process_decoder_only_prompt_async(
461
self,
462
- prompt: SingletonPrompt,
462
+ inputs: SingletonPromptInputs,
463
request_id: str,
464
lora_request: Optional[LoRARequest] = None,
465
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
466
) -> LLMInputs:
467
"""Async version of :meth:`_process_decoder_only_prompt`."""
468
prompt_comps = await self._extract_prompt_components_async(
469
- prompt,
469
+ inputs,
470
request_id=request_id,
471
lora_request=lora_request,
472
)
@@ -478,7 +478,7 @@ class InputPreprocessor:
478
479
def preprocess(
480
self,
481
- prompt: PromptType,
481
+ inputs: PromptInputs,
482
request_id: str,
483
lora_request: Optional[LoRARequest] = None,
484
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
@@ -488,17 +488,17 @@ class InputPreprocessor:
488
# Encoder-decoder model requires special mapping of
489
# input prompts to encoder & decoder
490
return self._process_encoder_decoder_prompt(
491
- prompt,
491
+ inputs,
492
request_id=request_id,
493
)
494
495
- if is_explicit_encoder_decoder_prompt(prompt):
495
+ if is_explicit_encoder_decoder_prompt(inputs):
496
raise ValueError("Cannot pass encoder-decoder prompt "
497
"to decoder-only models")
498
499
# Decoder-only operation
500
return self._process_decoder_only_prompt(
501
- prompt,
501
+ inputs,
502
request_id=request_id,
503
lora_request=lora_request,
504
prompt_adapter_request=prompt_adapter_request,
@@ -506,7 +506,7 @@ class InputPreprocessor:
506
507
async def preprocess_async(
508
self,
509
- prompt: PromptType,
509
+ inputs: PromptInputs,
510
request_id: str,
511
lora_request: Optional[LoRARequest] = None,
512
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
@@ -516,17 +516,17 @@ class InputPreprocessor:
516
# Encoder-decoder model requires special mapping of
517
# input prompts to encoder & decoder
518
return await self._process_encoder_decoder_prompt_async(
519
- prompt,
519
+ inputs,
520
request_id=request_id,
521
)
522
523
- if is_explicit_encoder_decoder_prompt(prompt):
523
+ if is_explicit_encoder_decoder_prompt(inputs):
524
raise ValueError("Cannot pass encoder-decoder prompt "
525
"to decoder-only models")
526
527
# Decoder-only operation
528
return await self._process_decoder_only_prompt_async(
529
- prompt,
529
+ inputs,
530
request_id=request_id,
531
lora_request=lora_request,
532
prompt_adapter_request=prompt_adapter_request,