endpoint_request_func.py 25.4 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
"""The request function for API endpoints."""

5
import io
6
7
8
9
10
import json
import os
import sys
import time
import traceback
11
from collections.abc import Awaitable
12
from dataclasses import dataclass, field
13
from typing import Any, Literal, Protocol
14
15

import aiohttp
16
import regex as re
17
18
19
20
21
from tqdm.asyncio import tqdm

AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60)


22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
class StreamedResponseHandler:
    """Handles streaming HTTP responses by accumulating chunks until complete
    messages are available."""

    def __init__(self):
        self.buffer = ""

    def add_chunk(self, chunk_bytes: bytes) -> list[str]:
        """Add a chunk of bytes to the buffer and return any complete
        messages."""
        chunk_str = chunk_bytes.decode("utf-8")
        self.buffer += chunk_str

        messages = []

        # Split by double newlines (SSE message separator)
        while "\n\n" in self.buffer:
            message, self.buffer = self.buffer.split("\n\n", 1)
            message = message.strip()
            if message:
                messages.append(message)

        # if self.buffer is not empty, check if it is a complete message
        # by removing data: prefix and check if it is a valid JSON
        if self.buffer.startswith("data: "):
            message_content = self.buffer.removeprefix("data: ").strip()
            if message_content == "[DONE]":
                messages.append(self.buffer.strip())
                self.buffer = ""
            elif message_content:
                try:
                    json.loads(message_content)
                    messages.append(self.buffer.strip())
                    self.buffer = ""
                except json.JSONDecodeError:
                    # Incomplete JSON, wait for more chunks.
                    pass

        return messages


63
64
65
@dataclass
class RequestFuncInput:
    """The input for the request function."""
66

67
    prompt: str | list[str]
68
69
70
71
    api_url: str
    prompt_len: int
    output_len: int
    model: str
72
73
74
75
76
    model_name: str | None = None
    logprobs: int | None = None
    extra_headers: dict | None = None
    extra_body: dict | None = None
    multi_modal_content: dict | list[dict] | None = None
77
    ignore_eos: bool = False
78
79
    language: str | None = None
    request_id: str | None = None
80
81
82
83
84


@dataclass
class RequestFuncOutput:
    """The output of the request function including metrics."""
85

86
87
88
89
90
    generated_text: str = ""
    success: bool = False
    latency: float = 0.0
    output_tokens: int = 0
    ttft: float = 0.0  # Time to first token
91
    itl: list[float] = field(default_factory=list)  # list of inter-token latencies
92
93
94
    tpot: float = 0.0  # avg next-token latencies
    prompt_len: int = 0
    error: str = ""
95
    start_time: float = 0.0
96
97


98
99
100
101
102
class RequestFunc(Protocol):
    def __call__(
        self,
        request_func_input: RequestFuncInput,
        session: aiohttp.ClientSession,
103
        pbar: tqdm | None = None,
104
    ) -> Awaitable[RequestFuncOutput]: ...
105
106


107
108
109
def _validate_api_url(
    api_url: str,
    api_name: str,
110
    expected_suffixes: str | set[str],
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
) -> None:
    if isinstance(expected_suffixes, str):
        expected_suffixes = {expected_suffixes}

    expected_suffixes = {*expected_suffixes, "profile"}

    if not api_url.endswith(tuple(expected_suffixes)):
        raise ValueError(f"{api_name} URL must end with one of: {expected_suffixes}.")


def _update_payload_common(
    payload: dict[str, Any],
    request_func_input: RequestFuncInput,
) -> None:
    if request_func_input.ignore_eos:
        payload["ignore_eos"] = request_func_input.ignore_eos
    if request_func_input.extra_body:
        payload.update(request_func_input.extra_body)


def _update_headers_common(
    headers: dict[str, Any],
    request_func_input: RequestFuncInput,
) -> None:
    if request_func_input.extra_headers:
        headers |= request_func_input.extra_headers
    if request_func_input.request_id:
        headers["x-request-id"] = request_func_input.request_id


141
142
async def async_request_openai_completions(
    request_func_input: RequestFuncInput,
143
    session: aiohttp.ClientSession,
144
    pbar: tqdm | None = None,
145
146
147
148
149
150
151
152
153
154
155
) -> RequestFuncOutput:
    """The async request function for the OpenAI Completions API.

    Args:
        request_func_input: The input for the request function.
        pbar: The progress bar to display the progress.

    Returns:
        The output of the request function.
    """
    api_url = request_func_input.api_url
156
    _validate_api_url(api_url, "OpenAI Completions API", "completions")
157

158
    payload = {
159
        "model": request_func_input.model_name
160
161
        if request_func_input.model_name
        else request_func_input.model,
162
163
164
165
166
167
168
169
170
171
        "prompt": request_func_input.prompt,
        "temperature": 0.0,
        "repetition_penalty": 1.0,
        "max_tokens": request_func_input.output_len,
        "logprobs": request_func_input.logprobs,
        "stream": True,
        "stream_options": {
            "include_usage": True,
        },
    }
172
173
174
175
176
177
    _update_payload_common(payload, request_func_input)

    headers = {
        "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}",
    }
    _update_headers_common(headers, request_func_input)
178
179
180
181
182
183

    output = RequestFuncOutput()
    output.prompt_len = request_func_input.prompt_len

    generated_text = ""
    st = time.perf_counter()
184
    output.start_time = st
185
186
    most_recent_timestamp = st
    try:
187
        async with session.post(url=api_url, json=payload, headers=headers) as response:
188
189
            if response.status == 200:
                first_chunk_received = False
190
191
192
                handler = StreamedResponseHandler()

                async for chunk_bytes in response.content.iter_any():
193
194
195
196
                    chunk_bytes = chunk_bytes.strip()
                    if not chunk_bytes:
                        continue

197
198
199
200
201
202
203
                    messages = handler.add_chunk(chunk_bytes)
                    for message in messages:
                        # NOTE: SSE comments (often used as pings) start with
                        # a colon. These are not JSON data payload and should
                        # be skipped.
                        if message.startswith(":"):
                            continue
204

205
                        chunk = message.removeprefix("data: ")
206

207
208
                        if chunk != "[DONE]":
                            data = json.loads(chunk)
209

210
211
212
213
214
215
216
217
218
219
220
221
222
                            # NOTE: Some completion API might have a last
                            # usage summary response without a token so we
                            # want to check a token was generated
                            if choices := data.get("choices"):
                                # Note that text could be empty here
                                # e.g. for special tokens
                                text = choices[0].get("text")
                                timestamp = time.perf_counter()
                                # First token
                                if not first_chunk_received:
                                    first_chunk_received = True
                                    ttft = time.perf_counter() - st
                                    output.ttft = ttft
223

224
225
                                # Decoding phase
                                else:
226
                                    output.itl.append(timestamp - most_recent_timestamp)
227
228
229
230

                                most_recent_timestamp = timestamp
                                generated_text += text or ""
                            elif usage := data.get("usage"):
231
                                output.output_tokens = usage.get("completion_tokens")
232
233
                if first_chunk_received:
                    output.success = True
234
235
                else:
                    output.success = False
236
237
                    output.error = (
                        "Never received a valid chunk to calculate TTFT."
238
239
                        "This response will be marked as failed!"
                    )
240
241
242
243
244
245
246
247
248
                output.generated_text = generated_text
                output.latency = most_recent_timestamp - st
            else:
                output.error = response.reason or ""
                output.success = False
    except Exception:
        output.success = False
        exc_info = sys.exc_info()
        output.error = "".join(traceback.format_exception(*exc_info))
249
250
251
252
253
254

    if pbar:
        pbar.update(1)
    return output


255
def _get_chat_content(
256
    request_func_input: RequestFuncInput,
257
258
259
    mm_position: Literal["first", "last"] = "last",
) -> list[dict[str, Any]]:
    text_contents = [{"type": "text", "text": request_func_input.prompt}]
260

261
    mm_contents = []
262
    if request_func_input.multi_modal_content:
263
264
        mm_content = request_func_input.multi_modal_content
        if isinstance(mm_content, list):
265
            mm_contents.extend(request_func_input.multi_modal_content)
266
        elif isinstance(mm_content, dict):
267
            mm_contents.append(request_func_input.multi_modal_content)
268
269
        else:
            raise TypeError(
270
                "multi_modal_content must be a dict or list[dict] for openai-chat"
271
            )
272
273
274
275
276
277
278
279
280
281

    if mm_position == "first":
        return mm_contents + text_contents

    return text_contents + mm_contents


async def async_request_openai_chat_completions(
    request_func_input: RequestFuncInput,
    session: aiohttp.ClientSession,
282
    pbar: tqdm | None = None,
283
284
285
286
287
288
289
    mm_position: Literal["first", "last"] = "last",
) -> RequestFuncOutput:
    api_url = request_func_input.api_url
    _validate_api_url(api_url, "OpenAI Chat Completions API", "chat/completions")

    content = _get_chat_content(request_func_input, mm_position=mm_position)

290
    payload = {
291
292
293
        "model": request_func_input.model_name
        if request_func_input.model_name
        else request_func_input.model,
294
        "messages": [
295
            {"role": "user", "content": content},
296
        ],
297
298
299
        "temperature": 0.0,
        "max_completion_tokens": request_func_input.output_len,
        "stream": True,
300
301
302
303
        "stream_options": {
            "include_usage": True,
        },
    }
304
305
    _update_payload_common(payload, request_func_input)

306
307
308
309
    headers = {
        "Content-Type": "application/json",
        "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}",
    }
310
    _update_headers_common(headers, request_func_input)
311
312
313
314
315
316
317

    output = RequestFuncOutput()
    output.prompt_len = request_func_input.prompt_len

    generated_text = ""
    ttft = 0.0
    st = time.perf_counter()
318
    output.start_time = st
319
320
    most_recent_timestamp = st
    try:
321
        async with session.post(url=api_url, json=payload, headers=headers) as response:
322
            if response.status == 200:
323
324
                handler = StreamedResponseHandler()
                async for chunk_bytes in response.content.iter_any():
325
326
327
328
                    chunk_bytes = chunk_bytes.strip()
                    if not chunk_bytes:
                        continue

329
330
331
332
333
334
335
336
337
                    messages = handler.add_chunk(chunk_bytes)
                    for message in messages:
                        # NOTE: SSE comments (often used as pings) start with
                        # a colon. These are not JSON data payload and should
                        # be skipped.
                        if message.startswith(":"):
                            continue

                        chunk = message.removeprefix("data: ")
338

339
340
341
                        if chunk != "[DONE]":
                            timestamp = time.perf_counter()
                            data = json.loads(chunk)
342

343
344
345
346
347
348
                            if choices := data.get("choices"):
                                content = choices[0]["delta"].get("content")
                                # First token
                                if ttft == 0.0:
                                    ttft = timestamp - st
                                    output.ttft = ttft
349

350
351
                                # Decoding phase
                                else:
352
                                    output.itl.append(timestamp - most_recent_timestamp)
353

354
355
                                generated_text += content or ""
                            elif usage := data.get("usage"):
356
                                output.output_tokens = usage.get("completion_tokens")
357

358
                            most_recent_timestamp = timestamp
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378

                output.generated_text = generated_text
                output.success = True
                output.latency = most_recent_timestamp - st
            else:
                output.error = response.reason or ""
                output.success = False
    except Exception:
        output.success = False
        exc_info = sys.exc_info()
        output.error = "".join(traceback.format_exception(*exc_info))

    if pbar:
        pbar.update(1)
    return output


async def async_request_openai_audio(
    request_func_input: RequestFuncInput,
    session: aiohttp.ClientSession,
379
    pbar: tqdm | None = None,
380
381
382
383
384
) -> RequestFuncOutput:
    # Lazy import without PlaceholderModule to avoid vllm dep.
    import soundfile

    api_url = request_func_input.api_url
385
    _validate_api_url(api_url, "OpenAI Audio API", {"transcriptions", "translations"})
386
387
388

    content = [{"type": "text", "text": request_func_input.prompt}]
    payload = {
389
390
391
392
393
394
395
        "model": request_func_input.model_name
        if request_func_input.model_name
        else request_func_input.model,
        "temperature": 0.0,
        "max_completion_tokens": request_func_input.output_len,
        "stream": True,
        "language": "en",
396
        # Flattened due to multipart/form-data
397
398
        "stream_include_usage": True,
        "stream_continuous_usage_stats": True,
399
    }
400
401
    _update_payload_common(payload, request_func_input)

402
403
404
    headers = {
        "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}",
    }
405
    _update_headers_common(headers, request_func_input)
406
407
408
409
410
411
412
413

    # Send audio file
    def to_bytes(y, sr):
        buffer = io.BytesIO()
        soundfile.write(buffer, y, sr, format="WAV")
        buffer.seek(0)
        return buffer

414
415
416
417
    mm_audio = request_func_input.multi_modal_content
    if not isinstance(mm_audio, dict) or "audio" not in mm_audio:
        raise TypeError("multi_modal_content must be a dict containing 'audio'")
    with to_bytes(*mm_audio["audio"]) as f:
418
419
420
421
        form = aiohttp.FormData()
        form.add_field("file", f, content_type="audio/wav")
        for key, value in payload.items():
            form.add_field(key, str(value))
422
423
424
425
426
427
428

        output = RequestFuncOutput()
        output.prompt_len = request_func_input.prompt_len

        generated_text = ""
        ttft = 0.0
        st = time.perf_counter()
429
        output.start_time = st
430
431
        most_recent_timestamp = st
        try:
432
433
434
            async with session.post(
                url=api_url, data=form, headers=headers
            ) as response:
435
                if response.status == 200:
436
437
438
                    handler = StreamedResponseHandler()

                    async for chunk_bytes in response.content.iter_any():
439
440
441
442
                        chunk_bytes = chunk_bytes.strip()
                        if not chunk_bytes:
                            continue

443
444
                        messages = handler.add_chunk(chunk_bytes)
                        for message in messages:
445
                            chunk = message.decode("utf-8").removeprefix("data: ")
446
447
448
449
450
                            if chunk != "[DONE]":
                                timestamp = time.perf_counter()
                                data = json.loads(chunk)

                                if choices := data.get("choices"):
451
                                    content = choices[0]["delta"].get("content")
452
453
454
455
456
457
458
459
                                    # First token
                                    if ttft == 0.0:
                                        ttft = timestamp - st
                                        output.ttft = ttft

                                    # Decoding phase
                                    else:
                                        output.itl.append(
460
461
                                            timestamp - most_recent_timestamp
                                        )
462
463
464
465

                                    generated_text += content or ""
                                elif usage := data.get("usage"):
                                    output.output_tokens = usage.get(
466
467
                                        "completion_tokens"
                                    )
468
469

                                most_recent_timestamp = timestamp
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486

                    output.generated_text = generated_text
                    output.success = True
                    output.latency = most_recent_timestamp - st
                else:
                    output.error = response.reason or ""
                    output.success = False
        except Exception:
            output.success = False
            exc_info = sys.exc_info()
            output.error = "".join(traceback.format_exception(*exc_info))

    if pbar:
        pbar.update(1)
    return output


487
async def _run_pooling_request(
488
    session: aiohttp.ClientSession,
489
490
491
    api_url: str,
    payload: dict[str, Any],
    headers: dict[str, Any],
492
    pbar: tqdm | None = None,
493
) -> RequestFuncOutput:
494
495
    output = RequestFuncOutput()
    st = time.perf_counter()
496
    output.start_time = st
497
    try:
498
        async with session.post(url=api_url, headers=headers, json=payload) as response:
499
            if response.status == 200:
500
                output.ttft = output.latency = time.perf_counter() - st
501
502
503
                data = await response.json()
                output.success = True
                output.generated_text = ""
504
                output.prompt_len = data.get("usage", {}).get("prompt_tokens", 0)
505
506
507
508
509
510
511
512
513
514
515
516
            else:
                output.success = False
                output.error = response.reason or ""
    except Exception as e:
        output.success = False
        output.error = str(e)

    if pbar:
        pbar.update(1)
    return output


517
518
519
async def async_request_openai_embeddings(
    request_func_input: RequestFuncInput,
    session: aiohttp.ClientSession,
520
    pbar: tqdm | None = None,
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
) -> RequestFuncOutput:
    api_url = request_func_input.api_url
    _validate_api_url(api_url, "OpenAI Embeddings API", "embeddings")

    payload = {
        "model": request_func_input.model_name
        if request_func_input.model_name
        else request_func_input.model,
        "input": request_func_input.prompt,
    }
    _update_payload_common(payload, request_func_input)

    headers = {
        "Content-Type": "application/json",
        "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}",
    }
    _update_headers_common(headers, request_func_input)

539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
    return await _run_pooling_request(
        session,
        api_url,
        payload=payload,
        headers=headers,
        pbar=pbar,
    )


async def async_request_vllm_rerank(
    request_func_input: RequestFuncInput,
    session: aiohttp.ClientSession,
    pbar: tqdm | None = None,
) -> RequestFuncOutput:
    api_url = request_func_input.api_url
    _validate_api_url(api_url, "vLLM score API", "rerank")

    assert (
        isinstance(request_func_input.prompt, list)
        and len(request_func_input.prompt) > 1
    )

    payload = {
        "model": request_func_input.model_name
        if request_func_input.model_name
        else request_func_input.model,
        "query": request_func_input.prompt[0],
        "documents": request_func_input.prompt[1:],
    }

    headers = {
        "Content-Type": "application/json",
        "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}",
    }
    _update_headers_common(headers, request_func_input)

    return await _run_pooling_request(
576
577
578
579
580
581
582
583
584
585
586
        session,
        api_url,
        payload=payload,
        headers=headers,
        pbar=pbar,
    )


async def async_request_openai_embeddings_chat(
    request_func_input: RequestFuncInput,
    session: aiohttp.ClientSession,
587
    pbar: tqdm | None = None,
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
    mm_position: Literal["first", "last"] = "last",
) -> RequestFuncOutput:
    api_url = request_func_input.api_url
    _validate_api_url(api_url, "OpenAI Embeddings API", "embeddings")

    content = _get_chat_content(request_func_input, mm_position=mm_position)

    payload = {
        "model": request_func_input.model_name
        if request_func_input.model_name
        else request_func_input.model,
        "messages": [
            {"role": "user", "content": content},
        ],
    }
    _update_payload_common(payload, request_func_input)

    headers = {
        "Content-Type": "application/json",
        "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}",
    }
    _update_headers_common(headers, request_func_input)

611
    return await _run_pooling_request(
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
        session,
        api_url,
        payload=payload,
        headers=headers,
        pbar=pbar,
    )


def _try_extract_request_idx(request_func_input: RequestFuncInput):
    if request_func_input.request_id:
        match = re.search(r"(\d+)$", request_func_input.request_id)
        if match:
            try:
                return int(match.group(1))
            except ValueError:
                pass

    return None


632
633
634
635
636
637
638
639
640
641
642
643
644
645
def _preprocess_clip(request_func_input: RequestFuncInput):
    if request_func_input.multi_modal_content:
        # Image input
        request_func_input.prompt = ""

    # max_model_len=77 is too short for most datasets,
    # so by default we truncate the prompt to max_model_len
    if request_func_input.extra_body is None:
        request_func_input.extra_body = {}
    if "truncate_prompt_tokens" not in request_func_input.extra_body:
        request_func_input.extra_body["truncate_prompt_tokens"] = -1


def _preprocess_vlm2vec(request_func_input: RequestFuncInput):
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
    if request_func_input.multi_modal_content:
        request_idx = _try_extract_request_idx(request_func_input)

        # Adjust the ratio manually if needed.
        use_image_only_prompt = request_idx is None or request_idx % 2 == 0

        if use_image_only_prompt:
            # Image input
            request_func_input.prompt = "Represent the given image."
        else:
            # Text+Image input
            request_func_input.prompt = (
                f"Represent the given image with the following question: "
                f"{request_func_input.prompt}"
            )

662
663
664
665

async def async_request_openai_embeddings_clip(
    request_func_input: RequestFuncInput,
    session: aiohttp.ClientSession,
666
    pbar: tqdm | None = None,
667
668
669
670
671
672
673
674
675
676
677
678
679
) -> RequestFuncOutput:
    _preprocess_clip(request_func_input)

    return await async_request_openai_embeddings_chat(
        request_func_input,
        session,
        pbar=pbar,
    )


async def async_request_openai_embeddings_vlm2vec(
    request_func_input: RequestFuncInput,
    session: aiohttp.ClientSession,
680
    pbar: tqdm | None = None,
681
682
683
) -> RequestFuncOutput:
    _preprocess_vlm2vec(request_func_input)

684
685
686
687
688
689
690
691
    return await async_request_openai_embeddings_chat(
        request_func_input,
        session,
        pbar=pbar,
        mm_position="first",
    )


692
693
694
async def async_request_infinity_embeddings(
    request_func_input: RequestFuncInput,
    session: aiohttp.ClientSession,
695
    pbar: tqdm | None = None,
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
) -> RequestFuncOutput:
    api_url = request_func_input.api_url
    _validate_api_url(api_url, "Infinity Embeddings API", "embeddings")

    payload = {
        "model": request_func_input.model_name
        if request_func_input.model_name
        else request_func_input.model,
    }

    if request_func_input.prompt:
        payload["input"] = request_func_input.prompt
    else:
        mm_content = request_func_input.multi_modal_content
        assert isinstance(mm_content, dict)

        mm_type = mm_content["type"]
        payload["input"] = mm_content[mm_type]["url"]
        payload["modality"] = mm_type.split("_", 1)[0]

    _update_payload_common(payload, request_func_input)

    headers = {
        "Content-Type": "application/json",
        "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}",
    }
    _update_headers_common(headers, request_func_input)

724
    return await _run_pooling_request(
725
726
727
728
729
730
731
732
733
734
735
        session,
        api_url,
        payload=payload,
        headers=headers,
        pbar=pbar,
    )


async def async_request_infinity_embeddings_clip(
    request_func_input: RequestFuncInput,
    session: aiohttp.ClientSession,
736
    pbar: tqdm | None = None,
737
738
739
740
741
742
743
744
745
746
) -> RequestFuncOutput:
    _preprocess_clip(request_func_input)

    return await async_request_infinity_embeddings(
        request_func_input,
        session,
        pbar=pbar,
    )


747
# TODO: Add more request functions for different API protocols.
748
ASYNC_REQUEST_FUNCS: dict[str, RequestFunc] = {
749
750
751
752
    "vllm": async_request_openai_completions,
    "openai": async_request_openai_completions,
    "openai-chat": async_request_openai_chat_completions,
    "openai-audio": async_request_openai_audio,
753
    "openai-embeddings": async_request_openai_embeddings,
754
755
756
    "openai-embeddings-chat": async_request_openai_embeddings_chat,
    "openai-embeddings-clip": async_request_openai_embeddings_clip,
    "openai-embeddings-vlm2vec": async_request_openai_embeddings_vlm2vec,
757
758
759
760
    # Infinity embedding server: https://github.com/michaelfeil/infinity
    "infinity-embeddings": async_request_infinity_embeddings,
    "infinity-embeddings-clip": async_request_infinity_embeddings_clip,
    # (Infinity embedding server does not support vlm2vec)
761
    "vllm-rerank": async_request_vllm_rerank,
762
}
763
764

OPENAI_COMPATIBLE_BACKENDS = [
765
766
767
    k
    for k, v in ASYNC_REQUEST_FUNCS.items()
    if v in (async_request_openai_completions, async_request_openai_chat_completions)
768
]