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 17, 01:45 UTC
[Dynamic Spec Decoding] Auto-disable by the running queue size (#4592)
Co-authored-by: Cade Daniel <edacih@gmail.com>
Cody Yu committed
May 8, 2024 at 14:44 UTC
f942efb5a3712498b8b583d2d9345f98d15f22f0
11 files changed
+227
-39
tests/samplers/test_rejection_sampler.py
+9
-4
@@ -42,9 +42,11 @@ def mock_causal_accepted_tensor(
42
@pytest.mark.parametrize(
43
"which_tokens_accepted",
44
["all_tokens_accepted", "no_tokens_accepted", "some_tokens_accepted"])
45
+@pytest.mark.parametrize("disable_bonus_tokens", [True, False])
46
@pytest.mark.parametrize("device", CUDA_DEVICES)
47
@torch.inference_mode()
47
-def test_correct_output_format(which_tokens_accepted: str, seed: int,
48
+def test_correct_output_format(which_tokens_accepted: str,
49
+ disable_bonus_tokens: bool, seed: int,
50
device: str):
51
"""Verify the output has correct format given predetermined accepted matrix.
52
"""
@@ -82,7 +84,8 @@ def test_correct_output_format(which_tokens_accepted: str, seed: int,
84
size=(batch_size, 1),
85
dtype=torch.int64)
86
85
- rejection_sampler = RejectionSampler()
87
+ rejection_sampler = RejectionSampler(
88
+ disable_bonus_tokens=disable_bonus_tokens)
89
rejection_sampler.init_gpu_tensors(rank=0)
90
output_token_ids = rejection_sampler._create_output( # pylint: disable=protected-access
91
accepted,
@@ -91,9 +94,11 @@ def test_correct_output_format(which_tokens_accepted: str, seed: int,
94
bonus_token_ids,
95
)
96
94
- # Bonus tokens are currently disabled. Verify they're set to -1.
97
+ expected_bonus_token_ids = bonus_token_ids.clone()
98
+ # If bonus tokens disabled. Verify they are set to -1.
99
# See https://github.com/vllm-project/vllm/issues/4212
96
- expected_bonus_token_ids = bonus_token_ids.clone() * 0 - 1
100
+ if disable_bonus_tokens:
101
+ expected_bonus_token_ids = expected_bonus_token_ids * 0 - 1
102
103
if which_tokens_accepted == "all_tokens_accepted":
104
# Expect all tokens to be equal to draft tokens.
tests/spec_decode/e2e/test_multistep_correctness.py
+34
@@ -536,6 +536,40 @@ def test_skip_speculation(baseline_llm_generator, test_llm_generator,
536
force_output_len=True)
537
538
539
+@pytest.mark.parametrize(
540
+ "common_llm_kwargs",
541
+ [{
542
+ "model": "JackFram/llama-160m",
543
+
544
+ # Skip cuda graph recording for fast test.
545
+ "enforce_eager": True,
546
+
547
+ # Required for spec decode.
548
+ "use_v2_block_manager": True
549
+ }])
550
+@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
551
+@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
552
+@pytest.mark.parametrize("test_llm_kwargs", [
553
+ {
554
+ "speculative_model": "JackFram/llama-68m",
555
+ "num_speculative_tokens": 5,
556
+ "speculative_disable_by_batch_size": 2,
557
+ },
558
+])
559
+@pytest.mark.parametrize("batch_size", [8])
560
+@pytest.mark.parametrize("output_len", [10])
561
+@pytest.mark.parametrize("seed", [1])
562
+def test_disable_speculation(baseline_llm_generator, test_llm_generator,
563
+ batch_size: int, output_len: int):
564
+ """Verify greedy equality when all sequences disable speculation.
565
+ """
566
+ run_greedy_equality_correctness_test(baseline_llm_generator,
567
+ test_llm_generator,
568
+ batch_size,
569
+ max_output_len=output_len,
570
+ force_output_len=True)
571
+
572
+
573
@pytest.mark.parametrize(
574
"common_llm_kwargs",
575
[{
tests/spec_decode/e2e/test_ngram_correctness.py
+1
-1
@@ -57,7 +57,7 @@ from .conftest import run_greedy_equality_correctness_test
57
@pytest.mark.parametrize("output_len", [
58
256,
59
])
60
-@pytest.mark.parametrize("batch_size", [1, 64])
60
+@pytest.mark.parametrize("batch_size", [1, 32])
61
@pytest.mark.parametrize("seed", [1])
62
def test_ngram_e2e_greedy_correctness(baseline_llm_generator,
63
test_llm_generator, batch_size: int,
tests/spec_decode/test_dynamic_spec_decode.py
new
+77
@@ -0,0 +1,77 @@
1
+from unittest.mock import MagicMock
2
+
3
+import pytest
4
+import torch
5
+
6
+from vllm.model_executor.layers.rejection_sampler import RejectionSampler
7
+from vllm.sequence import ExecuteModelRequest
8
+from vllm.spec_decode.metrics import AsyncMetricsCollector
9
+from vllm.spec_decode.multi_step_worker import MultiStepWorker
10
+from vllm.spec_decode.spec_decode_worker import SpecDecodeWorker
11
+from vllm.spec_decode.top1_proposer import Top1Proposer
12
+
13
+from .utils import create_batch, mock_worker
14
+
15
+
16
+@pytest.mark.parametrize('queue_size', [2, 4])
17
+@pytest.mark.parametrize('batch_size', [1, 2, 3, 6])
18
+@pytest.mark.parametrize('k', [1, 2, 5, 7, 10])
19
+@torch.inference_mode()
20
+def test_disable_spec_tokens(queue_size: int, batch_size: int, k: int):
21
+ """Verify that speculative tokens are disabled when the batch size
22
+ exceeds the threshold.
23
+ """
24
+ disable_by_batch_size = 3
25
+
26
+ draft_worker = mock_worker(cls=MultiStepWorker)
27
+ target_worker = mock_worker()
28
+ rejection_sampler = MagicMock(spec=RejectionSampler)
29
+ metrics_collector = MagicMock(spec=AsyncMetricsCollector)
30
+ worker = SpecDecodeWorker(proposer_worker=draft_worker,
31
+ scorer_worker=target_worker,
32
+ rejection_sampler=rejection_sampler,
33
+ metrics_collector=metrics_collector,
34
+ disable_by_batch_size=disable_by_batch_size)
35
+
36
+ exception_secret = 'artificial stop'
37
+ draft_worker.get_spec_proposals.side_effect = ValueError(exception_secret)
38
+
39
+ seq_group_metadata_list, _, _ = create_batch(batch_size, k)
40
+ execute_model_req = ExecuteModelRequest(
41
+ seq_group_metadata_list=seq_group_metadata_list,
42
+ num_lookahead_slots=k,
43
+ running_queue_size=queue_size)
44
+
45
+ with pytest.raises(ValueError, match=exception_secret):
46
+ worker.execute_model(execute_model_req=execute_model_req)
47
+
48
+ # When the batch size is larger than the threshold,
49
+ # we expect no speculative tokens (0).
50
+ expected_num_spec_tokens = None if queue_size < disable_by_batch_size else 0
51
+ assert seq_group_metadata_list[
52
+ 0].num_speculative_tokens == expected_num_spec_tokens
53
+
54
+ draft_worker.sampler_output.side_effect = ValueError(exception_secret)
55
+
56
+ proposer = Top1Proposer(
57
+ worker=draft_worker,
58
+ device='cpu', # not used
59
+ vocab_size=100, # not used
60
+ # Must be long enough to avoid being skipped due to length.
61
+ max_proposal_len=1024,
62
+ )
63
+
64
+ if queue_size < disable_by_batch_size:
65
+ # Should raise exception when executing the mocked draft model.
66
+ with pytest.raises(ValueError, match=exception_secret):
67
+ proposer.get_proposals(execute_model_req=ExecuteModelRequest(
68
+ seq_group_metadata_list=seq_group_metadata_list,
69
+ num_lookahead_slots=k), )
70
+ else:
71
+ # Should not execute the draft model because spec decode is disabled
72
+ # for all requests. Accordingly, the proposal length should be 0.
73
+ proposals = proposer.get_proposals(
74
+ execute_model_req=ExecuteModelRequest(
75
+ seq_group_metadata_list=seq_group_metadata_list,
76
+ num_lookahead_slots=k), )
77
+ assert proposals.proposal_lens.tolist() == [0] * batch_size
vllm/config.py
+24
-5
@@ -692,6 +692,7 @@ class SpeculativeConfig:
692
speculative_max_model_len: Optional[int],
693
enable_chunked_prefill: bool,
694
use_v2_block_manager: bool,
695
+ speculative_disable_by_batch_size: Optional[int],
696
ngram_prompt_lookup_max: Optional[int],
697
ngram_prompt_lookup_min: Optional[int],
698
) -> Optional["SpeculativeConfig"]:
@@ -720,6 +721,9 @@ class SpeculativeConfig:
721
use_v2_block_manager (bool): Whether vLLM is configured to use the
722
v2 block manager or not. Used for raising an error since the v2
723
block manager is required with spec decode.
724
+ speculative_disable_by_batch_size (Optional[int]): Disable
725
+ speculative decoding for new incoming requests when the number
726
+ of enqueue requests is larger than this value, if provided.
727
ngram_prompt_lookup_max (Optional[int]): Max size of ngram token
728
window, if provided.
729
ngram_prompt_lookup_min (Optional[int]): Min size of ngram token
@@ -730,7 +734,7 @@ class SpeculativeConfig:
734
the necessary conditions are met, else None.
735
"""
736
733
- if (speculative_model is None and num_speculative_tokens is None):
737
+ if speculative_model is None and num_speculative_tokens is None:
738
return None
739
740
if speculative_model is not None and num_speculative_tokens is None:
@@ -739,6 +743,12 @@ class SpeculativeConfig:
743
"num_speculative_tokens to be provided, but found "
744
f"{speculative_model=} and {num_speculative_tokens=}.")
745
746
+ if (speculative_disable_by_batch_size is not None
747
+ and speculative_disable_by_batch_size < 2):
748
+ raise ValueError("Expect the batch size threshold of disabling "
749
+ "speculative decoding is > 1, but got "
750
+ f"{speculative_disable_by_batch_size=}")
751
+
752
assert (speculative_model is not None
753
and num_speculative_tokens is not None)
754
@@ -807,6 +817,7 @@ class SpeculativeConfig:
817
draft_model_config,
818
draft_parallel_config,
819
num_speculative_tokens,
820
+ speculative_disable_by_batch_size,
821
ngram_prompt_lookup_max,
822
ngram_prompt_lookup_min,
823
)
@@ -876,8 +887,9 @@ class SpeculativeConfig:
887
draft_model_config: ModelConfig,
888
draft_parallel_config: ParallelConfig,
889
num_speculative_tokens: int,
879
- ngram_prompt_lookup_max: int,
880
- ngram_prompt_lookup_min: int,
890
+ speculative_disable_by_batch_size: Optional[int],
891
+ ngram_prompt_lookup_max: Optional[int],
892
+ ngram_prompt_lookup_min: Optional[int],
893
):
894
"""Create a SpeculativeConfig object.
895
@@ -886,12 +898,19 @@ class SpeculativeConfig:
898
draft_parallel_config: ParallelConfig for the draft model.
899
num_speculative_tokens: The number of tokens to sample from the
900
draft model before scoring with the target model.
901
+ speculative_disable_by_batch_size: Disable speculative
902
+ decoding for new incoming requests when the number of
903
+ enqueue requests is larger than this value.
904
+ ngram_prompt_lookup_max: Max size of ngram token window.
905
+ ngram_prompt_lookup_min: Min size of ngram token window.
906
"""
907
self.draft_model_config = draft_model_config
908
self.draft_parallel_config = draft_parallel_config
909
self.num_speculative_tokens = num_speculative_tokens
893
- self.ngram_prompt_lookup_max = ngram_prompt_lookup_max
894
- self.ngram_prompt_lookup_min = ngram_prompt_lookup_min
910
+ self.speculative_disable_by_batch_size = \
911
+ speculative_disable_by_batch_size
912
+ self.ngram_prompt_lookup_max = ngram_prompt_lookup_max or 0
913
+ self.ngram_prompt_lookup_min = ngram_prompt_lookup_min or 0
914
915
self._verify_args()
916
vllm/engine/arg_utils.py
+10
@@ -83,6 +83,7 @@ class EngineArgs:
83
speculative_model: Optional[str] = None
84
num_speculative_tokens: Optional[int] = None
85
speculative_max_model_len: Optional[int] = None
86
+ speculative_disable_by_batch_size: Optional[int] = None
87
ngram_prompt_lookup_max: Optional[int] = None
88
ngram_prompt_lookup_min: Optional[int] = None
89
@@ -467,6 +468,13 @@ class EngineArgs:
468
'draft model. Sequences over this length will skip '
469
'speculation.')
470
471
+ parser.add_argument(
472
+ '--speculative-disable-by-batch-size',
473
+ type=int,
474
+ default=EngineArgs.speculative_disable_by_batch_size,
475
+ help='Disable speculative decoding for new incoming requests '
476
+ 'if the number of enqueue requests is larger than this value.')
477
+
478
parser.add_argument(
479
'--ngram-prompt-lookup-max',
480
type=int,
@@ -547,6 +555,8 @@ class EngineArgs:
555
target_dtype=self.dtype,
556
speculative_model=self.speculative_model,
557
num_speculative_tokens=self.num_speculative_tokens,
558
+ speculative_disable_by_batch_size=self.
559
+ speculative_disable_by_batch_size,
560
speculative_max_model_len=self.speculative_max_model_len,
561
enable_chunked_prefill=self.enable_chunked_prefill,
562
use_v2_block_manager=self.use_v2_block_manager,
vllm/executor/gpu_executor.py
+2
@@ -93,6 +93,8 @@ class GPUExecutor(ExecutorBase):
93
spec_decode_worker = SpecDecodeWorker.create_worker(
94
scorer_worker=target_worker,
95
draft_worker_kwargs=draft_worker_kwargs,
96
+ disable_by_batch_size=self.speculative_config.
97
+ speculative_disable_by_batch_size,
98
)
99
100
assert self.parallel_config.world_size == 1, (
vllm/model_executor/layers/rejection_sampler.py
+9
-2
@@ -12,15 +12,21 @@ class RejectionSampler(nn.Module):
12
https://arxiv.org/pdf/2302.01318.pdf.
13
"""
14
15
- def __init__(self, strict_mode: bool = False):
15
+ def __init__(self,
16
+ disable_bonus_tokens: bool = True,
17
+ strict_mode: bool = False):
18
"""Create a rejection sampler.
19
20
Args:
21
+ disable_bonus_tokens: Whether or not to disable the bonus token.
22
+ Require when bonus tokens will cause corrupt KV cache for
23
+ proposal methods that require KV cache.
24
strict_mode: Whether or not to perform shape/device/dtype checks
25
during sampling. This catches correctness issues but adds
26
nontrivial latency.
27
"""
28
super().__init__()
29
+ self._disable_bonus_tokens = disable_bonus_tokens
30
self._strict_mode = strict_mode
31
32
# NOTE: A "bonus token" is accepted iff all proposal tokens are
@@ -312,7 +318,8 @@ class RejectionSampler(nn.Module):
318
# proposal methods that require KV cache. We can fix it by "prefilling"
319
# the bonus token in the proposer. The following issue tracks the fix.
320
# https://github.com/vllm-project/vllm/issues/4212
315
- output_with_bonus_tokens[:, -1] = -1
321
+ if self._disable_bonus_tokens:
322
+ output_with_bonus_tokens[:, -1] = -1
323
324
# Fill the recovered token ids.
325
output.mul_(~after_false_mask).add_(
vllm/sequence.py
+6
@@ -612,6 +612,12 @@ class SequenceGroupMetadata:
612
self._token_chunk_size = token_chunk_size
613
self.do_sample = do_sample
614
615
+ # The number of speculative tokens adopted in this request.
616
+ # None means specuative decoding is not used.
617
+ # Zero means speculative decoding is disabled for some reasons.
618
+ # TODO: We should maintain this states out of the sequence group.
619
+ self.num_speculative_tokens = None
620
+
621
if self._token_chunk_size is None:
622
if is_prompt:
623
self._token_chunk_size = list(seq_data.values())[0].get_len()
vllm/spec_decode/spec_decode_worker.py
+38
-21
@@ -1,5 +1,5 @@
1
from functools import cached_property
2
-from typing import List, Optional, Tuple
2
+from typing import Any, Dict, List, Optional, Tuple
3
4
import torch
5
@@ -54,7 +54,8 @@ class SpecDecodeWorker(LoraNotSupportedWorkerBase):
54
def create_worker(
55
cls,
56
scorer_worker: WorkerBase,
57
- draft_worker_kwargs,
57
+ draft_worker_kwargs: Dict[str, Any],
58
+ disable_by_batch_size: Optional[int],
59
) -> "SpecDecodeWorker":
60
61
ngram_prompt_lookup_max = (
@@ -62,7 +63,9 @@ class SpecDecodeWorker(LoraNotSupportedWorkerBase):
63
ngram_prompt_lookup_min = (
64
draft_worker_kwargs.pop("ngram_prompt_lookup_min"))
65
66
+ disable_bonus_tokens = True
67
if ngram_prompt_lookup_max > 0:
68
+ disable_bonus_tokens = False
69
proposer_worker = NGramWorker(**draft_worker_kwargs)
70
proposer_worker.set_ngram_window_size(ngram_prompt_lookup_min,
71
ngram_prompt_lookup_max)
@@ -75,9 +78,9 @@ class SpecDecodeWorker(LoraNotSupportedWorkerBase):
78
return SpecDecodeWorker(
79
proposer_worker,
80
scorer_worker,
78
- # TODO(cade) disable strict mode for speedup.
79
- rejection_sampler=RejectionSampler(strict_mode=True),
80
- )
81
+ disable_by_batch_size=disable_by_batch_size,
82
+ rejection_sampler=RejectionSampler(
83
+ disable_bonus_tokens=disable_bonus_tokens, ))
84
85
def __init__(
86
self,
@@ -85,6 +88,7 @@ class SpecDecodeWorker(LoraNotSupportedWorkerBase):
88
scorer_worker: WorkerBase,
89
rejection_sampler: RejectionSampler,
90
metrics_collector: Optional[AsyncMetricsCollector] = None,
91
+ disable_by_batch_size: Optional[int] = None,
92
):
93
"""
94
Create a SpecDecodeWorker.
@@ -97,11 +101,14 @@ class SpecDecodeWorker(LoraNotSupportedWorkerBase):
101
Worker.
102
rejection_sampler: A Torch module used to perform modified rejection
103
sampling for speculative decoding.
104
+ disable_by_batch_size: If the batch size is larger than this,
105
+ disable speculative decoding for new incoming requests.
106
metrics_collector: Helper class for collecting metrics; can be set
107
for testing purposes.
108
"""
109
self.proposer_worker = proposer_worker
110
self.scorer_worker = scorer_worker
111
+ self.disable_by_batch_size = disable_by_batch_size or float("inf")
112
self.rejection_sampler = rejection_sampler
113
114
self._metrics = AsyncMetricsCollector(
@@ -199,27 +206,41 @@ class SpecDecodeWorker(LoraNotSupportedWorkerBase):
206
"speculative decoding "
207
"requires non-None seq_group_metadata_list")
208
209
+ # When the batch size is too large, disable speculative decoding
210
+ # to stop trading off throughput for latency.
211
+ disable_all = (execute_model_req.running_queue_size >=
212
+ self.disable_by_batch_size)
213
+ if disable_all:
214
+ for seq_group_metadata in execute_model_req.seq_group_metadata_list:
215
+ # Once num_speculative_tokens is set to 0, the spec decode
216
+ # of this request will be disabled forever.
217
+ # TODO(comaniac): We currently store spec decoding specific
218
+ # state in the global data structure, but we should maintain
219
+ # this state within spec decode worker.
220
+ seq_group_metadata.num_speculative_tokens = 0
221
+
222
# If no spec tokens, call the proposer and scorer workers normally.
203
- # Used for prefill.
223
+ # This happens for prefill, or when the spec decode is disabled
224
+ # for this batch.
225
if execute_model_req.num_lookahead_slots == 0 or len(
226
execute_model_req.seq_group_metadata_list) == 0:
206
- return self._run_no_spec(execute_model_req)
227
+ return self._run_no_spec(execute_model_req,
228
+ skip_proposer=disable_all)
229
230
return self._run_speculative_decoding_step(execute_model_req)
231
232
@nvtx_range("spec_decode_worker._run_no_spec")
211
- def _run_no_spec(
212
- self,
213
- execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]:
214
- """Run a prefill step, without any speculation. The input is sent to the
215
- proposer and scorer model so that the KV cache is consistent between the
216
- two.
233
+ def _run_no_spec(self, execute_model_req: ExecuteModelRequest,
234
+ skip_proposer: bool) -> List[SamplerOutput]:
235
+ """Run a prefill step, without any speculation. The input is sent to
236
+ the proposer and scorer model so that the KV cache is consistent
237
+ between the two. When skip_proposer is True, the proposer model is
238
+ not called, meaning that the kv-cache in proposer for requests is not
239
+ updated, so they cannot enable spec decode in the rest decoding.
240
"""
218
- #logger.info("run proposer worker no spec")
219
-
220
- self.proposer_worker.execute_model(execute_model_req)
241
+ if not skip_proposer:
242
+ self.proposer_worker.execute_model(execute_model_req)
243
222
- #logger.info("run target worker no spec")
244
sampler_output = self.scorer_worker.execute_model(execute_model_req)
245
assert len(sampler_output) == 1
246
sampler_output = sampler_output[0]
@@ -244,22 +265,18 @@ class SpecDecodeWorker(LoraNotSupportedWorkerBase):
265
sequence.
266
"""
267
247
- #logger.info("get spec proposals")
268
# Generate proposals using draft worker.
269
proposals = self.proposer_worker.get_spec_proposals(execute_model_req)
270
251
- #logger.info("score proposals")
271
proposal_scores = self.scorer.score_proposals(
272
execute_model_req,
273
proposals,
274
)
275
257
- #logger.info("verify proposals")
276
accepted_token_ids, target_logprobs = self._verify_tokens(
277
execute_model_req.seq_group_metadata_list, proposal_scores,
278
proposals, execute_model_req.num_lookahead_slots)
279
262
- #logger.info("create output list")
280
return self._create_output_sampler_list(
281
execute_model_req.seq_group_metadata_list,
282
accepted_token_ids,
vllm/spec_decode/top1_proposer.py
+17
-6
@@ -56,7 +56,7 @@ class Top1Proposer(SpeculativeProposer):
56
proposal_lens,
57
nonzero_proposal_len_seqs,
58
nonzero_proposal_len_indices,
59
- ) = self._split_by_max_model_len(seq_group_metadata_list, proposal_len)
59
+ ) = self._split_by_proposal_len(seq_group_metadata_list, proposal_len)
60
61
if nonzero_proposal_len_seqs:
62
# Speculate tokens using the draft worker for the speculative
@@ -97,17 +97,27 @@ class Top1Proposer(SpeculativeProposer):
97
98
return proposals
99
100
- def _split_by_max_model_len(
100
+ def _split_by_proposal_len(
101
self,
102
seq_group_metadata_list: List[SequenceGroupMetadata],
103
proposal_len: int,
104
) -> Tuple[List[int], List[SequenceGroupMetadata], List[int]]:
105
- """Determine which sequences would exceed the max model length."""
105
+ """Split sequences by two groups:
106
+ 1. Sequences with non-zero proposal length.
107
+ 2. Sequences with zero proposal length (due to disabled speculation
108
+ or exceed the maximum model length).
109
+ """
110
111
proposal_lens: List[int] = []
112
nonzero_proposal_len_seqs: List[SequenceGroupMetadata] = []
113
nonzero_proposal_len_indices: List[int] = []
114
for i, seq_group_metadata in enumerate(seq_group_metadata_list):
115
+ # The speculative decoding for this request has been disabled
116
+ # (e.g. due to high traffic).
117
+ if seq_group_metadata.num_speculative_tokens == 0:
118
+ proposal_lens.append(0)
119
+ continue
120
+
121
seq_data = next(iter(seq_group_metadata.seq_data.values()))
122
seq_len = seq_data.get_len()
123
@@ -115,13 +125,14 @@ class Top1Proposer(SpeculativeProposer):
125
# are supported.
126
# If max_proposal_len is defined, then we shall no exccess this
127
# quota for nonzero_proposal
128
+ new_k = 0
129
if (self.max_proposal_len is None
130
or seq_len + proposal_len < self.max_proposal_len):
120
- proposal_lens.append(proposal_len)
131
+ new_k = proposal_len
132
nonzero_proposal_len_seqs.append(seq_group_metadata)
133
nonzero_proposal_len_indices.append(i)
123
- else:
124
- proposal_lens.append(0)
134
+ proposal_lens.append(new_k)
135
+ seq_group_metadata.num_speculative_tokens = new_k
136
137
return (
138
proposal_lens,