xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-09-06 06:16:59 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] V1 Memory Profiling: V0 Sampler Integration without Rejection Sampler (#13594)

Browse Source 
Signed-off-by: Jennifer Zhao <7443418+JenZhao@users.noreply.github.com>
Co-authored-by: Roger Wang <ywang@roblox.com>
This commit is contained in:
Jennifer Zhao 2025-02-22 00:08:29 -08:00 committed by GitHub
parent bb78fb318e
commit da31b5333e
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 26 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

28
vllm/v1/worker/gpu_model_runner.py
View File

@ -31,6 +31,7 @@ from vllm.v1.engine.mm_input_cache import MMInputCacheClient
from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheConfig, from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheConfig,
KVCacheSpec) KVCacheSpec)
from vllm.v1.outputs import LogprobsTensors, ModelRunnerOutput from vllm.v1.outputs import LogprobsTensors, ModelRunnerOutput
from vllm.v1.sample.metadata import SamplingMetadata
from vllm.v1.sample.rejection_sampler import INVALID_TOKEN_ID from vllm.v1.sample.rejection_sampler import INVALID_TOKEN_ID
from vllm.v1.spec_decode.ngram_proposer import NgramProposer from vllm.v1.spec_decode.ngram_proposer import NgramProposer
from vllm.v1.utils import bind_kv_cache from vllm.v1.utils import bind_kv_cache
@ -1305,11 +1306,34 @@ class GPUModelRunner(LoRAModelRunnerMixin):
if get_pp_group().is_last_rank: if get_pp_group().is_last_rank:
hidden_states = hidden_states[logit_indices] hidden_states = hidden_states[logit_indices]
logits = self.model.compute_logits(hidden_states, None) logits = self.model.compute_logits(hidden_states, None)
# TODO(woosuk): Consider the memory usage of the sampler. dummy_tensors = lambda v: torch.full(
(num_reqs, ), v, device=self.device)
dummy_metadata = SamplingMetadata(
temperature=dummy_tensors(0.5),
all_greedy=False,
all_random=False,
spec_token_ids=None,
top_p=dummy_tensors(0.9),
top_k=dummy_tensors(logits.size(1) - 1),
min_p=None,
generators={},
max_num_logprobs=None,
no_penalties=True,
prompt_token_ids=torch.ones_like(logits, dtype=torch.int64),
frequency_penalties=dummy_tensors(0.1),
presence_penalties=dummy_tensors(0.1),
repetition_penalties=dummy_tensors(0.1),
output_token_ids=[[] for _ in range(num_reqs)],
min_tokens={},
logit_bias=[None for _ in range(num_reqs)])
sampler_output = self.model.sample(
logits=logits, sampling_metadata=dummy_metadata)
else: else:
logits = None logits = None
sampler_output = None
dummy_metadata = None
torch.cuda.synchronize() torch.cuda.synchronize()
del hidden_states, logits del hidden_states, logits, sampler_output, dummy_metadata
self.encoder_cache.clear() self.encoder_cache.clear()
gc.collect() gc.collect()