[Core] Check for GPU<->CPU syncs during CI (#43107)
Signed-off-by: Nick Hill <nickhill123@gmail.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 4.8
parent
7b544ecb52
commit
615d4cfade
@@ -16,6 +16,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
|
||||
from vllm.sequence import IntermediateTensors
|
||||
from vllm.tasks import PoolingTask
|
||||
from vllm.transformers_utils.repo_utils import get_hf_file_bytes
|
||||
from vllm.utils.gpu_sync_debug import gpu_sync_allowed
|
||||
from vllm.v1.pool.metadata import PoolingMetadata
|
||||
|
||||
from ..layers.pooler import DispatchPooler
|
||||
@@ -110,19 +111,24 @@ class JinaForRankingPool(StepPool):
|
||||
prompt_token_ids = pooling_metadata.get_prompt_token_ids()
|
||||
|
||||
embeds_list = list[torch.Tensor | None]()
|
||||
for data, token_ids in zip(pooled_data_lst, prompt_token_ids):
|
||||
# for unfinished chunked prefill
|
||||
if data is None:
|
||||
embeds_list.append(None)
|
||||
else:
|
||||
docs_indexes = torch.where(torch.eq(token_ids, self.doc_token_id))[0]
|
||||
query_indexes = torch.where(torch.eq(token_ids, self.query_token_id))[0]
|
||||
# `torch.where` resolves the match count on the host.
|
||||
with gpu_sync_allowed():
|
||||
for data, token_ids in zip(pooled_data_lst, prompt_token_ids):
|
||||
# for unfinished chunked prefill
|
||||
if data is None:
|
||||
embeds_list.append(None)
|
||||
else:
|
||||
doc_match = torch.eq(token_ids, self.doc_token_id)
|
||||
docs_indexes = torch.where(doc_match)[0]
|
||||
query_indexes = torch.where(
|
||||
torch.eq(token_ids, self.query_token_id)
|
||||
)[0]
|
||||
|
||||
# The JinaForRanking model concatenates docs first, then query.
|
||||
# Let's stay consistent with this novel design.
|
||||
indexes = torch.cat([docs_indexes, query_indexes])
|
||||
embeds = self.projector(data[indexes])
|
||||
embeds_list.append(embeds)
|
||||
# The JinaForRanking model concatenates docs first, then query.
|
||||
# Let's stay consistent with this novel design.
|
||||
indexes = torch.cat([docs_indexes, query_indexes])
|
||||
embeds = self.projector(data[indexes])
|
||||
embeds_list.append(embeds)
|
||||
|
||||
return embeds_list
|
||||
|
||||
|
||||
Reference in New Issue
Block a user