xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-05-09 03:04:34 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Core][Bookkeeping] Update cu_num_accepted_tokens for all req_index (#27629)

Browse Source 
Signed-off-by: Jialin Ouyang <Jialin.Ouyang@gmail.com>
This commit is contained in:
Jialin Ouyang 2025-10-30 11:52:36 -07:00 committed by GitHub
parent ab98f6556f
commit 4574d48bab
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 9 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
vllm/v1/worker/gpu_model_runner.py
View File

@ -2323,11 +2323,19 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
else: else:
sampled_ids = valid_sampled_token_ids[req_idx] sampled_ids = valid_sampled_token_ids[req_idx]
num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
if cu_num_accepted_tokens is not None:
cu_num_accepted_tokens.append(
cu_num_accepted_tokens[-1] + num_sampled_ids
)
if not sampled_ids: if not sampled_ids:
continue continue
start_idx = self.input_batch.num_tokens_no_spec[req_idx] start_idx = self.input_batch.num_tokens_no_spec[req_idx]
end_idx = start_idx + len(sampled_ids) end_idx = start_idx + num_sampled_ids
assert end_idx <= self.max_model_len, ( assert end_idx <= self.max_model_len, (
"Sampled token IDs exceed the max model length. " "Sampled token IDs exceed the max model length. "
f"Total number of tokens: {end_idx} > max_model_len: " f"Total number of tokens: {end_idx} > max_model_len: "
@ -2343,11 +2351,6 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
req_state = self.requests[req_id] req_state = self.requests[req_id]
req_state.output_token_ids.extend(sampled_ids) req_state.output_token_ids.extend(sampled_ids)
if cu_num_accepted_tokens is not None:
cu_num_accepted_tokens.append(
cu_num_accepted_tokens[-1] + len(sampled_ids)
)
logprobs_lists = ( logprobs_lists = (
logprobs_tensors.tolists(cu_num_accepted_tokens) logprobs_tensors.tolists(cu_num_accepted_tokens)
if not self.use_async_scheduling and logprobs_tensors is not None if not self.use_async_scheduling and logprobs_tensors is not None