EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

[BugFix] Fix ascend scheduler bugs. (#822)

Browse Source 
This PR fixes two bugs in AscendScheduler:
1. When running with high concurrency, the length of running queue may
exceed the limit of max_num_seqs
2. When some requests are prempted and recomputing is activated, the
logic of computing new tokens is wrong.

Signed-off-by: whx-sjtu <2952154980@qq.com>
This commit is contained in:
whx
2025-05-12 21:15:17 +08:00
committed by GitHub
parent 701b0fd95e
commit 5998704c08
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm_ascend/core/scheduler.py
View File

@@ -74,7 +74,7 @@ class AscendScheduler(Scheduler):
# Schedule prefill requests first.
while self.waiting and token_budget > 0:
if len(scheduled_new_reqs) == self.max_num_running_reqs:
if len(self.running) == self.max_num_running_reqs:
break
request = self.waiting[0]
@@ -96,7 +96,7 @@ class AscendScheduler(Scheduler):
# Get already-cached tokens.
computed_blocks, num_computed_tokens = (
self.kv_cache_manager.get_computed_blocks(request))
num_new_tokens = request.num_prompt_tokens - num_computed_tokens
num_new_tokens = request.num_tokens - num_computed_tokens
if (0 < self.scheduler_config.long_prefill_token_threshold <
num_new_tokens):
num_new_tokens = (