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

[Bugfix][APC] Fix accuracy issue on prefix caching with AscendScheduler (#2714)

Browse Source 
### What this PR does / why we need it?
Fix accuracy issue on prefix caching with AscendScheduler

### How was this patch tested?
CI passed with `test_prefix_cache_with_ascend_scheduler`

- vLLM version: v0.10.1.1
- vLLM main:
6997a25ac6

---------

Signed-off-by: MengqingCao <cmq0113@163.com>
This commit is contained in:
Mengqing Cao
2025-09-04 08:22:46 +08:00
committed by GitHub
parent df88a2ecc8
commit 984bd7c13a
2 changed files with 7 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
.github/workflows/vllm_ascend_test.yaml vendored
View File

@@ -291,6 +291,6 @@ jobs:
pytest -sv tests/e2e/multicard/test_offline_inference_distributed.py::test_sp_for_qwen3_moe
#pytest -sv tests/e2e/multicard/test_pipeline_parallel.py
#pytest -sv tests/e2e/multicard/test_prefix_caching.py
pytest -sv tests/e2e/multicard/test_prefix_caching.py
pytest -sv tests/e2e/multicard/test_qwen3_moe.py
pytest -sv tests/e2e/multicard/test_torchair_graph_mode.py

9
vllm_ascend/core/scheduler.py
View File

@@ -67,7 +67,7 @@ class AscendScheduler(Scheduler):
preempted_reqs: list[Request] = []
if vllm_version_is("0.10.1.1") or vllm_version_is("0.10.1"):
req_to_new_block_ids: dict[str, list[int]] = {}
req_to_new_block_ids: dict[str, list[list[int]]] = {}
else:
req_to_new_blocks: dict[str, KVCacheBlocks] = {}
num_scheduled_tokens: dict[str, int] = {}
@@ -231,7 +231,9 @@ class AscendScheduler(Scheduler):
req_to_new_block_ids[request.request_id] = (
self.kv_cache_manager.get_block_ids(request.request_id))
else:
req_to_new_blocks[request.request_id] = new_blocks
req_to_new_blocks[
request.request_id] = self.kv_cache_manager.get_blocks(
request.request_id)
# Update request info.
num_scheduled_tokens[request.request_id] = num_new_tokens
token_budget -= num_new_tokens
@@ -354,7 +356,8 @@ class AscendScheduler(Scheduler):
# Get the longest common prefix among all requests in the running queue.
# This can be potentially used for cascade attention.
num_common_prefix_blocks = 0
num_common_prefix_blocks = [0] * len(
self.kv_cache_config.kv_cache_groups)
if self.running:
any_request = self.running[0]
num_common_prefix_blocks = (