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

drop ascend scheduler (#4498)

Browse Source 
Ascend scheduler was added for non chunk prefill case before, since that
the npu ops didn't work well with chunked prefill.

Now the ops with chunked prefill work better, it's time to remove the
ascend scheduler to use vLLM default scheduler.

- vLLM version: v0.11.2

---------

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2025-11-29 16:18:34 +08:00
committed by GitHub
parent 53a52d6614
commit f10acddb78
52 changed files with 85 additions and 2948 deletions
Download Patch File Download Diff File Expand all files Collapse all files

66
tests/e2e/multicard/test_prefix_caching.py
View File

@@ -1,6 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Compare the with and without prefix caching on V1 scheduler or AscendScheduler."""
"""Compare the with and without prefix caching on V1 scheduler."""
import pytest
@@ -84,67 +84,3 @@ def test_prefix_cache_with_v1_scheduler(model: str, max_tokens: int) -> None:
name_0="vllm_output",
name_1="prefix_cache_output",
)
@pytest.mark.skip(reason="Fix me, the accuracy is not correct")
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("max_tokens", [50])
def test_prefix_cache_with_ascend_scheduler(model: str,
max_tokens: int) -> None:
with VllmRunner(model,
additional_config={
'ascend_scheduler_config': {
'enabled': True,
},
},
enforce_eager=False,
max_model_len=2048,
tensor_parallel_size=2,
gpu_memory_utilization=0.7) as vllm_model:
vllm_output = vllm_model.generate_greedy(INPUT_PROMPTS, max_tokens)
with VllmRunner(model,
additional_config={
'ascend_scheduler_config': {
'enabled': True,
'enable_prefix_caching': True,
},
},
enforce_eager=False,
max_model_len=2048,
tensor_parallel_size=2,
gpu_memory_utilization=0.7) as vllm_model:
prefix_cache_output = vllm_model.generate_greedy(
INPUT_PROMPTS, max_tokens)
# TODO: enable apc and chunked prefill with ascend scheduler will lead accuracy problem.
# Disable it now. Fix it or drop the ascend scheduler in the future.
# with VllmRunner(model,
# additional_config={
# 'ascend_scheduler_config': {
# 'enabled': True,
# 'enable_prefix_caching': True,
# "enable_chunked_prefill": True,
# },
# },
# enforce_eager=True,
# max_model_len=2048,
# tensor_parallel_size=2,
# gpu_memory_utilization=0.7) as vllm_model:
# chunk_prefill_prefix_cache_output = vllm_model.generate_greedy(
# INPUT_PROMPTS, max_tokens)
check_outputs_equal(
outputs_0_lst=vllm_output,
outputs_1_lst=prefix_cache_output,
name_0="vllm_output",
name_1="prefix_cache_output",
)
# check_outputs_equal(
# outputs_0_lst=chunk_prefill_prefix_cache_output,
# outputs_1_lst=prefix_cache_output,
# name_0="chunk_prefill_prefix_cache_output",
# name_1="prefix_cache_output",
# )