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

21
tests/e2e/multicard/test_expert_parallel.py
View File

@@ -15,23 +15,14 @@ def test_e2e_ep_correctness(model_name):
max_tokens = 5
# FIXME: Really strange that chunked prefill might lead to different results, investigate further
with VllmRunner(
model_name,
tensor_parallel_size=2,
additional_config={"ascend_scheduler_config": {
"enabled": True
}},
enforce_eager=False) as vllm_model:
with VllmRunner(model_name, tensor_parallel_size=2,
enforce_eager=False) as vllm_model:
tp_output = vllm_model.generate_greedy(example_prompts, max_tokens)
with VllmRunner(
model_name,
tensor_parallel_size=2,
enable_expert_parallel=True,
additional_config={"ascend_scheduler_config": {
"enabled": True
}},
enforce_eager=False) as vllm_model:
with VllmRunner(model_name,
tensor_parallel_size=2,
enable_expert_parallel=True,
enforce_eager=False) as vllm_model:
ep_output = vllm_model.generate_greedy(example_prompts, max_tokens)
check_outputs_equal(

16
tests/e2e/multicard/test_fused_moe_allgather_ep.py
View File

@@ -49,13 +49,7 @@ def test_generate_with_allgather():
tensor_parallel_size=2,
max_model_len=1024,
dtype="auto",
enable_expert_parallel=True,
additional_config={
"ascend_scheduler_config": {
"enabled": True,
"chunked_prefill_enabled": False,
},
}) as vllm_model:
enable_expert_parallel=True) as vllm_model:
vllm_model.generate(example_prompts, sampling_params)
@@ -76,11 +70,5 @@ def test_generate_with_alltoall():
tensor_parallel_size=2,
max_model_len=1024,
dtype="auto",
enable_expert_parallel=True,
additional_config={
"ascend_scheduler_config": {
"enabled": True,
"chunked_prefill_enabled": False,
},
}) as vllm_model:
enable_expert_parallel=True) as vllm_model:
vllm_model.generate(example_prompts, sampling_params)

14
tests/e2e/multicard/test_offline_inference_distributed.py
View File

@@ -82,9 +82,6 @@ def test_models_distributed_DeepSeek_multistream_moe():
"enabled": True,
},
"enable_multistream_moe": True,
"ascend_scheduler_config": {
"enabled": True,
},
"refresh": True,
},
) as vllm_model:
@@ -154,14 +151,9 @@ def test_models_distributed_DeepSeek_W4A8DYNAMIC(model):
quantization="ascend",
enforce_eager=True,
enable_expert_parallel=True,
additional_config={
"torchair_graph_config": {
"enabled": False,
},
"ascend_scheduler_config": {
"enabled": True,
}
},
additional_config={"torchair_graph_config": {
"enabled": False,
}},
) as vllm_model:
vllm_model.generate_greedy(prompts, max_tokens)

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",
# )

9
tests/e2e/multicard/test_qwen3_next.py
View File

@@ -24,6 +24,7 @@ Run `pytest tests/e2e/multicard/test_qwen3_next.py`.
import os
from unittest.mock import patch
import pytest
from modelscope import snapshot_download # type: ignore
from tests.e2e.conftest import VllmRunner
@@ -63,6 +64,8 @@ def test_models_distributed_Qwen3_NEXT_TP4_FULL_DECODE_ONLY():
del vllm_model
@pytest.mark.skip(
reason="Qwen3-Next + MTP doesn't work with chunked prefill. Fix Me")
def test_models_distributed_Qwen3_NEXT_MTP_TP4_SIMILARITY():
example_prompts = [
"Hello, my name is",
@@ -89,12 +92,6 @@ def test_models_distributed_Qwen3_NEXT_MTP_TP4_SIMILARITY():
gpu_memory_utilization=0.8,
distributed_executor_backend="mp",
enforce_eager=True,
additional_config={
"ascend_scheduler_config": {
"enabled": True,
"enable_chunked_prefill": False
}
},
speculative_config={
"method": "qwen3_next_mtp",
"num_speculative_tokens": 1

12
tests/e2e/multicard/test_torchair_graph_mode.py
View File

@@ -44,9 +44,6 @@ def _deepseek_torchair_test_fixture(
kwargs = {}
if not use_v1_schduler:
kwargs = {
"ascend_scheduler_config": {
"enabled": True,
},
"refresh": True,
}
additional_config.update(**kwargs)
@@ -120,9 +117,6 @@ def _pangu_torchair_test_fixture(
# torchair is only work without chunked-prefill now
kwargs = {
"ascend_scheduler_config": {
"enabled": True,
},
"refresh": True,
}
additional_config.update(**kwargs)
@@ -185,9 +179,6 @@ def _qwen_torchair_test_fixture(
"torchair_graph_config": {
"enabled": False,
},
"ascend_scheduler_config": {
"enabled": True,
},
"refresh": True,
}
@@ -244,9 +235,6 @@ def _deepseek_v2_lite_torchair_test_fixure(
kwargs = {}
if not use_v1_schduler:
kwargs = {
"ascend_scheduler_config": {
"enable": True,
},
"refresh": True,
}
additional_config.update(**kwargs)