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

[Feat] allow using aclgraph in ray backend (#2589)

Browse Source 
### What this PR does / why we need it?

Allow using aclgraph in ray backend, for tp + pp + aclgraph in multi
machine

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.10.1.1
- vLLM main:
4ba0c587ba

Signed-off-by: withHades <244036962@qq.com>
This commit is contained in:
无脸男
2025-09-04 11:45:56 +08:00
committed by GitHub
parent aff5189c87
commit 0c0789be74
2 changed files with 0 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
tests/ut/test_platform.py
View File

@@ -363,36 +363,6 @@ class TestNPUPlatform(TestBase):
CUDAGraphMode.NONE,
)
@patch("vllm_ascend.utils.is_310p", return_value=False)
@patch("vllm_ascend.ascend_config.check_ascend_config")
@patch("vllm_ascend.ascend_config.init_ascend_config")
def test_check_and_update_config_disable_aclgraph_when_ray_enabled(
self, mock_init_ascend, mock_check_ascend, mock_is_310p):
mock_init_ascend.return_value = TestNPUPlatform.mock_vllm_ascend_config(
)
vllm_config = TestNPUPlatform.mock_vllm_config()
vllm_config.model_config.enforce_eager = False
vllm_config.compilation_config.level = CompilationLevel.PIECEWISE
vllm_config.parallel_config.distributed_executor_backend = "ray"
with self.assertLogs(logger="vllm", level="WARNING") as cm:
from vllm_ascend import platform
importlib.reload(platform)
self.platform.check_and_update_config(vllm_config)
print(30 * "=", f"cm.output: {cm.output}")
self.assertTrue(
"Ray distributed executor backend is not compatible with ACL Graph mode"
in cm.output[0])
self.assertEqual(
vllm_config.compilation_config.level,
CompilationLevel.NO_COMPILATION,
)
self.assertEqual(
vllm_config.compilation_config.cudagraph_mode,
CUDAGraphMode.NONE,
)
@patch("vllm_ascend.utils.is_310p", return_value=False)
@patch("vllm_ascend.ascend_config.check_ascend_config")
@patch("vllm_ascend.ascend_config.init_ascend_config")

6
vllm_ascend/platform.py
View File

@@ -185,12 +185,6 @@ class NPUPlatform(Platform):
"and use_cached_kv_cache_bytes in torchair_graph_config.")
delete_torchair_cache_file()
if parallel_config.distributed_executor_backend == "ray":
logger.warning(
"Ray distributed executor backend is not compatible with ACL Graph mode "
"right now. Setting CUDAGraphMode to NONE")
compilation_config.cudagraph_mode = CUDAGraphMode.NONE
# set cudaprah sizes before extending `compilation_config.splitting_ops`
vllm_config._set_cudagraph_sizes()