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

support async mtp (#4511)

Browse Source 
### What this PR does / why we need it?
this pr aims to support async_scheduling for mtp, which refer to vllm pr
https://github.com/vllm-project/vllm/pull/24799.
and this pr fix some synchronize problem in vllm-ascend.
### Does this PR introduce _any_ user-facing change?

### How was this patch tested?


- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

---------

Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
Co-authored-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
Ronald
2025-12-06 17:15:57 +08:00
committed by GitHub
parent f067623afd
commit 3480094d7c
8 changed files with 477 additions and 83 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
tests/ut/attention/test_attention_v1.py
View File

@@ -87,6 +87,7 @@ class TestAscendAttentionMetadataBuilder(TestBase):
self.mock_vllm_config.scheduler_config.decode_max_num_seqs = 10
self.mock_vllm_config.scheduler_config.chunked_prefill_enabled = False
self.mock_device = 'cpu:0'
torch.Tensor.pin_memory = lambda x: x # noqa
self.builder = AscendAttentionMetadataBuilder(None, None,
self.mock_vllm_config,
self.mock_device)

8
tests/ut/attention/test_mla_v1.py
View File

@@ -299,6 +299,7 @@ class TestAscendMLAMetadataBuilder(TestBase):
mock_vllm_config.scheduler_config.decode_max_num_seqs = 4
mock_vllm_config.scheduler_config.chunked_prefill_enabled = False
mock_device = 'cpu'
torch.Tensor.pin_memory = lambda x: x # noqa
mock_dcp.world_size = 1
dcp_group = MagicMock(spec=GroupCoordinator)
@@ -534,6 +535,7 @@ class TestAscendMLAMetadataBuilderBuild(TestBase):
mock_get_pcp_group):
mock_npu_available.return_value = False
mock_dcp_world_size.return_value = 1
torch.Tensor.pin_memory = lambda x: x # noqa
pcp_group = MagicMock(spec=GroupCoordinator)
pcp_group.world_size = 1
mock_get_pcp_group.return_value = pcp_group
@@ -599,6 +601,7 @@ class TestAscendMLAMetadataBuilderBuild(TestBase):
mock_get_pcp_group):
mock_npu_available.return_value = False
mock_dcp_world_size.return_value = 1
torch.Tensor.pin_memory = lambda x: x # noqa
pcp_group = MagicMock(spec=GroupCoordinator)
pcp_group.world_size = 1
mock_get_pcp_group.return_value = pcp_group
@@ -660,6 +663,8 @@ class TestAscendMLAMetadataBuilderBuild(TestBase):
mock_dcp_world_size,
mock_get_pcp_group):
mock_dcp_world_size.return_value = 1
torch.Tensor.pin_memory = lambda x: x # noqa
pcp_group = MagicMock(spec=GroupCoordinator)
pcp_group.world_size = 1
mock_get_pcp_group.return_value = pcp_group
@@ -713,6 +718,8 @@ class TestAscendMLAMetadataBuilderBuild(TestBase):
mock_dcp_world_size,
mock_get_pcp_group):
mock_dcp_world_size.return_value = 1
torch.Tensor.pin_memory = lambda x: x # noqa
pcp_group = MagicMock(spec=GroupCoordinator)
pcp_group.world_size = 1
mock_get_pcp_group.return_value = pcp_group
@@ -767,6 +774,7 @@ class TestAscendMLAMetadataBuilderBuild(TestBase):
mock_dcp_world_size,
mock_get_pcp_group):
mock_dcp_world_size.return_value = 1
torch.Tensor.pin_memory = lambda x: x # noqa
pcp_group = MagicMock(spec=GroupCoordinator)
pcp_group.world_size = 1
mock_get_pcp_group.return_value = pcp_group