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

[BugFix][DSv32] Fix DSA-CP PD role gating for deepseek v3.2 (v0.18.0) (#8291)

Browse Source 
### What this PR does / why we need it?
This PR backports the DSA-CP PD role gating fix to `releases/v0.18.0`.

The existing helper logic on the release branch does not handle the PD
mixed-role case correctly when deciding whether layer sharding or TP
`o_proj` handling should be enabled. Layer sharding should only run on
the P-side instance, while TP `o_proj` handling should stay enabled for
normal non-PD deployments and for the PD mixed-role (`kv_both`)
instance. This patch makes those conditions explicit and adds unit
coverage for the allowed and disallowed combinations, including the
DSA-CP-disabled path.

Such wrong condition lead to **vllm serve failures** in case: **FC1 +
PD-colocated KV pooling + no layer_sharding**, specifically causing:
1. insufficient Available KV cache memory
2. o_proj shape error in sfa_v1 attention module

### Does this PR introduce _any_ user-facing change?
No.

### How was this patch tested?
E2E test with dsv32 + FC1 + FULL_DECODE_ONLY +
kv_transfer_config(kv_both) + no layer_sharding

---------

Signed-off-by: chenchuw886 <chenchuw@huawei.com>
Co-authored-by: chenchuw886 <chenchuw@huawei.com>
This commit is contained in:
Frank Chen
2026-04-17 10:05:40 +08:00
committed by GitHub
parent 1dd1de8153
commit f85144cc57
2 changed files with 53 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

45
tests/ut/test_utils.py
View File

@@ -36,6 +36,7 @@ class TestUtils(TestBase):
from vllm_ascend import platform
importlib.reload(platform)
utils.enable_dsa_cp_with_layer_shard.cache_clear()
utils.enable_dsa_cp_with_o_proj_tp.cache_clear()
def test_nd_to_nz_2d(self):
# can be divided by 16
@@ -134,7 +135,9 @@ class TestUtils(TestBase):
def test_enable_dsa_cp_with_layer_shard_accepts_kv_producer(self):
mock_vllm_config = mock.MagicMock()
mock_vllm_config.kv_transfer_config = mock.MagicMock(kv_role="kv_producer")
mock_vllm_config.kv_transfer_config = mock.MagicMock(
kv_role="kv_producer", is_kv_producer=True, is_kv_consumer=False
)
with mock.patch("vllm.config.get_current_vllm_config", return_value=mock_vllm_config), \
mock.patch("vllm_ascend.utils.enable_dsa_cp", return_value=True):
@@ -142,7 +145,9 @@ class TestUtils(TestBase):
def test_enable_dsa_cp_with_layer_shard_rejects_kv_both(self):
mock_vllm_config = mock.MagicMock()
mock_vllm_config.kv_transfer_config = mock.MagicMock(kv_role="kv_both", is_kv_producer=True)
mock_vllm_config.kv_transfer_config = mock.MagicMock(
kv_role="kv_both", is_kv_producer=True, is_kv_consumer=True
)
with mock.patch("vllm.config.get_current_vllm_config", return_value=mock_vllm_config), \
mock.patch("vllm_ascend.utils.enable_dsa_cp", return_value=True):
@@ -156,6 +161,42 @@ class TestUtils(TestBase):
mock.patch("vllm_ascend.utils.enable_dsa_cp", return_value=True):
self.assertFalse(utils.enable_dsa_cp_with_layer_shard())
def test_enable_dsa_cp_with_layer_shard_rejects_when_dsa_cp_disabled(self):
with mock.patch("vllm_ascend.utils.enable_dsa_cp", return_value=False):
self.assertFalse(utils.enable_dsa_cp_with_layer_shard())
def test_enable_dsa_cp_with_o_proj_tp_accepts_missing_kv_transfer(self):
mock_vllm_config = mock.MagicMock()
mock_vllm_config.kv_transfer_config = None
with mock.patch("vllm.config.get_current_vllm_config", return_value=mock_vllm_config), \
mock.patch("vllm_ascend.utils.enable_dsa_cp", return_value=True):
self.assertTrue(utils.enable_dsa_cp_with_o_proj_tp())
def test_enable_dsa_cp_with_o_proj_tp_accepts_kv_both(self):
mock_vllm_config = mock.MagicMock()
mock_vllm_config.kv_transfer_config = mock.MagicMock(
kv_role="kv_both", is_kv_producer=True, is_kv_consumer=True
)
with mock.patch("vllm.config.get_current_vllm_config", return_value=mock_vllm_config), \
mock.patch("vllm_ascend.utils.enable_dsa_cp", return_value=True):
self.assertTrue(utils.enable_dsa_cp_with_o_proj_tp())
def test_enable_dsa_cp_with_o_proj_tp_rejects_single_role_pd(self):
mock_vllm_config = mock.MagicMock()
mock_vllm_config.kv_transfer_config = mock.MagicMock(
kv_role="kv_producer", is_kv_producer=True, is_kv_consumer=False
)
with mock.patch("vllm.config.get_current_vllm_config", return_value=mock_vllm_config), \
mock.patch("vllm_ascend.utils.enable_dsa_cp", return_value=True):
self.assertFalse(utils.enable_dsa_cp_with_o_proj_tp())
def test_enable_dsa_cp_with_o_proj_tp_rejects_when_dsa_cp_disabled(self):
with mock.patch("vllm_ascend.utils.enable_dsa_cp", return_value=False):
self.assertFalse(utils.enable_dsa_cp_with_o_proj_tp())
def test_vllm_version_is(self):
with mock.patch.dict(os.environ, {"VLLM_VERSION": "1.0.0"}):
with mock.patch("vllm.__version__", "1.0.0"):