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

[BugFix]Backport validate pd mode feature gates no fused mc2 v0.18.0 clean (#8583)

Browse Source 
### What this PR does / why we need it?
Backport validate pd mode feature gates no fused mc2 v0.18.0 clean
backport #8582
---------
Signed-off-by: wangxiaoteng <wangxiaoteng@huawei.com>
This commit is contained in:
wangxiaoteng888
2026-04-23 19:44:07 +08:00
committed by GitHub
parent 1fc7bc056d
commit c47371c1af
3 changed files with 0 additions and 141 deletions
Download Patch File Download Diff File Expand all files Collapse all files

128
tests/ut/test_platform.py
View File

@@ -541,134 +541,6 @@ class TestNPUPlatform(TestBase):
):
self.platform.check_and_update_config(vllm_config)
@patch("vllm_ascend.quantization.utils.maybe_auto_detect_quantization")
@patch("vllm_ascend.utils.get_ascend_device_type", return_value=AscendDeviceType.A3)
@patch("vllm_ascend.ascend_config.init_ascend_config")
@patch("vllm_ascend.core.recompute_scheduler.RecomputeSchedulerConfig.initialize_from_config")
def test_check_and_update_config_fused_mc2_rejects_pd_mixed_no_kv_transfer(
self, mock_init_recompute, mock_init_ascend, mock_soc_version, mock_auto_detect
):
mock_ascend_config = TestNPUPlatform.mock_vllm_ascend_config()
mock_ascend_config.recompute_scheduler_enable = False
mock_ascend_config.enable_mc2_hierarchy_comm = False
mock_init_ascend.return_value = mock_ascend_config
vllm_config = TestNPUPlatform.mock_vllm_config()
vllm_config.kv_transfer_config = None
vllm_config.parallel_config.decode_context_parallel_size = 1
vllm_config.parallel_config.prefill_context_parallel_size = 1
vllm_config.parallel_config.tensor_parallel_size = 1
vllm_config.scheduler_config = MagicMock()
mock_init_recompute.return_value = MagicMock()
from vllm_ascend import platform
importlib.reload(platform)
self.platform = platform.NPUPlatform()
with patch("vllm_ascend.platform.envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2", 1, create=True):
with pytest.raises(ValueError, match=r"VLLM_ASCEND_ENABLE_FUSED_MC2.*kv_role='kv_consumer'.*PD-mixed"):
with patch.object(platform.NPUPlatform, "_fix_incompatible_config"):
self.platform.check_and_update_config(vllm_config)
@patch("vllm_ascend.quantization.utils.maybe_auto_detect_quantization")
@patch("vllm_ascend.utils.get_ascend_device_type", return_value=AscendDeviceType.A3)
@patch("vllm_ascend.ascend_config.init_ascend_config")
@patch("vllm_ascend.core.recompute_scheduler.RecomputeSchedulerConfig.initialize_from_config")
def test_check_and_update_config_fused_mc2_rejects_pd_mixed_kv_both(
self, mock_init_recompute, mock_init_ascend, mock_soc_version, mock_auto_detect
):
mock_ascend_config = TestNPUPlatform.mock_vllm_ascend_config()
mock_ascend_config.recompute_scheduler_enable = False
mock_ascend_config.enable_mc2_hierarchy_comm = False
mock_init_ascend.return_value = mock_ascend_config
vllm_config = TestNPUPlatform.mock_vllm_config()
vllm_config.kv_transfer_config = MagicMock(kv_role="kv_both", engine_id="engine0")
vllm_config.parallel_config.decode_context_parallel_size = 1
vllm_config.parallel_config.prefill_context_parallel_size = 1
vllm_config.parallel_config.tensor_parallel_size = 1
vllm_config.scheduler_config = MagicMock()
mock_init_recompute.return_value = MagicMock()
from vllm_ascend import platform
importlib.reload(platform)
self.platform = platform.NPUPlatform()
with patch("vllm_ascend.platform.envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2", 1, create=True):
with pytest.raises(ValueError, match=r"VLLM_ASCEND_ENABLE_FUSED_MC2.*kv_role='kv_consumer'.*kv_role='kv_both'"):
with (
patch.object(platform.NPUPlatform, "_fix_incompatible_config"),
patch.object(platform, "check_kv_extra_config"),
):
self.platform.check_and_update_config(vllm_config)
@patch("vllm_ascend.quantization.utils.maybe_auto_detect_quantization")
@patch("vllm_ascend.utils.get_ascend_device_type", return_value=AscendDeviceType.A3)
@patch("vllm_ascend.ascend_config.init_ascend_config")
@patch("vllm_ascend.core.recompute_scheduler.RecomputeSchedulerConfig.initialize_from_config")
def test_check_and_update_config_fused_mc2_rejects_pd_disaggregated_kv_producer(
self, mock_init_recompute, mock_init_ascend, mock_soc_version, mock_auto_detect
):
mock_ascend_config = TestNPUPlatform.mock_vllm_ascend_config()
mock_ascend_config.recompute_scheduler_enable = False
mock_ascend_config.enable_mc2_hierarchy_comm = False
mock_init_ascend.return_value = mock_ascend_config
vllm_config = TestNPUPlatform.mock_vllm_config()
vllm_config.kv_transfer_config = MagicMock(kv_role="kv_producer", engine_id="engine0")
vllm_config.parallel_config.decode_context_parallel_size = 1
vllm_config.parallel_config.prefill_context_parallel_size = 1
vllm_config.parallel_config.tensor_parallel_size = 1
vllm_config.scheduler_config = MagicMock()
mock_init_recompute.return_value = MagicMock()
from vllm_ascend import platform
importlib.reload(platform)
self.platform = platform.NPUPlatform()
with patch("vllm_ascend.platform.envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2", 1, create=True):
with pytest.raises(ValueError, match=r"VLLM_ASCEND_ENABLE_FUSED_MC2.*kv_role='kv_consumer'.*kv_role='kv_producer'"):
with (
patch.object(platform.NPUPlatform, "_fix_incompatible_config"),
patch.object(platform, "check_kv_extra_config"),
):
self.platform.check_and_update_config(vllm_config)
@patch("vllm_ascend.quantization.utils.maybe_auto_detect_quantization")
@patch("vllm_ascend.utils.get_ascend_device_type", return_value=AscendDeviceType.A3)
@patch("vllm_ascend.ascend_config.init_ascend_config")
@patch("vllm_ascend.core.recompute_scheduler.RecomputeSchedulerConfig.initialize_from_config")
def test_check_and_update_config_fused_mc2_allows_pd_disaggregated_kv_consumer(
self, mock_init_recompute, mock_init_ascend, mock_soc_version, mock_auto_detect
):
mock_ascend_config = TestNPUPlatform.mock_vllm_ascend_config()
mock_ascend_config.recompute_scheduler_enable = False
mock_ascend_config.enable_mc2_hierarchy_comm = False
mock_init_ascend.return_value = mock_ascend_config
vllm_config = TestNPUPlatform.mock_vllm_config()
vllm_config.kv_transfer_config = MagicMock(kv_role="kv_consumer", engine_id="engine0")
vllm_config.parallel_config.decode_context_parallel_size = 1
vllm_config.parallel_config.prefill_context_parallel_size = 1
vllm_config.parallel_config.tensor_parallel_size = 1
vllm_config.scheduler_config = MagicMock()
mock_init_recompute.return_value = MagicMock()
from vllm_ascend import platform
importlib.reload(platform)
self.platform = platform.NPUPlatform()
with patch("vllm_ascend.platform.envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2", 1, create=True):
with (
patch.object(platform.NPUPlatform, "_fix_incompatible_config"),
patch.object(platform, "check_kv_extra_config"),
):
self.platform.check_and_update_config(vllm_config)
def test_update_block_size_for_backend_preserves_hybrid_block_size(self):
vllm_config = TestNPUPlatform.mock_vllm_config()
vllm_config.model_config.is_hybrid = True

2
vllm_ascend/envs.py
View File

@@ -94,8 +94,6 @@ env_variables: dict[str, Callable[[], Any]] = {
# Whether to anbale dynamic EPLB
"DYNAMIC_EPLB": lambda: os.getenv("DYNAMIC_EPLB", "false").lower(),
# Whether to enable fused MC2 (`dispatch_gmm_combine_decode` / `dispatch_ffn_combine`).
# Platform validation: only PD-disaggregated **decode** instances (`kv_role='kv_consumer'`).
# Not supported in PD-mixed mode (`kv_both` or no kv_transfer_config) or on prefill nodes (`kv_producer`).
# 0, or not set: default ALLTOALL and MC2 will be used.
# 1: ALLTOALL and MC2 might be replaced by `dispatch_ffn_combine` operator.
# `dispatch_ffn_combine` can be used only for moe layer with W8A8, EP<=32, non-mtp, non-dynamic-eplb.

11
vllm_ascend/platform.py
View File

@@ -448,17 +448,6 @@ class NPUPlatform(Platform):
if get_ascend_device_type() != AscendDeviceType._310P:
compilation_config.custom_ops = ["all"]
if envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2:
kv_transfer_config = vllm_config.kv_transfer_config
kv_role = getattr(kv_transfer_config, "kv_role", None)
if kv_transfer_config is None or kv_role != "kv_consumer":
raise ValueError(
"VLLM_ASCEND_ENABLE_FUSED_MC2 (fused mc2) only supports PD-disaggregated "
"decode nodes (D-side) with kv_role='kv_consumer'. It is not supported "
"in PD-mixed mode (no kv_transfer_config / kv_role='kv_both') nor on "
"prefill nodes (P-side) with kv_role='kv_producer'."
)
if envs_ascend.VLLM_ASCEND_BALANCE_SCHEDULING:
kv_transfer_config = vllm_config.kv_transfer_config
kv_role = getattr(kv_transfer_config, "kv_role", None)