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

[Bugfix] bugfix for moe_mlp in vllm-ascend/v0.11.0-dev (#4885)

Browse Source 
### What this PR does / why we need it?
This PR fixes a bug in the moe_mlp module by correcting the arguments
passed to the torch_npu.npu_dequant_swiglu_quant function.It properly
converts group_list from a cumulative sum to counts for the group_index
parameter.

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


- vLLM version: v0.12.0
- vLLM main: https://github.com/vllm-project/vllm/main

---------

Signed-off-by: tanqingshan (A)  <50050625@china.huawei.com>
Signed-off-by: tanqingshan (A) <50050625@china.huawei.com>
Co-authored-by: tanqingshan (A) <50050625@china.huawei.com>
Co-authored-by: Mercykid-bash <ruanche0218@gmail.com>
This commit is contained in:
Clorist33
2025-12-12 14:51:47 +08:00
committed by GitHub
parent 9c0ad46c1a
commit 4f0dddc9ee
5 changed files with 41 additions and 34 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
tests/ut/eplb/core/test_eplb_device_transfer_loader.py
View File

@@ -47,8 +47,8 @@ def test_generate_task_and_state_flow(mock_adaptor):
loader_obj.state = loader.ExpertWeightUpdateState.WAITING
loader_obj.generate_expert_d2d_transfer_task([], [], {}, 0)
assert loader_obj.comm_op_list is None
assert loader_obj.state == loader.ExpertWeightUpdateState.WAITING
assert not loader_obj.comm_op_list
assert loader_obj.state == loader.ExpertWeightUpdateState.READY
def test_asyn_transfer_and_update(mock_adaptor):

12
tests/ut/ops/test_expert_load_balancer.py
View File

@@ -82,7 +82,7 @@ class TestExpertLoadBalancer(TestBase):
)
self.assertEqual(expert_placement_map.shape,
(self.expert_load_balancer.layers_num,
self.expert_load_balancer.ranks_num, 10))
self.expert_load_balancer.ranks_num, 8))
self.assertTrue(torch.all(expert_placement_map >= -1))
def test_generate_log2phy_expert_map(self):
@@ -90,7 +90,7 @@ class TestExpertLoadBalancer(TestBase):
log2phy_map = self.expert_load_balancer.generate_log2phy_expert_map(
layer_id)
self.assertEqual(log2phy_map.shape,
(self.expert_load_balancer.ranks_num, 10))
(self.expert_load_balancer.ranks_num, 8))
self.assertTrue(torch.all(log2phy_map >= -1))
@mock.patch("torch_npu.npu._lazy_init")
@@ -101,7 +101,7 @@ class TestExpertLoadBalancer(TestBase):
rank_local_expert_num, rank_expert_map = self.expert_load_balancer.get_rank_placement_map(
layer_id, rank_id)
self.assertEqual(rank_local_expert_num, 5)
expected_tensor = torch.tensor([2, -1, 1, 3, -1, 4, -1, 0, -1, -1],
expected_tensor = torch.tensor([2, -1, 1, 3, -1, 4, -1, 0],
dtype=torch.int32).to(
rank_expert_map.device)
self.assertTrue(rank_expert_map.equal(expected_tensor))
@@ -109,7 +109,7 @@ class TestExpertLoadBalancer(TestBase):
rank_id = 1
rank_local_expert_num, rank_expert_map = self.expert_load_balancer.get_rank_placement_map(
layer_id, rank_id)
expected_tensor = torch.tensor([-1, 1, 4, -1, 2, -1, 0, 3, -1, -1],
expected_tensor = torch.tensor([-1, 1, 4, -1, 2, -1, 0, 3],
dtype=torch.int32).to(
rank_expert_map.device)
self.assertTrue(rank_expert_map.equal(expected_tensor))
@@ -119,7 +119,7 @@ class TestExpertLoadBalancer(TestBase):
rank_id = 0
log2phy_map = self.expert_load_balancer.get_rank_log2phy_map(
layer_id, rank_id)
expected_tensor = torch.tensor([2, 6, 1, 3, 7, 4, 5, 0, -1, -1],
expected_tensor = torch.tensor([2, 6, 1, 3, 7, 4, 5, 0],
dtype=torch.int32).to(
log2phy_map.device)
self.assertTrue(log2phy_map.equal(expected_tensor))
@@ -127,7 +127,7 @@ class TestExpertLoadBalancer(TestBase):
rank_id = 1
log2phy_map = self.expert_load_balancer.get_rank_log2phy_map(
layer_id, rank_id)
expected_tensor = torch.tensor([2, 6, 9, 3, 7, 4, 5, 8, -1, -1],
expected_tensor = torch.tensor([2, 6, 9, 3, 7, 4, 5, 8],
dtype=torch.int32).to(
log2phy_map.device)
self.assertTrue(log2phy_map.equal(expected_tensor))

5
tests/ut/ops/test_fused_ops.py
View File

@@ -293,13 +293,13 @@ class TestCumsumGroupList(TestBase):
def test_cumsum_group_list_with_type_0(self):
group_list = self.experts.cumsum(dim=0)
group_list_type = 0
result = cumsum_group_list(group_list, group_list_type)
result = cumsum_group_list(group_list, group_list_type, 0)
self.assertTrue(torch.equal(result, self.group_list))
def test_cumsum_group_list_with_type_1(self):
group_list = self.experts
group_list_type = 1
result = cumsum_group_list(group_list, group_list_type)
result = cumsum_group_list(group_list, group_list_type, 0)
self.assertTrue(torch.equal(result, self.group_list))
def test_cumsum_group_list_with_type_2(self):
@@ -312,6 +312,7 @@ class TestCumsumGroupList(TestBase):
group_list_type = 2
result = cumsum_group_list(group_list,
group_list_type,
0,
active_num=self.active_num,
expert_num=self.expert_num)
self.assertTrue(torch.equal(result, self.group_list))

3
tests/ut/ops/test_token_dispatcher.py
View File

@@ -130,7 +130,7 @@ class TestTokenDispatcherWithMC2(TestBase):
self.dispatcher.need_extra_args = True
self.dispatcher.enable_dispatch_v2 = True
self.dispatcher.output = torch.randint(0, 8, (10, 1))
self.dispatcher.moe_expert_num = len(self.dispatcher.expert_map)
kwargs = self.dispatcher.get_combine_mc_kwargs(hidden_states)
self.assertIn("tp_send_counts", kwargs)
@@ -148,6 +148,7 @@ class TestTokenDispatcherWithMC2(TestBase):
self.dispatcher.enable_dispatch_v2 = True
self.dispatcher.swiglu_out_scale = torch.randint(0, 8, (10, 1))
self.dispatcher.output = torch.randint(0, 8, (10, 1))
self.dispatcher.moe_expert_num = len(self.dispatcher.expert_map)
self.hidden_states = torch.randn(10, 128)
with patch("torch_npu.npu_moe_distribute_combine_v2",