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

[main] [refactor] refactor common_fused_moe.py (#2706)

Browse Source 
### What this PR does / why we need it?
1. Move prepare/finalize operation from moe_comm_method to
/ops/moe/fused_moe_prepare_and_finalize
2. Adapt to token_dispatcher in moe_comm_method
3. Move
moe_comm_method/experts_selector/token_dispatcher/fused_moe_prepare_and_finalize
to /ops/moe
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
e2e & ut

- vLLM version: v0.10.1.1
- vLLM main:
f4962a6d55

Signed-off-by: weichen <calvin_zhu0210@outlook.com>
Signed-off-by: Pr0Wh1teGivee <calvin_zhu0210@outlook.com>
Co-authored-by: weijinqian0 <12153182+weijinqian0@users.noreply.github.com>
This commit is contained in:
weichen
2025-09-08 20:09:50 +08:00
committed by GitHub
parent 1a82b16355
commit a041d4f328
21 changed files with 1052 additions and 932 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
tests/e2e/singlecard/ops/test_fused_moe.py
View File

@@ -28,9 +28,8 @@ import torch
import torch_npu
from vllm.model_executor.layers.activation import SiluAndMul
from vllm_ascend.ops.layers.experts_selector import select_experts
from vllm_ascend.ops.moe_dispatcher.token_dispatcher import \
TokenDispatcherWithAllGather
from vllm_ascend.ops.moe.experts_selector import select_experts
from vllm_ascend.ops.moe.token_dispatcher import TokenDispatcherWithAllGather
NUM_EXPERTS = [8, 64]
EP_SIZE = [1]
@@ -209,7 +208,7 @@ def test_select_experts(
dtype=torch.int32)
custom_routing_function.return_value = (mock_weights, mock_ids)
with patch("vllm_ascend.ops.layers.experts_selector._native_grouped_topk"
with patch("vllm_ascend.ops.moe.experts_selector._native_grouped_topk"
) as mock_native_grouped_topk:
mock_native_grouped_topk.side_effect = lambda x, num_groups, k: torch.randn_like(
x)

175
tests/e2e/singlecard/ops/test_moe_comm.py
View File

@@ -1,175 +0,0 @@
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
# Copyright 2023 The vLLM team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# This file is a part of the vllm-ascend project.
import gc
from types import SimpleNamespace
import pytest
import torch
from vllm.model_executor.layers.fused_moe.config import ( # isort: skip
FusedMoEConfig, FusedMoEParallelConfig)
from vllm_ascend.distributed.moe_comm_method import ( # isort: skip
AllGatherCommImpl, NativeAllGatherCommImpl)
@pytest.mark.parametrize("num_tokens", [16, 128])
@pytest.mark.parametrize("hidden_size", [64, 128])
@pytest.mark.parametrize("global_num_experts", [8, 16])
@pytest.mark.parametrize("num_local_experts", [4, 8])
@pytest.mark.parametrize("top_k_num", [2, 4])
@pytest.mark.parametrize("dtype", [torch.bfloat16, torch.float16])
@pytest.mark.parametrize("ep_rank", [0, 1])
@pytest.mark.parametrize("apply_a8_quantization", [False])
def test_all_gather_comm_impl(
num_tokens,
hidden_size,
global_num_experts,
num_local_experts,
top_k_num,
dtype,
ep_rank,
apply_a8_quantization,
mocker,
):
"""
Tests the AllGatherCommImpl against the NativeAllGatherCommImpl.
This test compares the outputs of the NPU-optimized AllGatherCommImpl
with a native PyTorch implementation (NativeAllGatherCommImpl) to ensure
correctness across various configurations.
"""
if top_k_num > global_num_experts:
pytest.skip("top_k_num cannot be greater than global_num_experts")
if num_local_experts > global_num_experts:
pytest.skip(
"num_local_experts cannot be greater than global_num_experts")
device = torch.device("npu")
# mock get_tensor_model_parallel_rank to return ep_rank
mocker.patch(
"vllm.model_executor.layers.fused_moe.config.get_tensor_model_parallel_rank",
return_value=ep_rank,
)
# make moe config
parallel_config = SimpleNamespace(
enable_expert_parallel=num_local_experts < global_num_experts)
moe_parallel_config: FusedMoEParallelConfig = FusedMoEParallelConfig.make(
tp_size_=max(2, global_num_experts // num_local_experts),
dp_size_=1,
vllm_parallel_config=parallel_config,
)
moe_config = FusedMoEConfig(
num_experts=global_num_experts,
experts_per_token=top_k_num,
hidden_dim=hidden_size,
num_local_experts=num_local_experts,
moe_parallel_config=moe_parallel_config,
in_dtype=dtype,
quant_config=None, # No quantization in this test
max_num_tokens=num_tokens,
)
# Instantiate implementations
native_impl = NativeAllGatherCommImpl(moe_config)
all_gather_impl = AllGatherCommImpl(moe_config)
# --- Input Data ---
hidden_states = torch.randn(num_tokens,
hidden_size,
device=device,
dtype=dtype)
topk_ids = torch.randint(0,
global_num_experts, (num_tokens, top_k_num),
device=device,
dtype=torch.int32)
topk_weights = torch.rand(num_tokens, top_k_num, device=device).to(dtype)
topk_weights = torch.nn.functional.softmax(topk_weights, dim=1)
num_experts = global_num_experts
expert_map = None
if num_local_experts < global_num_experts:
# Create a map where some experts are local and some are not
expert_map = torch.full((global_num_experts, ), -1, device=device)
expert_map[ep_rank * num_local_experts:(ep_rank + 1) *
num_local_experts] = torch.arange(num_local_experts,
device=device)
num_experts = num_local_experts
# --- Run Native Implementation (Golden Reference) ---
native_hidden_states_out = hidden_states.clone()
(
native_permuted_hidden,
native_expert_tokens,
_,
_,
) = native_impl.permute(hidden_states, topk_ids, topk_weights, expert_map,
num_experts, apply_a8_quantization)
# Simulate MLP output
native_mlp_output = torch.randn_like(native_permuted_hidden)
native_impl.unpermute(native_mlp_output, native_hidden_states_out)
# --- Run AllGather Implementation ---
all_gather_hidden_states_out = hidden_states.clone()
(
all_gather_permuted_hidden,
all_gather_expert_tokens,
_,
_,
) = all_gather_impl.permute(hidden_states, topk_ids, topk_weights,
expert_map, num_experts, apply_a8_quantization)
# Use the same simulated MLP output for a fair comparison
all_gather_mlp_output = native_mlp_output.clone()
all_gather_impl.unpermute(all_gather_mlp_output,
all_gather_hidden_states_out)
# --- Assertions ---
# Define tolerance based on dtype
atol = 1e-3 if dtype == torch.float16 else 1e-2
rtol = 1e-3 if dtype == torch.float16 else 1e-2
# 1. Compare expert_tokens from pre_process
assert torch.allclose(native_expert_tokens.to(
all_gather_expert_tokens.device),
all_gather_expert_tokens,
atol=atol,
rtol=rtol), "Expert tokens do not match."
# 2. Compare permuted_hidden_states from pre_process
num_valid_tokens = native_expert_tokens.sum()
assert torch.allclose(native_permuted_hidden[:num_valid_tokens].to(
all_gather_permuted_hidden.device),
all_gather_permuted_hidden[:num_valid_tokens],
atol=atol,
rtol=rtol), "Permuted hidden states do not match."
# 3. Compare final hidden_states from post_process
assert torch.allclose(native_hidden_states_out.to(
all_gather_hidden_states_out.device),
all_gather_hidden_states_out,
atol=atol,
rtol=rtol), "Final hidden states do not match."
gc.collect()
torch.npu.empty_cache()
torch.npu.reset_peak_memory_stats()

218
tests/ut/ops/test_fused_moe_prepare_and_finalize.py Normal file
View File

@@ -0,0 +1,218 @@
import unittest
from unittest.mock import MagicMock, patch
import torch
from vllm.model_executor.layers.fused_moe import FusedMoEConfig
from vllm_ascend.ops.moe.fused_moe_prepare_and_finalize import (
FusedMoEPrepareAndFinalizeWithAll2All,
FusedMoEPrepareAndFinalizeWithAllGather, FusedMoEPrepareAndFinalizeWithMC2)
class TestFusedMoEPrepareAndFinalize(unittest.TestCase):
def setUp(self):
# Mock FusedMoEConfig
self.moe_config = MagicMock(spec=FusedMoEConfig)
self.moe_config.tp_group = MagicMock()
self.moe_config.tp_group.device_group = MagicMock()
self.moe_config.dp_size = 1
self.moe_config.tp_size = 1
self.moe_config.ep_size = 1
self.moe_config.dp_group = MagicMock()
@patch(
"vllm_ascend.ops.moe.fused_moe_prepare_and_finalize.get_tensor_model_parallel_world_size",
return_value=1)
@patch(
"vllm_ascend.ops.moe.fused_moe_prepare_and_finalize.get_tensor_model_parallel_rank",
return_value=0)
@patch(
"vllm_ascend.ops.moe.fused_moe_prepare_and_finalize.get_forward_context"
)
def test_mc2_prepare_finalize(self, mock_get_forward_context, mock_tp_rank,
mock_tp_size):
mock_context = MagicMock()
mock_context.mc2_mask = torch.tensor([1, 0, 1])
mock_context.padded_num_tokens = 4
mock_get_forward_context.return_value = mock_context
layer = FusedMoEPrepareAndFinalizeWithMC2(self.moe_config)
hidden_states = torch.randn(3, 8)
router_logits = torch.randn(3, 2)
h_out, r_out, mask = layer.prepare(hidden_states, router_logits)
# Check padding and split
self.assertEqual(h_out.shape[0], 4)
self.assertEqual(r_out.shape[0], 4)
self.assertEqual(mask.tolist(), [1, 0, 1])
# Finalize
result = layer.finalize(h_out, reduce_results=False)
self.assertEqual(result.shape[0], 3)
@patch(
"vllm_ascend.ops.moe.fused_moe_prepare_and_finalize.get_tensor_model_parallel_world_size",
return_value=2)
@patch(
"vllm_ascend.ops.moe.fused_moe_prepare_and_finalize.get_tensor_model_parallel_rank",
return_value=0)
@patch(
"vllm_ascend.ops.moe.fused_moe_prepare_and_finalize.get_forward_context"
)
@patch("torch.distributed.all_gather")
def test_mc2_tp_split_allgather(self, mock_all_gather,
mock_get_forward_context, mock_tp_rank,
mock_tp_size):
mock_context = MagicMock()
mock_context.mc2_mask = torch.tensor([1, 0, 1, 0])
mock_context.padded_num_tokens = 4
mock_get_forward_context.return_value = mock_context
layer = FusedMoEPrepareAndFinalizeWithMC2(self.moe_config)
hidden_states = torch.randn(4, 8)
router_logits = torch.randn(4, 2)
h_out, r_out, mask = layer.prepare(hidden_states,
router_logits,
enable_shared_expert_dp=False,
replace_allreduce=False)
# With TP=2, should split into 2 parts
self.assertEqual(h_out.shape[0], 2)
# Mock all_gather behavior
def mock_all_gather_func(tensor_list, tensor, group=None):
tensor_list[0] = tensor
tensor_list[1] = tensor.clone()
mock_all_gather.side_effect = mock_all_gather_func
layer.split_hidden_states = [
torch.zeros_like(h_out),
torch.zeros_like(h_out)
]
final_result = layer.finalize(h_out, reduce_results=False)
# Should concat back to original size
self.assertEqual(final_result.shape[0], 4)
@patch(
"vllm_ascend.ops.moe.fused_moe_prepare_and_finalize.get_tensor_model_parallel_world_size",
return_value=1)
@patch(
"vllm_ascend.ops.moe.fused_moe_prepare_and_finalize.get_tensor_model_parallel_rank",
return_value=0)
def test_all2all_prepare_finalize(self, mock_tp_rank, mock_tp_size):
layer = FusedMoEPrepareAndFinalizeWithAll2All(self.moe_config)
hidden_states = torch.randn(3, 8)
router_logits = torch.randn(3, 2)
h_out, r_out, _ = layer.prepare(hidden_states, router_logits)
# Pad to tp_size=1, so no change
self.assertEqual(h_out.shape[0], 3)
result = layer.finalize(h_out, reduce_results=False)
self.assertEqual(result.shape[0], 3)
@patch(
"vllm_ascend.ops.moe.fused_moe_prepare_and_finalize.get_tensor_model_parallel_world_size",
return_value=2)
@patch(
"vllm_ascend.ops.moe.fused_moe_prepare_and_finalize.get_tensor_model_parallel_rank",
return_value=0)
@patch("torch.distributed.all_gather")
def test_all2all_tp_split_allgather(self, mock_all_gather, mock_tp_rank,
mock_tp_size):
layer = FusedMoEPrepareAndFinalizeWithAll2All(self.moe_config)
hidden_states = torch.randn(2, 8)
router_logits = torch.randn(2, 2)
h_out, r_out, _ = layer.prepare(hidden_states,
router_logits,
enable_shared_expert_dp=False,
replace_allreduce=False)
# Split due to TP=2
self.assertEqual(h_out.shape[0], 1)
# Mock all_gather
def mock_all_gather_func(tensor_list, tensor, group=None):
tensor_list[0] = tensor
tensor_list[1] = tensor.clone()
mock_all_gather.side_effect = mock_all_gather_func
layer.split_hidden_states = [
torch.zeros_like(h_out),
torch.zeros_like(h_out)
]
final_result = layer.finalize(h_out, reduce_results=False)
# Should concat back
self.assertEqual(final_result.shape[0], 2)
@patch("vllm_ascend.ops.moe.fused_moe_prepare_and_finalize.get_dp_group")
@patch(
"vllm_ascend.ops.moe.fused_moe_prepare_and_finalize.tensor_model_parallel_all_reduce"
)
@patch(
"vllm_ascend.ops.moe.fused_moe_prepare_and_finalize.get_forward_context"
)
def test_allgather_prepare_finalize(self, mock_get_forward_context,
mock_tp_all_reduce, mock_get_dp_group):
# Mock forward context
mock_context = MagicMock()
mock_context.max_tokens_across_dp = 6
mock_get_forward_context.return_value = mock_context
# Create a proper mock for DP group with working all_gather
mock_dp_group = MagicMock()
def mock_all_gather_func(tensor, dim):
# Simulate DP=2: repeat the tensor along the specified dimension
return torch.cat([tensor, tensor], dim=dim)
mock_dp_group.all_gather = mock_all_gather_func
mock_get_dp_group.return_value = mock_dp_group
self.moe_config.dp_size = 2
self.moe_config.tp_size = 1
self.moe_config.ep_size = 1
self.moe_config.dp_group = mock_dp_group
layer = FusedMoEPrepareAndFinalizeWithAllGather(self.moe_config)
hidden_states = torch.randn(3, 8)
router_logits = torch.randn(3, 2)
# Mock the gate function for rm_router_logits=False case
mock_gate = MagicMock()
mock_gate.return_value = (router_logits.repeat(2, 1), None)
h_out, r_out, _ = layer.prepare(hidden_states,
router_logits,
rm_router_logits=False,
gate=mock_gate)
# After all-gather with DP=2, should double the batch size
self.assertEqual(h_out.shape[0], 12)
self.assertEqual(r_out.shape[0], 12)
# Finalize with reduce_scatter
def mock_reduce_scatter_func(tensor, dim):
# Simulate reduce_scatter: take first half
return tensor[:3]
mock_dp_group.reduce_scatter = mock_reduce_scatter_func
result = layer.finalize(h_out, reduce_results=False)
self.assertEqual(result.shape[0], 3)
# Test with TP all-reduce
mock_tp_all_reduce.return_value = result
result_with_tp = layer.finalize(h_out, reduce_results=True)
self.assertEqual(result_with_tp.shape[0], 3)

24
tests/ut/ops/test_fused_ops.py
View File

@@ -22,14 +22,14 @@ import torch_npu
from pytest_mock import MockerFixture
from vllm.model_executor.layers.fused_moe import FusedMoEMethodBase
import vllm_ascend.ops.moe_dispatcher.token_dispatcher as token_dispatcher_module
import vllm_ascend.ops.moe.token_dispatcher as token_dispatcher_module
from tests.ut.base import TestBase
from vllm_ascend.ascend_forward_context import (FusedMoEState,
_get_fused_moe_state)
from vllm_ascend.ops.fused_moe import (AscendFusedMoE,
AscendUnquantizedFusedMoEMethod)
from vllm_ascend.ops.layers.experts_selector import select_experts
from vllm_ascend.ops.layers.moe_mlp import cumsum_group_list, unified_apply_mlp
from vllm_ascend.ops.moe.experts_selector import select_experts
from vllm_ascend.ops.moe.moe_mlp import cumsum_group_list, unified_apply_mlp
from vllm_ascend.utils import AscendSocVersion, adapt_patch
adapt_patch(True)
@@ -110,11 +110,11 @@ def mock_dist_env(mocker: MockerFixture):
captured_dispatchers[key] = mock_token_dispatcher_with_mc2
mock_register_token_dispatcher_patcher = patch(
'vllm_ascend.ops.moe_dispatcher.token_dispatcher._register_token_dispatcher',
'vllm_ascend.ops.moe.token_dispatcher._register_token_dispatcher',
side_effect=capture_register)
mock_get_token_dispatcher_patcher = patch(
'vllm_ascend.ops.moe_dispatcher.token_dispatcher.get_token_dispatcher',
'vllm_ascend.ops.moe.token_dispatcher.get_token_dispatcher',
side_effect=lambda name: captured_dispatchers.get(name))
default_mock_token_dispatcher = mock_token_dispatcher_with_allgather
@@ -158,7 +158,7 @@ def mock_dist_env(mocker: MockerFixture):
)), \
patch("vllm_ascend.utils.get_ascend_soc_version", return_value=AscendSocVersion.A3), \
patch.object(token_dispatcher_module, 'setup_token_dispatchers', mock_setup_token_dispatchers), \
patch('vllm_ascend.ops.layers.moe_mlp.get_forward_context',
patch('vllm_ascend.ops.moe.moe_mlp.get_forward_context',
return_value=mock_forward_context_obj):
yield {
@@ -562,8 +562,8 @@ class TestCumsumGroupList(TestBase):
class TestUnifiedApplyMLP(TestBase):
@patch('vllm_ascend.ops.layers.moe_mlp.get_forward_context')
@patch('vllm_ascend.ops.layers.moe_mlp.is_310p')
@patch('vllm_ascend.ops.moe.moe_mlp.get_forward_context')
@patch('vllm_ascend.ops.moe.moe_mlp.is_310p')
@patch('torch_npu.npu_grouped_matmul')
@patch('torch_npu.npu_dynamic_quant')
@patch('torch_npu.npu_dequant_swiglu_quant')
@@ -629,7 +629,7 @@ class TestUnifiedApplyMLP(TestBase):
self.assertEqual(result.dtype, torch.bfloat16)
@patch('vllm_ascend.ops.layers.moe_mlp.is_310p')
@patch('vllm_ascend.ops.moe.moe_mlp.is_310p')
@patch('torch_npu.npu_grouped_matmul')
@patch('torch_npu.npu_swiglu')
@patch('torch_npu.npu_dynamic_quant')
@@ -671,7 +671,7 @@ class TestUnifiedApplyMLP(TestBase):
self.assertEqual(result.shape, hidden_states.shape)
self.assertEqual(result.dtype, torch.float16)
@patch('vllm_ascend.ops.layers.moe_mlp.get_forward_context')
@patch('vllm_ascend.ops.moe.moe_mlp.get_forward_context')
@patch('torch_npu.npu_grouped_matmul')
@patch('torch_npu.npu_swiglu')
@patch('torch_npu.npu_dynamic_quant')
@@ -731,7 +731,7 @@ class TestUnifiedApplyMLP(TestBase):
self.assertEqual(result.shape, hidden_states.shape)
self.assertEqual(result.dtype, torch.bfloat16)
@patch('vllm_ascend.ops.layers.moe_mlp.is_310p')
@patch('vllm_ascend.ops.moe.moe_mlp.is_310p')
@patch('torch_npu.npu_grouped_matmul')
@patch('torch_npu.npu_swiglu')
@patch('torch_npu.npu_dynamic_quant')
@@ -776,7 +776,7 @@ class TestUnifiedApplyMLP(TestBase):
self.assertEqual(result.shape, hidden_states.shape)
self.assertEqual(result.dtype, torch.float16)
@patch("vllm_ascend.ops.layers.moe_mlp.get_forward_context")
@patch("vllm_ascend.ops.moe.moe_mlp.get_forward_context")
@patch("torch_npu.npu_grouped_matmul")
@patch("torch_npu.npu_swiglu")
@patch("torch_npu.npu_grouped_matmul_swiglu_quant")

212
tests/ut/ops/test_moe_comm_method.py Normal file
View File

@@ -0,0 +1,212 @@
from unittest.mock import MagicMock, patch
import torch
from vllm.model_executor.layers.fused_moe import FusedMoEConfig
from tests.ut.base import TestBase
from vllm_ascend.ops.moe.moe_comm_method import (AllGatherCommImpl,
AlltoAllCommImpl, MC2CommImpl)
class TestMoECommMethod(TestBase):
def setUp(self):
# Mock FusedMoEConfig
self.moe_config = MagicMock(spec=FusedMoEConfig)
self.moe_config.num_experts = 8
self.moe_config.num_local_experts = 2
self.moe_config.experts_per_token = 2
self.moe_config.tp_group = MagicMock()
self.moe_config.tp_group.device_group = MagicMock()
self.moe_config.dp_size = 1
self.moe_config.tp_size = 1
self.moe_config.ep_size = 1
self.moe_config.dp_group = MagicMock()
self.moe_config.num_global_redundant_experts = 0
@patch("vllm_ascend.ops.moe.moe_comm_method.get_forward_context")
@patch(
"vllm_ascend.ops.moe.moe_comm_method.FusedMoEPrepareAndFinalizeWithAllGather"
)
@patch("vllm_ascend.ops.moe.moe_comm_method.TokenDispatcherWithAllGather")
def test_all_gather_comm_impl(self, mock_token_dispatcher,
mock_prepare_finalize,
mock_get_forward_context):
# Mock forward context
mock_context = MagicMock()
mock_context.moe_comm_method = "all_gather"
mock_get_forward_context.return_value = mock_context
# Mock prepare finalize
mock_pf_instance = MagicMock()
mock_pf_instance.prepare.return_value = (torch.randn(4, 8),
torch.randn(4, 2), None)
mock_pf_instance.finalize.return_value = torch.randn(4, 8)
mock_prepare_finalize.return_value = mock_pf_instance
# Mock token dispatcher
mock_td_instance = MagicMock()
mock_token_dispatcher.return_value = mock_td_instance
# Create instance
comm_impl = AllGatherCommImpl(self.moe_config)
# Test prepare method
hidden_states = torch.randn(3, 8)
router_logits = torch.randn(3, 2)
h_out, r_out = comm_impl.prepare(hidden_states, router_logits)
# Verify prepare was called with correct arguments
mock_pf_instance.prepare.assert_called_once_with(
hidden_states, router_logits, False, False, False, None)
# Test finalize method
comm_impl.finalize(h_out, reduce_results=True)
mock_pf_instance.finalize.assert_called_once_with(h_out, True)
@patch("vllm_ascend.ops.moe.moe_comm_method.get_forward_context")
@patch(
"vllm_ascend.ops.moe.moe_comm_method.FusedMoEPrepareAndFinalizeWithMC2"
)
@patch("vllm_ascend.ops.moe.moe_comm_method.TokenDispatcherWithMC2")
def test_mc2_comm_impl(self, mock_token_dispatcher, mock_prepare_finalize,
mock_get_forward_context):
# Mock forward context
mock_context = MagicMock()
mock_context.moe_comm_method = "mc2"
mock_get_forward_context.return_value = mock_context
# Mock prepare finalize
mock_pf_instance = MagicMock()
mock_pf_instance.prepare.return_value = (torch.randn(4, 8),
torch.randn(4, 2),
torch.tensor([1, 0, 1, 0]))
mock_pf_instance.finalize.return_value = torch.randn(4, 8)
mock_prepare_finalize.return_value = mock_pf_instance
# Mock token dispatcher
mock_td_instance = MagicMock()
mock_token_dispatcher.return_value = mock_td_instance
# Create instance
comm_impl = MC2CommImpl(self.moe_config)
# Test prepare method
hidden_states = torch.randn(3, 8)
router_logits = torch.randn(3, 2)
h_out, r_out = comm_impl.prepare(hidden_states, router_logits)
# Verify prepare was called with correct arguments
mock_pf_instance.prepare.assert_called_once_with(
hidden_states, router_logits, False, False, False, None)
# Test finalize method
comm_impl.finalize(h_out, reduce_results=True)
mock_pf_instance.finalize.assert_called_once_with(h_out, True)
@patch("vllm_ascend.ops.moe.moe_comm_method.get_forward_context")
@patch(
"vllm_ascend.ops.moe.moe_comm_method.FusedMoEPrepareAndFinalizeWithAll2All"
)
@patch("vllm_ascend.ops.moe.moe_comm_method.TokenDispatcherWithAll2AllV")
def test_alltoall_comm_impl(self, mock_token_dispatcher,
mock_prepare_finalize,
mock_get_forward_context):
# Mock forward context
mock_context = MagicMock()
mock_context.moe_comm_method = "alltoall"
mock_get_forward_context.return_value = mock_context
# Mock prepare finalize
mock_pf_instance = MagicMock()
mock_pf_instance.prepare.return_value = (torch.randn(4, 8),
torch.randn(4, 2), None)
mock_pf_instance.finalize.return_value = torch.randn(4, 8)
mock_prepare_finalize.return_value = mock_pf_instance
# Mock token dispatcher
mock_td_instance = MagicMock()
mock_token_dispatcher.return_value = mock_td_instance
# Create instance
comm_impl = AlltoAllCommImpl(self.moe_config)
# Test prepare method
hidden_states = torch.randn(3, 8)
router_logits = torch.randn(3, 2)
h_out, r_out = comm_impl.prepare(hidden_states, router_logits)
# Verify prepare was called with correct arguments
mock_pf_instance.prepare.assert_called_once_with(
hidden_states, router_logits, False, False, False, None)
@patch("vllm_ascend.ops.moe.moe_comm_method.get_forward_context")
@patch(
"vllm_ascend.ops.moe.moe_comm_method.FusedMoEPrepareAndFinalizeWithAllGather"
)
@patch("vllm_ascend.ops.moe.moe_comm_method.TokenDispatcherWithAllGather")
@patch("vllm_ascend.ops.moe.moe_comm_method.unified_apply_mlp")
def test_fused_experts_method(self, mock_unified_apply_mlp,
mock_token_dispatcher, mock_prepare_finalize,
mock_get_forward_context):
# Mock forward context
mock_context = MagicMock()
mock_context.moe_comm_method = "all_gather"
mock_get_forward_context.return_value = mock_context
# Mock prepare finalize
mock_pf_instance = MagicMock()
mock_pf_instance.prepare.return_value = (torch.randn(4, 8),
torch.randn(4, 2), None)
mock_pf_instance.finalize.return_value = torch.randn(4, 8)
mock_prepare_finalize.return_value = mock_pf_instance
# Mock token dispatcher
mock_td_instance = MagicMock()
mock_td_instance.token_dispatch.return_value = {
"hidden_states": torch.randn(6, 8),
"group_list": torch.tensor([2, 2, 2]),
"group_list_type": 1
}
mock_td_instance.token_combine.return_value = torch.randn(4, 8)
mock_token_dispatcher.return_value = mock_td_instance
# Mock unified_apply_mlp
mock_unified_apply_mlp.return_value = torch.randn(6, 8)
# Create instance
comm_impl = AllGatherCommImpl(self.moe_config)
# Test fused_experts method
hidden_states = torch.randn(4, 8).contiguous()
w1 = torch.randn(16, 8).contiguous()
w2 = torch.randn(16, 8).contiguous()
topk_weights = torch.tensor([[0.5, 0.5], [0.3, 0.7], [0.8, 0.2],
[0.6, 0.4]])
topk_ids = torch.tensor([[0, 1], [1, 2], [2, 0], [1, 1]])
row_idx = torch.arange(4)
# Make sure tensors are contiguous and have correct strides
hidden_states = hidden_states.contiguous()
w1 = w1.contiguous()
w2 = w2.contiguous()
result = comm_impl.fused_experts(hidden_states=hidden_states,
w1=w1,
w2=w2,
topk_weights=topk_weights,
topk_ids=topk_ids,
row_idx=row_idx,
activation="silu")
# Verify result shape
self.assertEqual(result.shape, (4, 8))
# Verify token_dispatch was called
mock_td_instance.token_dispatch.assert_called_once()
# Verify unified_apply_mlp was called
mock_unified_apply_mlp.assert_called_once()
# Verify token_combine was called
mock_td_instance.token_combine.assert_called_once()

49
tests/ut/ops/test_token_dispatcher.py
View File

@@ -20,7 +20,8 @@ from unittest.mock import MagicMock, PropertyMock, patch
import torch
from tests.ut.base import TestBase
from vllm_ascend.ops.moe_dispatcher.token_dispatcher import (
from vllm_ascend.ops.moe.token_dispatcher import ( # isort: skip
AscendSocVersion, TokenDispatcherWithAll2AllV,
TokenDispatcherWithAllGather, TokenDispatcherWithMC2, _Dispatchers,
_register_token_dispatcher, get_token_dispatcher, setup_token_dispatchers)
@@ -34,7 +35,7 @@ class TestTokenDispatcherWithMC2(TestBase):
self.mc2_group.rank_in_group = 0
self.mc2_group.world_size = 8
self.mc2_group_patch = patch(
"vllm_ascend.ops.moe_dispatcher.token_dispatcher.get_mc2_group",
"vllm_ascend.ops.moe.token_dispatcher.get_mc2_group",
return_value=self.mc2_group)
self.mc2_group_patch.start()
@@ -52,7 +53,7 @@ class TestTokenDispatcherWithMC2(TestBase):
# Mock get_ascend_soc_version()
self.ascend_soc_version_patch = patch(
"vllm_ascend.ops.moe_dispatcher.token_dispatcher.get_ascend_soc_version",
"vllm_ascend.ops.moe.token_dispatcher.get_ascend_soc_version",
return_value=AscendSocVersion.A3)
self.ascend_soc_version_patch.start()
@@ -329,7 +330,7 @@ class TestTokenDispatcherWithAll2AllV(TestBase):
# Mock gather_from_sequence_parallel_region
patcher7 = patch(
'vllm_ascend.ops.moe_dispatcher.token_dispatcher.gather_from_sequence_parallel_region'
'vllm_ascend.ops.moe.token_dispatcher.gather_from_sequence_parallel_region'
)
self.mock_gather_from_sequence_parallel_region = patcher7.start()
self.addCleanup(patcher7.stop)
@@ -518,12 +519,8 @@ class TestDispatcherRegistry(TestBase):
self.assertIsNone(get_token_dispatcher("NonExistentDispatcher"))
@patch(
'vllm_ascend.ops.moe_dispatcher.token_dispatcher.TokenDispatcherWithAllGather'
)
@patch(
'vllm_ascend.ops.moe_dispatcher.token_dispatcher._register_token_dispatcher'
)
@patch('vllm_ascend.ops.moe.token_dispatcher.TokenDispatcherWithAllGather')
@patch('vllm_ascend.ops.moe.token_dispatcher._register_token_dispatcher')
def test_setup_token_dispatchers_ep_size_1_creates_allgather(
self, mock_register, mock_allgather_class):
kwargs = {"top_k": 2, "num_experts": 8}
@@ -537,12 +534,8 @@ class TestDispatcherRegistry(TestBase):
mock_allgather_class.assert_called_once_with(**kwargs)
mock_register.assert_called_once_with(mock_instance)
@patch(
'vllm_ascend.ops.moe_dispatcher.token_dispatcher.TokenDispatcherWithAll2AllV'
)
@patch(
'vllm_ascend.ops.moe_dispatcher.token_dispatcher._register_token_dispatcher'
)
@patch('vllm_ascend.ops.moe.token_dispatcher.TokenDispatcherWithAll2AllV')
@patch('vllm_ascend.ops.moe.token_dispatcher._register_token_dispatcher')
def test_setup_token_dispatchers_ep_size_2_creates_all2allv(
self, mock_register, mock_all2allv_class):
kwargs = {"top_k": 2, "num_experts": 16, "num_local_experts": 2}
@@ -556,15 +549,9 @@ class TestDispatcherRegistry(TestBase):
mock_all2allv_class.assert_called_once_with(**kwargs)
mock_register.assert_called_once_with(mock_instance)
@patch(
'vllm_ascend.ops.moe_dispatcher.token_dispatcher.TokenDispatcherWithAll2AllV'
)
@patch(
'vllm_ascend.ops.moe_dispatcher.token_dispatcher.TokenDispatcherWithMC2'
)
@patch(
'vllm_ascend.ops.moe_dispatcher.token_dispatcher._register_token_dispatcher'
)
@patch('vllm_ascend.ops.moe.token_dispatcher.TokenDispatcherWithAll2AllV')
@patch('vllm_ascend.ops.moe.token_dispatcher.TokenDispatcherWithMC2')
@patch('vllm_ascend.ops.moe.token_dispatcher._register_token_dispatcher')
def test_setup_token_dispatchers_ep_size_16_creates_all2allv_and_mc2(
self, mock_register, mock_mc2_class, mock_all2allv_class):
kwargs = {"top_k": 2, "num_experts": 32, "num_local_experts": 2}
@@ -584,15 +571,9 @@ class TestDispatcherRegistry(TestBase):
mock_register.assert_any_call(mock_all2allv_instance)
mock_register.assert_any_call(mock_mc2_instance)
@patch(
'vllm_ascend.ops.moe_dispatcher.token_dispatcher.TokenDispatcherWithAll2AllV'
)
@patch(
'vllm_ascend.ops.moe_dispatcher.token_dispatcher.TokenDispatcherWithMC2'
)
@patch(
'vllm_ascend.ops.moe_dispatcher.token_dispatcher._register_token_dispatcher'
)
@patch('vllm_ascend.ops.moe.token_dispatcher.TokenDispatcherWithAll2AllV')
@patch('vllm_ascend.ops.moe.token_dispatcher.TokenDispatcherWithMC2')
@patch('vllm_ascend.ops.moe.token_dispatcher._register_token_dispatcher')
def test_setup_token_dispatchers_ep_size_16_skips_if_exist(
self, mock_register, mock_mc2_class, mock_all2allv_class):
kwargs = {"top_k": 2, "num_experts": 32, "num_local_experts": 2}

6
tests/ut/quantization/test_w8a8.py
View File

@@ -5,8 +5,8 @@ import torch
from tests.ut.base import TestBase
from vllm_ascend.attention.attention_v1 import AscendAttentionState
from vllm_ascend.ops.layers.experts_selector import (_native_grouped_topk,
select_experts)
from vllm_ascend.ops.moe.experts_selector import (_native_grouped_topk,
select_experts)
from vllm_ascend.quantization.w8a8 import (AscendC8KVCacheMethod,
AscendW8A8FusedMoEMethod,
AscendW8A8LinearMethod,
@@ -784,7 +784,7 @@ class TestSelectExperts(TestBase):
self.assertEqual(ids.shape, (self.num_tokens, self.top_k))
self.assertEqual(ids.dtype, torch.int32)
@patch('vllm_ascend.ops.layers.experts_selector._native_grouped_topk')
@patch('vllm_ascend.ops.moe.experts_selector._native_grouped_topk')
def test_grouped_topk_with_correction_bias(self, mock_grouped_topk):
"""Test grouped topk with expert score correction bias"""
mock_grouped_topk.return_value = torch.ones(self.num_tokens,