Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
test_routers.py 5.72 KB
Newer Older
xingjinliang's avatar
升级megatron_v0.10  
xingjinliang committed
1
2
3
4
5 
# Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved.

import pytest
import torch
xingjinliang's avatar
同步最新代码  
xingjinliang committed
6 
from megatron.core.models.gpt.gpt_layer_specs import get_gpt_layer_local_spec
xingjinliang's avatar
升级megatron_v0.10  
xingjinliang committed
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29 
from megatron.core.transformer.moe.moe_layer import MoELayer
from megatron.core.transformer.moe.router import Router
from megatron.core.transformer.transformer_config import TransformerConfig
from megatron.training.initialize import _set_random_seed
from tests.unit_tests.test_utilities import Utils


class TestTop2Router:
    def setup_method(self, method):
        Utils.initialize_model_parallel(1, 1)
        _set_random_seed(seed_=123, data_parallel_random_init=False)
        print("done intializing")
        num_moe_experts = 4
        self.transformer_config = TransformerConfig(
            num_layers=2,
            hidden_size=12,
            num_attention_heads=4,
            num_moe_experts=num_moe_experts,
            use_cpu_initialization=True,
            moe_router_load_balancing_type="aux_loss",
            moe_router_topk=2,
            moe_aux_loss_coeff=0,
        )
xingjinliang's avatar
同步最新代码  
xingjinliang committed
30 
        transformer_layer_spec = get_gpt_layer_local_spec(
xingjinliang's avatar
升级megatron_v0.10  
xingjinliang committed
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86 
            num_experts=num_moe_experts, moe_grouped_gemm=False
        )
        self.sequential_mlp = MoELayer(
            self.transformer_config, transformer_layer_spec.submodules.mlp.submodules
        )
        self.router = self.sequential_mlp.router

    def teardown_method(self, method):
        Utils.destroy_model_parallel()

    @pytest.mark.internal
    def test_constructor(self):
        assert isinstance(self.router, Router)

        num_weights = sum([p.numel() for p in self.router.parameters()])
        assert num_weights == 12 * 4, num_weights

    @pytest.mark.internal
    @pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")
    @pytest.mark.internal
    @pytest.mark.parametrize("moe_router_pre_softmax", [(True), (False)])
    def test_router_forward(self, moe_router_pre_softmax):
        with torch.no_grad():
            self.router = self.router.cuda()
            self.router.config.moe_router_pre_softmax = moe_router_pre_softmax
            # [num tokens, hidden size]
            hidden_states = torch.randn((32, 2, self.router.config.hidden_size))
            hidden_states = hidden_states.cuda()
            scores, indices = self.router(hidden_states)

    @pytest.mark.internal
    @pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")
    @pytest.mark.internal
    def test_aux_loss(self):
        self.sequential_mlp = self.sequential_mlp.cuda()

        # Without aux loss
        hidden_states = torch.randn((32, 2, self.router.config.hidden_size))
        hidden_states = hidden_states.cuda()
        out = self.sequential_mlp(hidden_states)[0]
        out.sum().mul_(0).backward()
        assert self.sequential_mlp.router.weight.grad.abs().sum() == 0

        # With aux loss
        self.transformer_config.moe_aux_loss_coeff = 1
        out = self.sequential_mlp(hidden_states)[0]
        out.sum().mul_(0).backward()
        assert self.sequential_mlp.router.weight.grad.abs().sum() > 0

        # With Z loss
        self.transformer_config.moe_aux_loss_coeff = 0
        self.transformer_config.moe_z_loss_coeff = 1
        self.sequential_mlp.router.weight.grad.fill_(0)
        out = self.sequential_mlp(hidden_states)[0]
        out.sum().mul_(0).backward()
        assert self.sequential_mlp.router.weight.grad.abs().sum() > 0
xingjinliang's avatar
同步最新代码  
xingjinliang committed
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145 


class TestDeviceLimitedTop2Router:
    def setup_method(self, method):
        Utils.initialize_model_parallel(1, 1, expert_model_parallel_size=8)
        _set_random_seed(seed_=123, data_parallel_random_init=False)
        print("done intializing")
        num_moe_experts = 8
        self.transformer_config = TransformerConfig(
            num_layers=2,
            hidden_size=12,
            num_attention_heads=4,
            num_moe_experts=num_moe_experts,
            use_cpu_initialization=True,
            expert_model_parallel_size=8,
            moe_router_load_balancing_type="aux_loss",
            moe_router_topk_limited_devices=2,
            moe_router_pre_softmax=True,
            moe_router_topk=2,
            moe_aux_loss_coeff=0,
        )
        transformer_layer_spec = get_gpt_layer_local_spec(
            num_experts=num_moe_experts, moe_grouped_gemm=False
        )
        self.sequential_mlp = MoELayer(
            self.transformer_config, transformer_layer_spec.submodules.mlp.submodules
        )
        self.router = self.sequential_mlp.router

    def teardown_method(self, method):
        Utils.destroy_model_parallel()

    def test_constructor(self):
        assert isinstance(self.router, Router)

        num_weights = sum([p.numel() for p in self.router.parameters()])
        assert num_weights == 12 * 8, num_weights

    @pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")
    @pytest.mark.parametrize("moe_router_pre_softmax", [(True), (False)])
    def test_router_forward(self, moe_router_pre_softmax):
        with torch.no_grad():
            self.router = self.router.cuda()
            self.router.config.moe_router_pre_softmax = moe_router_pre_softmax
            if moe_router_pre_softmax:
                self.router.config.moe_router_topk_scaling_factor = 16.0
            # [num tokens, hidden size]
            hidden_states = torch.randn((32, 2, self.router.config.hidden_size))
            hidden_states = hidden_states.cuda()
            scores, indices = self.router(hidden_states)
            print(scores.shape, indices.shape)
            assert scores.shape == (64, 8)
            assert indices.shape == (64, 8)
            print(
                (indices == 0).sum(),
                (indices == 1).sum(),
                (indices == 2).sum(),
                (indices == 3).sum(),
            )