TransformerEngine-FL/.github/configs/metax.yml at b8da5252fe9a19b197f0c95ed89ffa505f260966 · BrianPei/TransformerEngine-FL · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
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
# CUDA Hardware Configuration for Megatron-LM-FL
# This file defines CI/CD settings for CUDA-based testing
# Test configurations are defined in tests/test_utils/config/platforms/cuda.yaml

hardware_name: metax
display_name: 'Metax Tests'

# Docker image for this hardware
# ci_image: cr.metax-tech.com/public-ai-release/maca/megatron-lm:0.12.0-maca.ai3.3.0.11-torch2.6-py312-ubuntu22.04-amd64
ci_image: localhost:5000/megatron-lm-with-te:v1

# Runner labels for this hardware
runner_labels:
  - self-hosted
  - Linux
  - X64
  - metax
  # - gpu-8
  - dev

# Container volumes (hardware-specific paths)
container_volumes:
  # - /home/flagscale_cicd/flask/static:/workspace/report
  # - /home/flagscale_cicd/flask/config:/workspace/config
  # - /home/flagscale_cicd/docker/docker_build/docker_data:/home/gitlab-runner/data
  # - /home/flagscale_cicd/docker/docker_build/docker_tokenizers:/home/gitlab-runner/tokenizers
  # - /home/flagscale_cicd/docker/docker_build/docker_data/Megatron-LM/datasets:/opt/data/datasets
  # - /home/flagscale_cicd/docker/docker_build/docker_tokenizers/Megatron-LM/tokenizers:/opt/data/tokenizers
  # --- 新增：Transformer Engine 开发专用路径 ---
  - /home/muxiuser/jinglong/TransformerEngine-FL:/workspace/TransformerEngine-FL  # 开发仓库
  - /home/muxiuser/jinglong:/opt/te_packages  # 存放编译好的 TE 包，供测试安装使用
  - /usr/local/maca:/usr/local/maca:ro  # [关键] 挂载宿主机的 MACA 驱动库（设为只读），确保算子能跑

# Container options (hardware-specific settings)
container_options: '--privileged --shm-size=500g --hostname megatron_cicd --user root --ulimit nofile=65535:65535 '

# Device types to run tests on
device_types:
  - C500

# Test matrix configuration
test_matrix:
  unit:
    devices:
      - C500
    # Ignored test files for unit tests
    # These files will be skipped when running pytest
    ignored_tests:
      - tests/unit_tests/data/test_preprocess_data.py
      - tests/unit_tests/dist_checkpointing/test_global_metadata_reuse.py
      - tests/unit_tests/dist_checkpointing/test_optimizer.py
      - tests/unit_tests/dist_checkpointing/test_nonpersistent.py
      - tests/unit_tests/dist_checkpointing/test_optimizer.py
      - tests/unit_tests/dist_checkpointing/test_safe_globals.py
      - tests/unit_tests/dist_checkpointing/models/test_moe_experts.py
      - tests/unit_tests/distributed/test_grad_sync_with_expert_parallel.py
      - tests/unit_tests/distributed/test_mcore_fully_sharded_data_parallel.py
      - tests/unit_tests/export/trtllm/test_distributed_fp8.py
      - tests/unit_tests/export/trtllm/test_single_device_fp8.py
      - tests/unit_tests/transformer/moe/test_a2a_token_dispatcher.py
      - tests/unit_tests/test_inference.py
      - tests/unit_tests/test_rl_utils.py
      - tests/unit_tests/models/test_gpt_model.py
      - tests/unit_tests/models/test_mamba_model.py
      - tests/unit_tests/post_training/test_modelopt_module_spec.py
      - tests/unit_tests/transformer/moe/test_aux_loss.py
      - tests/unit_tests/transformer/moe/test_moe_layer_discrepancy.py
      - tests/unit_tests/transformer/moe/test_routers.py
      - tests/unit_tests/transformer/test_attention.py
      - tests/unit_tests/transformer/test_attention_packed_seq.py
      - tests/unit_tests/transformer/test_cuda_graphs.py
      - tests/unit_tests/transformer/test_full_cuda_graph.py
      - tests/unit_tests/transformer/test_multi_latent_attention.py
      - tests/unit_tests/transformer/test_multi_token_prediction.py
      - tests/unit_tests/transformer/test_retro_attention.py
      - tests/unit_tests/transformer/test_transformer_block.py
      - tests/unit_tests/transformer/test_transformer_block_custom_pgs.py
      - tests/unit_tests/dist_checkpointing/test_local.py

  # functional:
  #   train:
  #     - device: C500
  #       task: train
  #       model: gpt
  #       case: all