[AMD] Track Miles MI355X daily images and enabled tests

[XinyuJiangCMU]

Scope

Track the Miles MI355X daily pipeline:

1. Build dated MI35X images from the latest Miles main.
2. Run the currently enabled ROCm 7.2 GPU tests on MI355X.

Live status

View	Purpose
Miles pipeline dashboard	Visual build → image → test history, refreshed hourly
ROCm 7.2 daily image	Image build used by the daily tests
ROCm 7.2 daily tests	Live queued/running/result status and logs

The dashboard reports suite-level status. GitHub Actions logs remain the source of truth for individual test results.

Daily pipeline

Step	Schedule	Output
ROCm 7.2 MI35X image	12:00 UTC	rocm/sgl-dev:miles-rocm720-mi35x-YYYYMMDD
ROCm 7.2 MI355X tests	17:30 UTC	2-GPU, 4-GPU, and 8-GPU suites

The test workflow selects the newest ROCm 7.2 dated image available.

Currently enabled GPU tests

The daily workflow currently selects 14 test files:

Suite	Enabled files	Runner pool
stage-c-2-gpu-mi350	1	linux-mi35x-gpu-2
stage-c-4-gpu-mi350	6	linux-mi35x-gpu-8
stage-c-8-gpu-mi350	7	linux-mi35x-gpu-8
Total	14

tests/e2e/
├── ckpt/                                      [8 GPU]
│   └── test_qwen3_4B_ckpt.py
├── long/                                      [2 GPU]
│   └── test_qwen2.5_0.5B_gsm8k_async.py
├── lora/                                      [4 GPU]
│   └── test_lora_qwen2.5_0.5B.py
├── megatron/                                  [4 GPU]
│   ├── test_mimo_7B_mtp_only_grad.py
│   └── test_qwen3_30B_A3B/
│       ├── test_baseline.py
│       └── test_r3_baseline.py
├── precision/                                 [4 GPU]
│   ├── test_hf_attention_cp_relayout.py
│   └── test_qwen3_5_cp_correctness.py
├── sglang_config/                             [8 GPU]
│   ├── test_sglang_config.py
│   ├── test_sglang_config_mixed_offload.py
│   └── test_sglang_config_mixed_offload_ft.py
└── short/                                     [8 GPU]
    ├── test_qwen2.5_0.5B_gsm8k_async_short.py
    ├── test_qwen2.5_0.5B_gsm8k_short.py
    └── test_run_megatron.py

The workflow invokes all three suites with --match-all-labels.

Pending review

• [AMD] Enable amd pr ci #2347

tests/e2e/                                      [4 GPU]
├── fsdp/test_qwen3_4B_fsdp_hybrid_shard_r2s2.py
├── megatron/
│   ├── model_scripts/
│   │   ├── test_deepseek_v4_flash_4layer_ci.py
│   │   ├── test_glm5_2_744b_a40b_5layer_ci.py
│   │   └── test_inkling_small_4layer_ci.py
│   ├── test_glm47_flash/test_amd_r3_mtp.py
│   ├── test_qwen3_30B_A3B/test_amd_moriep_fp8_bridge.py
│   ├── test_qwen3_4B_offload_disk_stream.py
│   └── test_qwen3_5_35B_A3B_mtp/test_amd_mtp1_spec_v2_r3.py

• [AMD CI] Enable verified ROCm tests #2480

tests/
├── e2e/                                        [2 GPU]
│   ├── fsdp/test_qwen3_0.6B_megatron_fsdp_align.py
│   ├── long/test_qwen2.5_0.5B_gsm8k.py
│   ├── sglang/test_session_server_multi_role/test_qwen3.py
│   └── short/test_qwen3_0.6B_fsdp_colocated_2xGPU.py
├── fast-gpu/test_run_megatron_worker_main.py    [2 GPU]
├── e2e/                                        [4 GPU]
│   ├── fsdp/test_qwen3_0.6B_fsdp_distributed.py
│   ├── megatron/
│   │   ├── test_qwen3_4B_offload_disk.py
│   │   ├── test_qwen3_4B_offload_disk_stream.py
│   │   ├── test_qwen3_4B_ppo.py
│   │   └── test_qwen3_4B_variable_gbs.py
│   └── sglang/
│       ├── test_chat_input_ids_equivalence.py
│       └── test_session_server_multi_role/test_qwen35.py
├── fast-gpu/test_det_process_group.py           [4 GPU]
└── e2e/                                        [8 GPU]
    ├── megatron/
    │   ├── test_qwen3_30B_A3B/test_disagg_broadcast.py
    │   ├── test_qwen3_30B_A3B/test_fully_async.py
    │   ├── test_qwen3_4B_disk_delta.py
    │   └── test_qwen3_4b_fully_async_eval.py
    └── short/
        ├── test_dumper.py
        ├── test_qwen2.5_0.5B_compact_rollout.py
        ├── test_qwen2.5_0.5B_fully_async_eval.py
        └── test_qwen2.5_0.5B_opd_sglang.py

GPU tests to enable next

This backlog only tracks GPU paths that have a plausible MI355X implementation. CPU tests and tests tied to NVFP4/Blackwell-only recipes, CUDA-only INT4 or modelopt kernels, or H200-specific FA3 parity are excluded.

tests/
├── e2e/
│   ├── fsdp/
│   │   ├── r3/
│   │   │   ├── test_glm47_flash_r3.py
│   │   │   ├── test_qwen3_30b_a3b_r3.py
│   │   │   └── test_qwen3_5_35b_a3b_r3.py
│   │   ├── test_qwen3_4B_fsdp_hybrid_shard_r2s4.py
│   │   ├── test_qwen3_4B_fsdp_true_on_policy.py
│   │   └── test_qwen3_vl_4B_fsdp.py
│   ├── ft/
│   │   ├── test_trainer_ft_deterministic_dp2_cp2_pp2.py
│   │   ├── test_trainer_ft_deterministic_dp2_cp2_real_rollout.py
│   │   ├── test_trainer_ft_deterministic_dp2_cp2_tp2_ep2.py
│   │   ├── test_trainer_ft_deterministic_dp4_cp2.py
│   │   ├── test_trainer_ft_no_failure_dp2_cp2_pp2.py
│   │   ├── test_trainer_ft_no_failure_dp2_cp2_real_rollout.py
│   │   ├── test_trainer_ft_no_failure_dp2_cp2_tp2_ep2.py
│   │   ├── test_trainer_ft_no_failure_dp4_cp2.py
│   │   ├── test_trainer_ft_with_failure_dp2_cp2_pp2.py
│   │   ├── test_trainer_ft_with_failure_dp2_cp2_real_rollout_dense.py
│   │   └── test_trainer_ft_with_failure_dp2_cp2_tp2_ep2.py
│   ├── long/
│   │   └── test_qwen3_0.6B_verifiers.py
│   ├── megatron/
│   │   ├── model_scripts/
│   │   │   ├── test_deepseek_v32_5layer_fp8.py
│   │   │   ├── test_glm5_1_744b_a40b_6layer_lora_ci.py
│   │   │   ├── test_glm5_2_744b_a40b_5layer_lora_ci.py
│   │   │   ├── test_glm5_744b_a40b_4layer_ci.py
│   │   │   ├── test_glm5_744b_a40b_4layer_r3.py
│   │   │   ├── test_gpt_oss_20b_moe_lora_ci.py
│   │   │   ├── test_inkling_small_4layer_lora_ci.py
│   │   │   └── test_nemotron_3_ultra_4layer_ci.py
│   │   ├── test_glm47_flash/
│   │   │   └── test_r3_mtp.py
│   │   ├── test_qwen3_30B_A3B/
│   │   │   ├── test_deepep_fp8.py
│   │   │   └── test_dp_attention.py       # AITER hipIpcGetMemHandle invalid argument
│   │   ├── test_qwen3_5_35B_A3B_mtp/
│   │   │   └── test_mtp0_spec_v2.py
│   │   ├── test_qwen3_30B_A3B_p2p.py
│   │   └── test_qwen3_5_35b_a3b_lora_ci.py
│   └── sglang/
│       └── test_session_server_multi_role/
│           ├── test_deepseekv4.py
│           ├── test_glm47.py
│           └── test_nemotron3.py
└── fast-gpu/
    └── test_fsdp_hybrid_shard.py

GPU tests intentionally skipped

These already carry disabled= on their register_cuda_ci, so upstream does not run them on any hardware either.

tests/
├── e2e/
│   ├── ckpt/
│   │   └── test_glm47_flash_ckpt.py
│   ├── ft/
│   │   ├── test_trainer_ft_random_dp2_cp2_real_rollout.py
│   │   ├── test_trainer_ft_random_dp2_cp2_tp2_ep2.py
│   │   └── test_trainer_ft_realistic_gsm8k.py
│   ├── megatron/
│   │   ├── test_glm47_flash/
│   │   │   └── test_r3_mtp_deepep.py
│   │   ├── test_qwen3_30B_A3B/
│   │   │   └── test_r3_deepep_fp8.py
│   │   ├── test_quick_start_glm4_9B.py
│   │   └── test_qwen3_5_35B_A3B_cp.py
│   ├── precision/
│   │   └── test_qwen3_0.6B_parallel_check.py
│   └── sglang/
│       ├── test_session_server_multi_role/
│       │   └── test_minimax_m27.py
│       └── test_r3_router_equivalence.py
└── fast-gpu/
    ├── test_mxfp8_quantizer.py
    └── test_semaphore.py

These tests are outside the MI355X enablement backlog because their current test contract is cuda-specific. They can be reconsidered only after the test or kernel path is redesigned for ROCm.

tests/
├── fast-gpu/
│   ├── test_nvfp4_quantizer.py
│   └── test_quantizer_ci
└── e2e/
    ├── megatron/
    │   ├── model_scripts/
    │   │   ├── test_deepseek_v32_5layer_mxfp8.py
    │   │   └── test_kimi_k25_2layer_ci.py
    │   ├── test_glm5_2_744b_a40b_5layer_nvfp4.py
    │   ├── test_joyai_llm_flash_mxfp8.py
    │   └── test_qwen3_30B_A3B/
    │       └── test_int4_rollout.py
    └── sglang/
        ├── test_session_v1_v2_parity.py
        └── test_session_server_multi_role/
            └── test_inkling.py

• NVFP4 and the B200 MXFP8 recipes require NVIDIA/Blackwell paths.
• The Kimi K2.5 and Qwen3 INT4 tests use the CUDA-only fake INT4 quantization extension.
• Session parity asserts an H200 device and FA3; the Inkling session test uses an NVFP4 checkpoint.

Last audited against Miles main on 2026-08-13.
Thanks to CI dashboard owners @michaelzhang-ai, @bingxche for making the Miles pipeline easy to track.