Skip to content

[Roadmap] Miles AMD Q3 Development #2025

Open

Description

@XinyuJiangCMU

Miles AMD Q3 Development Roadmap

1. Model support

1.1 DeepSeek-V4-Flash

Enablement

Bring DeepSeek-V4-Flash blockwise FP8 RL end to end on MI355X.

Performance Optimization

1.2 Kimi K3

Bring Kimi K3 RL online on MI355X, building on SGLang's ROCm 7.2 TP8 serving path.

2. Feature support

2.1 LoRA support

Make LoRA RL reliable on AMD, starting with existing Miles recipes and expanding to colocated memory-saving workflows.

  • Validate LoRA training, rollout, reward, and adapter weight synchronization on representative models.
  • Add reproducible MI355X recipes and multi-step correctness coverage.
  • Make LoRA compatible with torch_memory_saver on ROCm, including pause/resume and colocated rollout.
  • [AMD][LoRA] Add a gpt-oss-20b MoE LoRA RL launcher #3157

2.2 OPD

Enable On-Policy Distillation on AMD by validating Miles' existing teacher-student workflows on MI355X.

  • Validate the Qwen3-8B OPD recipes with both SGLang-served and Megatron-loaded teachers.
  • Verify teacher/student log-prob alignment, reverse-KL, reward, and multi-step training correctness.
  • Extend coverage to multi-teacher routing and colocated memory-saving workflows.

2.3 Speculative decoding (MTP)

2.4 Low precision training

blockwise FP8

Enable end-to-end blockwise FP8 training and rollout on ROCm gfx950, covering training kernels, model recipes, weight handling, and SGLang runtime correctness.

Other

2.5 True On-Policy training

Eliminate training-inference mismatch and align rollout and training log probabilities.

2.6 Torch Memory Saver

Enable reliable torch_memory_saver for colocated RL by fixing the ROCr VMM release regression

3. Stable daily build Docker

ROCm base images for MI350X / MI300, plus daily Miles ROCm image build automation (latest sglang-miles branch + Miles commit).

4. AMD CI Stabilization

Activity

  1. HaiShaw commented on Aug 6, 2026

    @HaiShaw

    @XinyuJiangCMU Thanks for putting this together!

  2. crystalhuanxie-del commented on Aug 11, 2026

    @crystalhuanxie-del

    @XinyuJiangCMU
    Hello, I’m interested in the Miles AMD development project and would like to contribute to it. However, I don’t know at which stage we are currently at. Can the task 2.5 be assigned to me, if there’s no specific person designated for it? Thank you!

  3. XinyuJiangCMU commented on Aug 13, 2026

    @XinyuJiangCMU
    ContributorAuthor

    Hi @crystalhuanxie-del
    Thanks for offering to help. AMD already has a true on-policy path using the Triton attention backend, with an open SGLang PR for it, but that work is quite old and hasn’t been updated for a while, so I’m not sure it still works as-is.

    For 2.5, I think batch-invariant ROCm kernels for RMSNorm, GEMM, and attention are still open areas. Feel free to pick one up and submit a PR.

  4. crystalhuanxie-del commented on Aug 13, 2026

    @crystalhuanxie-del

    Hi @crystalhuanxie-del Thanks for offering to help. AMD already has a true on-policy path using the Triton attention backend, with an open SGLang PR for it, but that work is quite old and hasn’t been updated for a while, so I’m not sure it still works as-is.

    For 2.5, I think batch-invariant ROCm kernels for RMSNorm, GEMM, and attention are still open areas. Feel free to pick one up and submit a PR.

    Okay! I will take it. Thanks for your explanation.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions