Add Vime ROCm support blog post#262
Conversation
Signed-off-by: pancake0003 <146360951+pancake0003@users.noreply.github.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 95ea49ad40
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| --ipc=host --network=host --device=/dev/kfd --device=/dev/dri \ | ||
| --security-opt seccomp=unconfined --group-add video --privileged \ | ||
| -e WANDB_API_KEY=$wandb_key vllm/vime-rocm | ||
| # wandb key is optional if you want to track with WandB |
There was a problem hiding this comment.
For readers who leave $wandb_key unset because this says the key is optional, the later ROCm launch script still enables W&B online mode (--use-wandb --wandb-mode online --wandb-key "${WANDB_API_KEY:-${wandb_key}}"), so a fresh container will try to authenticate/init W&B with an empty key before training starts. Either make the key required here or show the offline/disabled mode; otherwise the documented smoke test can fail before any rollout.
Useful? React with 👍 / 👎.
esmeetu
left a comment
esmeetu
left a comment
There was a problem hiding this comment.
Thanks for the detailed ROCm write-up and reproducible runbook. I found two additional claims that should be corrected before merging. The existing unresolved W&B-key thread also still applies: the launcher unconditionally enables W&B online mode, so the documentation should not call the key optional unless the script supports an offline/disabled path.
|
|
||
| `raw_reward` starts near 0 at step 0 and climbs gradually to around 0.5~0.6 by step 100. At initialization, the model has not yet been shaped by RL, so it approaches math problems with its base pretraining distribution; on competition-level problems from the dapo-math-17k dataset, a freshly initialized policy solves very few, yielding rewards close to zero. As training proceeds, the policy receives gradient signal from problems it gets partially or fully correct, and begins to favor reasoning patterns that the verifier rewards. With GRPO, the policy is generalizing across problem difficulty rather than overfitting to a narrow subset, which is the preferred learning dynamic for math RL. | ||
|
|
||
|
|
There was a problem hiding this comment.
The rising raw_reward here is measured on sampled training prompts, and the referenced ROCm launcher explicitly disables evaluation (EVAL_ARGS=()). It therefore shows training progress, but does not establish that the policy is generalizing rather than overfitting. Please either add held-out evaluation evidence or soften this to something like: “The rising training reward indicates optimization progress on the sampled training prompts; held-out evaluation is needed to assess generalization.”
| * R3 (Rollout Routing Replay) for AMD MoE workloads | ||
| * Asynchronous training pipelines | ||
| * Agentic RL for multi-turn tool calling and multi-agent settings | ||
|
|
There was a problem hiding this comment.
Could you clarify what remains future work here? The vime repository already includes scripts/run-qwen3-8B-async-amd.sh, an AMD-specific non-colocated async GRPO launcher using train_async.py with disjoint actor and rollout GPU pools. Either move basic async ROCm support into the supported list, or narrow this roadmap item to the remaining work, such as production hardening or performance optimization.
| <br> | ||
| <em>Throughput tending slightly updward with Qwen3-8B model on MI355X.</em> | ||
| </p> | ||
|
|
Signed-off-by: pancake0003 <146360951+pancake0003@users.noreply.github.com>
|
Thanks for the checks, fixed mentioned issues with new commit, please check. |
2 participants
Summary
_posts/2026-07-10-vime-rocm.mdassets/figures/2026-07-10-vime-rocm/on MI355X with figure captions
bundle exec jekyll serve