feat: tokenspeed support - #146
Conversation
Signed-off-by: Dogac Eldenk <dogacel@gmail.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: 42613a68a8
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".
| return engines | ||
|
|
||
| if engine_type not in ("hf", "sgl", "vllm", "trtllm"): | ||
| if engine_type not in ("hf", "sgl", "vllm", "trtllm", "tokenspeed"): |
There was a problem hiding this comment.
When a TokenSpeed run uses a DFlash or DSpark draft configuration, both normal training and offline materialization call _validate_and_configure_dflash, whose backend allowlist still excludes tokenspeed; the run therefore raises NotImplementedError before any TokenSpeed actor is created. Add this backend to that validation (or explicitly reject these algorithms in the new backend if unsupported) so the newly accepted engine type is usable with those training modes.
Useful? React with 👍 / 👎.
2 participants
Support training models with tokenspeed backend.
Tested both offline and online workload with Qwen3-8B, for more details check tokenspeed's main PR. [https://github.com/lightseekorg/tokenspeed/pull/798]