[Feature]: Allow ep sharding the MTP draft model (currently disabled as WAR)
Open
Nobody has claimed this yet.
feature request
Speculative Decoding
- Dominant language
- Python
- Stars
- 14.7k
- Forks
- 2.8k
- Avg merge
- 2d 23h
- Merged PRs (30d)
- 489
Description
🚀 The feature, motivation and pitch
Target and draft share a single process-wide MoeAlltoAll workspace + flag_val counter (singleton in MoeAlltoAll._WORKSPACE), so a misconfigured draft alltoall corrupts the workspace and the subsequent target alltoall calls hang or fault. As a WAR we are currently disabling ep sharding of the draft and replicating it.
See: https://github.com/nv-auto-deploy/TensorRT-LLM/blob/b9ee1dfae3a4c573e7c90f90067f9b371de248ed/tensorrt_llm/_torch/auto_deploy/transform/library/sharding.py#L1236
Basically, this MoE object should be one-per-graph and not one-per-runtime, so we'd probably need to add a dict there
Alternatives
No response
Additional context
No response
Before submitting a new issue...
- Make sure you already searched for relevant issues, and checked the documentation and examples for answers to frequently asked questions.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Assessment
This issue has not been assessed yet.