NVIDIA / NVIDIA/TensorRT-LLM

[AutoDeploy]: Investigate SuperV3 with MTP not scaling well

Open
#14,225 9 comments 0 reactions 1 assignee View on GitHub

Nobody has claimed this yet.

AutoDeploy feature request
Dominant language
Python
Stars
14.7k
Forks
2.8k
Avg merge
2d 23h
Merged PRs (30d)
489

Description

🚀 The feature, motivation and pitch

AD surpasses PT at WS=1 across all concurrencies — 18% faster at c=1, near-parity (1%) at c=64.
AD fails to scale at WS=4: PT leads by 20–35%. The gap is driven by poor WS=1→4 scaling on the AD side: PT scales 1.52× from WS=1 to WS=4 at c=1 (4.67ms→3.07ms), while AD only gains 1.04× (3.82ms→3.67ms).

Investigate and resolve the poor scaling on AutoDeploy side.
Baseline branch: nv-auto-deploy:gagam/super-mtp-perf-2-replay (see #13725)

Alternatives

No response

Additional context

See SuperV3 MTP ticket #12359
Scripts, configs and experiment data:
https://gitlab-master.nvidia.com/ghubaraagam/agent-reports/-/tree/main/260428_superv3_mtp?ref_type=heads

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

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.