AWS Trainium Frontier Competition

trn2.48xlarge instance rate

Two submissions ended in INFRA_FAILED today

Clarification — Phase 1 evaluation sequence length is fixed at 1024

There are no Capacity Blocks that meet your requirements in this region. Change the region or modify your requirements and try again.

invite links expired

Request for raw rank-1 traceback — no_checkpoint

Submission sub-b18271c7 REJECTED_BY_REVIEW

Submission sub-00a55580 — no_run_meta / worker startup failure

REJECTED_BY_REVIEW submission – requesting full reviewer feedback

Reviewing status

Issue accessing submission form / Attaching draft project to hackathon

Submission Issue Resolution (Sep 2–3) & Runner Update Changelog

Baseline compile hangs before neuronx-cc starts — possible beta eager runtime fork/exec deadlock

Submission platform

SAGEMAKER

Compilation time will be excluded from training budget

If you did not receive an email inviting you to the competition

Accidentally purchased a Capacity Block before credits became eligible

Github link

Private AMI AVAILABLE

Unable to create Capacity Block for trn2.3xlarge

Capacity Blocks: What to Expect and How to Get Capacity Blocks — Step-by-Step

Can train.py implement its own sequence packing rather than calling prepare.make_dataloader unchanged?

Compilation time in Phase 1

AWS Credits for capacity