Search results

FSDP collective desync (NCCL all-gather timeout) from per-rank sample skipping in DataFetcher

Status: Open.

#126 In lightseekorg/TorchSpec;

· chungen04 opened on Jun 26, 2026

All sglang TP workers process hidden states unnecessarily, causing training pipeline slowdown

Status: Open.

#122 In lightseekorg/TorchSpec;

· CH3CHOHCH3 opened on Jun 17, 2026

Support train Domino draft model?

Status: Open.

#114 In lightseekorg/TorchSpec;

· lengrongfu opened on Jun 1, 2026

Support for HF by the inference engine when using DFlash

Status: Open.

#105 In lightseekorg/TorchSpec;

· lixinqi7 opened on May 20, 2026

List out of index when debug_train_only

Status: Open.

#84 In lightseekorg/TorchSpec;

· Dogacel opened on Apr 27, 2026

[Feature] Support co-locate training and inference

enhancementNew feature or requestNew feature or request

high priority

Status: Open.

#81 In lightseekorg/TorchSpec;

· yubofredwang opened on Apr 22, 2026

Development Roadmap (2026 Q2)

high priority

Status: Open.

#34 In lightseekorg/TorchSpec;

· yubofredwang opened on Mar 7, 2026

refactor: extract shared metric aggregation logic in Eagle3Trainer

Status: Open.

#11 In lightseekorg/TorchSpec;

· cicirori opened on Feb 27, 2026