All pull requests

Pull requests

• KDA-evolved MoE training: faster and leaner FP8 and BF16 across all models and meshes

  #133·haok1402 opened Oct 23, 2026Collaborator·

  ·

• Add Qwen3-Omni GRPO and bounded HF rollout

  #132·kldxbjz opened Oct 22, 2026Contributor·

  ·

• Add native Qwen3-Omni multimodal and speech training

  #131·kldxbjz opened Oct 22, 2026Contributor·

  ·

• [Do not merge] Bit-exact validation test for HF streaming (depends on #109)

  #130·Rishith18 opened Oct 22, 2026·

  ·

• Support independent generation from native speech checkpoints

  #129·kldxbjz opened Oct 21, 2026Contributor·

  ·

• Add sequential HF policy and reference residency

  #128·kldxbjz opened Oct 21, 2026Contributor·

  ·

• Initialize full Qwen3 Omni speech components into native FSDP shards

  #127·kldxbjz opened Oct 21, 2026Contributor·

  ·

• Bound Omni HF rollout weight installation memory

  #126·kldxbjz opened Oct 21, 2026Contributor·

  ·

• Bound Omni checkpoint memory and validate released Thinker continuation

  #125·kldxbjz opened Oct 21, 2026Contributor·

  ·

• Add local Qwen3-Omni codec-policy GRPO and strict continuation

  #124·kldxbjz opened Oct 21, 2026Contributor·

  ·

• Add local HF rollout provider for native Qwen3-Omni GRPO

  #123·kldxbjz opened Oct 21, 2026Contributor·

  ·

• Add Qwen3-Omni speech generation and native post-training

  #122·kldxbjz opened Oct 20, 2026Contributor·

  ·

• Add strict full Qwen3-Omni component checkpoint conversion

  #121·kldxbjz opened Oct 20, 2026Contributor·

  ·

• Add native Qwen3-Omni Talker and first-code pipeline

  #120·kldxbjz opened Oct 20, 2026Contributor·

  ·

• Add native Qwen3-Omni Code2Wav decoder

  #119·kldxbjz opened Oct 20, 2026Contributor·

  ·

• Add native Omni Thinker multimodal pretraining, conversion and restart validation

  #118·kldxbjz opened Oct 20, 2026Contributor·

  ·

• Fix DeepSeek and Qwen3.5 routing replay

  #114·GaokaiZhang opened Oct 15, 2026Contributor·

  ·
  12

• Make failed offloads retryable and guard offloaded state

  #113·GaokaiZhang opened Oct 14, 2026Contributor·

  ·
  2

• Fix expert checkpoints for uneven FSDP shards and HSDP

  #111·GaokaiZhang opened Oct 6, 2026Contributor·

  ·
  7

• Bound memory usage during Hugging Face checkpoint import

  #110·akaashrp opened Oct 5, 2026Contributor·

  ·
  1

• Stream HuggingFace safetensors directly into FSDP-sharded training

  #109·Rishith18 opened Oct 2, 2026·

  ·
  11

• Add native Qwen3-Omni Thinker text pretraining

  #101·kldxbjz opened Sep 27, 2026Contributor·

  ·

• Prepare staged Omni data and connect it to pretraining

  #100·kldxbjz opened Sep 24, 2026Contributor·

  ·

• Scaffold native Qwen3-Omni correctness comparison

  #98·kldxbjz opened Sep 22, 2026Contributor·

  ·

• [Feature] Add Nemotron-CC v2 dataset support

  #83·akaashrp opened Aug 6, 2026Contributor·

  ·
  26