All pull requests

Pull requests

• [skycap] skycap check: report whether a record stays on one path, and why each fork happened

  #2440·jadenfix opened Oct 7, 2026Contributor·

  ·
  8

• [inference] Salt each LoRA adapter's prefix cache by load count instead of resetting every engine's cache on sync

  #2439·yapdianang opened Oct 7, 2026Contributor·

  ·
  8

• [tinker] Route each unpickled SDK client's requests to one engine

  #2438·yapdianang opened Oct 7, 2026Contributor·

  ·
  3

• [megatron] KDA context parallelism: opt-in all-gather exchange (SKYRL_KDA_CP_EXCHANGE=allgather)

  #2437·avigyabb opened Oct 7, 2026Collaborator·1 of 2 tasks·

  ·

• [skycap] Pull a run's records and step index from W&B

  #2436·kouroshHakha opened Oct 7, 2026Collaborator·

  ·
  15

• [megatron] mHC: save the bf16 input of the FP32 projection, not its upcast

  #2434·avigyabb opened Oct 6, 2026Collaborator·2 tasks done·

  ·
  5

• [megatron][vlm] GLM-5.3-Flash vision-language model on Megatron (CP=1, packing)

  #2432·xinyuangui2 opened Oct 6, 2026Collaborator·

  ·

• [skycap] Images in token mode

  #2431·xinyuangui2 opened Oct 6, 2026Collaborator·

  ·
  10

• [fix][train] Per-field TensorList padding so padded rows stay self-consistent

  #2430·RaghuHemadri opened Oct 6, 2026·

  ·

• [megatron] GLM-5.3-Flash DSA: query-chunked fused sparse attention (opt-in)

  #2423·avigyabb opened Oct 5, 2026Collaborator·2 tasks done·

  ·
  6

• [megatron] GLM-5.3-Flash DSA: shard the k-pool indexer's query rows across TP (opt-in)

  #2422·avigyabb opened Oct 5, 2026Collaborator·1 task done·

  ·
  4

• [sft] GLM-5.3-Flash long-context recipe and context-length sweep runner

  #2419·avigyabb opened Oct 5, 2026Collaborator·1 of 2 tasks·

  ·

• [megatron] Context parallelism for GLM-5.3-Flash (head-wise KDA, all-gather DSA)

  #2418·avigyabb opened Oct 5, 2026Collaborator·1 of 2 tasks·

  ·
  7

• [examples] GLM-5.3-Flash cluster setup for Anyscale: EFA and node-local uv caches

  #2417·avigyabb opened Oct 5, 2026Collaborator·2 tasks done·

  ·
  4

• [train] Forward NCCL_* and FI_* settings to the Ray runtime env

  #2416·avigyabb opened Oct 5, 2026Collaborator·1 task done·

  ·
  5

• [megatron] Opt-in node-deduplicated MoE all-to-all

  #2415·avigyabb opened Oct 5, 2026Collaborator·1 task done·

  ·
  9

• [megatron] MoE combine: exact bf16 reduce-scatter for 2-rank expert-TP groups

  #2414·avigyabb opened Oct 5, 2026Collaborator·1 task done·

  ·
  3

• [megatron] GLM-5.3-Flash DSA: score and select k-pool indexer pools in query chunks

  #2413·avigyabb opened Oct 5, 2026Collaborator·4 tasks done·

  ·
  15

• [megatron] Opt-in host offload of full-recompute checkpoint inputs

  #2412·avigyabb opened Oct 5, 2026Collaborator·2 tasks done·

  ·
  20

• [megatron] Full activation recompute for GLM-5.3-Flash's mHC layers

  #2411·avigyabb opened Oct 5, 2026Collaborator·1 task done·

  ·
  5

• [train][StepWise] Add opt-in trajectory-level GSPO ratio for step-wise training

  #2409·RaghuHemadri opened Oct 5, 2026·

  ·

• [fully-async] Isolate prefix cache across weight versions for OpenAI-compatible requests

  #2408·RaghuHemadri opened Oct 5, 2026·

  ·

• [skyrl-train] Honor max_tokens_per_microbatch in loss-path forward

  #2407·RaghuHemadri opened Oct 5, 2026·

  ·
  1

• [megatron][vlm] Context parallelism for Qwen3-VL / Qwen3.5-VL

  #2404·xinyuangui2 opened Oct 4, 2026Collaborator·

  ·
  10

• Enable Custom Quantization

  #2399·benji-cannot-code opened Oct 4, 2026Contributor·

  ·
  1