Repository navigation

All pull requests

[tinker] Support retaining only the K most recent checkpoints
#2677Diferjfj opened Nov 8, 2026
7
[fix][train] Preserve prompt groups during replacement sampling
#2676Hang-get opened Nov 8, 2026
4
[megatron] Cover MLA attention projections in LoRA target_modules="all-linear"
#2674Diferjfj opened Nov 8, 2026
7
[train] Partial rollout: stale_group_policy wait and drop
#2673xinyuangui2 opened Nov 8, 2026Collaborator
[docs] Add a supported algorithms overview
#2672Diferjfj opened Nov 8, 2026
1
[train] Colocated partial rollout as a streaming trainer
#2670xinyuangui2 opened Nov 8, 2026Collaborator
[train] Split the fully-async trainer into a streaming package with a hookable base loop
#2669xinyuangui2 opened Nov 8, 2026Collaborator
1
[megatron] Full activation recompute for GLM-5.3-Flash's mHC layers (re-land #2411 on main)
#2667avigyabb opened Nov 8, 2026Collaborator
4
[weight sync] Colocated sync instrumentation and byte-parity harness
#2666erictang000 opened Nov 6, 2026Collaborator3 tasks done
[examples] DeepSeek-V4.1-Flash long-horizon LoRA RL: gsm8k_long_horizon env, async multi-turn recipe, long_horizon launcher and benches [enablement]
#2665avigyabb opened Nov 6, 2026Collaborator
[fully-async] Simulated training: load a LoRA adapter on every simulated weight sync (SKYRL_SIM_LORA_ADAPTER) [enablement]
#2664avigyabb opened Nov 6, 2026Collaborator
[inference] Name the positions and values when sample-support capture sees non-finite sampled logprobs [enablement]
#2663avigyabb opened Nov 6, 2026Collaborator
[inference] Session-sticky direct routing (SKYRL_SESSION_STICKY_DIRECT=1) [optimization]
#2662avigyabb opened Nov 6, 2026Collaborator
[vllm] DeepSeek-V4.1 LoRA under data-parallel attention + expert parallelism (int64 offsets, global LoRA metadata, wait-mode pause) [enablement][stacked on #2660]
#2661avigyabb opened Nov 6, 2026Collaborator
[vllm] DeepSeek-V4.1: fused MoE-LoRA via vLLM's two-kernel path, with tuned MoE-LoRA and dense LoRA tiles [optimization][stacked on #2659]
#2660avigyabb opened Nov 6, 2026Collaborator
[vllm] DeepSeek-V4.1: LoRA MoE on FlashInfer's TRT-LLM MXFP4 kernel instead of Marlin [optimization]
#2659avigyabb opened Nov 6, 2026Collaborator
[config] Allow ignore_eos with sample-support capture [enablement]
#2658avigyabb opened Nov 6, 2026Collaborator
[r3] Reject out-of-range replayed expert ids with a clear error [enablement]
#2657avigyabb opened Nov 6, 2026Collaborator
[infra] Forward MBRIDGE_STATE_*, VLLM_TUNED_CONFIG_FOLDER, VLLM_PP_LAYER_PARTITION and NCCL buffer settings to the Ray runtime env [enablement]
#2656avigyabb opened Nov 6, 2026Collaborator
[train] Step timing breakdown: forward/backward vs optimizer, colocated offload/backload, engine sleep, weight-sync stages [optimization]
#2655avigyabb opened Nov 6, 2026Collaborator
[train] Skip the trainer's between-phase empty_cache under expandable segments (Megatron) [optimization]
#2654avigyabb opened Nov 6, 2026Collaborator
[megatron] Fused LM-head autotune: share one tuning decision above 64k tokens [optimization]
#2653avigyabb opened Nov 6, 2026Collaborator
[megatron] DeepSeek-V4: fused clamped-SwiGLU expert activation by default [optimization]
#2652avigyabb opened Nov 6, 2026Collaborator
[megatron] DeepSeek-V4.1: context parallelism (packed THD, contiguous layout) [enablement][stacked on #2650]
#2651avigyabb opened Nov 6, 2026Collaborator
[megatron] DeepSeek-V4.1 LoRA: frozen base weights stored as MXFP8, decompressed one layer at a time [enablement][stacked on #2649]
#2650avigyabb opened Nov 6, 2026Collaborator

Pull requests