Repository Issue Activity (beta)

#8765 [rl] Fixed-budget value of dynamic environment sampling curricula in SkyRL

#8731 [canary-tpu] grug MoE train crashes: megablox GMM output block (512,32) invalid for (524288,128)

#8790 Test 4x long-context skew at qk=1.75

#8323 [infra] Finish the fork-migration rollout and per-fork follow-ups

#8781 [iris] Interactive band does not preempt batch under Kueue TAS lack-of-fit

#8784 [finelog] Phantom nested-partition catalog row retries archive upload forever

#8783 [finelog] Aborted migration's objects are never garbage-collected

#3316 Experiment: parallel-attn-mlp v6e-4 sweep for non-NaN hyperparameters

#6081 [iris] Make reconcile batch apply structurally order-independent (per-task update coalescing)

#6227 [Epic] Levanter TPU RL rollout backend parity

#6228 [levanter] Complete dense Qwen3-8B TPU parity matrix

#8780 Finelog: Don't build indices etc for forwarding servers

#8774 [iris] Preempted Ray head cannot reform its gang: workers stay terminally succeeded

#6217 [Iris scheduler] Higher Queues Should Preempt Across Node Types for Reserved Compute

#7463 [RFC] Find faster CPU graders for MATH, GSM8K, and HumanEval

#7519 [gpu] Validate JAX-Toolbox on cw-us-east-08a GB200

#7344 grug 512-GPU (8-rack GB200) run silently hangs mid-training (collective wedge, no error)

#8233 Next hero run burndown

#7429 [canary-grug-multislice] Megascale init timed out after 3600s — 2nd v6e-8 slice never provisioned (us-east5-b out of capacity)

#8533 [experiment][data] Student-relative scoring of agentic traces as a pre-SFT proxy for data value

#8744 [xla-fork] Keep the bazel cache small enough to be retained

#7856 Aug Hero Run LR Sweep

#7024 [grug] Track JaxPP pipeline-parallel MoE training

#1898 [Epic] Inference To Dos

#3037 [Experiment] Reproduce NAMO weight-decay ablation claim in Marin (50k-step, GPT2-small-equivalent)