Issues

• Confident-target logprobs, loss and gradients round to exactly 0 in the LM head and RL logprobs

  PrimeIntellect-ai/prime-rl#4271 · Filter by author catherinelee274 opened on Oct 26, 2026

• SFT with LoRA doesn't produce a vLLM-loadable adapter unless online evals are configured

  PrimeIntellect-ai/prime-rl#4203 · Filter by author magnusahmad opened on Oct 22, 2026

• Full CPU offload crashes when the GPU has no numa_node entry in sysfs (WSL2)

  PrimeIntellect-ai/prime-rl#4201 · Filter by author Arsh624 opened on Oct 22, 2026

• Unify completion and compact logprob transport

  PrimeIntellect-ai/prime-rl#4175 · Filter by author hallerite opened on Oct 20, 2026

• SFT: skip ignored tokens before the LM-head projection

  PrimeIntellect-ai/prime-rl#4136 · Filter by author aarjunsrinivasan opened on Oct 17, 2026

• GRPO gives non-zero advantages to groups whose rewards are all equal

  PrimeIntellect-ai/prime-rl#4131 · Filter by author RaphaelFakhri opened on Oct 17, 2026
  1

• Trainer import fails on hosts whose CUDA driver is older than the installed cutlass flavor: the optional flash_attn.cute import only catches ImportError

  PrimeIntellect-ai/prime-rl#4050 · Filter by author hazelxue opened on Oct 13, 2026

• CUDA device ordinals can select the wrong NVML GPU for NUMA binding

  PrimeIntellect-ai/prime-rl#3964 · Filter by author IshaGarg1103 opened on Oct 10, 2026

• RL trainer's logged loss is not the objective it optimizes: divided by micro-batches and dp_cp size

  PrimeIntellect-ai/prime-rl#3937 · Filter by author awesome-pro opened on Oct 8, 2026
  1

• LoRA on a MoE model with expert parallelism raises ImportError: TOKEN_GROUP_ALIGN_SIZE_M was removed by #3411

  PrimeIntellect-ai/prime-rl#3936 · Filter by author awesome-pro opened on Oct 8, 2026

• [Feature Request] Lossless sparse delta synchronization for remote vLLM deployments

  PrimeIntellect-ai/prime-rl#3884 · Filter by author GengLUO opened on Oct 5, 2026

• Qwen3.5 checkpointed backward fails when FLA tensor caching changes SAC operation order

  PrimeIntellect-ai/prime-rl#3882 · Filter by author cdreetz opened on Oct 5, 2026

• NIXL: trainer hangs when the orchestrator's "complete" notification is not delivered

  PrimeIntellect-ai/prime-rl#3742 · Filter by author yafshar opened on Oct 1, 2026

• Qwen3.5 0.8B/2B/4B: lm_head never loaded under impl=custom (tied embeddings), training silently gets zero gradients

  PrimeIntellect-ai/prime-rl#3700 · Filter by author AryanMadhavVerma opened on Sep 27, 2026
  1

• solve_rates assumes binary rewards; expose a correctness threshold?

  PrimeIntellect-ai/prime-rl#3566 · Filter by author erkancamli opened on Sep 17, 2026
  2

• CPU-offloaded checkpoint restore loses optimizer group counters and settings

  PrimeIntellect-ai/prime-rl#3561 · Filter by author llucid-97 opened on Sep 16, 2026

• Hosted Training: eval rollouts ignore enable_thinking = false

  PrimeIntellect-ai/prime-rl#3532 · Filter by author zafstojano opened on Sep 11, 2026
  1

• Unable to run deepswe environment eval run in hosted and local

  PrimeIntellect-ai/prime-rl#3530 · Filter by author TheSeriousProgrammer opened on Sep 11, 2026

• Consider removing TARGET_LAG now that off-policyness includes in-queue staleness

  PrimeIntellect-ai/prime-rl#3310 · Filter by author mikasenghaas opened on Aug 19, 2026

• Verify evals on Papers with Code

  PrimeIntellect-ai/prime-rl#3281 · Filter by author NielsRogge opened on Aug 16, 2026

• [Feature Request] CriPO: criterion-level credit assignment for rubric-based RL

  PrimeIntellect-ai/prime-rl#3278 · Filter by author Mantissagithub opened on Aug 15, 2026
  2

• Custom training-example selection: what should I build on?

  PrimeIntellect-ai/prime-rl#3173 · Filter by author rawwerks opened on Aug 1, 2026
  3

• [Feature Request] Include advantage values in rollout trace records (traces.jsonl)

  PrimeIntellect-ai/prime-rl#3170 · Filter by author Alex-Shan opened on Jul 31, 2026

• README trainer validation command fails because torchrun environment variables are not set

  PrimeIntellect-ai/prime-rl#3146 · Filter by author KoyunChen opened on Jul 28, 2026

• FFT resume fails on Qwen3.5-4B: missing optimizer state for untrained visual params

  PrimeIntellect-ai/prime-rl#2676 · Filter by author mikasenghaas opened on May 30, 2026