reef-nccl-ep-qwen3-8b-lora-coevolve-m7-multisize

reef-nccl-ep-qwen3-8b-lora-coevolve-m7-multisize

Qwen3-8B LoRA adapters from the first MULTI-SIZE RL training run (reef TTTD recipe) of an alternating harness/model co-evolution loop on a GPU-kernel-configuration tuning task. Each training step queries a different message size (schedule cycling 8192, 2048, 512, 4096, 1024, 1536 tokens); the adopted size-conditional instruction plus one set of LoRA weights internalizes the best known configuration for every size.

Partial checkpoint set (honest note): the run's storage retention policy pruned intermediate checkpoints under disk pressure while the run was live. Steps 0, 1, 2, 3, 5, 6, 7, 11, 13, 14 survive (10 of 15); steps 4, 8, 9, 10, 12 were lost. Step 15 never committed (the run deadlocked at its final step and was cancelled; steps 0-14 committed cleanly). The full per-step reward record survives in the training log and is reproduced below.

Per-step mean rollout rewards (task reward = measured bandwidth score at the queried size; per-size best known scores in brackets):

On the first pass through the six-size schedule (steps 0-5) the model is below the best known score at three sizes. By the second pass (steps 6-11) every size is at its best known score, and the third partial pass (steps 12-14) sustains it: one model holds all six size regimes simultaneously.