Round 1: please rerun valid r40q4 after releasing prior GGUF runners

#1
by boweizh1204 - opened

Hi Ralph team β€” please rerun this unchanged shakedown submission after fixing the per-submission GGUF cleanup:

  • Hotkey: 5CPABbwvKphDTGP5qsAdpBNACNkoh3yhEBAtuwCYdYw9PZto
  • Artifact: hf://boweizh1204/r40q4@c0fa327a81b9d6b279aca815da178d0ac4dab186
  • Model SHA-256: 9ffb6e4e6343e671948ff3ce7338c8c0eea23aed6f5edd2364208732db095648
  • Validator rejection: scoring raised ValueError: Failed to load model from file: out/artifacts/boweizh1204_r40q4_main/model.gguf

The full immutable HF file was downloaded and matches that SHA-256 byte-for-byte. It also loads successfully (LOAD_OK) with the validator's exact llama_cpp==0.3.34 using n_ctx=8192, n_batch=512, n_threads=8, and n_gpu_layers=-1 on CPU. This rules out a corrupt/truncated GGUF and shows the failure is specific to the validator's CUDA/runtime state.

The published commitment/scoring order has nine successfully scored submissions before this one. In the current scoring loop, every challenger is retained through registry[sub.model_id] = runner; after scoring, GGUFStudentRunner._llm remains live and is never closed. A load of this model reports approximately 4.29 GiB weights + 1.125 GiB KV + 0.559 GiB compute per runner. Keeping nine of those alongside the BF16 parent and observer makes the tenth CUDA load order-dependent and is consistent with the generic llama-cpp ValueError seen here.

Suggested fix: do not retain challenger runners in registry; close/unload each challenger in a finally block after its score/canary/freeze work, retaining only actual incumbent runners. Then rerun round 1. The existing commitment, reveal, and artifact are unchanged and should not require resubmission.

References:

Sign up or log in to comment