Rank-128 LoRA adapter (bf16) extracted with mergekit from Goedel-LM/Goedel-Prover-V2-32B against the base model Qwen/Qwen3-32B, then sanitized for vLLM serving.

The raw mergekit extraction included full-rank modules_to_save tensors (embed_tokens, lm_head, and norm layers) that vLLM's LoRA runtime does not support. This upload contains only the pure low-rank lora_A/lora_B weights (448 pairs: 64 layers x q/k/v/o/gate/up/down projections), with modules_to_save: null in adapter_config.json. No resize_token_embeddings() call is needed to load this adapter.

This adapter has not been benchmarked yet (the intended minif2f evaluation did not produce results). It passes structural validation (only lora_A/lora_B keys, complete A/B pairs, correct Qwen3-32B shapes, no NaN/Inf) and loads and generates through vLLM's LoRA runtime.