Collections

Experimental models on a forked vLLM 0.30 (V1, eager) + OrcaSAQ2. Not stock `vllm serve`. Now: EXL3 5bpw HybridAgent-8B.

Qwen3-4B and Qwen2.5-Coder-7B fine-tuned to repair buggy Python, with FocusTracer dynamic slices. Follow-up to the Qwen2.5-3B ablation.