iamPi/qwen36-35b-a3b-top3-qlora_e2

• data: vuhaian/top3_lastdance, loss on the last assistant turn only
• adapter: r=32, alpha=64, on the attention/linear-attention mixers and the shared expert. The 256 routed experts are 3D nn.Parameter tensors, which PEFT cannot target, so they stay frozen.
• base quantised to NF4 on its nn.Linear modules only (2.36B of 35.1B params); the experts remain bf16.
• lr 5e-5 cosine, warmup 0.03, packed to 16,384 tokens, global batch 16 packs.