update: Eagle3 finetune ep9 weights (KL loss, post-norm data, 29K samples, 10 epochs)

config: Eagle3DeepseekV2ForCausalLM + MLA params, aux_layers=[2,30,58]

Eagle3 finetune ep6 (kimi_k2 MLA): accept=62.8%, 256.5 tok/s@c=1, +22% vs NVIDIA

Add Eagle3 v2 epoch 7 checkpoint (trained on K2.5 API log data)

initial commit