Kimi-K3-4layer (CI slice)

A 4-layer bf16 slice of moonshotai/Kimi-K3, built for SkyRL's Megatron <> vLLM roundtrip logprob tests with examples/train/kimi_k3/make_slice.py. Not a usable language model.

• Layers 0-3 of the release: KDA (dense MLP), KDA (MoE) x2, NoPE-MLA (MoE).
• First 128 of 896 routed experts per MoE layer (router rows truncated to match; top-16 routing).
• Routed experts dequantized from MXFP4 to bf16 (exact); quantization_config removed.
• attn_res_block_size = 2 (release: 12) so the attention-residual bank spans two blocks.
• Everything else (attention, shared experts, embeddings, LM head, vision tower) copied unchanged.

Subject to the license of the original model.