obalcells/kimi-k2-thinking-mo-midtrained-lora

Kimi K2 Thinking — midtrained model organism LoRA

LoRA adapter (r=32) for the midtrained stage of an eval-aware model organism trained on Kimi K2 Thinking. Part of the SDF-honeypots project (safety-research/sdf-honeypots).

Training: Converted from Tinker checkpoint to PEFT format via scripts/convert_tinker_moe_adapter.py (3D unfuse: w1→gate_proj, w2→down_proj, w3→up_proj for MoE layers).

Base model: moonshotai/Kimi-K2-Thinking (DeepSeek-V3 architecture, 61 layers, MLA attention, INT4 compressed-tensors MoE).

Adapter size: 39GB

Load with vLLM:

vllm serve moonshotai/Kimi-K2-Thinking --tensor-parallel-size 8 \
  --enable-lora --max-lora-rank 64 \
  --enforce-eager --trust-remote-code
# then POST /v1/load_lora_adapter with this repo path

Known serving gotcha: set TORCH_COMPILE_DISABLE=1 to avoid UntypedStorage errors with --enable-lora on the FusedMoE layers.