laguna-xs2-dense-k8-cuda-sft-v2

Laguna-XS.2 → Dense (K=8) · CUDA-SFT-extended (follow-up SFT)

A ~3.0 B dense CUDA-kernel model — a follow-up SFT on top of laguna-xs2-dense-k8-cuda-sft, trained on more CUDA / C++ kernel code.

Lineage

poolside/Laguna-XS.2 (33B/3B-active MoE, 256 experts)
   → densify (K=8 dense SwiGLU) → DO-ACP warm-start
   → reconstruction-pretrain (kernel mixture, "V2")
   → SFT (SakanaAI CUDA, level_1+2, 400 steps)        = laguna-xs2-dense-k8-cuda-sft
   → SFT-extended (level_1+2+3, +500 steps)           = THIS MODEL
   → RFT/GRPO (verifiable reward)                      = next (laguna-xs2-dense-k8-cuda-rft)

Why a follow-up SFT (rationale)

The first SFT (400 steps, level_1+2) produced a model that emits working CUDA on simple ops (ReLU/Tanh ~3/4 at pass@k) but showed two gaps:

• Thin C++ idiom coverage — it botches more involved C++/CUDA constructs (e.g. float4* v = float4* ptr; instead of reinterpret_cast<float4*>(ptr)), so vectorized kernels fail to compile.
• Limited CUDA breadth — harder ops (Sigmoid/GeLU/Softmax) compile/verify inconsistently.

This follow-up extends the SFT with more CUDA + C++ kernel data (Sakana level_1+2+3, +500 steps from the previous checkpoint) to broaden C++/CUDA coverage before RL. It is the mid checkpoint in a 3-way comparison: SFT → SFT-extended (this) → SFT-extended-RFT.

Training

Evaluation

Benchmarked 3-way (SFT / SFT-extended / SFT-extended-RFT) on KernelBench-Lite L1 (10 elementwise ops, K=4, subprocess-isolated compile+correctness vs PyTorch eager). Results table: github.com/Tyronita/laguna-dense-cuda-kernels.

Intended use

Research base for RFT (RL on verified compile+correctness+speedup). Kernels are not verified at generation time — compile & check before use, and isolate execution (a bad kernel corrupts the CUDA context).