Genghan/PTXBench-Qwen3.6-27B-s6

PTXBench Qwen3.6-27B-s6

This private repository contains the final PEFT LoRA adapter for PTXBench Qwen3.6-27B-s6. It was trained from Qwen/Qwen3.6-27B with rank 32 on the s6 configuration of Genghan/PTXBench-Qwen3.6-27B-SFT.

Qwen3.6 represents each linear-attention Q/K/V input projection as one fused in_proj_qkv module in Transformers. The release adapter therefore rewrites each independently trained rank-32 Q/K/V triplet as one mathematically equivalent block-diagonal rank-96 adapter. All other modules remain rank 32. See FUSION_PROVENANCE.json for the conversion details.

Field	Value
Internal recipe	fixit-v6
Training rows	258
Epochs	5
Learning rate	4.65e-4
Maximum length	65,536
Release-tested reconstruction base revision	6a9e13bd6fc8f0983b9b99948120bc37f49c13e9

Load without merging

from peft import PeftModel
from transformers import AutoModelForMultimodalLM

base = AutoModelForMultimodalLM.from_pretrained(
    "Qwen/Qwen3.6-27B",
    revision="6a9e13bd6fc8f0983b9b99948120bc37f49c13e9",
    torch_dtype="auto",
    device_map="auto",
    token=True,
)
model = PeftModel.from_pretrained(base, "Genghan/PTXBench-Qwen3.6-27B-s6", token=True)

Run python merge_adapter.py to create a standalone merged model directory. The merge requires the full base-model download and substantial CPU memory and disk space; it does not require a GPU.

This checkpoint is intended for research on CUDA kernel generation and repair. It inherits the limitations of the base model and may emit incorrect or unsafe CUDA. Generated kernels must be compiled and evaluated in an isolated environment before use.