Gptqmodel
Monthly
Out-of-bounds read in ModelCloud GPTQModel ≤7.2.0 allows an attacker who controls a GPTQ checkpoint file to trigger unauthorized GPU device memory reads during model inference. The Triton dequantization kernel in `tritonv2.py` indexes `scales`/`qzeros` buffers using the checkpoint-supplied `g_idx` tensor without any upper-bound guard, so a crafted entry ≥ `num_groups` reads past the allocated buffer on the device. A public proof-of-concept exists (GitHub issue #2949), and vendor-released fix v7.3.0 is available.
Out-of-bounds read in ModelCloud GPTQModel ≤7.2.0 allows an attacker who controls a GPTQ checkpoint file to trigger unauthorized GPU device memory reads during model inference. The Triton dequantization kernel in `tritonv2.py` indexes `scales`/`qzeros` buffers using the checkpoint-supplied `g_idx` tensor without any upper-bound guard, so a crafted entry ≥ `num_groups` reads past the allocated buffer on the device. A public proof-of-concept exists (GitHub issue #2949), and vendor-released fix v7.3.0 is available.