GPT-QModel v7.3.6

What's Changed

• feat: support tile-misaligned GPTQ Marlin shapes by @ZX-ModelCloud in #3028
• Fix QQQ workspace lifetime and reduce packing memory by @yujiongzhang in #3029
• [MODEL] add HunyuanOCR quantization support by @ZX-ModelCloud in #3030
• [MODEL] Add locateanything quantization support by @ZX-ModelCloud in #3032
• feat: support tile-misaligned 4-bit AWQ Marlin shapes by @ZX-ModelCloud in #3033
• Fix CUDA JIT cache reuse and update CUTLASS by @Qubitium in #3039
• fix(tests): drop the extra self argument that makes TestKernelOutput.setUp raise TypeError by @Anai-Guo in #3037
• fix(tests): update run_layer_stage callers to the layer_names kwarg by @Anai-Guo in #3038
• fix(pack): forward quant_result to make_quant in pack_model by @Anai-Guo in #3034
• [MODEL] support Qwen3.8 Flash Next quantization by @ZX-ModelCloud in #3040
• fix(marlin): validate devices and register runtime buffers by @ZX-ModelCloud in #3041
• fix(awq): stop variable-length calibration collapsing to the last batch by @Leonccaa in #3036
• bump package dependencies to latest stable pypi versions by @Qubitium in #3042
• Reuse ABI-compatible JIT extension caches across Python/torch envs by @Qubitium in #3043
• Bump version from 7.3.5 to 7.3.6 by @Qubitium in #3044

New Contributors

• @yujiongzhang made their first contribution in #3029
• @Leonccaa made their first contribution in #3036

Full Changelog: v7.3.5...v7.3.6