Release GPT-QModel v7.5.0 · ModelCloud/GPTQModel

What's Changed

[MODEL] support K2 Horizon quantization by @ZX-ModelCloud in #3065
fix(qlinear): harden empty-input and CUDA dispatch contracts by @ZX-ModelCloud in #3069
Avoid full model scans during per-module finalization by @peter-positron in #3068
Scope banner Git lookup to the package checkout by @peter-positron in #3066
Align GPTQ keep masks with captured tensor devices by @peter-positron in #3067
Align version with 7.5.0 development changelog by @Qubitium in #3070
refactor(machete): harden runtime cache by @ZX-ModelCloud in #3071
fix: make kernel selection device-aware by @ZX-ModelCloud in #3072
test(kernels): expand edge-case coverage by @ZX-ModelCloud in #3073
[MODEL] add qwen_drive quantization support by @ZX-ModelCloud in #3074
Defer InternVL torchvision imports until image preprocessing by @peter-positron in #3078
fix(looper): import Sequence used in paroquant_processor annotations by @Anai-Guo in #3080
Fix partial packing words in the Torch fallback by @peter-positron in #3075
Validate offload metadata before saving meta tensors by @peter-positron in #3076
Validate required qweight tensors in GPTQ safetensors checkpoints by @peter-positron in #3077
Apply model replay hooks during Paro layer preparation by @peter-positron in #3079
[MODEL] add nanbeige quantization support by @ZX-ModelCloud in #3081
[MODEL] add diffusion_gemma quantization support by @ZX-ModelCloud in #3082
docs: finalize 7.5.0 release notes by @Qubitium in #3083

Full Changelog: v7.4.0...v7.5.0