v7.10.0

GPT-QModel 7.10.0 🚀

• Models: Added Ricursive Ripple (ricursive) quantization support for the 65B preview checkpoint, including load-time Transformers 5 generation/masking patches, per-head Q/K norm and gated-attention handling, and v_proj Hessian isolation. Fixed pre-lm-head norm resolution across multimodal model definitions.
• Performance: Accelerated Ricursive Ripple decode by up to 10x per row at batch 16 with fused RMSNorm and native GQA attention, and added chunked MLP prefill to bound activation memory. The CUDA expandable-segments allocator is now installed before Torch initializes, so wide-layer multi-GPU runs actually use it.
• Checkpointing: Added gptqmodel.reshard() for per-layer safetensors sharding with one global shard, byte-for-byte tensor copies, preserved metadata, and config/tokenizer copying. Added a one-time warning when quantization shards are not layer-wise.
• Multi-GPU: Honored explicit device lists for quantization placement, routed cross-device moves per direction to avoid asymmetric ROCm SIGSEGVs, and documented cudaMallocAsync for wide-layer Hessian allocation.
• ROCm: Made AMD device families explicit and enforced kernel support contracts; gated W4A FP8/NVFP4 and ExLlamaV2/VPTQ off unsupported HIP paths with fast failures; fixed QQQ kernel launches to use the tensor's device context.
• Quantization: Added method-specific integer bit-width validation across GPTQ, AWQ, QQQ, ParoQuant, FP8, bitsandbytes, EXL3, GGUF, and VPTQ; EXL3 now rejects fractional trellis widths; MoE routing-bypass batch sizes are validated.
• Correctness: Fixed Ricursive Ripple cached decode, mask alignment, per-row positions, and labeled-loss handling, plus EXL3 tile-scratch bounds and sign-vector precision.

Full list of changes

What's Changed

• feat(models): add Ricursive Ripple (ricursive) support by @Qubitium in #3535
• fix(qqq): run the CUDA kernel in the tensor's device context by @ZX-ModelCloud in #3536
• fix(models): resolve pre-lm-head norm candidates consistently by @ZX-ModelCloud in #3537
• fix(loader): honour an explicit device list for quantization placement by @Qubitium in #3538
• fix(alloc): install the CUDA allocator default before Torch initialises it by @Qubitium in #3539
• docs(alloc): document the async allocator for wide-layer multi-GPU runs by @Qubitium in #3540
• fix(p2p): route cross-device moves per direction to avoid ROCm SIGSEGV by @ZX-ModelCloud in #3542
• perf(ricursive): 10x faster decode via fused RMSNorm and native GQA attention by @Qubitium in #3544
• feat(reshard): add per-layer checkpoint sharding with one global shard by @ZX-ModelCloud in #3543
• feat(lazy_turtle): warn once when quantization shards are not layer-wise by @ZX-ModelCloud in #3541
• fix(ricursive): correct multi-token cached decode and honour logits_to_keep by @Qubitium in #3545
• fix(exl3): bound tile scratch and keep sign vectors fp16 on load by @Qubitium in #3546
• fix(ricursive): rename MLP chunk env var to RIPPLE_MLP_CHUNK by @Qubitium in #3549
• fix(ricursive): mask from actual cache positions, keep labeled loss by @Qubitium in #3551
• chore(looper): remove the per-run processor log on success by @Lumea-art in #3548
• fix(rocm): make AMD device families explicit and enforce kernel support contracts by @ZX-ModelCloud in #3550
• fix(moe): reject a routing-bypass batch size that cannot batch by @Lumea-art in #3547
• fix(ricursive): align the mask to the written key span and forward cache positions by @Qubitium in #3552
• fix(exl3): reject a fractional bits — EXL3 has no fractional trellis by @Qubitium in #3553
• fix(w4a): gate the FP8/NVFP4 kernels off ROCm by @ZX-ModelCloud in #3554
• Validate integer quant bit widths per method by @Qubitium in #3555
• fix(exllamav2): fail fast on ROCm instead of attempting a JIT build by @ZX-ModelCloud in #3556
• fix(tests): make extension-load and AWQ kernel tests device-aware by @ZX-ModelCloud in #3557
• release: bump to 7.10.0 and add README release notes by @Qubitium in #3559

Full Changelog: v7.9.0...v7.10.0