Sync wiki with main through eb90dbdd5
Covers merged PRs #3478, #3483, #3484, #3485:
- #3478 ROCm RDNA4: gfx950/gfx120*/gfx125* fused GEMV+small-M GEMM
with per-arch launch tables in utils/arch.py, ROCm>=10.1 floor,
rocBLAS default on CDNA+RDNA4, AWQ GEMV/QQQ/VPTQ_CUDA/EXL3 explicit
ROCm rejections, ParoQuant Triton + FP8 ROCm support, FP4 not
advertised on ROCm, compute_device_filter honored by replicas and
resolve_pack_device (Installation, Inference-Backends,
Kernel-Backends-Reference, FP8-and-NVFP4, AWQ, QQQ, ParoQuant,
VPTQ, EXL3, Memory-Offload, Troubleshooting)
- #3485 Qwen-Omni transformers 5.22 thinker.model.language_model
layout: dual module_tree variants + thinker_* helpers (Multimodal)
- #3483 bounded remote image fetches: load_image_from_url timeouts,
retries, byte/pixel caps, ImageFetchError, env overrides
(Multimodal, Troubleshooting)
- #3484 MoE early-stop: run_parallel skips empty batches instead of
raising (Multimodal, Memory-Offload)
Sync wiki with main through 44f5c0a
Cover flint (#3458) and raon1_mini (#3460) model registrations:
add both to the MODEL_MAP registry table and the MoE definition
reference ('num_experts' count field, GateUpDownMoELifecycleHooks),
note Flint's Defuser-split stacked experts alongside Qwen3-VL MoE,
and re-pin the touched pages' source links to revision 44f5c0a.
Rename project references: GPTQModel -> GPT-QModel
Sync wiki with main through c1976918
Add workflow SVGs and refresh documentation navigation and accuracy
Full documentation site: concepts, all quant methods, config reference, lifecycle, backends, internals