ggml Files

## Overview

This release adds tensor-split support to the multi-backend (meta) backend with improved split-state propagation, reworks the Metal kernels into per-op sources with parallel compilation, and fixes ggml_clamp to be a proper non-in-place op. It also brings new ops (POOL_1D, PAD_REFLECT_1D), Q2_K SYCL kernels, MoE bias fusion on OpenCL, and assorted fixes across the CUDA, Metal, SYCL, Vulkan, OpenCL and WebGPU backends.

### API changes

• ggml_clamp is now a non-in-place op returning a new tensor; the previous in-place (view) behavior is available via the new ggml_clamp_inplace (llama.cpp#27644)

### Core changes

• Fix ggml_clamp to be a proper non-in-place op, add ggml_clamp_inplace, and allow in-place clamp in the allocator (llama.cpp#27644)
• Optimize CPU concat by replacing per-element copies with row-level memcpy (llama.cpp#24575)
• Meta backend: add tensor-split mode, including delayed all-reduce for independent shared-expert branches (llama.cpp#26490)
• Meta backend: fix tensor split-state propagation for mul_mat, reshape, flash-attn and other ops (llama.cpp#27574)
• Shorten virtual device naming in the CUDA and Metal backends (llama.cpp#27608)

### Backend changes

#### CUDA

• Add POOL_1D op support (llama.cpp#27573)
• Shorten virtual device naming (llama.cpp#27608)

#### Metal

• Split the monolithic kernel source into per-op files and compile them in parallel (llama.cpp#26561)
• Per-device tuned (Q, NE) flash-attention vector sizes (llama.cpp#26570)
• Shorten virtual device naming (llama.cpp#27608)

#### SYCL

• Re-add Q2_K reordered MMVQ and ESIMD kernels (llama.cpp#27490)
• Mark TQ2_0 as not supported (llama.cpp#27660)

#### Vulkan

• Add PAD_REFLECT_1D op (llama.cpp#26586)

#### OpenCL

• Fuse MoE per-expert bias adds into the GLU and combine epilogues (opt-out via env vars) (llama.cpp#26431)

#### WebGPU

• Fix handling of infinity values in ARGSORT and TOP_K (llama.cpp#27538)
• Reorder includes to avoid a V/K macro conflict in the KV-overlap case (llama.cpp#27545)

More info

• Releases and versioning of ggml-org projects

Changelog since v0.21.0

34dc0e55 ggml : bump version to 0.22.0 (#1607) 5ef9ad90 sync : llama.cpp e19a660e sycl : mark tq2_0 as not supported (llama/27660) 29f05607 webgpu : fix handling of infinity values during ARGSORT and TOP_K (llama/27538) d737987d metal : per-device tuned (Q, NE) for flash-attn vec (llama/26570) 674b15cf metal: per-op source split + parallel compile (llama/26561) 12c8fb07 ggml : shorten virtual device naming in CUDA and Metal (llama/27608) 03555f63 webgpu : reorder includes since V that appears in common_decls.tmpl may be defined as K in flash_attn_decls.tmpl if KV_OVERLAP (llama/27545) 709321d3 ggml : fix ggml_clamp (llama/27644) 3526cdc0 Deepseek 4: -sm tensor (llama/26490) 70ff0f7f Fix meta tensor split state propagation (llama/27574) 0aab8701 cuda : add POOL_1D support (llama/27573) 6af3397b vulkan : added the PAD_REFLECT_1D operation (llama/26586) 47c985d7 ggml: optimize concat op by replacing per-element memcpy with row-level memcpy (llama/24575) 04f16ef3 sycl : add Q2_K reordered MMVQ and ESIMD kernels (again) (llama/27490) 4a57f1a6 opencl: fold the gpt-oss MoE per-expert bias adds into the epilogue (op/kernel fusion) (llama/26431) d99724f2 readme : update following llama.cpp style (#1599) 43724a63 ci : run builds only on source and build file changes (#1600)