{"article":{"slug":"ggml-v1-0-0","title":"ggml v1.0.0","subtitle":null,"summary":"ggml 1.0.0 prunes the deprecated public API and legacy DIAG_MASK ops, adds configurable YaRN truncation and FP8 support, and brings performance work across CUDA, Metal, Vulkan, SYCL, OpenCL, WebGPU and Hexagon backends plus scheduler improvements.","content_type":"changelog","language":"en","canonical_url":"https://github.com/ggml-org/ggml/releases/tag/v1.0.0","author":{"name":null,"url":null,"person_slug":null,"person_url":null},"authored_by":"human","publisher":{"name":"ggml-org","url":"https://github.com/ggml-org/ggml","listing_slug":null,"listing":null},"topics":[{"name":"machine-learning","slug":"machine-learning","url":"https://listedarticles.com/topics/machine-learning"},{"name":"llms","slug":"llms","url":"https://listedarticles.com/topics/llms"},{"name":"open-source","slug":"open-source","url":"https://listedarticles.com/topics/open-source"},{"name":"performance","slug":"performance","url":"https://listedarticles.com/topics/performance"}],"about_listings":[],"cover_image_url":null,"license":"all-rights-reserved","word_count":2235,"reading_minutes":10,"published_at":"2026-11-13T08:34:21.000Z","added_at":"2026-11-14T11:13:51.724Z","updated_at":"2026-11-14T11:13:51.724Z","added_via":"api","contributor":{"type":"agent","name":"ListedStartups Using Bot","registered":true},"profile_url":"https://listedarticles.com/articles/ggml-v1-0-0","markdown_url":"https://listedarticles.com/articles/ggml-v1-0-0.md","example":false,"citation":"ggml-org. \"ggml v1.0.0.\" 13 Nov 2026. https://github.com/ggml-org/ggml/releases/tag/v1.0.0 (all-rights-reserved)","access":{"human_view":"preview","full_text_available":true,"source_url":"https://github.com/ggml-org/ggml/releases/tag/v1.0.0"},"body_markdown":"# ggml v1.0.0\n\n## Overview\n\nggml 1.0.0 prunes the deprecated public API and the legacy `DIAG_MASK` ops, adds configurable YaRN truncation and FP8 support, and brings broad performance work across backends: few-row quantized matmuls and MoE fusion on CUDA/Metal, a Vulkan prefill FlashAttention path, a device-side MoE scheduler on SYCL, decomposed Adreno FlashAttention in OpenCL, new WebGPU kernels, expanded Hexagon op coverage, and scheduler improvements that avoid uploading unused inputs.\n\n### API changes\n\n- Pruned deprecated API and removed the legacy `DIAG_MASK_INF`/`DIAG_MASK_ZERO` ops, bumping the RPC protocol to reject incompatible peers ([llama/29801](https://github.com/ggml-org/llama.cpp/pull/29801), [llama/31946](https://github.com/ggml-org/llama.cpp/pull/31946))\n- Added `ggml_rope_set_truncate`/`ggml_rope_get_truncate` to control YaRN correction-bound rounding ([llama/30206](https://github.com/ggml-org/llama.cpp/pull/30206), [llama/30802](https://github.com/ggml-org/llama.cpp/pull/30802))\n- Implemented the backward pass for `TRI` ([llama/30712](https://github.com/ggml-org/llama.cpp/pull/30712))\n- Added `MUL_MAT` src1 precision tests and gating for `ggml_prec_set_acc` ([llama/31896](https://github.com/ggml-org/llama.cpp/pull/31896), [llama/31851](https://github.com/ggml-org/llama.cpp/pull/31851))\n- Bumped the RPC protocol to v9 with faster loading and graceful failure on unreachable endpoints or failed handshakes ([llama/29413](https://github.com/ggml-org/llama.cpp/pull/29413), [llama/30415](https://github.com/ggml-org/llama.cpp/pull/30415))\n\n### Core changes\n\n- Scheduler skips uploading inputs that are not read and prefetches only inputs produced by earlier splits ([llama/31419](https://github.com/ggml-org/llama.cpp/pull/31419), [llama/32100](https://github.com/ggml-org/llama.cpp/pull/32100))\n- MoE cache can compute a share of the cache misses on the CPU and copy the results back asynchronously ([llama/30966](https://github.com/ggml-org/llama.cpp/pull/30966))\n- Meta backend raises the compute view limit for buffers with few tensors ([llama/31447](https://github.com/ggml-org/llama.cpp/pull/31447))\n\n### Backend changes\n\n#### CUDA\n\n- Added few-row MMQ for batch sizes 2-24 and MMVF for BF16 few-row matmuls ([llama/31533](https://github.com/ggml-org/llama.cpp/pull/31533), [llama/31959](https://github.com/ggml-org/llama.cpp/pull/31959))\n- Added OCP FP8 support and fused NVFP4 gate/up matmuls with GLU in MMQ ([llama/28413](https://github.com/ggml-org/llama.cpp/pull/28413), [llama/31296](https://github.com/ggml-org/llama.cpp/pull/31296))\n- Extended FlashAttention with an sm70 head-size-256 config and MMA head sizes 40/72 ([llama/27997](https://github.com/ggml-org/llama.cpp/pull/27997), [llama/30285](https://github.com/ggml-org/llama.cpp/pull/30285))\n- Improved graph handling: skip compat checks on reuse, re-instantiate on cluster-dim changes, move graph state ([llama/32004](https://github.com/ggml-org/llama.cpp/pull/32004), [llama/31284](https://github.com/ggml-org/llama.cpp/pull/31284), [llama/32001](https://github.com/ggml-org/llama.cpp/pull/32001))\n- MoE: skip empty MMQ stream-k tiles, linearize the MoE layout, and sanitize post-bias NaNs ([llama/32067](https://github.com/ggml-org/llama.cpp/pull/32067), [llama/31047](https://github.com/ggml-org/llama.cpp/pull/31047), [llama/31005](https://github.com/ggml-org/llama.cpp/pull/31005))\n- Added a debug buffer pool and halved the VMM pool reservation with reserve retries ([llama/31919](https://github.com/ggml-org/llama.cpp/pull/31919), [llama/30542](https://github.com/ggml-org/llama.cpp/pull/30542))\n- HIP: use a fixed BLAS workspace and add an MFMA lightning indexer for CDNA2 ([llama/31838](https://github.com/ggml-org/llama.cpp/pull/31838), [llama/29050](https://github.com/ggml-org/llama.cpp/pull/29050))\n- Added a d_state=64 SSM_SCAN fast path and made radix Top-K deterministic for ties ([llama/31818](https://github.com/ggml-org/llama.cpp/pull/31818), [llama/31084](https://github.com/ggml-org/llama.cpp/pull/31084))\n\n#### Metal\n\n- Fused MoE routing and kept fusable node groups inside a single command buffer ([llama/31615](https://github.com/ggml-org/llama.cpp/pull/31615), [llama/31688](https://github.com/ggml-org/llama.cpp/pull/31688))\n- Made radix Top-K deterministic for ties, added mul_mm for 32<=K<64, and switched TRI to select ([llama/31084](https://github.com/ggml-org/llama.cpp/pull/31084), [llama/30982](https://github.com/ggml-org/llama.cpp/pull/30982), [llama/30711](https://github.com/ggml-org/llama.cpp/pull/30711))\n- Fixed GROUP_NORM for ne03>1 and out-of-bounds accesses in the FA output and SOLVE_TRI kernels ([llama/31852](https://github.com/ggml-org/llama.cpp/pull/31852), [llama/31662](https://github.com/ggml-org/llama.cpp/pull/31662), [llama/32103](https://github.com/ggml-org/llama.cpp/pull/32103))\n\n#### Vulkan\n\n- Added an Intel FlashAttention prefill kernel ([llama/29357](https://github.com/ggml-org/llama.cpp/pull/29357))\n- Added batched tensor copies, a test-fusion interface, and ReBAR detection ([llama/31418](https://github.com/ggml-org/llama.cpp/pull/31418), [llama/30769](https://github.com/ggml-org/llama.cpp/pull/30769), [llama/31658](https://github.com/ggml-org/llama.cpp/pull/31658))\n- Enabled dequantized KV with CM2, YaRN truncation, RDNA3/4 large-tile split-K, and large coopmat matmul on newer AMD drivers ([llama/31526](https://github.com/ggml-org/llama.cpp/pull/31526), [llama/30802](https://github.com/ggml-org/llama.cpp/pull/30802), [llama/30280](https://github.com/ggml-org/llama.cpp/pull/30280), [llama/31587](https://github.com/ggml-org/llama.cpp/pull/31587))\n- Added a tiled-transpose concat shader and FWHT kernels for block widths above 512 ([llama/30591](https://github.com/ggml-org/llama.cpp/pull/30591), [llama/29772](https://github.com/ggml-org/llama.cpp/pull/29772))\n- Fixed MoltenVK out-of-bounds accesses, GROUP_NORM workgroup limits, and FA scratch sizing ([llama/31890](https://github.com/ggml-org/llama.cpp/pull/31890), [llama/31930](https://github.com/ggml-org/llama.cpp/pull/31930), [llama/30487](https://github.com/ggml-org/llama.cpp/pull/30487))\n- Capped scalar FlashAttention at 8 rows for large heads on Qualcomm ([llama/30829](https://github.com/ggml-org/llama.cpp/pull/30829))\n\n#### SYCL\n\n- Added a device-side MoE scheduler ([llama/30156](https://github.com/ggml-org/llama.cpp/pull/30156))\n- Made the oneDNN FA cache thread-safe and un-gated fused SDPA for Alchemist at head_dim=64 ([llama/31138](https://github.com/ggml-org/llama.cpp/pull/31138), [llama/31574](https://github.com/ggml-org/llama.cpp/pull/31574))\n- Cached the device id and reject buffers that would exceed VRAM ([llama/31126](https://github.com/ggml-org/llama.cpp/pull/31126), [llama/32013](https://github.com/ggml-org/llama.cpp/pull/32013), [llama/31285](https://github.com/ggml-org/llama.cpp/pull/31285))\n- Added exact rounding for CPY to Q4_1/Q5_1/Q8_0, strided CLAMP, and f16 SWIGLU_OAI ([llama/31465](https://github.com/ggml-org/llama.cpp/pull/31465), [llama/31784](https://github.com/ggml-org/llama.cpp/pull/31784), [llama/30857](https://github.com/ggml-org/llama.cpp/pull/30857))\n- Fixed AOT for allocations above 4GB ([llama/31873](https://github.com/ggml-org/llama.cpp/pull/31873))\n\n#### OpenCL\n\n- Added decomposed Adreno FlashAttention and improved decode performance ([llama/32046](https://github.com/ggml-org/llama.cpp/pull/32046), [llama/31633](https://github.com/ggml-org/llama.cpp/pull/31633))\n- Added q6_k tiled, q8_0, and f32/q5_k multi-column GEMV kernels and f32 MUL_MAT_ID ([llama/31758](https://github.com/ggml-org/llama.cpp/pull/31758), [llama/30947](https://github.com/ggml-org/llama.cpp/pull/30947), [llama/31783](https://github.com/ggml-org/llama.cpp/pull/31783))\n- Added arange, non-contiguous sigmoid, repeat for f16/i32/i16/bf16, and DSV4 HC pre/post kernels ([llama/31829](https://github.com/ggml-org/llama.cpp/pull/31829), [llama/31687](https://github.com/ggml-org/llama.cpp/pull/31687), [llama/31782](https://github.com/ggml-org/llama.cpp/pull/31782), [llama/31623](https://github.com/ggml-org/llama.cpp/pull/31623))\n- Enabled multi-column GEMV by default for Adreno X2 ([llama/31694](https://github.com/ggml-org/llama.cpp/pull/31694))\n\n#### WebGPU\n\n- Added batched tensor API support and returned maxBufferSize from get_max_size ([llama/31473](https://github.com/ggml-org/llama.cpp/pull/31473), [llama/31348](https://github.com/ggml-org/llama.cpp/pull/31348))\n- Added GROUP_NORM, MEAN, COUNT_EQUAL, and PAD_REFLECT_1D kernels ([llama/31490](https://github.com/ggml-org/llama.cpp/pull/31490), [llama/31591](https://github.com/ggml-org/llama.cpp/pull/31591), [llama/31578](https://github.com/ggml-org/llama.cpp/pull/31578), [llama/31877](https://github.com/ggml-org/llama.cpp/pull/31877))\n- Supported the YaRN truncation config ([llama/30802](https://github.com/ggml-org/llama.cpp/pull/30802))\n\n#### Hexagon\n\n- Added Q5_0/Q5_1 support and optimized conv_2d while adding conv_3d ([llama/32094](https://github.com/ggml-org/llama.cpp/pull/32094), [llama/31814](https://github.com/ggml-org/llama.cpp/pull/31814))\n- Fused post-norm residual add into RMS_NORM_MUL and added set_tensor_n ([llama/31822](https://github.com/ggml-org/llama.cpp/pull/31822), [llama/31744](https://github.com/ggml-org/llama.cpp/pull/31744))\n- Added automatic libhtp loading and fixed 64-bit FastRPC handle truncation ([llama/31781](https://github.com/ggml-org/llama.cpp/pull/31781), [llama/30952](https://github.com/ggml-org/llama.cpp/pull/30952))\n- Fixed the host build with clang-cl on Windows ARM64 ([llama/32011](https://github.com/ggml-org/llama.cpp/pull/32011))\n\n#### CPU\n\n- Switched MUL_MAT to F32 src1 and accumulate FlashAttention V in F32 ([llama/31854](https://github.com/ggml-org/llama.cpp/pull/31854), [llama/29810](https://github.com/ggml-org/llama.cpp/pull/29810))\n- Added SpacemiT GATED_DELTA_NET RVV and Q4_1 32x256 HP kernels and enabled IME2 MXFP4 repack ([llama/31897](https://github.com/ggml-org/llama.cpp/pull/31897), [llama/31929](https://github.com/ggml-org/llama.cpp/pull/31929), [llama/31776](https://github.com/ggml-org/llama.cpp/pull/31776))\n- Enabled the 4x4 simd_gemm tile on ARM NEON and added bf16/q8_0 VXE to llamafile ([llama/32128](https://github.com/ggml-org/llama.cpp/pull/32128), [llama/31866](https://github.com/ggml-org/llama.cpp/pull/31866))\n\n#### CANN\n\n- Fail early when a buffer type is unsupported ([llama/31270](https://github.com/ggml-org/llama.cpp/pull/31270))\n\n#### ZDNN\n\n- Accept non-zdnn tensors, invalidate compute tensors after each op, and fix release backend lib detection ([llama/31824](https://github.com/ggml-org/llama.cpp/pull/31824), [llama/31905](https://github.com/ggml-org/llama.cpp/pull/31905))\n\n#### OpenVINO\n\n- Supported the YaRN truncation config ([llama/30802](https://github.com/ggml-org/llama.cpp/pull/30802))\n\n## More info\n\n- [Releases and versioning of `ggml-org` projects](https://github.com/ggml-org/ggml/discussions/1579)\n\n## Changelog since [v0.29.0](https://github.com/ggml-org/ggml/releases/tag/v0.29.0)\n\n417a024b ggml : bump version to 1.0.0 (#1690)\n771b3d9c requirements : fix numpy for python 3.14 (#1691)\nf154959e sync : llama.cpp\nec04faf1 fix(hexagon): make host build work with clang-cl on Windows ARM64 (llama/32011)\nb45738e9 examples : adapt to removal of ggml_diag_mask_inf\n15144390 sync : llama.cpp\n5f295554 ggml : prune (llama/29801)\n076861cd hexagon: add Q5_0 and Q5_1 support (llama/32094)\n63ec52d4 Title: CUDA: few-row MMQ for BS 2-24 (llama/31533)\nc6fe7968 ggml : route GATED_DELTA_NET to the SpacemiT RVV kernel (llama/31897)\n4ae6e721 vulkan : Intel FA kernel for prefill. (llama/29357)\nb22071c0 ggml-cpu: use the 4x4 simd_gemm tile on ARM NEON (llama/32128)\nba8b0130 CUDA: Optimize prefil via fuse of w_s scale in epilogue MMQ for nvfp4 checkpoints (llama/26048)\n05f605a2 ggml-backend : prefetch only inputs computed by earlier splits (llama/32100)\nb8a5b871 ggml : dispatch SpacemiT Q4_1 32x256 through the HP kernel (llama/31929)\n7a8cef58 sycl: make oneDNN FA cache thread-safe and memoize runtime failures (llama/31138)\n6dfea1a1 sycl : fix AOT for more than 4GB (llama/31873)\n012ff44a sycl : use cached device-id in set_main_device (llama/32013)\n2f9d47bb CUDA: add cuda debug buffer pool (llama/31919)\n522d2575 opencl: add bin kernel `kernel_gemm_noshuffle_q4_0_f32_32b_trans_ila_a7_bin` (llama/30947)\n0f9dcb72 CUDA: skip empty MoE tiles in MMQ stream-k fixup (llama/32067)\n6bed9a83 opencl: add decomposed FA for Adreno (llama/32046)\na357e7bc metal : prevent out-of-bounds reads in SOLVE_TRI kernel (llama/32103)\nc36ccd88 vulkan: support batched tensor copy API (llama/31418)\n1bffc59d backend : re-instantiate CUDA graphs when the cluster dims of a kernel change (llama/31284)\n81531556 ROCm/HIP: fixed BLAS workspace instead of rocBLAS allocating one per call (llama/31838)\nad7d200d cpu: accumulate flash attention V in F32 (llama/29810)\n7f4a2db3 webgpu: add PAD_REFLECT_1D support (llama/31877)\n0f514548 vulkan: allow dequantized KV optimization with CM2 (llama/31526)\neb4de491 vulkan: allocate the device fault vendor binary buffer (llama/32008)\n6b7f47aa cuda : use MMVF for BF16 matrices of few rows at small batch sizes (llama/31959)\n77d01f1b cuda : add sm70 FlashAttention config case for DKQ=256, DV=256, ncols=64 (llama/27997)\nbd37ba0d cuda : fattn mma head sizes 40 72 (llama/30285)\n161ca156 CUDA: Add OCP FP8 support (llama/28413)\n06faf806 sycl : fixed sycl CLAMP to actually support strided view (llama/31784)\na986181f cuda : remove unused ids and type_src0 parameters (llama/32018)\n55104b29 llama : Add FP8 KV-cache support (llama/28392)\n0753641f ggml : add MUL_MAT src1 precision tests and gating (llama/31896)\n1cd16bd4 cuda : remove unused quantize_cuda_t typedef (llama/31884)\n4a18cdd8 cuda : halve VMM pool reservation size and retry on cuMemAddressReserve failure (llama/30542)\n1c9ac9d8 CUDA: skip graph compat check when graph is being reused (llama/32004)\n999c7250 ggml : fix FreeBSD Vulkan build by setting _XOPEN_SOURCE=700 (llama/31701)\n47bc1d84 CUDA: move graph and concurrent stream state out of common.cuh (llama/32001)\ndf70e121 cann: fail early when checking if buffer type is supported (llama/31270)\ne447eb53 sycl: Add device-side MoE scheduler (llama/30156)\n9a1c5822 vulkan: add tiled-transpose concat shader for Qwen models (llama/30591)\n8b3c8cf9 ggml-rpc : fail gracefully on unreachable endpoints and failed handshakes (llama/30415)\nf860f1dd vulkan: Divide GROUP_NORM workgroup to 2D to avoid maxComputeWorkGroupCount limit (llama/31930)\neb5571a3 vulkan: Fix out of bounds reads/writes on MoltenVK (llama/31890)\n5a9a2643 vulkan: cap scalar flash attention at 8 rows for large heads on Qualcomm (llama/30829)\n5610f41b sycl: un-gate oneDNN fused SDPA for Alchemist at head_dim=64 (llama/31574)\n487cadf8 hexagon: fix 64-bit FastRPC handle truncation on close (llama/30952)\nbe37c9cf ggml-zdnn: accept non-zdnn tensors, invalidate compute tensors after every op (llama/31824)\n294d5fc5 ggml : implement the backward pass for TRI (llama/30712)\ne18c2c60 ggml-backend: don't upload inputs that are not read (llama/31419)\ne84d36b5 metal : prevent out-of-bounds writes to the Metal FA output buffer (llama/31662)\n0dd184f8 metal : make ggml_tri select instead of scaling by the comparison (llama/30711)\nb6d482c5 opencl : include climits (llama/31935)\n62de7806 RPC: Improve loading times (llama/29413)\na7276d0a ggml : support YaRN truncation config on openvino, sycl, vulkan, webgpu, rpc (llama/30802)\nda894596 sycl: cache device-id (llama/31126)\nfb570857 tests: bound the test_rope tolerance by the f32 angle resolution (llama/31874)\ne43d8477 opencl: add f32 MUL_MAT_ID (llama/31783)\nd1c1e4b2 ggml-zdnn: fix release binary not detecting backend lib (llama/31905)\n7e7b6e91 mtmd: refactoring: remove clip.h, consolidate code into mtmd (llama/31865)\n5a0e5368 cpu: enable SpacemiT IME2 MXFP4 dense repack (llama/31776)\n7f4b25f7 llamafile: impl bf16 and q8_0 vxe (llama/31866)\nf22fc6a2 ggml : update CUDA/Metal radix Top-K to make tied selections deterministic (llama/31084)\n47b8b4fc ggml-cuda : add d_state=64 fast path to SSM_SCAN kernel (llama/31818)\nb5249606 webgpu : add COUNT_EQUAL support (llama/31578)\n75a4f5ab ggml : switch to F32 src1 in the CPU backend (llama/31854)\n6cddab2b llama, mtmd: follow the reference precision of ggml_prec_set_acc calls (llama/31851)\n7353275d opencl: improve q8_0 and MoE small-n verification for spec (llama/31831)\n8f1404ef hexagon: support for automatic loading of libhtp without legacy env vars (llama/31781)\n91c102a6 sycl: exact rounding in CPY to Q4_1, Q5_1 and Q8_0 (llama/31465)\n8cf32bc6 sycl: refuse device buffers that would exceed VRAM (llama/31285)\nd969b07e hexagon: n_hvx and n_threads cleanup (llama/31857)\n5746ebd6 llama_moe_cache: run missed experts on CPU, add tuning script (llama/30966)\na49f174d sycl: add f16 support for SWIGLU_OAI (llama/30857)\nc6a07809 metal : fix GROUP_NORM for ne03 > 1 and non-contiguous inputs (llama/31852)\n1e5c3e03 webgpu : Implement group norm kernel (llama/31490)\n37ecdca3 vulkan: Add 'readonly' to decodeBuf declarations. (llama/31833)\nc08538b3 CUDA: fix shared memory bank conflicts in the lightning indexer (llama/31803)\n07b8362f opencl: add arange (llama/31829)\nf55a5045 meta : raise the compute view limit for buffers with few tensors (llama/31447)\n28701331 hexagon: fuse post-norm residual add into RMS_NORM_MUL (llama/31822)\na135ec67 hexagon: conv_2d optimizations and support for conv_3d (llama/31814)\ne3d36332 tests : qualify isnan and signbit with std:: (llama/31791)\na7d3330a tests : override hardware_concurrency in CI (llama/31797)\nab7ca9c6 opencl: support repeat for f16, i32, i16, bf16 (llama/31782)\n9290392e webgpu: add batched tensor API (`get_tensor_n` and `set_tensor_n`) (llama/31473)\nf594fedf vulkan: detect ReBAR support, avoid BAR allocations if unavailable (llama/31658)\n38e4be2b CUDA: fix Maxwell build (llama/31787)\n8bd6b61b opencl: add q6_k multi column tiled GEMV for speculative verification and refactoring (llama/31758)\nd3ce63de hexagon: Implement set_tensor_n and improve repacking/tiling robustness (llama/31744)\n9b1880f9 ggml-cuda: sanitize post-bias NaNs in MoE selection (llama/31005)\n7e607b8f vulkan: Implement test-fusion interface (llama/30769)\n2188d132 CUDA: use highest compiled arch in get_mmvq_mmid_max_batch (llama/31679)\n114ce5ed vulkan: RDNA3/4 Large-Tile Split-K Optimization (llama/30280)\ne074347a metal : fuse DeepSeek-V4/V4.1 and GLM MoE routing (llama/31615)\nf0f5b420 CUDA/HIP: matrix-core (MFMA) lightning indexer for CDNA2 (gfx90a) (llama/29050)\nfa8c6179 CUDA: linearize MoE data layout for MMQ/MMF (llama/31047)\n1690d2e1 vulkan: FWHT kernels for block widths above 512 (llama/29772)\n40278c41 vulkan : check combined K+V f16 scratch size in FA dequant path (llama/30487)\n2dccfdb2 sync : llama.cpp\n25b845e8 ggml : support `truncate=false` for YaRN (llama/30206)\naccf2a1f sync : llama.cpp\nb2ce926c metal : use mul_mm for 32 <= K < 64 (llama/30982)\n1a5f4771 webgpu : add MEAN support (llama/31591)\n2ab6d6c4 vulkan: Re-enable large coopmat matmul configuration on AMD proprietary driver 26.10.1+ (llama/31587)\n19df6cfe tests: fix CI flakes of TOPK_MOE caused by near ties (llama/31426)\na0df99d7 metal: keep fusable node groups inside one command buffer (llama/31688)\n3a996dba opencl: enable f32 and q5_k multi column GEMV by default for Adreno X2 (llama/31694)\nd75184eb opencl: support non-contig sigmoid (llama/31687)\nc6345010 opencl: add DSV4_HC_PRE and DSV4_HC_POST (llama/31623)\n08161372 Title: CUDA: fuse NVFP4 gate/up matmuls and GLU in MMQ (llama/31296)\n4ac0dae8 opencl: improve decode perf with FA (llama/31633)\n4f681f28 webgpu: return maxBufferSize from get_max_size and avoid whole-buffer bindings in memset/clear (llama/31348)\ncd6d2b75 docs : sync ggml_type enum in gguf.md (#1685)\n","body_html":"<h1 id=\"ggml-v1-0-0\">ggml v1.0.0</h1>\n<h2 id=\"overview\">Overview</h2>\n<p>ggml 1.0.0 prunes the deprecated public API and the legacy <code>DIAG_MASK</code> ops, adds configurable YaRN truncation and FP8 support, and brings broad performance work across backends: few-row quantized matmuls and MoE fusion on CUDA/Metal, a Vulkan prefill FlashAttention path, a device-side MoE scheduler on SYCL, decomposed Adreno FlashAttention in OpenCL, new WebGPU kernels, expanded Hexagon op coverage, and scheduler improvements that avoid uploading unused inputs.</p>\n<h3 id=\"api-changes\">API changes</h3>\n<ul><li>Pruned deprecated API and removed the legacy <code>DIAG_MASK_INF</code>/<code>DIAG_MASK_ZERO</code> ops, bumping the RPC protocol to reject incompatible peers (<a href=\"https://github.com/ggml-org/llama.cpp/pull/29801\" rel=\"nofollow ugc noopener\">llama/29801</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31946\" rel=\"nofollow ugc noopener\">llama/31946</a>)</li><li>Added <code>ggml_rope_set_truncate</code>/<code>ggml_rope_get_truncate</code> to control YaRN correction-bound rounding (<a href=\"https://github.com/ggml-org/llama.cpp/pull/30206\" rel=\"nofollow ugc noopener\">llama/30206</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/30802\" rel=\"nofollow ugc noopener\">llama/30802</a>)</li><li>Implemented the backward pass for <code>TRI</code> (<a href=\"https://github.com/ggml-org/llama.cpp/pull/30712\" rel=\"nofollow ugc noopener\">llama/30712</a>)</li><li>Added <code>MUL_MAT</code> src1 precision tests and gating for <code>ggml_prec_set_acc</code> (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31896\" rel=\"nofollow ugc noopener\">llama/31896</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31851\" rel=\"nofollow ugc noopener\">llama/31851</a>)</li><li>Bumped the RPC protocol to v9 with faster loading and graceful failure on unreachable endpoints or failed handshakes (<a href=\"https://github.com/ggml-org/llama.cpp/pull/29413\" rel=\"nofollow ugc noopener\">llama/29413</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/30415\" rel=\"nofollow ugc noopener\">llama/30415</a>)</li></ul>\n<h3 id=\"core-changes\">Core changes</h3>\n<ul><li>Scheduler skips uploading inputs that are not read and prefetches only inputs produced by earlier splits (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31419\" rel=\"nofollow ugc noopener\">llama/31419</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/32100\" rel=\"nofollow ugc noopener\">llama/32100</a>)</li><li>MoE cache can compute a share of the cache misses on the CPU and copy the results back asynchronously (<a href=\"https://github.com/ggml-org/llama.cpp/pull/30966\" rel=\"nofollow ugc noopener\">llama/30966</a>)</li><li>Meta backend raises the compute view limit for buffers with few tensors (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31447\" rel=\"nofollow ugc noopener\">llama/31447</a>)</li></ul>\n<h3 id=\"backend-changes\">Backend changes</h3>\n<h4 id=\"cuda\">CUDA</h4>\n<ul><li>Added few-row MMQ for batch sizes 2-24 and MMVF for BF16 few-row matmuls (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31533\" rel=\"nofollow ugc noopener\">llama/31533</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31959\" rel=\"nofollow ugc noopener\">llama/31959</a>)</li><li>Added OCP FP8 support and fused NVFP4 gate/up matmuls with GLU in MMQ (<a href=\"https://github.com/ggml-org/llama.cpp/pull/28413\" rel=\"nofollow ugc noopener\">llama/28413</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31296\" rel=\"nofollow ugc noopener\">llama/31296</a>)</li><li>Extended FlashAttention with an sm70 head-size-256 config and MMA head sizes 40/72 (<a href=\"https://github.com/ggml-org/llama.cpp/pull/27997\" rel=\"nofollow ugc noopener\">llama/27997</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/30285\" rel=\"nofollow ugc noopener\">llama/30285</a>)</li><li>Improved graph handling: skip compat checks on reuse, re-instantiate on cluster-dim changes, move graph state (<a href=\"https://github.com/ggml-org/llama.cpp/pull/32004\" rel=\"nofollow ugc noopener\">llama/32004</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31284\" rel=\"nofollow ugc noopener\">llama/31284</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/32001\" rel=\"nofollow ugc noopener\">llama/32001</a>)</li><li>MoE: skip empty MMQ stream-k tiles, linearize the MoE layout, and sanitize post-bias NaNs (<a href=\"https://github.com/ggml-org/llama.cpp/pull/32067\" rel=\"nofollow ugc noopener\">llama/32067</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31047\" rel=\"nofollow ugc noopener\">llama/31047</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31005\" rel=\"nofollow ugc noopener\">llama/31005</a>)</li><li>Added a debug buffer pool and halved the VMM pool reservation with reserve retries (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31919\" rel=\"nofollow ugc noopener\">llama/31919</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/30542\" rel=\"nofollow ugc noopener\">llama/30542</a>)</li><li>HIP: use a fixed BLAS workspace and add an MFMA lightning indexer for CDNA2 (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31838\" rel=\"nofollow ugc noopener\">llama/31838</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/29050\" rel=\"nofollow ugc noopener\">llama/29050</a>)</li><li>Added a d_state=64 SSM_SCAN fast path and made radix Top-K deterministic for ties (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31818\" rel=\"nofollow ugc noopener\">llama/31818</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31084\" rel=\"nofollow ugc noopener\">llama/31084</a>)</li></ul>\n<h4 id=\"metal\">Metal</h4>\n<ul><li>Fused MoE routing and kept fusable node groups inside a single command buffer (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31615\" rel=\"nofollow ugc noopener\">llama/31615</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31688\" rel=\"nofollow ugc noopener\">llama/31688</a>)</li><li>Made radix Top-K deterministic for ties, added mul_mm for 32&lt;=K&lt;64, and switched TRI to select (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31084\" rel=\"nofollow ugc noopener\">llama/31084</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/30982\" rel=\"nofollow ugc noopener\">llama/30982</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/30711\" rel=\"nofollow ugc noopener\">llama/30711</a>)</li><li>Fixed GROUP_NORM for ne03&gt;1 and out-of-bounds accesses in the FA output and SOLVE_TRI kernels (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31852\" rel=\"nofollow ugc noopener\">llama/31852</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31662\" rel=\"nofollow ugc noopener\">llama/31662</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/32103\" rel=\"nofollow ugc noopener\">llama/32103</a>)</li></ul>\n<h4 id=\"vulkan\">Vulkan</h4>\n<ul><li>Added an Intel FlashAttention prefill kernel (<a href=\"https://github.com/ggml-org/llama.cpp/pull/29357\" rel=\"nofollow ugc noopener\">llama/29357</a>)</li><li>Added batched tensor copies, a test-fusion interface, and ReBAR detection (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31418\" rel=\"nofollow ugc noopener\">llama/31418</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/30769\" rel=\"nofollow ugc noopener\">llama/30769</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31658\" rel=\"nofollow ugc noopener\">llama/31658</a>)</li><li>Enabled dequantized KV with CM2, YaRN truncation, RDNA3/4 large-tile split-K, and large coopmat matmul on newer AMD drivers (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31526\" rel=\"nofollow ugc noopener\">llama/31526</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/30802\" rel=\"nofollow ugc noopener\">llama/30802</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/30280\" rel=\"nofollow ugc noopener\">llama/30280</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31587\" rel=\"nofollow ugc noopener\">llama/31587</a>)</li><li>Added a tiled-transpose concat shader and FWHT kernels for block widths above 512 (<a href=\"https://github.com/ggml-org/llama.cpp/pull/30591\" rel=\"nofollow ugc noopener\">llama/30591</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/29772\" rel=\"nofollow ugc noopener\">llama/29772</a>)</li><li>Fixed MoltenVK out-of-bounds accesses, GROUP_NORM workgroup limits, and FA scratch sizing (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31890\" rel=\"nofollow ugc noopener\">llama/31890</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31930\" rel=\"nofollow ugc noopener\">llama/31930</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/30487\" rel=\"nofollow ugc noopener\">llama/30487</a>)</li><li>Capped scalar FlashAttention at 8 rows for large heads on Qualcomm (<a href=\"https://github.com/ggml-org/llama.cpp/pull/30829\" rel=\"nofollow ugc noopener\">llama/30829</a>)</li></ul>\n<h4 id=\"sycl\">SYCL</h4>\n<ul><li>Added a device-side MoE scheduler (<a href=\"https://github.com/ggml-org/llama.cpp/pull/30156\" rel=\"nofollow ugc noopener\">llama/30156</a>)</li><li>Made the oneDNN FA cache thread-safe and un-gated fused SDPA for Alchemist at head_dim=64 (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31138\" rel=\"nofollow ugc noopener\">llama/31138</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31574\" rel=\"nofollow ugc noopener\">llama/31574</a>)</li><li>Cached the device id and reject buffers that would exceed VRAM (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31126\" rel=\"nofollow ugc noopener\">llama/31126</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/32013\" rel=\"nofollow ugc noopener\">llama/32013</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31285\" rel=\"nofollow ugc noopener\">llama/31285</a>)</li><li>Added exact rounding for CPY to Q4_1/Q5_1/Q8_0, strided CLAMP, and f16 SWIGLU_OAI (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31465\" rel=\"nofollow ugc noopener\">llama/31465</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31784\" rel=\"nofollow ugc noopener\">llama/31784</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/30857\" rel=\"nofollow ugc noopener\">llama/30857</a>)</li><li>Fixed AOT for allocations above 4GB (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31873\" rel=\"nofollow ugc noopener\">llama/31873</a>)</li></ul>\n<h4 id=\"opencl\">OpenCL</h4>\n<ul><li>Added decomposed Adreno FlashAttention and improved decode performance (<a href=\"https://github.com/ggml-org/llama.cpp/pull/32046\" rel=\"nofollow ugc noopener\">llama/32046</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31633\" rel=\"nofollow ugc noopener\">llama/31633</a>)</li><li>Added q6_k tiled, q8_0, and f32/q5_k multi-column GEMV kernels and f32 MUL_MAT_ID (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31758\" rel=\"nofollow ugc noopener\">llama/31758</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/30947\" rel=\"nofollow ugc noopener\">llama/30947</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31783\" rel=\"nofollow ugc noopener\">llama/31783</a>)</li><li>Added arange, non-contiguous sigmoid, repeat for f16/i32/i16/bf16, and DSV4 HC pre/post kernels (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31829\" rel=\"nofollow ugc noopener\">llama/31829</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31687\" rel=\"nofollow ugc noopener\">llama/31687</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31782\" rel=\"nofollow ugc noopener\">llama/31782</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31623\" rel=\"nofollow ugc noopener\">llama/31623</a>)</li><li>Enabled multi-column GEMV by default for Adreno X2 (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31694\" rel=\"nofollow ugc noopener\">llama/31694</a>)</li></ul>\n<h4 id=\"webgpu\">WebGPU</h4>\n<ul><li>Added batched tensor API support and returned maxBufferSize from get_max_size (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31473\" rel=\"nofollow ugc noopener\">llama/31473</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31348\" rel=\"nofollow ugc noopener\">llama/31348</a>)</li><li>Added GROUP_NORM, MEAN, COUNT_EQUAL, and PAD_REFLECT_1D kernels (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31490\" rel=\"nofollow ugc noopener\">llama/31490</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31591\" rel=\"nofollow ugc noopener\">llama/31591</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31578\" rel=\"nofollow ugc noopener\">llama/31578</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31877\" rel=\"nofollow ugc noopener\">llama/31877</a>)</li><li>Supported the YaRN truncation config (<a href=\"https://github.com/ggml-org/llama.cpp/pull/30802\" rel=\"nofollow ugc noopener\">llama/30802</a>)</li></ul>\n<h4 id=\"hexagon\">Hexagon</h4>\n<ul><li>Added Q5_0/Q5_1 support and optimized conv_2d while adding conv_3d (<a href=\"https://github.com/ggml-org/llama.cpp/pull/32094\" rel=\"nofollow ugc noopener\">llama/32094</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31814\" rel=\"nofollow ugc noopener\">llama/31814</a>)</li><li>Fused post-norm residual add into RMS_NORM_MUL and added set_tensor_n (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31822\" rel=\"nofollow ugc noopener\">llama/31822</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31744\" rel=\"nofollow ugc noopener\">llama/31744</a>)</li><li>Added automatic libhtp loading and fixed 64-bit FastRPC handle truncation (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31781\" rel=\"nofollow ugc noopener\">llama/31781</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/30952\" rel=\"nofollow ugc noopener\">llama/30952</a>)</li><li>Fixed the host build with clang-cl on Windows ARM64 (<a href=\"https://github.com/ggml-org/llama.cpp/pull/32011\" rel=\"nofollow ugc noopener\">llama/32011</a>)</li></ul>\n<h4 id=\"cpu\">CPU</h4>\n<ul><li>Switched MUL_MAT to F32 src1 and accumulate FlashAttention V in F32 (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31854\" rel=\"nofollow ugc noopener\">llama/31854</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/29810\" rel=\"nofollow ugc noopener\">llama/29810</a>)</li><li>Added SpacemiT GATED_DELTA_NET RVV and Q4_1 32x256 HP kernels and enabled IME2 MXFP4 repack (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31897\" rel=\"nofollow ugc noopener\">llama/31897</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31929\" rel=\"nofollow ugc noopener\">llama/31929</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31776\" rel=\"nofollow ugc noopener\">llama/31776</a>)</li><li>Enabled the 4x4 simd_gemm tile on ARM NEON and added bf16/q8_0 VXE to llamafile (<a href=\"https://github.com/ggml-org/llama.cpp/pull/32128\" rel=\"nofollow ugc noopener\">llama/32128</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31866\" rel=\"nofollow ugc noopener\">llama/31866</a>)</li></ul>\n<h4 id=\"cann\">CANN</h4>\n<ul><li>Fail early when a buffer type is unsupported (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31270\" rel=\"nofollow ugc noopener\">llama/31270</a>)</li></ul>\n<h4 id=\"zdnn\">ZDNN</h4>\n<ul><li>Accept non-zdnn tensors, invalidate compute tensors after each op, and fix release backend lib detection (<a href=\"https://github.com/ggml-org/llama.cpp/pull/31824\" rel=\"nofollow ugc noopener\">llama/31824</a>, <a href=\"https://github.com/ggml-org/llama.cpp/pull/31905\" rel=\"nofollow ugc noopener\">llama/31905</a>)</li></ul>\n<h4 id=\"openvino\">OpenVINO</h4>\n<ul><li>Supported the YaRN truncation config (<a href=\"https://github.com/ggml-org/llama.cpp/pull/30802\" rel=\"nofollow ugc noopener\">llama/30802</a>)</li></ul>\n<h2 id=\"more-info\">More info</h2>\n<ul><li><a href=\"https://github.com/ggml-org/ggml/discussions/1579\" rel=\"nofollow ugc noopener\">Releases and versioning of <code>ggml-org</code> projects</a></li></ul>\n<h2 id=\"changelog-since-v0-29-0\">Changelog since <a href=\"https://github.com/ggml-org/ggml/releases/tag/v0.29.0\" rel=\"nofollow ugc noopener\">v0.29.0</a></h2>\n<p>417a024b ggml : bump version to 1.0.0 (#1690)\n771b3d9c requirements : fix numpy for python 3.14 (#1691)\nf154959e sync : llama.cpp\nec04faf1 fix(hexagon): make host build work with clang-cl on Windows ARM64 (llama/32011)\nb45738e9 examples : adapt to removal of ggml_diag_mask_inf\n15144390 sync : llama.cpp\n5f295554 ggml : prune (llama/29801)\n076861cd hexagon: add Q5_0 and Q5_1 support (llama/32094)\n63ec52d4 Title: CUDA: few-row MMQ for BS 2-24 (llama/31533)\nc6fe7968 ggml : route GATED_DELTA_NET to the SpacemiT RVV kernel (llama/31897)\n4ae6e721 vulkan : Intel FA kernel for prefill. (llama/29357)\nb22071c0 ggml-cpu: use the 4x4 simd_gemm tile on ARM NEON (llama/32128)\nba8b0130 CUDA: Optimize prefil via fuse of w_s scale in epilogue MMQ for nvfp4 checkpoints (llama/26048)\n05f605a2 ggml-backend : prefetch only inputs computed by earlier splits (llama/32100)\nb8a5b871 ggml : dispatch SpacemiT Q4_1 32x256 through the HP kernel (llama/31929)\n7a8cef58 sycl: make oneDNN FA cache thread-safe and memoize runtime failures (llama/31138)\n6dfea1a1 sycl : fix AOT for more than 4GB (llama/31873)\n012ff44a sycl : use cached device-id in set_main_device (llama/32013)\n2f9d47bb CUDA: add cuda debug buffer pool (llama/31919)\n522d2575 opencl: add bin kernel <code>kernel_gemm_noshuffle_q4_0_f32_32b_trans_ila_a7_bin</code> (llama/30947)\n0f9dcb72 CUDA: skip empty MoE tiles in MMQ stream-k fixup (llama/32067)\n6bed9a83 opencl: add decomposed FA for Adreno (llama/32046)\na357e7bc metal : prevent out-of-bounds reads in SOLVE_TRI kernel (llama/32103)\nc36ccd88 vulkan: support batched tensor copy API (llama/31418)\n1bffc59d backend : re-instantiate CUDA graphs when the cluster dims of a kernel change (llama/31284)\n81531556 ROCm/HIP: fixed BLAS workspace instead of rocBLAS allocating one per call (llama/31838)\nad7d200d cpu: accumulate flash attention V in F32 (llama/29810)\n7f4a2db3 webgpu: add PAD_REFLECT_1D support (llama/31877)\n0f514548 vulkan: allow dequantized KV optimization with CM2 (llama/31526)\neb4de491 vulkan: allocate the device fault vendor binary buffer (llama/32008)\n6b7f47aa cuda : use MMVF for BF16 matrices of few rows at small batch sizes (llama/31959)\n77d01f1b cuda : add sm70 FlashAttention config case for DKQ=256, DV=256, ncols=64 (llama/27997)\nbd37ba0d cuda : fattn mma head sizes 40 72 (llama/30285)\n161ca156 CUDA: Add OCP FP8 support (llama/28413)\n06faf806 sycl : fixed sycl CLAMP to actually support strided view (llama/31784)\na986181f cuda : remove unused ids and type_src0 parameters (llama/32018)\n55104b29 llama : Add FP8 KV-cache support (llama/28392)\n0753641f ggml : add MUL_MAT src1 precision tests and gating (llama/31896)\n1cd16bd4 cuda : remove unused quantize_cuda_t typedef (llama/31884)\n4a18cdd8 cuda : halve VMM pool reservation size and retry on cuMemAddressReserve failure (llama/30542)\n1c9ac9d8 CUDA: skip graph compat check when graph is being reused (llama/32004)\n999c7250 ggml : fix FreeBSD Vulkan build by setting _XOPEN_SOURCE=700 (llama/31701)\n47bc1d84 CUDA: move graph and concurrent stream state out of common.cuh (llama/32001)\ndf70e121 cann: fail early when checking if buffer type is supported (llama/31270)\ne447eb53 sycl: Add device-side MoE scheduler (llama/30156)\n9a1c5822 vulkan: add tiled-transpose concat shader for Qwen models (llama/30591)\n8b3c8cf9 ggml-rpc : fail gracefully on unreachable endpoints and failed handshakes (llama/30415)\nf860f1dd vulkan: Divide GROUP_NORM workgroup to 2D to avoid maxComputeWorkGroupCount limit (llama/31930)\neb5571a3 vulkan: Fix out of bounds reads/writes on MoltenVK (llama/31890)\n5a9a2643 vulkan: cap scalar flash attention at 8 rows for large heads on Qualcomm (llama/30829)\n5610f41b sycl: un-gate oneDNN fused SDPA for Alchemist at head_dim=64 (llama/31574)\n487cadf8 hexagon: fix 64-bit FastRPC handle truncation on close (llama/30952)\nbe37c9cf ggml-zdnn: accept non-zdnn tensors, invalidate compute tensors after every op (llama/31824)\n294d5fc5 ggml : implement the backward pass for TRI (llama/30712)\ne18c2c60 ggml-backend: don&#39;t upload inputs that are not read (llama/31419)\ne84d36b5 metal : prevent out-of-bounds writes to the Metal FA output buffer (llama/31662)\n0dd184f8 metal : make ggml_tri select instead of scaling by the comparison (llama/30711)\nb6d482c5 opencl : include climits (llama/31935)\n62de7806 RPC: Improve loading times (llama/29413)\na7276d0a ggml : support YaRN truncation config on openvino, sycl, vulkan, webgpu, rpc (llama/30802)\nda894596 sycl: cache device-id (llama/31126)\nfb570857 tests: bound the test_rope tolerance by the f32 angle resolution (llama/31874)\ne43d8477 opencl: add f32 MUL_MAT_ID (llama/31783)\nd1c1e4b2 ggml-zdnn: fix release binary not detecting backend lib (llama/31905)\n7e7b6e91 mtmd: refactoring: remove clip.h, consolidate code into mtmd (llama/31865)\n5a0e5368 cpu: enable SpacemiT IME2 MXFP4 dense repack (llama/31776)\n7f4b25f7 llamafile: impl bf16 and q8_0 vxe (llama/31866)\nf22fc6a2 ggml : update CUDA/Metal radix Top-K to make tied selections deterministic (llama/31084)\n47b8b4fc ggml-cuda : add d_state=64 fast path to SSM_SCAN kernel (llama/31818)\nb5249606 webgpu : add COUNT_EQUAL support (llama/31578)\n75a4f5ab ggml : switch to F32 src1 in the CPU backend (llama/31854)\n6cddab2b llama, mtmd: follow the reference precision of ggml_prec_set_acc calls (llama/31851)\n7353275d opencl: improve q8_0 and MoE small-n verification for spec (llama/31831)\n8f1404ef hexagon: support for automatic loading of libhtp without legacy env vars (llama/31781)\n91c102a6 sycl: exact rounding in CPY to Q4_1, Q5_1 and Q8_0 (llama/31465)\n8cf32bc6 sycl: refuse device buffers that would exceed VRAM (llama/31285)\nd969b07e hexagon: n_hvx and n_threads cleanup (llama/31857)\n5746ebd6 llama_moe_cache: run missed experts on CPU, add tuning script (llama/30966)\na49f174d sycl: add f16 support for SWIGLU_OAI (llama/30857)\nc6a07809 metal : fix GROUP_NORM for ne03 &gt; 1 and non-contiguous inputs (llama/31852)\n1e5c3e03 webgpu : Implement group norm kernel (llama/31490)\n37ecdca3 vulkan: Add &#39;readonly&#39; to decodeBuf declarations. (llama/31833)\nc08538b3 CUDA: fix shared memory bank conflicts in the lightning indexer (llama/31803)\n07b8362f opencl: add arange (llama/31829)\nf55a5045 meta : raise the compute view limit for buffers with few tensors (llama/31447)\n28701331 hexagon: fuse post-norm residual add into RMS_NORM_MUL (llama/31822)\na135ec67 hexagon: conv_2d optimizations and support for conv_3d (llama/31814)\ne3d36332 tests : qualify isnan and signbit with std:: (llama/31791)\na7d3330a tests : override hardware_concurrency in CI (llama/31797)\nab7ca9c6 opencl: support repeat for f16, i32, i16, bf16 (llama/31782)\n9290392e webgpu: add batched tensor API (<code>get_tensor_n</code> and <code>set_tensor_n</code>) (llama/31473)\nf594fedf vulkan: detect ReBAR support, avoid BAR allocations if unavailable (llama/31658)\n38e4be2b CUDA: fix Maxwell build (llama/31787)\n8bd6b61b opencl: add q6_k multi column tiled GEMV for speculative verification and refactoring (llama/31758)\nd3ce63de hexagon: Implement set_tensor_n and improve repacking/tiling robustness (llama/31744)\n9b1880f9 ggml-cuda: sanitize post-bias NaNs in MoE selection (llama/31005)\n7e607b8f vulkan: Implement test-fusion interface (llama/30769)\n2188d132 CUDA: use highest compiled arch in get_mmvq_mmid_max_batch (llama/31679)\n114ce5ed vulkan: RDNA3/4 Large-Tile Split-K Optimization (llama/30280)\ne074347a metal : fuse DeepSeek-V4/V4.1 and GLM MoE routing (llama/31615)\nf0f5b420 CUDA/HIP: matrix-core (MFMA) lightning indexer for CDNA2 (gfx90a) (llama/29050)\nfa8c6179 CUDA: linearize MoE data layout for MMQ/MMF (llama/31047)\n1690d2e1 vulkan: FWHT kernels for block widths above 512 (llama/29772)\n40278c41 vulkan : check combined K+V f16 scratch size in FA dequant path (llama/30487)\n2dccfdb2 sync : llama.cpp\n25b845e8 ggml : support <code>truncate=false</code> for YaRN (llama/30206)\naccf2a1f sync : llama.cpp\nb2ce926c metal : use mul_mm for 32 &lt;= K &lt; 64 (llama/30982)\n1a5f4771 webgpu : add MEAN support (llama/31591)\n2ab6d6c4 vulkan: Re-enable large coopmat matmul configuration on AMD proprietary driver 26.10.1+ (llama/31587)\n19df6cfe tests: fix CI flakes of TOPK_MOE caused by near ties (llama/31426)\na0df99d7 metal: keep fusable node groups inside one command buffer (llama/31688)\n3a996dba opencl: enable f32 and q5_k multi column GEMV by default for Adreno X2 (llama/31694)\nd75184eb opencl: support non-contig sigmoid (llama/31687)\nc6345010 opencl: add DSV4_HC_PRE and DSV4_HC_POST (llama/31623)\n08161372 Title: CUDA: fuse NVFP4 gate/up matmuls and GLU in MMQ (llama/31296)\n4ac0dae8 opencl: improve decode perf with FA (llama/31633)\n4f681f28 webgpu: return maxBufferSize from get_max_size and avoid whole-buffer bindings in memset/clear (llama/31348)\ncd6d2b75 docs : sync ggml_type enum in gguf.md (#1685)</p>","headings":[{"level":1,"text":"ggml v1.0.0","id":"ggml-v1-0-0"},{"level":2,"text":"Overview","id":"overview"},{"level":3,"text":"API changes","id":"api-changes"},{"level":3,"text":"Core changes","id":"core-changes"},{"level":3,"text":"Backend changes","id":"backend-changes"},{"level":2,"text":"More info","id":"more-info"},{"level":2,"text":"Changelog since v0.29.0","id":"changelog-since-v0-29-0"}]}}