sync : ggml #4079
ggerganov
merged
94 commits into
master
from
sync-ggml-26-09-23
HIP: fattn-mma: use fp32 accumulation on MFMA devices (llama/28576)
8245ab5c
cuda : enable i16 and i32 for DUP (llama/28897)
604ce9ba
metal : add FA kernels for HSK=96, HSV=64 (MiniCPM3) (llama/28599)
7811e766
opencl: add generic ssm_scan (llama/28881)
1e010e13
OpenVINO: optimize stateful decode and GPU MoE inference (llama/28638)
331102cd
vulkan: support sparse Flash Attention (llama/28105)
bbba9c77
cuda: support row-contiguous SUM_ROWS (llama/26308)
6f3b8df6
rpc : hash-cache only weights (llama/28789)
debd5418
opencl: choose the MoE expert matmul by batch size for speculative de…
8a268e56
HIP: Enable AllReduce for ROCm (llama/27825)
ab86a282
hex-cpy: use dma if src and dst are contiguous (llama/28906)
663fbdd7
hexagon: add back missing contiguous fast-path and hvx_copy_uu for ea…
b09da3ab
metal: fix NaN in mul_mm_id when activations exceed f16 range (llama/…
d7b4425d
vulkan: make MUL_MAT_ID BN/2 tail unconditional (llama/28923)
664d7278
HIP: broaden MoE ncols_opt tile heuristic on RDNA3.5 architecture (ll…
bedc0158
qwen4exp: add hc ops (llama/28901)
d8bb428e
rpc : invalidate cached compute graph when a referenced buffer is fre…
7cdbc81a
spacemit : fix wrong transpose function for int16 data (llama/25161)
b8de1249
CUDA/HIP: improve access patterns in im2col (llama/28013)
ada4e50e
hexagon: accept the zeroed rope probe in supports_op (llama/28995)
c7eb1023
hexagon: Support for K-Quants Q4_K and Q6_K (llama/28994)
0b1d7e8a
vulkan: work around NV bug with argsort_large.comp (llama/28975)
007b813a
Fix function signature for `ggml_backend_sycl_split_buffer_type` (lla…
4b384f79
vulkan: support qwen4exp hc ops (llama/28988)
6182e7a9
vulkan: fix buffer_reference alignment in im2col shaders (llama/28996)
00d6373d
opencl: fix various warnings (llama/28984)
71c60ca0
sycl: ssm_conv: fuse the SiLU epilogue into the ssm_conv kernel (llam…
ea57fef3
vulkan: skip unneeded MoE work in mul_mm coopmat1 path (llama/25483)
f3e74c43
sycl : fix the B70 mem allocate error when >19.3GB (llama/28953)
e75bc29d
gguf : align the data section relative to the GGUF start, not the fil…
d1f6be42
vulkan: split buffers and debug code into separate files, add shared …
2e9b3027
openvino : Update OpenVINO to 2026.4;fix clangd,MSVC warnings; (llama…
8df72174
vulkan: raise the hoisted row-id limit for mul_mat_id from 256 to 512…
e1430a2c
vulkan: add IQ3_S MMQ matmul kernels (llama/28822)
6df62265
ggml : handle graph buffer reservation failure (llama/26070)
67d41b06
ggml-webgpu: fix supports_op condition for GET_ROWS (llama/28978)
595ba7a8
ggml-cpu: add F16 input to the FWHT (llama/27779)
5ab3eafc
opencl: add bin kernel `kernel_gemm_noshuffle_q6_k_f32_32b_trans_ila_…
1b410326
hexagon: HMX flash-attention head_dim padding (support DK=DV=72) (lla…
1b8fa2df
hexagon: im2col update (llama/29103)
ba559d48
hexagon: add ROLL op support (llama/29105)
6268533a
opencl: add support for bin kernel `flash_attn_f32_f16_bin` (llama/29…
145fb2df
cuda : fix CUB argsort corruption caused by in-place keys (llama/28389)
91ac4e72
metal : support qwen4exp hc ops (llama/29000)
f28c20b5
metal : fix FA support checks (llama/29122)
4d69a3eb
metal : add MoE and SSM_CONV fusion optimizations (llama/28948)
5abf283f
hexagon: enable support for TOP_K op (llama/29113)
6ff9477b
hexagon: add support for GEGLU_QUICK (llama/29114)
6490d228
hexagon: enable I32 GET_ROWS (llama/29116)
57dab1bd
metal: add F16 input to the FWHT (llama/29094)
6f784084
CUDA: enable sparse fa for qwen4 (llama/28770)
e2d4ad1e
metal : support arbitrary hc in dsv4_hc_pre (llama/29169)
c5f3e52c
CUDA: tune FA for Gemma 4 on Ampere or newer (llama/29152)
f73f729c
webgpu : add fused gdn + cpy (llama/28976)
2b72de19
metal : fix deprecation warnings from macOS 27 SDK (llama/29136)
742d1e96
cuda : tune MMVQ to MMQ crossover for SM70 (Volta) (llama/28912)
9a08af87
hexagon: overhaul of buffer and DMA handling to support 64bit mapping…
ad7fbadc
ggml-cpu: ARM Repack kernels for Q1_0 (llama/23492)
a28a5da8
sycl : coalesce MKL-FA softmax loads instead of one work-item per row…
fb6f1a37
ggml-metal : simplify fusion pattern op list declaration (llama/29206)
9ebc4501
CUDA: Follow up of #25635, refactoring FA shared smem swizzle (llama/…
67864cd7
sycl : pinned memory use right device context instead of 0 (llama/28895)
3cb8360f
sycl : support gated DSV4_HC_PRE and optional HC_POST comb matrix (ll…
329344f5
ggml : fix dimension and stride truncation in ggml_permute (llama/29227)
73e32344
cuda : accelerate conv2d with implicit GEMM (llama/29135)
9c0d0989
ggml-cuda : convert contiguous tensors four elements at a time (llama…
f912142b
cuda: fix sm_70 tile compilation error (llama/29224)
c0cd0c7a
metal : fix mask bounds in flash attention block pre-pass (llama/29220)
f981add3
hexagon: new HMX-optimized GATED_DELTA_NET (llama/29199)
87d71af1
opencl: add A8 Q4_0 non-MoE dp4a binary kernel (llama/29055)
9aeb8174
Performance tune for gemma4-26b-a4b flash attention shape. (llama/28450)
2e8ae14e
ggml : IQ1_M build prefix sums once per block (llama/28706)
54da0136
metal : gate mul_mm_id src1 rescale behind ggml_prec (llama/29029)
71cb54df
vulkan: add Intel Xe flash attention optimization kernels (2/3, Xe-LP…
dc682804
opencl: add bin kernel `kernel_gemm_noshuffle_q4_k_q8_1_dp4a_ila_a8_b…
bd5b1661
HIP : optimize IQ2/IQ3 (`__vsub4` `__vcmpne4`) using SWAR (llama/27962)
637e2208
hex-dma: introduce direct-mapped DMA cache that is better suited for …
c650515a
vulkan: hide internal symbols to prevent duplicate-dlopen state destr…
3de8d470
sycl : support op get_rows_back, only support fp32/fp16 (llama/25266)
7670dbac
sycl: extend MMVQ GLU fusion, add rms_norm+scale and ssm_conv+silu fu…
69fe6264
sycl : support new UT case for mul_mat_hadamard fp16 (llama/29218)
8725aba0
cuda: top-k MoE should always fire (llama/28432)
259d7732
ggml-meta: resolve multi buffer views (llama/29266)
4175763a
sycl : fix compile warnings
31bd3a9e
common : fix for two functions when top_k exceeds the vocabulary size…
d8afa650
ggml : bump version to 0.25.0 (ggml/1635)
cf511d90
sync : ggml
a0622519
danbev
approved these changes
on 2026-09-23
vulkan: add IQ4_XS MMQ/MMV matmul kernels (llama/28415)
ff21d19c
metal : key the fa-vec tuned table by family instead of SKU (llama/29…
bdb954d4
CUDA: enable sparse-fa for dsv4 prefill (again) (llama/29298)
4edbc0b5
metal: add the missing f32 x bf16 mul_mv variants (llama/28741)
882ac361
CUDA: add a reserve to avoid spurious warning on older GCC builds (ll…
9619f20d
ggml : bump version to 0.25.1 (ggml/1637)
210bc823
sync : ggml
de89e82a
Login to write a write a comment.
Login via GitHub