Patrick BuckleyandJohannes Gäßler 69f0d907ee ggml-cuda: native bf16 flash attention for vec kernel (llama/20525) ...

* ggml-cuda: native bf16 flash attention for vec and tile kernels

mma kernel still converts bf16 to fp16 before launch, native mma bf16 todo

* ggml-cuda: address code owner review feedback

reverted tile kernel changes to avoid larger refactor

* fix ci failures on turing and hip

* fix bf16 vec kernel compile on hip v_dot2 platforms

* add comments

---------

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>

2026-03-29 15:04:36 +03:00

..

ggml-blas

ggml-blas: set mkl threads from thread context (llama/20602)

2026-03-29 15:04:36 +03:00

ggml-cann

CANN: add BF16 support for core operators (llama/20152)

2026-03-29 15:04:36 +03:00

ggml-cpu

ggml-cpu: add always_inline to tinyBLAS_PPC accumulator saves (llama/20791)

2026-03-29 15:04:36 +03:00

ggml-cuda

ggml-cuda: native bf16 flash attention for vec kernel (llama/20525)

2026-03-29 15:04:36 +03:00

ggml-hexagon

hexagon: add Matrix Extensions (HMX) for Hexagon NPU backend (llama/20693)

2026-03-29 15:04:36 +03:00

ggml-hip

ggml-cuda: native bf16 flash attention for vec kernel (llama/20525)

2026-03-29 15:04:36 +03:00

ggml-metal

metal : add FA specialization for HSK = 320, HSV = 256 (llama/20549)

2026-03-16 13:10:15 +02:00

ggml-musa

ggml-cuda: native bf16 flash attention for vec kernel (llama/20525)

2026-03-29 15:04:36 +03:00

ggml-opencl

opencl: fix l2_norm (llama/20480)

2026-03-16 13:10:15 +02:00

ggml-openvino

ggml : add OpenVINO backend (llama/15307)

2026-03-16 13:10:15 +02:00

ggml-rpc

fix(rpc): prevent division by zero in deserialize_tensor (llama/20712)

2026-03-29 15:04:36 +03:00

ggml-sycl

ehance UPSCALE to support all UT cases (llama/20637)

2026-03-29 15:04:36 +03:00

ggml-virtgpu

ggml-virtgpu: improve the reliability of the code (llama/19846)

2026-02-27 20:57:58 +02:00

ggml-vulkan

Add shader count for Intel Arc Pro B60 (llama/20818)

2026-03-29 15:04:36 +03:00

ggml-webgpu

ggml webgpu: ops support for qwen3.5 (SET, TRI_SOLVE, SSM_CONV, GATED_DELTA_NET) + GET_ROWS optimization (llama/20687)

2026-03-29 15:04:36 +03:00

ggml-zdnn

ggml-zdnn : mark zDNN buffers as non-host (llama/18967)

2026-01-30 15:56:40 +02:00

ggml-zendnn

ggml-zendnn: update code for latest ZenDNN API (llama/19923)

2026-02-27 20:57:58 +02:00

CMakeLists.txt

ggml : add OpenVINO backend (llama/15307)

2026-03-16 13:10:15 +02:00

ggml-alloc.c

ggml : make ggml_is_view as API (llama/19539)

2026-02-27 20:57:58 +02:00

ggml-backend-dl.cpp

hexagon: enable offloading to Hexagon on Windows on Snapdragon (llama/19150)

2026-01-30 15:56:40 +02:00

ggml-backend-dl.h

hexagon: enable offloading to Hexagon on Windows on Snapdragon (llama/19150)

2026-01-30 15:56:40 +02:00

ggml-backend-impl.h

llama: use host memory if device reports 0 memory (llama/18587)

2026-01-14 09:11:59 +02:00

ggml-backend-reg.cpp

ggml : add OpenVINO backend (llama/15307)

2026-03-16 13:10:15 +02:00

ggml-backend.cpp

llama : disable graph reuse with pipeline parallelism (llama/20463)

2026-03-16 13:10:15 +02:00

ggml-common.h

ggml : add NVFP4 quantization type support (llama/19769)

2026-03-16 13:10:15 +02:00

ggml-impl.h

ggml : add NVFP4 quantization type support (llama/19769)

2026-03-16 13:10:15 +02:00

ggml-opt.cpp

…

ggml-quants.c

ggml : guard against sumq2 being 0 in IQ4_NL (llama/20460)

2026-03-16 13:10:15 +02:00

ggml-quants.h

ggml : add NVFP4 quantization type support (llama/19769)

2026-03-16 13:10:15 +02:00

ggml-threading.cpp

…

ggml-threading.h

…

ggml.c

ggml : restore ggml_type_sizef() to aboid major version bump (ggml/1441)

2026-03-18 15:18:24 +02:00

ggml.cpp

…

gguf.cpp

gguf : sync (ggml/0)

2026-02-27 20:57:58 +02:00