Vulkan perfetto#15859

Closed

walidbr wants to merge 87 commits intoggml-org:masterfrom

walidbr:vulkan_perfetto

Commits on Sep 7, 2025

Added perfetto support (CPU + Vulkan) backends
walidbr
committed
Added perfetto support (CPU + Vulkan) backends
walidbr
committed
Merge branch 'master' into vulkan_perfetto
walidbr
committed
server : removed obsolete doc (#15670)

l29ah
authored and
walidbr
committed
CANN: FIx compiler warnings (#15661)

noemotiovon
authored and
walidbr
committed
vulkan: Skip syncing for prealloc_y when it is reused (#15544)

jeffbolznv
authored and
walidbr
committed
CUDA: use FP32 arithmetic for conv2d (#15683)

JohannesGaessler
authored and
walidbr
committed
llama: use FA + max. GPU layers by default (#15434)

JohannesGaessler
authored and
walidbr
committed
Update build.md to remove MSVC arm64 notes (#15684)

slaren
authored and
walidbr
committed
ggml: update kleidiai to v1.13.0 (#15663)

chaxu01
authored and
walidbr
committed
vulkan: clamp matmul and FA results to the max finite value (#15652)

jeffbolznv
authored and
walidbr
committed
vulkan: Allow fallback to sysmem memory when vidmem is full (#15649)

jeffbolznv
authored and
walidbr
committed
vulkan : remove unused portability_enumeration_ext variable (#15679)

danbev
authored and
walidbr
committed
vulkan: mul_mat_id coopmat2 optimizations (#15546)

jeffbolznv
authored and
walidbr
committed
vulkan: handle large sizes for get_rows (#15686)

jeffbolznv
authored and
walidbr
committed
ci : explicitly set fa off or on (#15692)

CISC
authored and
walidbr
committed
llama : separate compute buffer reserve from fattn check (#15696)

slaren
authored and
walidbr
committed
llama : fix fattn reserve call n_seqs parameter (#15699)

slaren
authored and
walidbr
committed
metal : fix checks for available FA kernels (#15700)

ggerganov
authored and
walidbr
committed
server : enable /slots by default and make it secure (#15630)

ggerganov
authored and
walidbr
committed
sampling : optimize samplers by reusing bucket sort (#15665)

authored and
walidbr
committed
CANN: fix RoPE cache issue on multi-device (#15629)

hipudding
authored and
walidbr
committed
CANN: Optimize MUL_MAT_ID (#15658)

hipudding
authored and
walidbr
committed
CUDA: fix build error from ambiguous __half conversions in conv2d (#15690)

qnixsynapse
authored and
walidbr
committed
docs : add Hunyuan to models section (#15707)

DamonFool
authored and
walidbr
committed
ggml : WebGPU add TRANSPOSE and RESHAPE to supported ops (#15695)

authored and
walidbr
committed
Vulkan: Add Integer Dot Product mul_mat_vec shader for legacy quants (#14903)

0cc4m
authored and
walidbr
committed
convert : remove redundant code (#15708)

DamonFool
authored and
walidbr
committed
ggml: aarch64: Implement SVE F16 kernels for vector functions (#15115)

authored and
walidbr
committed
ggml: SVE support for exponential functions (#15145)

authored and
walidbr
committed
vulkan: disable large mmv subgroups on older Nvidia GPUs (#15717)

0cc4m
authored and
walidbr
committed
vulkan: add missing clamps in new mul_mat_id paths (#15702)

jeffbolznv
authored and
walidbr
committed
vulkan: use memory budget extension to read memory usage (#15545)

giladgd
authored and
walidbr
committed
ggml-backend: raise GGML_MAX_SPLIT_INPUTS (#15722)

JohannesGaessler
authored and
walidbr
committed
CANN: Support ext_factor in rope (#15710)

hipudding
authored and
walidbr
committed
CANN: Support eager execution mode under ACL graph compilation (#15712)

noemotiovon
authored and
walidbr
committed
opencl: add attn sinks support for FA kernels (#15706)

rmatif
authored and
walidbr
committed
vulkan: Fix macro parameter order for f32 matmul shaders (#15716)

jeffbolznv
authored and
walidbr
committed
CANN: Resolve soft_max precision issue (#15730)

hipudding
authored and
walidbr
committed
vulkan: fix shaders gen when no integer dot is available (#15740)

0cc4m
authored and
walidbr
committed
llama: -fa 1/0/-1 aliases for -fa on/off/auto (#15746)

JohannesGaessler
authored and
walidbr
committed
chore: Update `.clang-format` to use `BinPackArguments=true` (#15744)

ORippler
authored and
walidbr
committed
fix: resolve unsigned int initialization warning for n_dims/size in gguf.cpp (#15754)

skrandy
authored and
walidbr
committed
CANN: Fix type float_t to float (#15736)

noemotiovon
authored and
walidbr
committed
CANN: Mask unsupported TRANSPOSE_1D operator (#15733)

hipudding
authored and
walidbr
committed
model-conversion : add missing curl script [no ci] (#15761)

danbev
authored and
walidbr
committed
ggml-cpu : optimize RVV kernels (#15720)

xctan
authored and
walidbr
committed
CANN: Add RoPE contiguous check for 310I DUP device (#15735)

hipudding
authored and
walidbr
committed
model-conversion : remove hardcoded /bin/bash shebangs [no ci] (#15765)

danbev
authored and
walidbr
committed
llama : fix incorrect model type for Gemma 270M (#15764)

danbev
authored and
walidbr
committed
sampling : optimize dist sampler (#15704)

ggerganov
authored and
walidbr
committed
model-conversion : fix pyright errors (#15770)

danbev
authored and
walidbr
committed
CUDA: Optimize `rms_norm_f32` kernel and its fused variants, giving 1-6% perf E2E (#15715)

authored and
walidbr
committed
ggml vulkan: add hardsigmoid and hardswish operations (#15762)

relent95
authored and
walidbr
committed
vulkan : update ggml_vk_instance_validation_ext_available (#15666)

danbev
authored and
walidbr
committed
vulkan: don't use std::string in load_shaders, to improve compile time (#15724)

jeffbolznv
authored and
walidbr
committed
vulkan: fix mmv subgroup16 selection (#15775)

0cc4m
authored and
walidbr
committed
CANN: fix acl_rstd allocation size in ggml_cann_rms_norm (#15760)

authored and
walidbr
committed
opencl: add hs=40 to FA (#15758)

rmatif
authored and
walidbr
committed
CANN: Fix precision issue on 310I DUO multi-devices (#15784)

hipudding
authored and
walidbr
committed
ggml: add ops for WAN video model (cuda && cpu) (#15669)

authored and
walidbr
committed
Document the new max GPU layers default in help (#15771)

ericcurtin
authored and
walidbr
committed
server: add exceed_context_size_error type (#15780)

Xuan-Son Nguyen
authored and
walidbr
committed
CANN: Refactor ND to NZ workspace to be per-device (#15763)

authored and
walidbr
committed
llama : set n_outputs to 1 to avoid 0 outputs mean-pooling (#15791)

authored and
walidbr
committed
metal : Add template specialization for mul_mm_id w/ ne20 == 10 (#15799)

gabe-l-hart
authored and
walidbr
committed
llama : add support for EmbeddingGemma 300m (#15798)

danbev
authored and
walidbr
committed
scripts : add Jinja tester PySide6 simple app (#15756)

authored and
walidbr
committed
chat : nemotron thinking & toolcalling support (#15676)

pwilkin
authored and
walidbr
committed
chat : fixed crash when Hermes 2 <tool_call> had a newline before it (#15639)

authored and
walidbr
committed
model-conversion : add --embeddings flag to modelcard.template [no ci] (#15801)

danbev
authored and
walidbr
committed
kv-cache : fix SWA checks + disable cacheless iSWA (#15811)

ggerganov
authored and
walidbr
committed
gguf: gguf_writer refactor (#15691)

Green-Sky
authored and
walidbr
committed
tests : add --list-ops and --show-coverage options (#15745)

danbev
authored and
walidbr
committed
CUDA: fastdiv, launch bounds for mmvq + q8_1 quant (#15802)

JohannesGaessler
authored and
walidbr
committed
Implement --log-colors with always/never/auto (#15792)

ericcurtin
authored and
walidbr
committed
Thinking model disabled assistant prefill (#15404)

gabe-l-hart
authored and
walidbr
committed
ci : exempt correct research label (#15825)

CISC
authored and
walidbr
committed
aLoRA Support (#15327)

authored and
walidbr
committed
ggml-cpu: drop support for nnpa intrinsics (#15821)

taronaeo
authored and
walidbr
committed
ggml-cpu: document use of "free" memory [no ci] (#15834)

JohannesGaessler
authored and
walidbr
committed
server : implement prompt processing progress report in stream mode (#15827)

authored and
walidbr
committed
server : speed up tests (#15836)

Xuan-Son Nguyen
authored and
walidbr
committed
kleidiai: generalize compute_forward_kv_cache to compute_forward_fp16 (#15817)

chaxu01
authored and
walidbr
committed
CUDA: faster tile FA (Pascal/AMD), headsize 256 (#15769)

JohannesGaessler
authored and
walidbr
committed
Added perfetto support (CPU + Vulkan) backends
walidbr
committed
Added vulkan build workflow, x86
walidbr
committed