sync : ggml #3912
ggerganov
merged
30 commits into
master
from
sync-ggml-26-06-26
mtmd, arg: fix utf8 handling on windows (llama/24779)
9e3f81e1
ggml-webgpu: add adapter toggles for F16 on Vulkan + NVIDIA
7163f689
ggml : optimize AMX (llama/24806)
a32ed1c1
fix(hexagon): use padded stride for ssm-conv weights (llama/24470)
44aff5f6
support bf16 on bin_bcast OP and unary OPs (llama/24838)
bf0da8c0
opencl: q8_0 gemv precision improvement (llama/24923)
2c3dd7b2
ggml-webgpu: improve MTP inference by using mat-vec path for small ba…
44d75b60
vulkan: link ggml-cpu when GGML_VULKAN_CHECK_RESULTS / RUN_TESTS are …
c793db07
vulkan: make mul_mm ALIGNED a spec constant (llama/24689)
67fd5c79
vulkan: support CONV_3D (llama/24612)
f8f62c78
vulkan: Support GET_ROWS_BACK (llama/24883)
d0aaddd3
vulkan: support all backend tests for SQR/SQRT/SIN/COS/CLAMP/LEAKY_RE…
f55f4eb8
vulkan: Apply bias before softmax in FA, to avoid overflow (llama/24909)
376093c3
vulkan: fail the build when a shader fails to compile (llama/24450)
7f3f9fd1
vulkan: allow reducing the graph submission batches to avoid timeouts…
21bedb35
hexagon: MUL_MAT and MUL_MAT_ID rework : 32x32 tiled weight repack, k…
393a2c80
opencl: support non-contig rows in norm (llama/24965)
3d57322a
sycl : fix the failed UT cases of conv_3d (llama/24900)
bb6b2ae8
sycl : support --split-mode tensor (llama/24152)
c1e9f248
ggml : address integer overflows in binary ops CUDA implementation (l…
26fef3fa
CUDA: Various fixes to `cpy.cu` (llama/25000)
f8b9ba4c
opencl: flush profiling batch at shutdown for incomplete batches (lla…
8005ef82
CUDA: batch out_prod broadcast (dps2>1) path with cublasSgemmBatched …
8be1e63e
sycl : clamp softmax input to avoid underflow (llama/24941)
9f0a6b6a
ggml-cpu: fix SVE leftover path in ggml_vec_dot_f32 (llama/24699)
96e90a85
CUDA: add cublasSgemmBatched mapping for HIP/MUSA vendor headers (lla…
c3281af6
vulkan: Workaround compiler bug in conv2d coopmat2 path (llama/24924)
325c37a4
ggml : bump version to 0.15.3 (ggml/1550)
8b2bb5cb
sync : ggml
b5f276c4
talk-llama : sync llama.cpp
fa343d2d
danbev
approved these changes
on 2026-06-26
Login to write a write a comment.
Login via GitHub