Repository navigation

All pull requests

Labels (15)Milestones (0)New pull request

Open

(57)

Closed

(886)

examples: fix greedy sampling at --temp 0 for GPT-2/GPT-J
#1678ariannamethod opened Nov 2, 2026Contributor
CUDA/HIP: add J=128 tiles for Q4_K, Q6_K, Q4_0 and Q4_1 on gfx1030
#1677q760440238 opened Nov 1, 2026
gguf: fix tensor shape overflow check mismatch and add regression test (#1668)
#1673imgolamrabbani opened Oct 26, 2026
vulkan: fix struct padding in VkDeviceQueueInfo2 for translation layers
#1672key2 opened Oct 26, 2026
ggml-rpc : fix tensor bounds checks, recv_msg length_error, and destr…
#1671imgolamrabbani opened Oct 24, 2026
ggml : fix argument order in ggml_rms_norm_back comment
#1656BruceXSK opened Oct 13, 2026
cmake: avoid linking MATH_LIBRARY-NOTFOUND
#1647hoult14 opened Sep 29, 2026
mnist : make mnist_model move-only
#1640165hh opened Sep 24, 2026
add getexecpath for get_executable_path() on OpenBSD
#1631rfht opened Sep 18, 2026
fix(cuda): include <cuda/execution> and <cuda/stream_ref> for CUB top-k path (CCCL 3.2+)
#1625ousamabenyounes opened Sep 11, 2026
simple : temporary change to trigger Vulkan crash on macOS
#1623ggerganov opened Sep 9, 2026Member
1
Fix builds on s390x
#1620WhyNotHugo opened Sep 6, 2026
9
vulkan: only request 32-wide FA subgroups where size control can grant them
#1619thewh1teagle opened Sep 4, 2026Contributor
vulkan: fall back to vkGetDeviceQueue when vkGetDeviceQueue2 returns null
#1616thewh1teagle opened Sep 4, 2026Contributor
vulkan: load the Vulkan loader at runtime instead of linking it
#1614thewh1teagle opened Sep 4, 2026Contributor
ggml-opencl: guard F16 KQ/KQV path with use_adreno_kernels (fix SD3.5 NaN)
#1613Hmission opened Sep 2, 2026
ggml-opencl: Mali tiled GEMM half-precision variant (2.86x on Mali-G925)
#1612Hmission opened Sep 2, 2026
sgemm: handle m % 4 != 0 instead of declining the whole matmul
#1604femelo opened Aug 21, 2026
sgemm: write the operand addresses so GCC will strength-reduce them on aarch64
#1603femelo opened Aug 21, 2026
sgemm: don't pick a register tile GCC will spill on aarch64
#1602femelo opened Aug 21, 2026
Fix Qualcomm HTP and Vulkan execution on Windows ARM64
#1601roozbehid opened Aug 21, 2026
Add WIN_PART/WIN_UNPART + head-size-32 attention for SAM3
#1590greenjava opened Aug 18, 2026
fix(metal): do not latch has_error on transient GPU failures
#1585olxqnz5et7g2rv opened Aug 16, 20262 tasks
Implement GGML_OP_IM2COL_3D for Metal backend
#1574venomega opened Jul 27, 2026
opencl: add Adreno xmem SDPA path
#1562happyyzy opened Jul 11, 2026Contributor

Navigation Menu

Repository navigation

All pull requests

examples: fix greedy sampling at --temp 0 for GPT-2/GPT-J

CUDA/HIP: add J=128 tiles for Q4_K, Q6_K, Q4_0 and Q4_1 on gfx1030

gguf: fix tensor shape overflow check mismatch and add regression test (#1668)

vulkan: fix struct padding in VkDeviceQueueInfo2 for translation layers

ggml-rpc : fix tensor bounds checks, recv_msg length_error, and destr…

ggml : fix argument order in ggml_rms_norm_back comment

cmake: avoid linking MATH_LIBRARY-NOTFOUND

mnist : make mnist_model move-only

add getexecpath for get_executable_path() on OpenBSD

fix(cuda): include <cuda/execution> and <cuda/stream_ref> for CUB top-k path (CCCL 3.2+)

simple : temporary change to trigger Vulkan crash on macOS

Fix builds on s390x

vulkan: only request 32-wide FA subgroups where size control can grant them

vulkan: fall back to vkGetDeviceQueue when vkGetDeviceQueue2 returns null

vulkan: load the Vulkan loader at runtime instead of linking it

ggml-opencl: guard F16 KQ/KQV path with use_adreno_kernels (fix SD3.5 NaN)

ggml-opencl: Mali tiled GEMM half-precision variant (2.86x on Mali-G925)

`sgemm`: handle `m % 4 != 0` instead of declining the whole matmul

`sgemm`: write the operand addresses so GCC will strength-reduce them on aarch64

`sgemm`: don't pick a register tile GCC will spill on aarch64

Fix Qualcomm HTP and Vulkan execution on Windows ARM64

Add WIN_PART/WIN_UNPART + head-size-32 attention for SAM3

fix(metal): do not latch has_error on transient GPU failures

Implement GGML_OP_IM2COL_3D for Metal backend

opencl: add Adreno xmem SDPA path

Pull requests