Lumpiasty

Lumpiasty pushed to fable5/vram-ondemand at Lumpiasty/llama.cpp

2026-07-25 22:32:35 +00:00

961cd62ebc server: also free the compute-graph scheduler on KV eviction

Lumpiasty pushed to fable5/vram-ondemand at Lumpiasty/llama.cpp

2026-07-25 22:07:08 +00:00

9573505011 server: restore weights/KV on wake, not at decode, so KV eviction is safe

Lumpiasty pushed to main at Lumpiasty/nixos-config

2026-07-25 17:38:56 +00:00

d07be06293 add user to cdrom group

Lumpiasty pushed to main at Lumpiasty/nixos-config

2026-07-25 17:38:23 +00:00

dd446c742e Remove ntfsplus module

Lumpiasty pushed to main at Lumpiasty/nixos-config

2026-07-25 17:26:33 +00:00

be49c9858d remove python-lsp-server

26370a3d19 remove patool workaround

0e8cd39223 update networkmanager to 1.58.0

7b02a1d7f6 remove kde gear

bc72ae3d99 remove oh-my-pi

Compare 6 commits »

Lumpiasty pushed to main at Lumpiasty/nixos-config

2026-07-25 16:24:13 +00:00

ab184e2005 update nixos

42f554b371 exclude kdev-python due to build failure

1a75b67cc8 exclude calligra due to build failure

b7fa5d89c5 remove workaround for fixed openldap issue

3dbd5ca53b work around build failure of python.patool

Compare 8 commits »

Lumpiasty pushed to fable5/vram-ondemand at Lumpiasty/llama.cpp

2026-07-25 12:43:03 +00:00

deee33503b llama: guard buffer iteration against released device buffers

f2d79d89f5 server: coordinate model load with the VRAM arbiter to avoid load-time OOM

bf68c5446e server: keep the sleep-wake path active when the VRAM arbiter is enabled

7398c40eda server: optionally evict the KV cache too (Phase 2 of VRAM sharing)

be7f3b3172 server: on-demand VRAM sharing to time-share one GPU between models

Compare 6 commits »

Lumpiasty created branch fable5/vram-ondemand in Lumpiasty/llama.cpp

2026-07-25 12:43:03 +00:00

Lumpiasty pushed to fresh-start at Lumpiasty/klaster

2026-07-23 17:07:08 +00:00

d4931833f0 Merge pull request 'Update debian Docker tag to v13.6' (#365) from renovate/debian-13.x into fresh-start

e32beb1002 Update debian Docker tag to v13.6

Compare 2 commits »

Lumpiasty merged pull request Lumpiasty/klaster#365

2026-07-23 17:07:08 +00:00

Update debian Docker tag to v13.6

Lumpiasty pushed to fresh-start at Lumpiasty/klaster

2026-07-22 22:32:09 +00:00

6524a14745 increase amdgpu timeout

Lumpiasty pushed to fable5/prefetch-experts at Lumpiasty/llama.cpp

2026-07-22 19:26:29 +00:00

e5eb5edb58 docs(readme): document GCN mask_opt and fix Vulkan serving guidance

Lumpiasty pushed to fable5/prefetch-experts at Lumpiasty/llama.cpp

2026-07-22 19:04:34 +00:00

4442815c02 ggml-vulkan: enable flash-attn mask_opt for GCN large head sizes

891760faba docs(readme): document Vulkan behavior of the MoE-offload flags

1bc7c581d8 ggml-vulkan: don't assert compute_ctx empty before perf timestamp

3972da9cc9 ggml-vulkan: pre-stage host weights when import is unavailable

42e52045dc ggml-vulkan: tiled transpose fast-path for concat with transposed source

Compare 32 commits »

Lumpiasty pushed to fable5/prefetch-experts at Lumpiasty/llama.cpp

2026-07-22 18:40:54 +00:00

d6c2fd9209 ggml-vulkan: enable flash-attn mask_opt for GCN large head sizes

Lumpiasty pushed to fresh-start at Lumpiasty/klaster

2026-07-21 15:44:11 +00:00

858e121c55 Merge pull request 'Update gitea.lumpiasty.xyz/lumpiasty/supervisord:latest Docker digest to c7adeea' (#367) from renovate/gitea.lumpiasty.xyz-lumpiasty-supervisord-latest into fresh-start

edd56e84d7 Update gitea.lumpiasty.xyz/lumpiasty/supervisord:latest Docker digest to c7adeea

Compare 2 commits »

Lumpiasty merged pull request Lumpiasty/klaster#367

2026-07-21 15:44:10 +00:00

Update gitea.lumpiasty.xyz/lumpiasty/supervisord:latest Docker digest to c7adeea

Lumpiasty pushed to fresh-start at Lumpiasty/klaster

2026-07-21 15:33:14 +00:00

87b6003b23 add ca-certificates to the supervisord image

Lumpiasty pushed to fable5/prefetch-experts at Lumpiasty/llama.cpp

2026-07-21 15:17:30 +00:00

9a06d73363 docs(readme): document Vulkan behavior of the MoE-offload flags

Lumpiasty pushed to fable5/prefetch-experts at Lumpiasty/llama.cpp

2026-07-21 15:08:23 +00:00

da2888cbab ggml-vulkan: don't assert compute_ctx empty before perf timestamp

20e3cfa700 ggml-vulkan: pre-stage host weights when import is unavailable

7ac4d75e29 ggml-vulkan: tiled transpose fast-path for concat with transposed source

b82ca44fde ggml-vulkan: pin mmap CPU weights for faster H2D uploads

81edb9bde8 docs(readme): add usage + benchmark instructions for the MoE-offload optimizations

Compare 226 commits »

Lumpiasty pushed to fable5/prefetch-experts at Lumpiasty/llama.cpp

2026-07-17 01:04:18 +00:00

50eb8ade68 ggml-vulkan: pre-stage host weights when import is unavailable