-
f2bb4c7f32
llama: skip the compute-buffer size check when the context is cold
fable5/vram-ondemand-rebase-20260909
LumpiastyandClaude Opus 5
2026-09-11 03:26:15 +02:00
-
48f0221c98
server: replace the VRAM doorbell with a want-lock
LumpiastyandClaude Opus 5
2026-09-10 17:45:18 +02:00
-
f9a5c231ed
scripts: add the RX 580 benchmark harness
LumpiastyandClaude Opus 5
2026-09-10 17:06:01 +02:00
-
c233ce9b51
ggml-vulkan: skip the concat transpose fast path below one tile
LumpiastyandClaude Opus 5
2026-09-10 17:06:01 +02:00
-
f9d41c711f
docs(readme): Polaris MoE profile - GCN mask_opt, the ids-readback win, tuning and dead ends
Lumpiasty
2026-09-09 00:41:59 +02:00
-
04e9aec9bd
server: on-demand mmproj - free encoder VRAM on the text path
Lumpiasty
2026-07-26 14:56:56 +02:00
-
6564834854
mtmd: release/restore the encoder weights from VRAM on demand
Lumpiasty
2026-07-26 14:56:43 +02:00
-
c49d19cc64
llama: free backend compute scratch on cold; guard memory_breakdown
Lumpiasty
2026-07-26 14:56:05 +02:00
-
06a29462c1
ggml: add ggml_backend_free_scratch to drop Vulkan compute prealloc
Lumpiasty
2026-07-26 14:55:50 +02:00
-
ec691cd027
llama: evict recurrent/SSM state on device release
Lumpiasty
2026-07-26 14:55:39 +02:00
-
0d68afc4c1
server: also free the compute-graph scheduler on KV eviction
Lumpiasty
2026-07-26 00:32:33 +02:00
-
4597050be9
server: restore weights/KV on wake, not at decode, so KV eviction is safe
Lumpiasty
2026-07-26 00:07:01 +02:00
-
857fbbfb39
llama: guard buffer iteration against released device buffers
Lumpiasty
2026-07-25 14:36:17 +02:00
-
35d35d5c31
server: coordinate model load with the VRAM arbiter to avoid load-time OOM
Lumpiasty
2026-07-25 14:36:16 +02:00
-
ad657136a2
server: keep the sleep-wake path active when the VRAM arbiter is enabled
Lumpiasty
2026-07-25 02:32:16 +02:00
-
811b4fab77
server: optionally evict the KV cache too (Phase 2 of VRAM sharing)
Lumpiasty
2026-07-25 02:02:18 +02:00
-
811eb55e12
server: on-demand VRAM sharing to time-share one GPU between models
Lumpiasty
2026-07-24 22:32:29 +02:00
-
786b4077a4
ggml-vulkan: tiled transpose fast-path for concat with transposed source
Lumpiasty
2026-07-15 23:44:42 +02:00
-
6d483cc3c1
vulkan: record why flash-attn shmem staging stays off for GCN
LumpiastyandClaude Opus 5
2026-07-27 23:56:14 +02:00
-
c2d0492791
ggml-vulkan: enable flash-attn mask_opt for GCN large head sizes
Lumpiasty
2026-07-22 20:40:34 +02:00
-
dbc449afd0
sched: skip the MoE routing-ids readback when the batch uses every expert
LumpiastyandClaude Opus 5
2026-07-27 23:56:14 +02:00
-
f3f1a8f276
llama: disable lazy tensor loading by default on iGPUs (#28326)
Ruben Ortlam
2026-09-08 18:05:09 +02:00
-
9113cc1880
ggml : fix msvc+clang ggml_vld1q_u32 (#28284)
Sarah Wu
2026-09-08 07:40:26 -07:00
-
d4389a4dd9
Revert "ggml-cuda : restore prop.integrated on HIP builds (#24233)" (#28604)
uvos
2026-09-08 16:19:53 +02:00
-
5d806aa257
server : apply checkpoint min-step eviction only when the checkpoint list is full (#28302)
Foad Abo Dahood
2026-09-08 16:01:03 +03:00
-
88ada91c18
metal : fix idle threads in mul_mv_iq3_xxs for ne00 < 1024 (#28086)
Foad Abo Dahood
2026-09-08 15:54:42 +03:00
-
415e909d84
spec: single device drafter should create meta backend wrapper (#28390)
Aman Gupta
2026-09-08 20:44:33 +08:00
-
03fa73cb27
ci : disable npm gha cache (#28600)
Sigbjørn Skjæret
2026-09-08 14:06:15 +02:00
-
1744c6bde8
ci : add PYTEST_WORKERS=1 to fix server-self-hosted job (#28603)
Daniel Bevenius
2026-09-08 13:36:03 +02:00
-
ca86fb222e
llama : add missing headers (#28566)
Pepper Gray
2026-09-08 12:59:53 +02:00
-
64e9bceb2c
vulkan : fuse UNARY(GELU|SIGMOID|SILU|SOFTPLUS) + MUL (#27220)
Ankit Khandelwal
2026-09-08 13:05:02 +05:30
-
f014bfef8b
Fix Vulkan-Hpp handle usage on 32-bit targets. (#22892)
miyan
2026-09-08 15:34:12 +08:00
-
895c045fd1
chat : split specialized parsers into common/parsers (#27764)
Piotr Wilkin (ilintar)
2026-09-08 08:29:37 +02:00
-
7d701b5929
opencl: properly handle non-contiguous inputs to conv2d (#28503)
lhez
2026-09-07 23:26:34 -07:00
-
5a6caa05fc
ggml : update ggml_prec specification (#26675)
Georgi Gerganov
2026-09-08 09:06:24 +03:00
-
9dcf84e5ae
model : support Kimi-K3 recurrent-state rollback (#28466)
Frank Dai
2026-09-07 20:31:48 -07:00
-
050dde50c9
hexagon: add RELU and LEAKY_RELU ops (#28585)
Todor Boinovski
2026-09-07 17:04:25 -07:00
-
67672dc5b7
ci : bump ty to 0.0.78 (#28548)
Sigbjørn Skjæret
2026-09-07 21:10:06 +02:00
-
f114f91f9e
tests : initialize the L2_NORM batch array (#28553)
Pascal
2026-09-07 19:54:13 +02:00
-
e71b80510c
Revert "CUDA: size routed MoE MMQ N-tiles from typical expert width on RDNA3 (#24546)" (#28551)
Piotr Wilkin (ilintar)
2026-09-07 16:28:19 +02:00
-
ccc3646c63
nix : update deprecated expressions (#28145)
Zhaolun Yin
2026-09-07 14:59:45 +01:00
-
c0b1871bc7
webgpu: format the GET_ROWS case block (#28542)
Pascal
2026-09-07 15:55:14 +02:00
-
160bd031b2
server: fix LRU hang on multiple requests same model (#28539)
Xuan-Son NguyenandPascal
2026-09-07 15:50:46 +02:00
-
dbeb37548e
sycl: add a batched L2_NORM kernel (#28222)
Titaniumtown
2026-09-07 06:24:14 -07:00
-
7a333e7240
vulkan: add DeepSeek-V4 hyper-connection fused ops (DSV4_HC_COMB/PRE/POST) (#26578)
Kevin HopperandKevin Hopper
2026-09-07 08:24:03 -05:00
-
0c963452ea
CUDA: size routed MoE MMQ N-tiles from typical expert width on RDNA3 (#24546)
ravel7524andCarl Philipp Klemm
2026-09-07 09:22:42 -04:00
-
4735997382
ggml: add gfx90c HIP support (#26454)
AuroraRAS
2026-09-07 22:21:42 +09:00
-
d23c47f2a9
convert : refactor Hy4-preview conversion - move HC tensor mapping to the global map (#28451)
fairydreamingandStanisław Szymczyk
2026-09-07 15:20:58 +02:00
-
73ab7599b5
CUDA: branchless Q4_K/Q5_K unpack to speed up mmvq, L2 prefetch on DGX Spark (#26705)
Pranesh Gonegandlaandpraneshgo
2026-09-07 11:36:58 +00:00
-
0cae43063c
vulkan: support type-aligned GET_ROWS (#28253)
Jeff BolzandAiChiTuDouPian
2026-09-07 05:22:10 -05:00
-
1173700b9c
examples : print ggml_version and ggml_commit in test-cmake [no ci] (#28538)
Daniel Bevenius
2026-09-07 12:11:40 +02:00
-
5202104b59
caps : recheck typed content if template checks for string (#28511)
Sigbjørn Skjæret
2026-09-07 09:14:32 +02:00
-
9a7570587c
convert : write explicit recurrent_layers for Qwen3-Next / Qwen3.5 (#28208)
DevVexusandClaude Opus 5
2026-09-07 02:12:50 -05:00
-
b74f590eaf
ggml-cuda: fix divergent barrier in f16 flash attention (#27870)
Siavash Norouzi
2026-09-06 23:23:21 -07:00
-
992cb503cd
ggml: allow backend inputs to not create another split (#28387)
Aman Gupta
2026-09-07 14:10:40 +08:00
-
9ac8c408a3
vulkan: rms_norm fusion opportunities (#28024)
Jeff Bolz
2026-09-07 01:08:28 -05:00
-
2092353c8b
ci : add container image checking and tagging (wip) (#28394)
Daniel Bevenius
2026-09-07 07:23:39 +02:00
-
8fe90e1fbf
vulkan: add TQ1_0 support (mm, mat-vec, mat-vec-id, dequant, get_rows) (#27765)
Anjielon
2026-09-07 06:35:30 +02:00
-
465e49b9ce
convert : add
--fuse-qkv flag to fuse Q/K/V into QKV during HF-to-GGUF conversion (#22780)
PikaPikachu
2026-09-07 00:47:05 +08:00
-
5fdfa62829
models : fix GDN normalization from
max to rsqrt (#28068)
Daniel HanandGeorgi Gerganov
2026-09-06 09:46:21 -07:00
-
3ad1ba7336
[Model] Support for Spark2_5ForCausalLM implementation (#27868)
2026-09-06 23:43:58 +08:00
-
d03efa5d53
opencl: properly choose weights pack for q4_K, q5_K mul_mat (#28402)
lhez
2026-09-06 08:33:08 -07:00
-
73a43d1f69
cuda: fixes races in mmid and mmf (#28475)
Aman Gupta
2026-09-06 19:45:01 +08:00
-
9e0e220594
grammar : fix max repetition threshold (#28469)
Aldehir Rojas
2026-09-06 03:59:10 -05:00
-
0afb805b19
ui: Improve Chat Messages rendering performance (#28460)
Aleksander Grygier
2026-09-06 10:52:40 +02:00
-
7620399f58
common: add --log-jsonl (#28437)
Xuan-Son Nguyen
2026-09-06 08:21:22 +02:00
-
c457e3bf7f
ui : embed assets directly with CMake (#28445)
Adrien Gallouët
2026-09-06 07:49:39 +02:00
-
971595d669
metal : add remaining fa-vec tunings for M2 Max (#28458)
Niklas Wenzel
2026-09-06 07:37:01 +02:00
-
74a7c897f0
Github: limit blank issues to maintainers (#28435)
Johannes Gäßler
2026-09-05 22:42:35 +02:00
-
6a1a922d26
metal : fix memory leak in early return (#28399)
Niklas Wenzel
2026-09-05 12:19:47 +02:00
-
4d9176092d
sycl : fix test-backend-ops CI break && restore Kronecker product FWHT support (#28016) (#28254)
Jingxin (Philip) Li
2026-09-05 10:37:12 +08:00
-
cd8cdf397d
sycl: attribute device allocations by site (GGML_SYCL_MEMTRACE) (#27631)
Nick Farrell
2026-09-05 12:36:02 +10:00
-
427291b5b3
metal : add remaining fa-vec tunings for M3 (#28396)
Isaac
2026-09-05 00:08:34 +05:30
-
85d5703a3b
ui : fix MCP image attachments not displayed in tool block (#25789) (#28089)
nachobh
2026-09-04 19:53:16 +02:00
-
1548a240e3
opencl: extend the elementwise and data‐movement op coverage (#27633)
Hongqiang Wang
2026-09-04 10:12:26 -07:00
-
4acf4a4cb8
opencl: add Adreno xmem SDPA path (#26331)
yzyyzyhhhandhappyyzy
2026-09-05 01:12:05 +08:00
-
8b4b3558f1
ci : move more jobs to ccache-buckets (#28375)
Sigbjørn Skjæret
2026-09-04 15:50:33 +02:00
-
1863ac0333
ui: export conversations from database instead of cached store (#27432)
Tom Tan
2026-09-04 06:13:10 -07:00
-
49c0dc82b8
model : add Tencent Hy 4 (hy_v4) preview architecture support (#28127)
2026-09-04 20:31:36 +08:00
-
5266f24da7
llama.cpp : bump version to 0.4.0 (#28386)
Georgi Gerganov
2026-09-04 15:22:38 +03:00
-
64a155d242
sync : ggml (#28379)
Georgi Gerganov
2026-09-04 14:39:19 +03:00
-
163a40796f
model, mtmd: fix gemma4 vision handling (#28335)
Xuan-Son Nguyen
2026-09-04 12:23:27 +02:00
-
8f83678fd8
metal : add remaining fa-vec tunings for M3 Max (#28373)
Niklas Wenzel
2026-09-04 11:46:31 +02:00
-
86b351fd64
ggml : replace compile definitions with version.h.in (#28364)
Daniel Bevenius
2026-09-04 10:28:23 +02:00
-
d509cb1e86
Don't use npx inside a package.json script (#28270)
Evan Huus
2026-09-04 04:27:56 -04:00
-
4cbe8b070b
ggml : don't crash when backend search path can't be read (#28271)
Adrien Gallouët
2026-09-04 09:24:06 +02:00
-
24f5bf8a41
ggml : remove GGML_CUDA_PEER_MAX_BATCH_SIZE (#28177)
Adrien Gallouët
2026-09-04 09:22:01 +02:00
-
a529af96e2
docs : update maintainer PRs link and regenerate AUTHORS (#28365)
Georgi Gerganov
2026-09-04 10:20:49 +03:00
-
38521ec33f
vendor: update BoringSSL to 0.20260903.0 (#28354)
Alessandro de Oliveira Faria (A.K.A.CABELO)
2026-09-04 04:10:26 -03:00
-
0ef4d560e1
ci : disable failing openvino tests (#28347)
Ravi Panchumarthy
2026-09-03 23:18:11 -07:00
-
c390d0abbc
common : make build info output stream configurable (#28322)
Adrien Gallouët
2026-09-04 08:13:20 +02:00
-
832fd6f174
ggml-cpu(s390x) : fix q5_1 uninitialized v_acc (#28332)
Aaron Teo
2026-09-04 13:55:50 +08:00
-
9a4843cf2f
src : add n_expert_used_max function (#28323)
Daniel Bevenius
2026-09-04 06:36:51 +02:00
-
6703d7894c
sycl: fuse rms_norm+mul+add and add+add residual chains (#27610)
Frosty40
2026-09-03 23:05:40 -05:00
-
f9f09f02cc
SYCL: Refactor GGML_SYCL_ENABLE_MKL_FA to global var (#26863)
Ozymandias_EBON
2026-09-03 21:45:53 -05:00
-
d230ddd763
llama: fix whole source code rebuilt on each new commit (#28278)
Xuan-Son Nguyen
2026-09-03 23:53:04 +02:00
-
c5a5535e6e
common/json-schema : fix GBNF grammar generation for empty object schemas (#28279)
Sergey Sklyarov
2026-09-03 22:37:50 +02:00
-
95ef7fc160
opencl: quant lm_head / decode GEMV and medium-batch GEMM optimizations (speculative decoding/MTP) (#26477)
Hongqiang Wang
2026-09-03 09:46:19 -07:00
-
8c1a25166b
tune MMVQ to MMQ crossover for SM87 (#28285)
kbenkhaled
2026-09-03 12:40:42 -04:00
-
d30500b83b
snapdragon: ci updates to use new run script (#28293)
Max Krasnyansky
2026-09-03 08:59:11 -07:00