Repository navigation

Sync fork to upstream master (+604 commits) — Gemma4 verified - #29

Closed

lalalune wants to merge 623 commits into

mainfrom

eliza/upstream-merge-m1

Closed

Sync fork to upstream master (+604 commits) — Gemma4 verified#29
lalalune wants to merge 623 commits into
mainfrom
eliza/upstream-merge-m1

This pull request is big! We're only showing the most recent 250 commits

Commits on Jun 5, 2026
Commits on Jun 6, 2026
Commits on Jun 7, 2026
- docker : bump cuda13 to 13.3.0 (#24228)
  CISC
  authored
  3f7c79d
- convert : fix Gemma4 with no audio encoder (#24242)
  CISC
  authored
  f71af35
- arg: Skip mmproj download when user supplied mmproj (#24239)
  konradmb
  authored
  465b1f0
- spec : fix vocab compatibility check (#24256)
  CISC
  authored
  8a091c4
- llama : add Gemma4 MTP (#23398)
  am17an
  authored
  04eb4c4
- kv-cache: follow the source cache size when sharing cells (#24267)
  ServeurpersoCom
  authored
  f0156d1
- kv-cache : avoid kv cells copies (#24277)
  ggerganov
  authored
  379ac66
- convert : fix conversion for Mistral-Medium-3.5-128B (#24268)
  dfriehs
  authored
  8a963fc
- common : relax sampler name matching (#23744)
  
  ddh0
  and
  CISC
  authored
  9e3b928
Commits on Jun 8, 2026
Commits on Jun 9, 2026
- ggml-webgpu: Add clang-format job (#24308)
  reeselevine
  authored
  3ac3c20
- Remove case for GGML_TYPE_Q4_K in mvvq.cu (#23528)
  ravel7524
  authored
  e3471b3
- ggml-cpu : fix rms_norm_back wrong output under in-place aliasing (#24305)
  
  devYRPauli
  and
  ggerganov
  authored
  fd3271e
- models : fix plamo2 attention_key/value_length regression (#24317)
  CISC
  authored
  f0152ef
- server : do not clear slots without unified KV cache (#24190)
  
  authored
  961e9a3
- ggml : add GGML_OP_COL2IM_1D (#24206)
  ServeurpersoCom
  authored
  2602169
- ui: fix mobile chat form overflow and bust stale bundle cache (#24158)
  ServeurpersoCom
  authored
  efbacf8
- server: log prompts to directory (#22031)
  
  jacekpoplawski
  and
  ngxson
  authored
  1e91256
- mtmd: refactor video subproc handling (#24316)
  
  ngxson
  and
  Noeda
  authored
  9682e35
- ui: Fix excessive style recalculation on hover (#24243)
  ntowle
  authored
  ae735b1
- vulkan: add `v_dot2_f32_f16` support in matrix-matrix multiplication and Flash Attention (#24123)
  0cc4m
  authored
  b4e3dc6
- vulkan: reduce iq1 shared memory usage for mul_mm (#24287)
  jeffbolznv
  authored
  d6d0ce8
- mtmd: build_vit batching (#24352)
  sfallah
  authored
  49f3542
- ui: add opt-in run_javascript frontend tool (#24244)
  ServeurpersoCom
  authored
  4836095
- ci : fix windows release (#24369)
  CISC
  authored
  e25a32e
- graph: Fix granite speech model inference by applying embedding scale when deepstack is not used (#24357)
  
  arnu515
  and
  ngxson
  authored
  d73cd07
- webui: implement pinned conversations support (#21387)
  
  remeh
  and
  ServeurpersoCom
  authored
  76da245
Commits on Jun 10, 2026
- speculative : fix "ngram-map-k4v" name in logging (#24253)
  ddh0
  authored
  d2e22ed
- ci : bump komac version (#24396)
  CISC
  authored
  039e20a
- CUDA: Fix ssm_scan_f32 data-races (#24360)
  ORippler
  authored
  fb83cc9
- chat: fix LFM2/LFM2.5 ignoring json_schema (#24377)
  tdakhran
  authored
  d2462f8
- Remove padding and multiple D2D copies for MTP (#24086)
  gaugarg-nv
  authored
  e95dae1
- vendor : update LibreSSL to 4.3.2 (#24397)
  angt
  authored
  ac4cdde
Commits on Jun 11, 2026
- server : skip checkpoints beyond pos_next (#24411)
  
  aldehir
  and
  ggerganov
  authored
  db94854
- vocab : refactor normalizer flags into options struct, add strip_accents (#24371)
  
  o7si
  and
  CISC
  authored
  68f3066
- vocab : adopt leading TemplateProcessing special token as BOS (#24428)
  o7si
  authored
  1bfbdb1
- server: skip unused log lines on router mode (#24463)
  ngxson
  authored
  18ef86e
- vulkan: use medium matmul tile on Asahi Linux (#24306)
  xingjianll
  authored
  1af154a
- vulkan: add fast path for contiguous buffer transfers (#23973)
  winstonma
  authored
  fdc3db9
- ggml : bump version to 0.15.0 (ggml/1539)
  ggerganov
  committed
  17e59d6
- sync : ggml
  ggerganov
  committed
  263cc04
- vulkan: ifdef eMesaHoneykrisp (build fix) (#24479)
  jeffbolznv
  authored
  4c65955
- docker : support specifying the GCC version for CUDA (#24447)
  wencan
  authored
  1593d56
Commits on Jun 12, 2026
- opencl: add q5_0/q5_1 gemm and gemv kernels for Adreno (#24319)
  
  shaofeiqi
  and
  lhez
  authored
  ba1df05
- [SYCL] Fix CI build & release for SYCL backend (#24387)
  arthw
  authored
  099ea76
- ggml: support concat for scalar types at cuda backend (#24011)
  zihaomu
  authored
  85f99dc
- spec: add EAGLE3 speculative decoding support (#18039)
  
  authored
  88a3927
- UI/jpeg exif orientation (#24196)
  ServeurpersoCom
  authored
  6471e3c
- vendor : update cpp-httplib to 0.47.0 (#24395)
  angt
  authored
  70b54e1
- ggml : bump version to 0.15.1 (ggml/1541)
  ggerganov
  committed
  e08c226
- sync : ggml
  ggerganov
  committed
  f532be8
- fit : avoid including llama-ext.h in fit.h (#24506)
  ggerganov
  authored
  02182fc
- ui: PWA support (#23871)
  allozaur
  authored
  f7ca93d
- vulkan: add pipeline barriers for memcpy read operations (#23770)
  0cc4m
  authored
  3e7bd4f
- server : fix reasoning budget WebUI precedence over model.ini (#24517)
  ggerganov
  authored
  ebc1077
- ci : unbreak release (#24544)
  CISC
  authored
  cd50446
- ci : unbreak release harder (#24545)
  CISC
  authored
  f58bad4
- mtmd: add batching API (#24384)
  ngxson
  authored
  e37abd6
Commits on Jun 13, 2026
- fix sycl links in release notes (#24527)
  muhammad-salem
  authored
  c34b922
- fit : wrap llama_device_memory_data (#24522)
  ggerganov
  authored
  d8a24cc
- server: clean up static assets handling (#24550)
  
  ngxson
  and
  aldehir
  authored
  57fe1f0
- ui: keep original file name and path (#24568)
  ngxson
  authored
  597b667
- vulkan: support non-contig unary/glu ops (#24215)
  jeffbolznv
  authored
  1a7718b
- jinja : fix split and replace with empty first arg (#24574)
  CISC
  authored
  341babc
- ui: build-time gzip compression (#24571)
  ngxson
  authored
  e8067a8
- jinja : fix negative step slice with start/stop values (#24580)
  CISC
  authored
  f05cf46
- Add arch support for cohere2-MoE (#24260)
  
  michaelw9999
  and
  CISC
  authored
  4988f6e
- ui : fix llama-ui-embed crash when no asset dir is given (#24597)
  aldehir
  authored
  53bd47e
Commits on Jun 14, 2026
Commits on Jun 15, 2026
Commits on Jun 16, 2026
Commits on Jun 17, 2026
Commits on Jun 18, 2026
Commits on Jun 19, 2026
- ggml-cpu: support K tails in power10 Q8/Q4 MMA matmul (#24753)
  
  shalinib-ibm
  and
  taronaeo
  authored
  8141e73
- server : consolidate slot selection into get_available_slot (#24755)
  ggerganov
  authored
  80452d6
- pi : remove docs from system prompt (#24791)
  ggerganov
  authored
  5bd21b8
- ggml : bump version to 0.15.2 (ggml/1548)
  ggerganov
  committed
  1868af1
- sync : ggml
  ggerganov
  committed
  5fd2dc2
- server: fix non-bound n_discard value (ctx shifting) (#24786)
  
  ngxson
  and
  ggerganov
  authored
  159d093
- spec: support eagle3 for qwen3.5 & 3.6 (#24593)
  
  ruixiang63
  and
  ggerganov
  authored
  b14e3fb
- mtmd: several bug fixes (#24784)
  ngxson
  authored
  e2e7a9b
- docker : build the UI (#24794)
  aldehir
  authored
  38724ab
- server: add --agent arg, remove redundant webui naming compat (#24801)
  ngxson
  authored
  8c2d6f6
- vendor : update cpp-httplib to 0.48.0 (#24787)
  cabelo
  authored
  0d2d9cc
- arg: Add comment line support to --api-key-file (#23168)
  kucharskim
  authored
  fabde3b
- server: remove all internal mentions about "webui" (#24817)
  ngxson
  authored
  175147e
- mtmd, arg: fix utf8 handling on windows (#24779)
  ngxson
  authored
  e475fa2
- server : optimize get_token_probabilities (#24796)
  angt
  authored
  4b48a53
- server: refactor child --> router communication (#24821)
  ngxson
  authored
  2b686a9
- ggml-webgpu: add adapter toggles for F16 on Vulkan + NVIDIA
  yomaytk
  authored
  f449e05
Commits on Jun 20, 2026
Commits on Jun 21, 2026
- common/peg : refactor until gbnf grammar generation (#24839)
  aldehir
  authored
  063d9c1
- spec : Support Step3.5/3.7 flash mtp3 (#24340)
  
  forforever73
  and
  am17an
  authored
  d789527
- minor : clean-up whitespaces (#24862)
  ggerganov
  authored
  8a118ee
- server: real-time model load progress tracking via /models/sse (#24828)
  ngxson
  authored
  d6d8995
- server: add "verbose" field to schema (#24864)
  ngxson
  authored
  bfa3219
- mtmd: add load progress callback (#24865)
  ngxson
  authored
  2f89acc
- jinja : implement call statement (#24847)
  CISC
  authored
  bf53382
- mtmd: fix mtmd_get_memory_usage (#24867)
  ngxson
  authored
  0d135df
- server: refactor batch construction (#24843)
  ngxson
  authored
  bddfd2b
- server: fix report progress for loading spec models, add "stages" list (#24870)
  ngxson
  authored
  7c082bc
- common/peg : implement ac parser for stricter grammar generation (#24869)
  aldehir
  authored
  52b3df0
Commits on Jun 22, 2026
- docs/android.md: Add dependency `libandroid-spawn` for building in termux (#21812)
  aafsmarak
  authored
  0ef6f06
- server: fix edit_file crash on append at end of file (line_start -1) (#24893)
  ServeurpersoCom
  authored
  d0f9d2e
- sampling : remove unconditional softmax+sort in top-n-sigma sampler (#22645)
  TimNN
  authored
  37957e8
- [SYCL] support bf16 on bin_bcast OP and unary OPs (#24838)
  arthw
  authored
  f8cc15f
- ui: model status and load progress via /models/sse feed (#24878)
  ServeurpersoCom
  authored
  099b579
- server: refactor/generalize input file schema (#24299)
  ngxson
  authored
  6ee0f65
- server: (router) move model downloading to dedicated process (#24834)
  ngxson
  authored
  721354f
- ui: Prioritize favorite models in model selection (#24766)
  mahdiou
  authored
  9c0ac88
- kv-cache : SWA checkpoints store only non-masked cells (#23981)
  
  ggerganov
  authored and
  lalalune
  committed
  3e81729
- feat(elizainference): multi-backend streaming-LLM FFI seam (M3) + LiteRT/MLX scaffolds (M4/M5)
  
  lalalune
  and
  claude
  committed
  c849143
- Merge remote-tracking branch 'upstream/master' into eliza/upstream-merge-m1
  lalalune
  committed
  33e7b00
- fix(merge): reconcile eliza customizations with upstream hparams refactor
  
  lalalune
  and
  claude
  committed
  04ac76a
- fix(merge): reconcile fused vision FFI with upstream mtmd/clip API changes
  
  lalalune
  and
  claude
  committed
  d0f701b
- build(cuda): Gemma-clean CUDA backend — take upstream ggml-cuda over the merged tree
  
  lalalune
  and
  claude
  committed
  8cd30ba
- build(vulkan): Gemma-clean Vulkan backend — take upstream ggml-vulkan over merged tree
  
  lalalune
  and
  claude
  committed
  7524637
Commits on Jun 24, 2026
- feat(kokoro): CrispASR ggml forward + raw-text G2P with misaki normalization
  
  lalalune
  and
  claude
  committed
  a11adb5
- fix(kokoro): force CPU backend for the StyleTTS2 forward (Vulkan/Mali garbles audio)
  
  lalalune
  and
  claude
  committed
  dc8bdc8
- fix(mtmd): drop bogus attn_mask upload for QWEN3A projector
  
  lalalune
  and
  claude
  committed
  f7406ff
Commits on Jun 25, 2026
- fix(kokoro): accept mainline 'kokoro.*' GGUF tensor names in the fused loader
  
  lalalune
  and
  claude
  committed
  126a5e4
- feat(kokoro): vendor CrispStrobe converters that produce the loader's bert.* GGUF (#9588)
  
  lalalune
  and
  claude
  committed
  021c565
- fix(kokoro): map predictor F0/N AdaIN norm tensors (fixes synthesis segfault) (#9588)
  
  lalalune
  and
  claude
  committed
  0048b9e
- fix(kokoro): cast F32 decoder F0/N conv kernels to F16 for the im2col path (#9588)
  
  lalalune
  and
  claude
  committed
  5ffdcd0
- fix(kokoro): cast all F32 conv kernels to F16 for the im2col path (#9588)
  
  lalalune
  and
  claude
  committed
  1503e35
- feat(asr): cut ASR over to Gemma-4 native (audio mmproj over shared model), remove Qwen3-ASR
  
  lalalune
  and
  claude
  committed
  788571c
- fix(kokoro): converter emits >=2-D weights as F16 (fixes mainline-format noise) (#9588)
  
  lalalune
  and
  claude
  committed
  05c29c3
- feat(litert): re-target litert-backend to the stable litert_lm_* C API + C-API embedding proof harness
  
  lalalune
  and
  claude
  committed
  988cd9d
- feat(litert): LiteRT-LM ASR path (Gemma USM audio) — C-API + proof harness
  
  lalalune
  and
  claude
  committed
  f159e15
- feat(litert): wire Gemma USM ASR live into eliza_inference_asr_transcribe (warm engine)
  
  lalalune
  and
  claude
  committed
  fd8a8a5

Commits

Commits on Jun 5, 2026

Commits on Jun 6, 2026

Commits on Jun 7, 2026

Commits on Jun 8, 2026

Commits on Jun 9, 2026

Commits on Jun 10, 2026

Commits on Jun 11, 2026

Commits on Jun 12, 2026

Commits on Jun 13, 2026

Commits on Jun 14, 2026

Commits on Jun 15, 2026

Commits on Jun 16, 2026