fd9bd632 (HEAD -> dev, origin/dev) llama-bench: fix device-info
e08e473c llama-bench:add device-info
f86112d6 add dump_pos,fix pos_tensor of pre_by_embeds
d1bc743c set u_batch = max_seq_len
2f201160 delete useless code
f131bf19 fix patch of qwen2.5vl,n_patch of minicpm = 64,fix n_ctx_slot,comment out M-ROPE
ebce9e1a (origin/runtime_replace) Refactor cal-llm integration by removing runtime capacity loading and adjusting request handling
e850ff4d Enhance cal-llm integration by adding runtime checks for CALRT_LIBRARY and updating link settings for Linux and Windows
9985688f Refactor CalrtEngineConfig initialization by removing redundant parameters
33d96dae add llama-build.sh
19210246 (refs/stash) WIP on dev: d028722a replace cparam.n_ubatch with n_ubatch()
5af229fe index on dev: d028722a replace cparam.n_ubatch with n_ubatch()
37f068d8 release
13508e5e reduce cmake log
b6b79194 reduce cmake log
38ca6895 Add support for cal-llm profile dumping to JSONL file
46847c24 Simplify CMakeLists.txt
d028722a replace cparam.n_ubatch with n_ubatch()
63442fde fix ubatch and cmakelist
ddd60c92 Add cal-llm backend profiling support with command-line option
fc3f4a9f add dump and rt case generation
c931ef31 calrt: fix prefill_by_ids.fix create_buf
27161300 fix n_ctx_slot error by adding runtime capacity loading for cal-llm
a4374124 run llama-passkey successfully, fix llama-bench, copy 3 times input to infer qwen2.5vl
b35bda12 Implement vocab-only model initialization and enhance cal-llm backend error handling
81a2cbd4 Enhance CMake configuration for cal-llm integration by updating paths and adding library properties
3b14ed55 - Removed references to CALRT from clip.h, mtmd-helper.cpp, mtmd-helper.h, mtmd.cpp, and mtmd.h. - Updated server.cpp to conditionally include CAL-LLM support. - Introduced option LLAMA_SERVER_USE_CAL_LLM in CMakeLists.txt to enable CAL-LLM integration. - Refactored server context to handle CAL-LLM requests and responses. - Updated image processing logic to remove CALRT-specific implementations. - Ensured compatibility with existing llama functionality while integrating new CAL-LLM features.
b2003d6c try passkey and embedding
7800c772 try passkey and embedding
dde8b822 clip:add minicpm patch;
fabcc7da calrt: fix encode dump
1e978796 Merge remote-tracking branch 'origin/dev-ben' into dev
f584311c calrt: fix encode dump
1bec0db5 (origin/dev-ben) bench: support 2 calbins
465df20c support one calbin
a339954c (tag: v0.3.0) Merge branch 'dev' into dev-cli
8e76fc33 finish dev of llama-cli
eeeef3f6 add unknown time info: prefill and decode
431ad6f1 update calrt_infer
efa1876b (tag: v0.2.2) update version print
4a2a3181 add calrt-version
e3ea5541 add calrt-version
0b8e4394 to debug
5eb46fff clip: add image_to_patches; calrt: vision copy data from bf16 to fp32
005112c5 clip: add image_to_patches; calrt: vision copy data from bf16 to fp32
50967064 rm debug code
5d219bfc calrt: recover base = batch_index * dict_len
6787a53a calrt:input of multimodel from fp32 to bf16
e1a76abf calrt:add CalcoreRT version print
99dd308a (tag: v0.2.1) server: fix max_seq_len
06c7852e Merge branch 'dev-ot' into dev
ed62eb33 try server-test:not success
015e988a little change
96b4e267 merge dev-ot
99962021 (origin/dev-ot) calrt: add onetoken slice and untile_prefill
9af6682e args: add -ca
e24f775e calrt: add copy_data_to_ibuf for vision model
795179be (origin/test-cicd) add ci yaml
a206460e calrt: support mixture of prefill_by_ids and prefill_by_embeds
d8b2aaa0 server: fix context-shift
b6f29ec8 server: fix context-shift
6d55eae7 (tag: v0.2.0) add annotations
af8055f5 add annotations
6f54682d Merge branch 'dev' of http://192.168.10.26:1000/yunzhe/llama.cpp into dev
42a18167 fix prompt_cache issue
a2d5bf36 server: comment fixed time
bd3a57a1 Merge branch 'dev' of http://192.168.10.26:1000/yunzhe/llama.cpp into dev
fcfcdf80 tmp fix one run >4K problem
74d437ab Merge branch 'dev' of http://192.168.10.26:1000/yunzhe/llama.cpp into dev
5813c943 add --device-info
1f0d5bb0 add --device-info
0cd44929 rm calculet0.txt
b8153b6b server: fix limit tokens to max_seq_len
82842cfc server: fix limit tokens to max_seq_len
622a2299 server: comment circle print
5f47a738 Merge branch 'dev' of http://192.168.10.26:1000/yunzhe/llama.cpp into dev
ecf01a17 fix n_parallel problem
207ca38c delet extra printf
99d2078e fix buffer resize problem
9626239f calrt: fix MapBuf
da724f3e calrt: add MapBuf
18814f9b server : reset inference time
bbee06d6 calrt: add test time print
39edc6ec (tag: v0.1.1) add t_incopy t_outcopy
2006831f (tag: v0.1.0) v0.1.0
e126c21e sever: fix commit id
5603e050 calrt: add t_incopy and t_outcopy
1e8a16cd calrt: fix slice and add infer/copy time
893e52bd delete unnecessary commit id
398ecf71 rm llama-server_old.cpp
6d5e8390 server: printf calrt commit ID
e4eaab70 calrt: prefil to prefill
aeacc23a calrt: add slice only on decode
cbaa7492 reduce warning
3c08ebf0 calrt: comment LOG_ERROR in untile_decode_cpy
613bc2b9 calrt: fix max_seq_len judgment
4e194ecc calrt: fix ubatch.embd cpy
16cbf81a calrt: fix model_name of untile_cpy
8b4e17d9 merge origin dev
6dacdd2c calrt: add multimodal
3e6ad644 calrt: add multimodal
41e6636e (tag: v0.0.1) server: correct the max_seq_len judgment
d0a1b2c7 calrt: comment out dump
b37b7d97 Merge branch 'dev' of http://192.168.10.26:1000/yunzhe/llama.cpp into dev
07817cef fix kv issue
7ddafbdc calrt: commented out cal_ctx->encode(batch)
059009ee calrt: add multimodel
1b073661 Merge branch 'dev' of http://192.168.20.229:1000/yunzhe/llama.cpp into dev
6295273c arg: update hf_repo url to https://download.calculet.tech:9443
b2c45d8d clart: past_kv_len = n_tokens
c6ea3f0d CMakeLists: STATIC to SHARED
f8fba66b fix calrt_install include path
db4a61d2 adapt to host-rt install structure
e672e02c Merge branch 'dev' of http://192.168.10.26:1000/yunzhe/llama.cpp into dev
27b82f64 adapt to calrt_objs target
7a3a9a0e calrt_infer: fix past_kv_len
583c387e caserver: lim generated tokens to n_ctx_per_seq
0edc14a6 caserver: limit generated tokens to n_ctx_per_seq
aa0a17d7 calrt: add bf16_to_fp32
fa332e77 Add bf16_to_fp32
aa54a7c6 Add bf16_to_fp32
84f2c0b5 add bf16_to_fp32 in untile_decode/prefill_cpy
0bd95719 fix get_model_by_name
365eb0b7 comment out pos buffer file
886cd1fb fix compile problem
9328f213 add prefill-only mode
49fea70a adapt to new calbin test_case
74a118df sync with calrt update
98e8f9ac add timestamp
8c8152a0 adapt to calrt csr
5d238793 add timestamp for one decode process
e6285a74 adapt to calrt
6b1ecdbb add kv_tensor for pld test
240d9bb7 adapt to calrt modification
7eeb241a fix byte_size error
dfb62500 add elf in CMakelist
e31bce25 dump src tensor
6444e227 test golden case
d81d7b19 Merge branch 'master' into dev
e492f5dc redo untile logic
1f5accb8 Fix garbled output with REPACK at high thread counts (#16956)
2759ccdb CUDA: avoid mul + bias fusion when doing fusion (#16935)
c5023daf opencl: support imrope (#16914)
e7da30b5 fix: Viewing multiple PDF attachments (#16974)
ed8aa633 model-conversion : pass config to from_pretrained (#16963)
48bd2650 server : add props.model_alias (#16943)
622cd010 ggml: CUDA: add head size 72 for flash-attn (#16962)
070ff4d5 mtmd: add --image-min/max-tokens (#16921)
bf7b0c97 mtmd: pad mask for qwen2.5vl (#16954)
fcfce040 ggml : LoongArch fixes (#16958)
ee3a5a10 sync: minja (glm 4.6 & minmax m2 templates) (#16949)
7e994168 SYCL: optimized repeat_back kernel (3× fewer asm instructions, 2× faster)Feature/sycl repeat back opt (#16869)
bcfa8762 feat(webui): improve LaTeX rendering with currency detection (#16508)
a2054e3a test-backend-ops : fix segfault in moe-expert-reduce test in support mode and coverage (#16936)
dd528680 ci : disable failing riscv cross build (#16952)
6b9a5242 model: add Janus Pro for image understanding (#16906)
2f966b8e clip : use FA (#16837)
cd5e3b57 server : support unified cache across slots (#16736)
87c9efc3 common : move gpt-oss reasoning processing to init params (#16937)
76af40aa docs: remove llama_sampler_accept reference in sampling sample usage (#16920)
7db35a79 CUDA: add FLOOR, CEIL, ROUND, TRUNC unary ops (#16917)
a864132b devops: fix failing s390x docker build (#16918)
d38d9f08 ggml: add s390x cpu-feats (#16774)
7fd205a8 scripts : add script to bench models (#16894)
2f68ce7c webui: auto-refresh /props on inference start to resync model metadata (#16784)
e4a71599 webui: add HTML/JS preview support to MarkdownContent with sandboxed iframe (#16757)
dd5e8cab vendor : update cpp-httplib to 0.27.0 (#16846)
cf659bbb mtmd: refactor preprocessing + support max/min pixels (#16878)
d8b860a2 Add a setting to display message generation statistics (#16901)
1ae74882 webui: recognize AsciiDoc files as valid text files (#16850)
961660b8 common : allow --system-prompt-file for diffusion-cli (#16903)
74fef412 codeowners : update after refactor (#16905)
5d8bb900 vulkan: Fix multi_add invalid descriptor usage (#16899)
2e76e013 vulkan: fuse mul_mat+add and mul_mat_id+add_id (#16868)
d3dc9dd8 CUDA: Remove unneded bias/gate dims in fused mmvq (#16858)
bea04522 refactor : llama-model.cpp (#16252)
0de0a015 model : Minimax M2 (#16831)
e58d5856 model : add Granite Hybrid nano types (#16896)
31c511a9 CUDA: Volta tensor core support for MMF (#16843)
6d39015a sync : ggml
4146d6a1 CUDA: add expert reduce kernel (#16857)
8da3c0e2 batch : fix consistency checks for the input positions (#16890)
c22473b5 server : don't print user inputs to console (#16871)
0f715b4e server : fix typos in server.cpp comments [no ci] (#16883)
d2d931f1 vulkan: disable spirv-opt for rope shaders (#16872)
2976b037 vulkan: Fix crash when FP16 mul_mat accumulation is not supported (#16796)
d2a2673d vulkan: fix shmem overrun in mmq id shader (#16873)
2e05afd2 add pcie support
13002a08 ggml-hexagon: respect input size when getting/setting tensor data (#16836)
6eb208d1 ci : enable free-disk-space on cuda docker build (#16877)
9984cbb6 opencl: fix boundary handling for mul_mm (#16875)
ce18efea convert : update transformers requirements (#16866)
16724b5b server : bump request URI max length to 32768 (#16862)
b52edd25 server : remove n_past (#16818)
517b7170 cpu: introduce chunking for repack matmuls and enable matmul-id chunking on ARM64 (#16833)
835e918d common: fix typo in cli help text (#16864)
d261223d model: add support for qwen3vl series (#16780)
dcca0d3a cpu: introduce chunking for flash attention (#16829)
bacddc04 model: Add support for CogVLM model (#15002)
229bf686 cuda : fix argsort with 64k+ rows (#16849)
d7395115 llama : use std::abs instead of abs (#16853)
052df28b vulkan: Handle argsort with a large number of rows (#16851)
8b11deea Hide latency of bias and gate-loading (#16847)
b9ce9401 vulkan: Fuse rope+set_rows (#16769)
3464bdac llama: fix ASAN error with M-RoPE (#16848)
e3af5563 llama: store mrope data in KV cell (#16825)
10fcc412 vulkan: Update topk_moe fusion to handle gpt's late softmax (#16656)
bcf5bda6 Vulkan MMQ Integer Dot Refactor and K-Quant support (#16536)
3eb2be1c Hexagon Op queue & dispatch optimizations (#16820)
e41bcce8 CUDA: use fastdiv in set-rows (#16834)
144a4ce8 vendor : sync minja (#16500)
f549b000 vulkan: Call ggml_vk_buffer_write_2d from ggml_vk_buffer_copy (#16793)
9a3ea685 CUDA: Fix bug in topk-moe for gpt-oss (#16821)
338074c3 sycl: add RMS_NORM_BACK operation support (#16808)
851553ea cuda: add SET operation support (#16804)
85a7d867 memory : remove KV cache size padding (#16812)
a8ca18b4 llama-bench : clarify benchmarked parts of the computation (#16823)
8284efc3 initialise buffer.device in ggml_hexagon_session (#16816)
1c1409e1 embedding: add raw option for --embd-output-format (#16541)
7a0e900e llama: consistent ctx <-> buf order for KV cache (#16746)
280d97be grammar : support array references in json schema (#16792)
3479efd1 CANN: Improve device ID handling and aclnnArange checks (#16752)
463bbf20 CUDA: add unused vars to mmvf and mmvq (#16807)
ad8d36be sycl: add SSM_CONV operation support (#16800)
c053e18a chat: Add LFM2 tool handling (#16763)
e1ab0848 mtmd : fix idefics3 preprocessing (#16806)
5a4ff43e llama : disable pipeline parallelism if compute buffer allocation fails (#16748)
10640e31 ggml : fix interpolate with align-corners and ne=1 (#16700)
80d28f10 HIP: fix AMDGPU_TARGETS, update documentation (#16803)
c55d53ac model : add LightOnOCR-1B model (#16764)
945501f5 llama: fix leaked buffers for mmap + split files (#16765)
75cbdd3f test-backend-ops: print failed tests at the end (#16785)
2b9bd9bf sycl: add ROLL operation support (#16665)
59fc1ec8 sycl: add REPEAT_BACK operation support (#16734)
75d33b93 CUDA: support for weight clamp in top-k norm (#16702)
3470a5c8 ggml-alloc : make gallocr prefer chunks that allow memory reuse (#16788)
bd562fe4 cuda : use fast copy when src and dst are of different type and contiguous (#16789)
bbac6a26 ggml: fix cuda kernel launch configuration for k_compute_batched_ptrs to support large batch (#16744)
73a48c97 convert : enable expert group selection for all models with it (#16691)
f696428c graph : add clamping to ffn_moe_weights_sum to avoid div-by-zero (#16655)
7cce4f81 model : set res->t_embd in SmallThinker models (#16782)
8d886282 docs : add Jamba to Text-only models list (#16778)
f77c13b9 CUDA: General GEMV fusion (#16715)
3cfa9c3f vulkan: deduplicate Microsoft Direct3D12 devices (#16689)
5d195f17 convert : handle mmproj filename/path properly (#16760)
226f295f model : set res->t_embd in PLaMo2 models (#16766)
f90b4a8e vulkan: delete dead code (#16732)
8423d019 vulkan: Optimize SSM_SCAN (#16645)
5cca2542 convert : avoid dequantizing mxfp4 for GPT-OSS (#16756)
55945d2e ggml: fix CUDA grid launch condition for large block_nums.y in binbcast (#16742)
0bcb40b4 CUDA: use CUB for arbitary size argsort (#16754)
69e9ff01 webui: support q URL parameter (#16728)
5a91109a model-conversion : add trust_remote_code for orig model run [no ci] (#16751)
f8f071fa convert : handle pre-quantized models (#14810)
0bf47a1d server: add memory breakdown print (#16740)
dd62dcfa convert : Make mistral-common dependency optional (#16738)
d0660f23 mtmd-cli : allow using --jinja (#16718)
fe6a9882 Manually link -lbsd to resolve flock symbol on AIX (#16610)
061f0eff ggml-cuda: use passed ops instead of hardcoded ops (#16712)
8cf6b42d server : send partial stop string when <EOG> is reached (#15007)
9de9672a sycl: use async memory allocation to fix crashes during graph recording (#16644)
63d2fc46 Add experimental ggml-hexagon backend for the Hexagon NPU (#16547)
a2e0088d Revert "ggml : Leverage the existing GGML_F32_VEC helpers to vectorize ggml_v…" (#16723)
9b9201f6 webui: introduce OpenAI-compatible model selector in JSON payload (#16562)
19a5a3ed ggml : Leverage the existing GGML_F32_VEC helpers to vectorize ggml_vec_set_f32 for faster fills (#16522)
d8eaa26e tests : fix test-thread-safety when compiling with multiple backends (#16699)
9285325c CUDA: fix bug in topk-moe softmax (#16711)
03792ad9 CUDA: topk-moe: add optional parameter for gpt-oss (#16649)
51d1a8c9 CUDA: better error for FA kernel with 0 occupancy (#16643)
4926419c ggml: add ggml_can_fuse_subgraph (#16662)
6ea37f57 opencl: fix warnings and clean up profiling (#16688)
fb349848 vulkan: Handle FA with all -inf mask values (#16447)
6de8ed75 sycl : add PAD_REFLECT_D1 operator support (#16145)
84bf3c67 model : add BailingMoeV2 support (#16063)
c9c1972e Handle legacy 'context' attachments (#16687)
b617cfd2 ggml-alloc : fix leak when reusing a tensor with a larger size (#16679)
79068501 Prevent premature submission on IME input (#16673)
0e4a0cf2 Import/Export UX improvements (#16619)
13f2cfad Enable per-conversation loading states to allow having parallel conversations (#16327)
06332e28 llama-batch: fix build fails with `-Werror=missing-braces` (#16614)
72d53e6c readme: update bindings (#16651)
2330de7b SYCL: Add support for FLOOR,CEIL,ROUND and TRUNC unary operators (#16613)
7062dd84 llama-context: only warn on pooling_type when user specified (#16674)
c20c0a5c kv: turn on common-prefix mtmd: fix output size bug
0398752d model : add Granite Hybrid types (#16635)
4f73d0a9 ci : fix binaries release failure for s390x (binaries may not work yet) (#16664)
cec5edbc ci : avoid manual updates of docs/ops.md (#16663)
fcb235b4 ci: include s390x release binaries (#16648)
55754beb CODEOWNERS: update for ggml-cuda/mmf (#16660)
ee09828c HIP: fix GPU_TARGETS (#16642)
e56abd20 vulkan: Implement topk_moe fused shader, ported from CUDA (#16641)
38355c6c CUDA: use registers instead of smem in topk-moe (#16647)
81387858 opencl: transposed gemm/gemv moe kernel with mxfp4,f32 (#16602)
66b0dbcb llama-model: fix insonsistent ctxs <-> bufs order (#16581)
41386cf3 rpc : report actual free memory (#16616)
3d4e86bb vulkan: Add State Space Model (SSM) Operations Support (#16463)
342c728d ggml : fix SpaceMit IME array out-of-bounds in task assignment (#16629)
ababae7e webui: reorganize settings layout (#16607)
b1949159 vulkan: fix debug build (add_rms_len/data not found) (#16624)
9ad4f193 metal : add `CONV_TRANSPOSE_2D` (#16542)
79967ec5 grammar : use int64_t to avoid int overflows in int schema to grammar conversion logic (#16626)
ceff6bb2 SYCL SET operator optimized for F32 tensors (#16350)
1bb4f433 mtmd : support home-cooked Mistral Small Omni (#14928)
683fa6ba fix: added a normalization step for MathJax-style \[\] and \(\) delimiters (#16599)
b22572e9 sycl : add ARANGE operator (#16362)
62425b3a kv: let seq_rm free whole seq since current hw-op cannot support common-prefix
7a50cf38 CANN: format code using .clang-format (#15863)
76a63a29 kv: fix common-prefix bug by implementing llama_memory_seq_rm
6f5d9246 common : Update the docs on -t --threads (#16236)
adc9b60f ggml-cpu: replace putenv with setenv for const-correctness (#16573)
ee50ee1e SYCL: Add GGML_OP_MEAN operator support (#16009)
7adc79c0 gguf-py : add support for endian conversion of BF16 data (#16594)
466c1911 cpu : add FLOOR, CEIL, ROUND and TRUNC unary operators (#16083)
0cb7a068 opencl: add q8_0 mm support (#16469)
d93f8439 opencl: fix FA for f32 (#16584)
f9fb33f2 Add server-driven parameter defaults and syncing (#16515)
f4ce81c4 metal: optimise `GGML_OP_SUM` (#16559)
17304cbc server : fix img token logs (#16595)
3e3cb19f llama-quant: add support for mmproj (#16592)
5acd4554 CUDA: Changing the CUDA scheduling strategy to spin (#16585)
554fd578 server : fix mtmd checkpoints (#16591)
fa882fd2 metal : avoid using Metal's gpuAddress property (#16576)
ffa05903 vulkan: Add ACC_TYPE_VEC2 implementation (#16203)
120bf704 CUDA + openCL: fix bug in accessing rms_norm->src while doing fusion (#16577)
4258e0cf vulkan: Support FA with K/V in F32 (#16543)
7ea15bb6 vulkan: Improve build time for MSVC (#16545)
9c7185dd CUDA: enable FA for FP32 KV cache (#16546)
1ee9d0b4 CUDA: use fastdiv + ggml_cuda_mad for mmvf (#16557)
48e2fa9f CUDA: add fp kernel for larger batch size MoE (#16512)
5b6913c4 cuda : remove legacy copy-op pointer indirection code (#16485)
bc07349a server : dynamic token limit for prompt cache (#16560)
e60f241e metal : FA support F32 K and V and head size = 32 (#16531)
e38b7c6e graph : support cacheless embeddings with FA and iSWA (#16528)
5016b728 opencl: fix build targeting CL 2 (#16554)
7049736b CUDA: fix numerical issues in tile FA kernel (#16540)
01d2bdc2 ggml : fix build broken with -march=armv9-a on MacOS (#16520)
56fc38b9 CANN: fix CPU memory leak in CANN backend (#16549)
1fb9504e fix: add remark plugin to render raw HTML as literal text (#16505)
b6721707 add different path for unified and seperate compilation
3f750f8d metal: add support for opt_step_sgd (#16539)
c515fc57 ggml : fix scalar path for computing norm (#16558)
f9bc66c3 CANN: Update several operators to support FP16 data format (#16251)
a31cf36a metal : add opt_step_adamw and op_sum (#16529)
81d54bbf webui: remove client-side context pre-check and rely on backend for limits (#16506)
c7be9feb [SYCL] fix UT fault cases: count-equal, argsort, pad OPs (#16521)
8415f61e ci : add Vulkan on Ubuntu with default packages build (#16532)
2c301e91 common : handle unicode during partial json parsing (#16526)
4b2dae38 common : update presets (#16504)
41aac5c6 ggml : Fix FP16 ELU positive branch (#16519)
a2fba89a hparams : add check for layer index in is_recurrent (#16511)
20cc625e ggml: Correct SVE implementation in ggml_vec_dot_f16_unroll (#16518)
11f0af55 CUDA: faster tile FA, add oob checks, more HSs (#16492)
a3cb0474 metal : fix mul-mm condition + fix mul-mv permuted kernels (#16494)
4a8fbe0a feat: render user content as markdown option (#16358)
31d0ff18 server / ranking : add sorting and management of top_n (#16403)
97870e64 cuda : avoid initializing unused devices (#16510)
477a66b0 convert : correctly handle LLaMA tokenizer for Jamba (#16470)
e60f01d9 server : fix division by zero when reporting stats (#16501)
81086cd6 vocab : mark EOT token for Granite models (#16499)
68ee98ae server : return HTTP 400 if prompt exceeds context length (#16486)
cdb6da46 server : log requests to /v1/completions (#16495)
6d69ab3f cmake : Dont define XOPENSOURCE on AIX (#16481)
1faa13a1 webui: updated the chat service to only include max_tokens in the req… (#16489)
1deee0f8 cpu : optimize the ggml NORM operation (#15953)
d00cbea6 server : host-memory prompt caching (#16391)
8328fd4b No markdown in cot (#16483)
56b47958 model-conversion : add support for SentenceTransformers (#16387)
2c0d875a ci: add ARM64 Kleidiai build and test support (#16462)
aa4711d3 CANN: Improve ACL graph matching (#16166)
d80d6d24 kleidiai: kernel interface refactoring (#16460)
b2602137 [SYCL] refactor soft_max, add soft_max_back (#16472)
e08db425 model: EmbeddingGemma Adding Support for SentenceTransformers Dense Modules (#16367)
12bbc3fa refactor: centralize CoT parsing in backend for streaming mode (#16394)
9d088284 Disable CUDA host buffers on integrated GPUs (#16308)
d2ee056e server : fix cancel pending task (#16467)
b2c08c9e metal : mark FA blocks (#16372)
7fdd16b4 server : improve context checkpoint logic (#16440)
74b8fc17 ggml webgpu: profiling, CI updates, reworking of command submission (#16452)
aeaf8a36 llama : support LiquidAI LFM2-MoE hybrid model (#16464)
df1b612e server : add `/v1/health` endpoint (#16461)
4e0388aa webui : added download action (#13552) (#16282)
ef4c5b87 presets : fix pooling param for embedding models (#16455)
c61ae20d rpc : update documentation (#16441)
0123ff38 memory : use sequential equal splits for recurrent modules (#16442)
0a319bb7 metal : add support for non-padded FA KV (#16148)
1d6092fc tests : add -INF blocks to the KQ mask in the FA tests (#16380)
8ae32dc9 metal : various optimizations + refactoring (#16446)
3df2244d llama : add --no-host to disable host buffers (#16310)
c08002a1 chat : Granite Docling stopping (#16438)
3a002afa ci : refactor sdk caching to minimize storage (#16414)
a23b9bdb ggml : fix unaligned access in AMX code (#16315)
04e632a4 ci : remove missing reranker model files (#16444)
a80ff183 ggml-cpu : fix leftover handling in ggml_vec_scale_f32 for SVE (#16443)
1d49ca37 nix : removed metal for nix (#16118)
c5fef0fc server: update readme to mention n_past_max metric (#16436)
ca71fb9b model : Granite docling + Idefics3 preprocessing (SmolVLM) (#16206)
35266573 ggml webgpu: actually add softmax, fix rms_norm offset (#16400)
86df2c9a vulkan: use a more appropriate amount of threads when generating shaders (#16418)
f3928396 rpc : check src buffer when copying tensor (#16421)
898acba6 rpc : add support for multiple devices (#16276)
e29acf74 vulkan : incremental shader builds (#16341)
128d522c chat : support Magistral thinking (#16413)
f6dcda39 server : context checkpointing for hybrid and recurrent models (#16382)
606a73f5 metal : fix loop bound in ggml_mem_ranges (#16412)
946f71ed llama : fix shapes for bert/mpt q/k norm (#16409)
638d3302 ggml : fix graph reallocation with multiple chunks (#16396)
84c8e305 Fix missing messages on sibling navigation (#16408)
2aaf0a2a vulkan: Replace uses of maxMemoryAllocationSize and VK_WHOLE_SIZE (#16354)
0e1f8385 vulkan: Fix FA coopmat1 invalid array indexing (#16365)
ad126479 ci : change macos-13 to macos-15-intel (#16401)
77233277 Capture model name only after first token (streaming) or completed request (#16405)
e308efda vulkan: in flash attention, bounds check against nem1 (don't rely on GGML_KQ_MASK_PAD) (#16316)
136bda78 webui : Fix messages payload sent to chat completions (#16402)
5113efd3 fix: track viewportHeight via window.innerHeight to avoid unwanted scrolling (#16356)
d64c8104 test-barrier : do not use more threads than physically available (#16389)
ef07a409 ggml webgpu: add support for soft_max, optimize rms_norm (#16357)
34fcc5a4 model : Apertus model implementation (#15852)
91a2a565 musa: update compile flags (#16265)
72ee736c ci : fix ubuntu-latest-cmake-rpc (disable ccache) (#16388)
f09aefaa ci: update vulkan ci (#16294)
bbd32bc0 ci : fix clean-up of old logs (#16381)
2be72c2b SYCL: Update to oneAPI 2025.2 (#16371)
95ce0985 HIP: add IMbackK to codeowner (#16375)
c8dedc99 CI: reenable cdna in rocm docker builds (#16376)
e95fec64 HIP: Disable ROCWMMA fattn on CDNA when compiled against ROCWMMA 2.0.0 (#16221)
ded67b94 llama : parameter conversion and loading fixes for PLaMo2 variants (#16075)
1fe4e38c ci: Properly install rocwmma for hip builds (#16305)
4201deae common: introduce http.h for httplib-based client (#16373)
76479927 Conversation action dialogs as singletons from Chat Sidebar + apply conditional rendering for Actions Dropdown for Chat Conversation Items (#16369)
2a9b6338 Improve code block color theming (#16325)
1104ca1a ci : use registry cache for docker builds (#16366)
4f157592 Add optional setting for showing "Model used:" information (#16337)
132d6735 vulkan: make ggml_vk_default_dispatcher support older vulkan headers (#16345)
aa9538a6 webui: Remove running `llama-server` within WebUI `dev.sh` script (#16363)
e74c92e8 model : support GLM 4.6 (make a few NextN/MTP tensors not required) (#16359)
b2ba81db ci : fix ccache key for ubuntu-cpu-cmake (#16355)
bf6f3b3a common : disable progress bar without a tty (#16352)
7c156df4 opencl: support pad_ext (#15888)
16b0ca0d Chatapi ignore empty sampling (#16330)
8d78cd26 ggml webgpu: support for rope,div,sub,glu,scale,cont operators (#16187)
d1c84a66 opencl: support ne3 in get_rows (#15866)
364a7a6d common : remove common_has_curl() (#16351)
2df5bcf3 ci : disable ccache for android (#16348)
075c0156 ggml : bump version to 0.9.4 (ggml/1363)
98c8269a server: adapt to calrt
2790ecc3 adapt to calrt
f528e97d add calrt mtmd support
a0143103 cuda : Enable CUDA Graph usage for Nemotron Nano v2 (NemotronH) (#16328)
35fb8249 metal : dynamic simdgroups for MV kernels (#16340)
3c62aed8 common : simplify etag tracking by removing json (#16342)
f1eb1cb1 kleidiai : fix work size and threads sync for fp16 (#16246)
de41f2b7 codeowners: add codeowners for opencl backend (#16344)
a74a0d69 tests: override test_set_rows::max_nmse_err to allow for occasional rounding differences (#16295)
5f7e166c Fix thinking blocks with quotes + add handling `[THINK]...[/THINK]` blocks (#16326)
d72f5f7b ci : add AMD runners and workflows (#16249)
b77e6c18 ggml: riscv: add riscv spacemit backend (#15288)
2ddd3f23 sync : ggml
4d3d455d sync : whisper.cpp (ggml/1359)
c9b1c064 ggml : remove -dev suffix from release version (ggml/1355)
b6ae75af ggml : bump version to 0.9.3 (ggml/1353)
b6dff20e ggml : prepare for development of 0.9.2-dev
2db78c75 ggml : bump version to 0.9.1
02463ab2 ggml-backend : add root cause in error message if loading backend library fails (#16172)
adc76347 ggml : check cuda and metal argsort limits and add test (#16323)
3a2bdcda Improve Mobile UI for dialogs and action dropdowns (#16222)
66bb7985 fix: preserved zero values in chat settings inputs and textareas by switching to nullish coalescing for field values and default placeholders (#16312)
2f61c0f5 llama-cli: prevent spurious assistant token (#16202)
3ffd0fae perplexity : show more kl-divergence data (#16321)
a4a0aa5e ggml : fix dependencies for ggml_set_rows (#16318)
92cd103f vulkan: Fix validation failure in quantized flash attention (#16292)
b887d2f3 ggml : fix GGML_F32_VEC_FMA argument order in ggml_vec_mad1_f32 (#16307)
bd0af02f common : fix reasoning before forced tool call via tool_choice = required (#16264)
d9e0e7c8 ci : fix musa docker build (#16306)
0124ac98 devops: switch to using ubuntu-22.04-s390x image (#16302)
2811c652 Fixed a few typos in the README of the LLaMA.cpp HTTP Server [no ci] (#16297)
d8359f5f vulkan: 64-bit im2col (#16135)
6a2c6145 metal : extend mat-mat multiplication support (#16225)
3b53634f metal : fuse non-sequential nodes (#16102)
1384abf8 vulkan: handle mat_mul with A matrix > 4GB (#16176)
e6d65fb0 vulkan: support arbitrary KV dimension in flash attention (#16160)
8656f5de vulkan : make the vulkan.hpp dynamic dispatcher instance private (#16224)
4807e8f9 Show message actions by default (#16289)
c0bfc57a CUDA: mul_mat_id for mmf for bs <= 64 for f16 and bs <= 32 for f32 (#16277)
75a3a6c2 CUDA: refactor and deduplicate vector FA kernels (#16208)
0499b29c vulkan: throw system error instead of SIGABRT during init on older devices (#16156)
234e2ff8 server : remove old LLAMA_SERVER_SSL (#16290)
3f81b4e9 vulkan: support GET_ROWS for k-quants (#16235)
ace6a545 build : add LLAMA_OPENSSL option (#16287)
72b24d96 model : make minicpm embedding_scale, residual_scale and logit_scale optional with legacy defaults (#16273)
624207e6 devops: add s390x & ppc64le CI (#15925)
807e8c6d Enhance text file detection logic for file attachments (#16199)
1a189278 Allow viewing conversations even when llama server is down (#16255)
e0539eb6 webui: switch to hash-based routing (alternative of #16079) (#16157)
5d0a40f3 Always show message actions for mobile UI + improvements for user message sizing (#16076)
d12a9836 codeowners : add rgerganov as owner of RPC [no ci] (#16279)
cc1cfa27 mtmd : fix uninitialized variable in bicubic_resize (#16275)
54dbc370 metal : report OOM errors (#16274)
b995a107 common : use cpp-httplib as a cURL alternative for downloads (#16185)
4710dd31 build : fix build-ios-device (#16257)
9b265118 ggml-cpu: implement MXFP4 SIMD for s390x (#16193)
