{"entries":[{"name":"cogvlm.cpp","mode":"0100644","size":3027,"lastCommit":{"hash":"1e645345702154ba4813d3d9bbdbd97718de82c0","oid":"1e645345702154ba4813d3d9bbdbd97718de82c0","parents":["cd708db0ccd59e0770b5ff655db55c995b6fd174"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-03-19T13:11:39+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-03-19T13:11:39+01:00"},"message":"mtmd: add clip_graph::build_mm() (#20751)\n\n* clip: add build_mm()\n\n* apply to all models\n\n* add TODO for bias overload","tree":"199f6c4e93e7126cb52a6a91ed19e502fd794cfa"}},{"name":"conformer.cpp","mode":"0100644","size":9668,"lastCommit":{"hash":"f5c6ae18278b75712fc992587d9ea527d4ea2218","oid":"f5c6ae18278b75712fc992587d9ea527d4ea2218","parents":["5a69c974392020e514c3b2b2910bb92f847cb4c9"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-06-06T11:06:51+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-06T11:06:51+02:00"},"message":"mtmd, server: add \"placeholder bitmap\" for counting tokens , add */input_tokens API (#23913)\n\n* mtmd: add \"placeholder bitmap\" for counting tokens w/o preprocessing\n\n* fast path skip preproc for placeholder\n\n* fix build\n\n* correct the api\n\n* add server endpoint + tests\n\n* add object name\n\n* update docs\n\n* add proxy handling\n\n* fix build\n\n* fix audio input path\n\n* use is_placeholder in process_mtmd_prompt()\n\n* nits\n\n* nits (2)\n\n* docs: clarify chat/completions/input_tokens is not official\n\n* fix merge problem","tree":"f14db5652990fbaf3c81bc03a91f0bc27b98c8bd"}},{"name":"deepseekocr.cpp","mode":"0100644","size":18629,"lastCommit":{"hash":"9731ad3f29da96f588711a0d1eb08cf210721e16","oid":"9731ad3f29da96f588711a0d1eb08cf210721e16","parents":["af5172627d3513a7efed526b206dca9cd6536452"],"author":{"name":"Saba Fallah","email":"sabafallah@gmail.com","when":"2026-08-18T19:05:19+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-18T19:05:19+02:00"},"message":"mtmd: deepseek-ocr SAM ggml_conv_2d with the im2col kept in F32 (#26727)","tree":"0270adf4ab3bd40ee89b767e40832031b9e134d8"}},{"name":"deepseekocr2.cpp","mode":"0100644","size":3172,"lastCommit":{"hash":"717dad5c8e9652cb58393de8287dd6668cb8c26d","oid":"717dad5c8e9652cb58393de8287dd6668cb8c26d","parents":["9a688e51e601199ad530115b9ce0bf6a7e71d75f"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-08-05T13:34:52+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-05T13:34:52+02:00"},"message":"mtmd: support multi-row batching for deepseek-ocr (#26154)\n\n* mtmd: support multi-row batching for deepseek-ocr\n\n* mtmd: weave deepseek-ocr rows in one shot instead of per row (#26615)\n\n---------\n\nCo-authored-by: Saba Fallah ","tree":"97810ea9d8240c53dbcf2e84b17daf06427ca6ac"}},{"name":"dots3note.cpp","mode":"0100644","size":2437,"lastCommit":{"hash":"54ee5ee643f29abba6852903ddfdb688c2361b5b","oid":"54ee5ee643f29abba6852903ddfdb688c2361b5b","parents":["3a653fea932e6a026e14cc0fb6507a986679f688"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-08-22T10:35:50+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-22T10:35:50+02:00"},"message":"mtmd: support dots3-note vision+audio (#27524)\n\n* text: conversion\n\n* init impl\n\n* mtmd: conversion\n\n* impl mtmd cpp\n\n* Update gguf-py/gguf/tensor_mapping.py\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"e6eabd88f99116b32a04cde3a7d0fc0d35238806"}},{"name":"dotsocr.cpp","mode":"0100644","size":1699,"lastCommit":{"hash":"501aeed18f54716eacc7bd6397376b7829602084","oid":"501aeed18f54716eacc7bd6397376b7829602084","parents":["0ec191e1d761a4daec873dfd73278ee356944fe5"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-04-09T12:16:38+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-04-09T12:16:38+02:00"},"message":"mtmd: support dots.ocr (#17575)\n\n* convert gguf\n\n* clip impl\n\n* fix conversion\n\n* wip\n\n* corrections\n\n* update docs\n\n* add gguf to test script","tree":"9212a689a9609e2f94e4ec1c0c344735e9b36ddd"}},{"name":"exaone4_5.cpp","mode":"0100644","size":6409,"lastCommit":{"hash":"f5c6ae18278b75712fc992587d9ea527d4ea2218","oid":"f5c6ae18278b75712fc992587d9ea527d4ea2218","parents":["5a69c974392020e514c3b2b2910bb92f847cb4c9"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-06-06T11:06:51+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-06T11:06:51+02:00"},"message":"mtmd, server: add \"placeholder bitmap\" for counting tokens , add */input_tokens API (#23913)\n\n* mtmd: add \"placeholder bitmap\" for counting tokens w/o preprocessing\n\n* fast path skip preproc for placeholder\n\n* fix build\n\n* correct the api\n\n* add server endpoint + tests\n\n* add object name\n\n* update docs\n\n* add proxy handling\n\n* fix build\n\n* fix audio input path\n\n* use is_placeholder in process_mtmd_prompt()\n\n* nits\n\n* nits (2)\n\n* docs: clarify chat/completions/input_tokens is not official\n\n* fix merge problem","tree":"f14db5652990fbaf3c81bc03a91f0bc27b98c8bd"}},{"name":"gemma4a.cpp","mode":"0100644","size":12908,"lastCommit":{"hash":"547765a93e5ad7b4e8ca84d78f6d83f36ad8ee25","oid":"547765a93e5ad7b4e8ca84d78f6d83f36ad8ee25","parents":["9e209c5aee8effac146463c8dc32984a4b4d2672"],"author":{"name":"Stephen Cox","email":"scox@ict.co","when":"2026-04-13T00:15:26+12:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-04-12T14:15:26+02:00"},"message":"mtmd: add Gemma 4 audio conformer encoder support (#21421)\n\n* mtmd: add Gemma 4 audio conformer encoder support\n\nAdd audio processing for Gemma 4 E2B/E4B via a USM-style Conformer.\n\nArchitecture:\n- 12-layer Conformer: FFN → Self-Attention → Causal Conv1D → FFN → Norm\n- Subsampling Conv Projection: 2x Conv2D(stride=2) with LayerNorm\n- Full self-attention with sinusoidal RPE and sliding window mask (24)\n- Logit softcapping at 50.0, ClippableLinear clamping\n- Output: 1024 → 1536 → RMSNorm → multimodal embedder\n\nMel preprocessing (dedicated mtmd_audio_preprocessor_gemma4a):\n- HTK mel scale, 128 bins, magnitude STFT, mel_floor=1e-3\n- Standard periodic Hann window (320 samples), zero-padded to FFT size\n- Semicausal left-padding (frame_length/2 samples)\n- Frame count matched to PyTorch (unfold formula)\n- No pre-emphasis, no Whisper-style normalization\n- Mel cosine similarity vs PyTorch: 0.9998\n\nKey fixes:\n- Tensor loading dedup: prevent get_tensor() from creating duplicate\n entries in ctx_data. Fixed with std::set guard.\n- ClippableLinear clamp_info loading moved after per-layer tensors.\n- Sliding window mask (24 positions) matching PyTorch context_size.\n- Skip Whisper normalization for Gemma4 mel output.\n\nTested on E2B and E4B with CPU and Vulkan backends.\nTranscribes: \"Glad to see things are going well and business is starting\nto pick up\" (matching ground truth).\n\nRef: #21325","tree":"e13047a27afae3a73383d6f471a185dae0c0ca80"}},{"name":"gemma4ua.cpp","mode":"0100644","size":485,"lastCommit":{"hash":"a731805cedc83c0514cbd808a2e38ec46c759cc2","oid":"a731805cedc83c0514cbd808a2e38ec46c759cc2","parents":["ee4cf705bb6cdc66ed0db7343e665ee58c7e5825"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-06-03T17:10:35+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-03T17:10:35+02:00"},"message":"mtmd, model: allow skip build_vit() (#24077)\n\n* add model\n\n* nits","tree":"7a7382d113f84687d60939a704628389279f35d4"}},{"name":"gemma4uv.cpp","mode":"0100644","size":2819,"lastCommit":{"hash":"a731805cedc83c0514cbd808a2e38ec46c759cc2","oid":"a731805cedc83c0514cbd808a2e38ec46c759cc2","parents":["ee4cf705bb6cdc66ed0db7343e665ee58c7e5825"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-06-03T17:10:35+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-03T17:10:35+02:00"},"message":"mtmd, model: allow skip build_vit() (#24077)\n\n* add model\n\n* nits","tree":"7a7382d113f84687d60939a704628389279f35d4"}},{"name":"gemma4v.cpp","mode":"0100644","size":4888,"lastCommit":{"hash":"b21e4de74567f5eef213765c9476a843c2e43f0d","oid":"b21e4de74567f5eef213765c9476a843c2e43f0d","parents":["d9f918d2d06079b4336688e819eee821c8a9cd9e"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-08-22T16:33:47+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-22T16:33:47+02:00"},"message":"mtmd: use ggml_rope_set_offset (#27521)\n\n* mtmd: use ggml_rope_set_offset\n\n* add comment","tree":"612d00b0e6ee334c61cbdff8c05eabf45eb8a98a"}},{"name":"glm4v.cpp","mode":"0100644","size":4592,"lastCommit":{"hash":"f5c6ae18278b75712fc992587d9ea527d4ea2218","oid":"f5c6ae18278b75712fc992587d9ea527d4ea2218","parents":["5a69c974392020e514c3b2b2910bb92f847cb4c9"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-06-06T11:06:51+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-06T11:06:51+02:00"},"message":"mtmd, server: add \"placeholder bitmap\" for counting tokens , add */input_tokens API (#23913)\n\n* mtmd: add \"placeholder bitmap\" for counting tokens w/o preprocessing\n\n* fast path skip preproc for placeholder\n\n* fix build\n\n* correct the api\n\n* add server endpoint + tests\n\n* add object name\n\n* update docs\n\n* add proxy handling\n\n* fix build\n\n* fix audio input path\n\n* use is_placeholder in process_mtmd_prompt()\n\n* nits\n\n* nits (2)\n\n* docs: clarify chat/completions/input_tokens is not official\n\n* fix merge problem","tree":"f14db5652990fbaf3c81bc03a91f0bc27b98c8bd"}},{"name":"granite-speech.cpp","mode":"0100644","size":12964,"lastCommit":{"hash":"a3900a669419e38f4e5f13a1a773857a38fd4cb1","oid":"a3900a669419e38f4e5f13a1a773857a38fd4cb1","parents":["7c908502ea0868e6ae913f79ba84ba844a5b386a"],"author":{"name":"Gabe Goodhart","email":"ghart@us.ibm.com","when":"2026-06-23T04:03:31-06:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-23T12:03:31+02:00"},"message":"model: Granite Speech Plus (#24818)\n\n* feat: Add conversion support for Granite Speech Plus\n\nBranch: GraniteSpeechPlus\nAI-usage: full (Bob, OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* feat: Extend granite_speech to support plus multi-layer concatenation\n\nBranch: GraniteSpeechPlus\nAI-usage: draft (Bob, OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* fix(conversion): Fix plural naming for feature_layers for audio\n\nBranch: GraniteSpeechPlus\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix(mtmd): Align feature_layer usage and naming everywhere\n\nBranch: GraniteSpeechPlus\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* style: Use fstring for log\r\n\r\nSigned-off-by: Gabe Goodhart \n\nCo-authored-by: Xuan-Son Nguyen \n\n---------\n\nCo-authored-by: Xuan-Son Nguyen ","tree":"f1f36d1b5f2a694c51450fe1c4d7699567a1e8ac"}},{"name":"granite4-vision.cpp","mode":"0100644","size":15827,"lastCommit":{"hash":"9d57ce456c94d241dde672b2db9cf18879766568","oid":"9d57ce456c94d241dde672b2db9cf18879766568","parents":["16d222fc5ead59d20039501a37251c9ed457a454"],"author":{"name":"Hemanth Battu","email":"56206750+hbattu73@users.noreply.github.com","when":"2026-08-14T15:25:26-07:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-15T00:25:26+02:00"},"message":"mtmd: fix Granite4 Vision image sequence assembly (#26653)\n\n* mtmd: fix granite 4v grid assembly\n\n(cherry picked from commit 91f82eb1b489bff0dc3f649edf2cb9e0b7f656e9)\n\n* mtmd: fix truncation for scaled image height and width before unpad\n\nSigned-off-by: Hemanth Battu \n\n* mtmd: remove MTMD_DUMP_EMBD debug scaffolding\n\nSigned-off-by: Hemanth Battu \n\n* clean up comments, clarify about anyres_info excluded from serialization\n\n* add_newline is now dead code\n\n---------\n\nSigned-off-by: Hemanth Battu \nCo-authored-by: Xuan Son Nguyen \nCo-authored-by: Hemanth Battu ","tree":"1210db073bc7451c953630c6c04f34352cacf8fa"}},{"name":"hunyuanvl.cpp","mode":"0100644","size":2470,"lastCommit":{"hash":"6a257d44633d4a752183ed778b88d2924d0a6b9d","oid":"6a257d44633d4a752183ed778b88d2924d0a6b9d","parents":["3a479c9132072815cb70a443b4efa45bb66b3f59"],"author":{"name":"wendadawen","email":"130649302+wendadawen@users.noreply.github.com","when":"2026-05-21T06:35:37+08:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-05-21T00:35:37+02:00"},"message":"mtmd, model : merge HunyuanOCR into HunyuanVL and fix OCR vision precision (#23329)\n\n- HunyuanOCR shares the same HF arch and vision layout as HunyuanVL butwas split into a separate path that skipped the +0.1 bilinear sampler used by the HF reference.\n- Collapse OCR into the HUNYUANVL projector + HUNYUAN_VL text arch","tree":"a5beb5d6663d90063f27c2c0eacb09090c9b5cc2"}},{"name":"internvl.cpp","mode":"0100644","size":2684,"lastCommit":{"hash":"db52540f730de39efcf7172d4ab1f79bb50556e2","oid":"db52540f730de39efcf7172d4ab1f79bb50556e2","parents":["3a3edc9ac65cca79584ca497be41d70c75a58ba8"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-06-19T01:16:16+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-19T01:16:16+02:00"},"message":"mtmd: add batching support for internvl (#24775)","tree":"0c6d7c42ab7e4b615c17071cc72602f63603f8b6"}},{"name":"kimik25.cpp","mode":"0100644","size":3675,"lastCommit":{"hash":"f5c6ae18278b75712fc992587d9ea527d4ea2218","oid":"f5c6ae18278b75712fc992587d9ea527d4ea2218","parents":["5a69c974392020e514c3b2b2910bb92f847cb4c9"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-06-06T11:06:51+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-06T11:06:51+02:00"},"message":"mtmd, server: add \"placeholder bitmap\" for counting tokens , add */input_tokens API (#23913)\n\n* mtmd: add \"placeholder bitmap\" for counting tokens w/o preprocessing\n\n* fast path skip preproc for placeholder\n\n* fix build\n\n* correct the api\n\n* add server endpoint + tests\n\n* add object name\n\n* update docs\n\n* add proxy handling\n\n* fix build\n\n* fix audio input path\n\n* use is_placeholder in process_mtmd_prompt()\n\n* nits\n\n* nits (2)\n\n* docs: clarify chat/completions/input_tokens is not official\n\n* fix merge problem","tree":"f14db5652990fbaf3c81bc03a91f0bc27b98c8bd"}},{"name":"kimivl.cpp","mode":"0100644","size":2090,"lastCommit":{"hash":"e39a2ce66d0a61915f22097e5453e291618b3518","oid":"e39a2ce66d0a61915f22097e5453e291618b3518","parents":["a8c7f33d792c7ba8ecb889752f80945373e2ddfc"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2025-12-12T21:14:48+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2025-12-12T21:14:48+01:00"},"message":"clip: move model cgraphs into their own files (#17965)\n\n* clip: move model cgraphs into their own files\n\n* more explicit enums\n\n* fix linux build\n\n* fix naming\n\n* missing headers\n\n* nits: add comments for contributors","tree":"a129286f23312a890576038045fb6e0874706779"}},{"name":"llama4.cpp","mode":"0100644","size":3697,"lastCommit":{"hash":"1e645345702154ba4813d3d9bbdbd97718de82c0","oid":"1e645345702154ba4813d3d9bbdbd97718de82c0","parents":["cd708db0ccd59e0770b5ff655db55c995b6fd174"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-03-19T13:11:39+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-03-19T13:11:39+01:00"},"message":"mtmd: add clip_graph::build_mm() (#20751)\n\n* clip: add build_mm()\n\n* apply to all models\n\n* add TODO for bias overload","tree":"199f6c4e93e7126cb52a6a91ed19e502fd794cfa"}},{"name":"llava.cpp","mode":"0100644","size":17347,"lastCommit":{"hash":"a3900a669419e38f4e5f13a1a773857a38fd4cb1","oid":"a3900a669419e38f4e5f13a1a773857a38fd4cb1","parents":["7c908502ea0868e6ae913f79ba84ba844a5b386a"],"author":{"name":"Gabe Goodhart","email":"ghart@us.ibm.com","when":"2026-06-23T04:03:31-06:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-23T12:03:31+02:00"},"message":"model: Granite Speech Plus (#24818)\n\n* feat: Add conversion support for Granite Speech Plus\n\nBranch: GraniteSpeechPlus\nAI-usage: full (Bob, OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* feat: Extend granite_speech to support plus multi-layer concatenation\n\nBranch: GraniteSpeechPlus\nAI-usage: draft (Bob, OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* fix(conversion): Fix plural naming for feature_layers for audio\n\nBranch: GraniteSpeechPlus\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix(mtmd): Align feature_layer usage and naming everywhere\n\nBranch: GraniteSpeechPlus\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* style: Use fstring for log\r\n\r\nSigned-off-by: Gabe Goodhart \n\nCo-authored-by: Xuan-Son Nguyen \n\n---------\n\nCo-authored-by: Xuan-Son Nguyen ","tree":"f1f36d1b5f2a694c51450fe1c4d7699567a1e8ac"}},{"name":"mimo-audio.cpp","mode":"0100644","size":9637,"lastCommit":{"hash":"1cbfd1988311775425d36c0ce066590f7d3049cf","oid":"1cbfd1988311775425d36c0ce066590f7d3049cf","parents":["0e4a0362239713ea95a6864a17a8de4b0ad90d62"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-07-27T23:17:09+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-07-27T23:17:09+02:00"},"message":"mtmd: support MiMo-V2.5 audio input (RVQ-based model) (#26190)\n\n* gguf converter for mimo audio\n\n* fix conv\n\n* cpp impl\n\n* nits\n\n* nits 2","tree":"09f06f2c45edae28ee29260a671a340995c3156d"}},{"name":"mimovl.cpp","mode":"0100644","size":9432,"lastCommit":{"hash":"f5c6ae18278b75712fc992587d9ea527d4ea2218","oid":"f5c6ae18278b75712fc992587d9ea527d4ea2218","parents":["5a69c974392020e514c3b2b2910bb92f847cb4c9"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-06-06T11:06:51+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-06T11:06:51+02:00"},"message":"mtmd, server: add \"placeholder bitmap\" for counting tokens , add */input_tokens API (#23913)\n\n* mtmd: add \"placeholder bitmap\" for counting tokens w/o preprocessing\n\n* fast path skip preproc for placeholder\n\n* fix build\n\n* correct the api\n\n* add server endpoint + tests\n\n* add object name\n\n* update docs\n\n* add proxy handling\n\n* fix build\n\n* fix audio input path\n\n* use is_placeholder in process_mtmd_prompt()\n\n* nits\n\n* nits (2)\n\n* docs: clarify chat/completions/input_tokens is not official\n\n* fix merge problem","tree":"f14db5652990fbaf3c81bc03a91f0bc27b98c8bd"}},{"name":"minicpmv.cpp","mode":"0100644","size":14742,"lastCommit":{"hash":"de699957b92f490efebad149665b0dccf127eaff","oid":"de699957b92f490efebad149665b0dccf127eaff","parents":["ddd4ec1428a6201e18975ea52b07c71e0f9aef26"],"author":{"name":"tc-mb","email":"157115220+tc-mb@users.noreply.github.com","when":"2026-08-01T19:38:36+08:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-01T13:38:36+02:00"},"message":"mtmd: add minicpmv46 downsample (#25993)\n\n* add minicpmv46 downsample\n\nSigned-off-by: tc-mb \n\n* put downsample mode inside gguf.\n\nSigned-off-by: tc-mb \n\n* build mtmd_image_preprocessor_llava_uhd\n\nSigned-off-by: tc-mb \n\n* fix code\n\nSigned-off-by: tc-mb \n\n* add convert\n\nSigned-off-by: tc-mb \n\n* add 4x ignore vit merger\n\nSigned-off-by: tc-mb \n\n---------\n\nSigned-off-by: tc-mb ","tree":"33cc5cb0f52b1cb3861cc9e2992d92792c0e7dee"}},{"name":"minimax-m3.cpp","mode":"0100644","size":3056,"lastCommit":{"hash":"b21e4de74567f5eef213765c9476a843c2e43f0d","oid":"b21e4de74567f5eef213765c9476a843c2e43f0d","parents":["d9f918d2d06079b4336688e819eee821c8a9cd9e"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-08-22T16:33:47+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-22T16:33:47+02:00"},"message":"mtmd: use ggml_rope_set_offset (#27521)\n\n* mtmd: use ggml_rope_set_offset\n\n* add comment","tree":"612d00b0e6ee334c61cbdff8c05eabf45eb8a98a"}},{"name":"mobilenetv5.cpp","mode":"0100644","size":16548,"lastCommit":{"hash":"1e645345702154ba4813d3d9bbdbd97718de82c0","oid":"1e645345702154ba4813d3d9bbdbd97718de82c0","parents":["cd708db0ccd59e0770b5ff655db55c995b6fd174"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-03-19T13:11:39+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-03-19T13:11:39+01:00"},"message":"mtmd: add clip_graph::build_mm() (#20751)\n\n* clip: add build_mm()\n\n* apply to all models\n\n* add TODO for bias overload","tree":"199f6c4e93e7126cb52a6a91ed19e502fd794cfa"}},{"name":"models.h","mode":"0100644","size":18051,"lastCommit":{"hash":"54ee5ee643f29abba6852903ddfdb688c2361b5b","oid":"54ee5ee643f29abba6852903ddfdb688c2361b5b","parents":["3a653fea932e6a026e14cc0fb6507a986679f688"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-08-22T10:35:50+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-22T10:35:50+02:00"},"message":"mtmd: support dots3-note vision+audio (#27524)\n\n* text: conversion\n\n* init impl\n\n* mtmd: conversion\n\n* impl mtmd cpp\n\n* Update gguf-py/gguf/tensor_mapping.py\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"e6eabd88f99116b32a04cde3a7d0fc0d35238806"}},{"name":"muse-glimmer.cpp","mode":"0100644","size":4235,"lastCommit":{"hash":"62bf73d25c53b8161f8a22894d4f90c4aebbd7d0","oid":"62bf73d25c53b8161f8a22894d4f90c4aebbd7d0","parents":["a52077c4cabb4f3c0298329c9d2dd1324d5604cb"],"author":{"name":"Pedro Cuenca","email":"pedro@huggingface.co","when":"2026-08-10T13:07:27+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-10T13:07:27+02:00"},"message":"model: Muse Glimmer Support (#26841)\n\n* Get started with Onyx\n\n* Add architecture\n\n* Skip keys handled in super()\n\n* Loading tensors\n\n* Shorten\n\n* Graph\n\n* Apply suggestion from @pcuenca\n\n* Remove norm now embedding in transformers weights\n\n* Add eot\n\n* Explicit output_multiplier\n\n* Handle post_norm_eps\n\n* No super call; unhardcode eot.\n\nThe pattern `self._set_vocab_gpt2()` seems preferred throughout the\ncodebase, and it allows `set_vocab()` to be called from a different part\nof the Python class hierarchy: the drafter model converter that we may\nneed eventually.\n\n* Register for drafting\n\n* DFlash: inherit rope type from the linked target.\n\nAnother option would be to store it in the gguf file itself.\n\n* mmproj conversion\n\nNote: some fields to be renamed after the implementation works. We are\nkeeping compatibility with the reference Meta gguf for testing purposes.\n\n* \"clip\" header declarations\n\n* Load mmproj\n\n* Pre-processing\n\n* Graph\n\n* Go back to using delimiters.\n\nOtherwise our generations are worse.\n\nTransformers does not use them. We need to trace inputs to verify\nwhether they are equivalent.\n\n* downsample_factor -> merge_size\n\n* Add vision graph\n\nlol, forgot from a previous commit\n\n* Additional renames, align with llama.cpp / transformers\n\n* Prefer _size instead of independent _h and _w\n\n* Fix token layout\n\nCo-authored-by: Young Han \n\n* onyx: bring the chat parser onto the onyx branch\n\ncommon/chat.cpp on this branch has no Onyx handling, so a converted model\nserves malformed chat: the assistant preamble leaks into content\n(\"to=self<|message|>...\") and tool calls fail with\n\n HTTP 500 \"The model produced output that does not match the expected\n peg-native format\"\n\ncommon_chat_params_init_onyx exists on onyx-fair-patch, added there by\n8bb73dd3d. It was never on this branch, so this is not a regression --\nthe two lines developed independently.\n\nThe code here is taken verbatim from that commit. It is the clean side of\n`git merge origin/onyx-fair-patch`: chat.cpp is one of the files that\nmerges without conflict. The full merge is not viable -- it produces 13\nconflicts, including add/add on conversion/onyx.py and src/models/onyx.cpp\nwhere the q_norm-folding and metadata-scale approaches contradict each\nother, and #4/#7 are stacked on this branch's side of that.\n\nVerified on this branch: builds with 0 errors, converts an Onyx checkpoint,\nand serving it gives \"4\" for \"What is 2+2?\" plus a correct\nget_weather {\"city\":\"Paris\"} tool call, where the unported branch gives the\ntwo failures above.\n\nNo converter or runtime changes are included, so this should not interact\nwith the q_norm work.\n\nCo-authored-by: Beto de Paola \n\n* Less params, bilinear pos-emb interpolation as a graph op instead of CPU\n\n* Map to symbolic V_MMPROJ instead of strings\n\n* Make a couple params explicit\n\n* Patchify via build_inp()\n\n* No param for rope_theta\n\n* Small cleanup\n\n* Restore blank line\n\n* Unpermute, to adapt to the latest transformers checkpoint\n\n* Apply norm after token embeddings\n\nThis follows the latest transformers approach.\n\n* Remove duplicated function\n\n* build_vit\n\n* onyx: use the model rope theta on sliding-window layers\n\n* DFlash: conversion from transformers drafter\n\n* Revert rope_type derivation from target\n\nNOTE: this breaks compatibility with Meta's distributed DFlash GGUFs, as\nthe Q/K are stored in \"NEOX\" (rotated half) format, like in\ntransformers.\n\n* Apply suggestion from @pcuenca\n\n* Set model type\n\n* Remove comment that will become obsolete\n\n* Hardcode post_norm_rms_eps instead of new param\n\n* Derive SWA+RoPE pattern from gguf array or scalar\n\n* Fix model type <-> number of layers\n\n* Reorder\n\n* Rename\n\n* Fix typo\n\n* DFlash: seed the draft KV cache from multimodal embedding batches\n\n`common_speculative_impl_draft_dflash::process()` returned early on any batch carrying embeddings, so an image prefill never had its target-layer features fused through the DFlash encoder and injected into the draft's KV cache. That left a hole spanning the image's positions, and the next injection at a post-image position failed to initialize its batch:\n\n```\ndecoding image batch 1/1, n_tokens_batch = 256\ndecode: failed to initialize batch\nllama_decode: failed to decode, ret = -1\nprocess: llama_decode(ctx_dft) failed rc=-1 (n_tokens=17, offset=0)\nsrv decode: failed to process speculative batch\n```\n\nEvery image request with `--spec-type draft-dflash` failed with HTTP 500. Text-only was unaffected, since those batches carry token ids and were let through.\n\nRestore the earlier condition, which admits a batch that is either tokens or embeddings and skips only the degenerate neither/both cases. The rest of `process()` is already layout-agnostic -- it gathers features via `llama_get_embeddings_layer_inp()` and indexes `batch_in.pos[]` / `batch_in.seq_id[]`, none of which assume token ids -- so this is the whole fix.\n\nValidated against `muse-glimmer-30B-bf16.gguf` + `mmproj-muse-glimmer-30B-bf16.gguf` + a DFlash draft head, on an image describe-the-shapes request:\n\n- before: HTTP 500, `failed to process speculative batch`\n- after: HTTP 200, draft acceptance 0.34012 (167 accepted / 491 generated), mean len 3.04\n\nOutput equivalence holds, which is the property that matters: at temperature 0 the drafted response is byte-identical to the same request served with no draft attached (1213/1213 chars), so the draft is drafting correctly through the image context rather than merely not crashing.\n\n* Conversion: prefer rewrite to mapping\n\n* Revert \"Conversion: prefer rewrite to mapping\"\n\nThis reverts commit a92d0ac584d315e876741e85b6dad3dbc8b23bf7.\n\n* fix lint\n\n* sliding_window metadata is not optional\n\n* disable state save/load\n\n* Apply suggestion from @pcuenca\n\n---------\n\nCo-authored-by: Young Han \nCo-authored-by: Beto de Paola \nCo-authored-by: Daniel Han \nCo-authored-by: ruanrms \nCo-authored-by: Xuan Son Nguyen \nCo-authored-by: Sigbjørn Skjæret ","tree":"7bfbb0b0e9f1665cd305d193682413b50a79735f"}},{"name":"nemotron-v2-vl.cpp","mode":"0100644","size":1156,"lastCommit":{"hash":"01d8eaa28d57bfc6d06e30072085ed0ef12e06c5","oid":"01d8eaa28d57bfc6d06e30072085ed0ef12e06c5","parents":["1725e316c1a780759ec134ca5a2999f4d53ce273"],"author":{"name":"Anav Prasad","email":"anavp@nvidia.com","when":"2026-02-14T05:07:00-08:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-02-14T14:07:00+01:00"},"message":"mtmd : Add Nemotron Nano 12B v2 VL support (#19547)\n\n* nemotron nano v2 vlm support added\n\n* simplified code; addressed reviews\n\n* pre-downsample position embeddings during GGUF conversion for fixed input size","tree":"d2d6f93a488b6ecbe9d1a876f200fdfbe67a9a7f"}},{"name":"paddleocr.cpp","mode":"0100644","size":1766,"lastCommit":{"hash":"237958db339300bdd8028608cc08b2ba2685ec33","oid":"237958db339300bdd8028608cc08b2ba2685ec33","parents":["abb9f3c42b5e6acee9e8e37836ef691d1a41bdb8"],"author":{"name":"megemini","email":"megemini@outlook.com","when":"2026-02-20T00:05:25+08:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-02-19T17:05:25+01:00"},"message":"model: Add PaddleOCR-VL model support (#18825)\n\n* support PaddleOCR-VL\n\n* clip: update PaddleOCR model loader parameters to prevent OOM during warmup\n\n* [update] add paddleocr vl text model instead of ernie4.5\n\n* [update] restore change of minicpmv\n\n* [update] format\n\n* [update] format\n\n* [update] positions and patch merge permute\n\n* [update] mtmd_decode_use_mrope for paddleocr\n\n* [update] image min/max pixels\n\n* [update] remove set_limit_image_tokens\n\n* upate: preprocess without padding\n\n* clean up\n\n* Update convert_hf_to_gguf.py\n\nCo-authored-by: Sigbjørn Skjæret \n\n* Update convert_hf_to_gguf.py\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Xuan Son Nguyen \nCo-authored-by: Sigbjørn Skjæret ","tree":"ff877cca21e1c8a1025869f759281419b9681b2b"}},{"name":"parakeet.cpp","mode":"0100644","size":19704,"lastCommit":{"hash":"7e1e28cae36d41fe7bbe9dae7c9625de6565c063","oid":"7e1e28cae36d41fe7bbe9dae7c9625de6565c063","parents":["6e2bc65fb2a822a8082a7515858a7c180115dd42"],"author":{"name":"Daniel Bevenius","email":"daniel.bevenius@gmail.com","when":"2026-07-28T17:20:25+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-07-28T17:20:25+02:00"},"message":"mtmd : add Nemotron 3 Nano Omni support (parakeet) (#22520)\n\n* mtmd : add Nemotron 3 Nano Omni support (parakeet)\n\nThis commit adds support for the subsampling and encoder part of\nNemotron Nemo 3 omni model.\n\nThe Parakeet subsampling/encoder were taken from parakeet.cpp which\nis currently a pull request against whisper.cpp. I've tried to copy the\ncode a close as possible to hopefully enable easy patching between the\nthese two project later.\n\nRefs: https://github.com/ggml-org/whisper.cpp/pull/3735\n\n* mtmd : generate rel pos tensor in graph instead of in conversion [no ci]\n\nThis commit removes the generation of the relative positional tensor in\nthe model conversion script and instead computes it in the encoder\ngraph. This is only done for the window of positions required for the\ncurrent audio sample.\n\n* mtmd : add clip_get_model to clip API [no ci]\n\nThis commit adds a function to get access to the clip_model. It also\nremoves the two functions clip_get_mel_filter_tensor, and\nclip_get_window_tensor(const struct clip_ctx * ctx) which can now use\nclip_get_model to access the model tensors that it needs.\n\n* mtmd : read mel_filters and window into hparams\n\n* mtmd : use set_input_f32 lambda [no ci]\n\n* mtmd : add better asserts for mel_filters and hann window [no ci]\n\n* mtmd : add missing size_t cast\n\n* mtmd : change type of pad to size_t\n\n* mtmd : zero initialize samples_padded\n\n* mtmd : remove unsued ctx member from parakeet preprocessor\n\n* mtmd : make log_mel_spectrogram_parakeet_worker_thread private static\n\n* mtmd : sync/update parakeeet impl with latest whisper.cpp\n\nThis commit updates the parakeet code in mtmd to reflect the latest\nupdates to parakeet.cpp in whisper.cpp.\n\nA follow up commit will address the currently hardcoded dw_pad and see\nif we can add n_conv_kernel as a model metadata field.\n\n* mtmd : add audio_conv_kernel_size to model conversion\n\nThis commit updates the model conversion to read the conv_kernel_size\nfield from the sound_config section of the models config.json file.\nIt then uses this field instead of the hardcoded values in parakeet.cpp.\n\n* mtmd : cleanup [no ci]\n\n* conversion : call super().filter_tensors [no ci]\n\n* do not discard result of super filter_tensors\n\n* mtmd : use build_mm instead of ggml_mul_mat\n\n* mtmd : use build_ffn\n\n* mtmd : move and reuse get_vector lambda\n\n* mtmd : use build_inp_raw for parakeet\n\n* mtmd : throw exception in get_scalar instead of assert\n\n* mtmd : fix std::min call\n\n* mtmt : use .c_str in throw clause in get_vector\n\n* mtmd : check for F32 type and non-empty tensor in get_vector\n\nThe get_vector lambda is used by get_scalar but also standalone to read\nin the mel_filters and the window data. Therefor we are not checking\nfor 1D tensors but allowing multiple dimensions. We do have a check in\nget_scalar to verify the size of the vector.\n\n* mtmd : replace hardcoded 1101 for n_tokens_real\n\n* mtmd : assert subsampling_factor is 8\n\nThis commit adds an assert of the parakeet subsampling factor to check\nthat it is 8.\n\nThe motivation for this is that this model currently has three\nconvolutions with a stride of 2. If the underlying model updates the\nsubsampling factor these convolution operations will need to be updated\nand this will produce and error if this occurs.\n\n* mtmd : remove unused ggml_tensors attn_pos_w and mm_norm_w\n\n* mtmd : remove single thread path\n\nThis commit removes the single thread path which was a left over from\nthe original parakeet.cpp where n_threads is configurable.\n\n* fix some security issues\n\n---------\n\nCo-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>\nCo-authored-by: Xuan Son Nguyen ","tree":"0c8f6bdaa3f79a116b002671f6ccaf4248c8b36c"}},{"name":"pixtral.cpp","mode":"0100644","size":3512,"lastCommit":{"hash":"b11f7c16bc1c3b320c5cd427d545ee6fa6bd535c","oid":"b11f7c16bc1c3b320c5cd427d545ee6fa6bd535c","parents":["f818065d753700341b5355ce36ca676ecfb5d139"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-06-26T08:43:29+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-26T08:43:29+02:00"},"message":"mtmd: add more validations (#25013)\n\n* mtmd: add more validations\n\n* fix\n\n* refactor a bit\n\n* type check for get_arr_int","tree":"58537532778ee154c64e4b37c70965b8381bb273"}},{"name":"pockettts-gen.cpp","mode":"0100644","size":13527,"lastCommit":{"hash":"6e62ba538478202094edc6c100c782719e310aa3","oid":"6e62ba538478202094edc6c100c782719e310aa3","parents":["8d274dd7c6233ed73c7509cc2a8be9960f7df7d5"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-08-11T14:18:30+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-11T14:18:30+02:00"},"message":"mtmd: support pocket-tts (#26871)\n\n* adapt the api\n\n* text model ok\n\n* working impl, need verify and clean up\n\n* mtmd: build the pocket-tts transposed convolutions as GEMM + col2im\n\nggml_conv_transpose_1d has no grouped mode, so the depthwise upsample\nwas built as one convolution and one concat per channel, which floods\nthe graph with small nodes and makes kernel launches dominate the\ndecoder.\n\nFold both cases into the column form the seanet decoder already needs:\nthe general case reshapes the kernel to [IC, K * OC] and matmuls it\nwith the input, the depthwise case batches a matmul over the channels\nso a step scales its own kernel. A single col2im_1d then scatter-adds\nthe columns back to the signal, with the same shape as before, so the\noverlap-add tail, the streaming state and the bias are untouched.\n\nGeneration time per frame drops by 80% on CUDA and by 50% on CPU. The\noutput matches the previous implementation sample for sample, with a\ncorrelation of 0.999994 and identical frame counts.\n\n* flow_temp + frames_after_eos\n\n* chunking\n\n* mtmd: carry the remaining pocket-tts per-pack settings\n\nThe language packs also tune the end-of-speech padding and the padding\nof short prompts, next to the temperature already carried in the\nmmproj: french_24l asks for 8 tail frames instead of the guessed 3,\nenglish_2026-01 asks for short prompts to be padded with spaces.\n\nWrite both in the mmproj as clip.gen.audio.frames_after_eos and\nclip.gen.audio.pad_short_text, keyed on the pack in the conversion\nscript like the temperature. The loader keeps them optional, so a\nmmproj without them behaves as before. Map semicolons to commas for\nevery pack instead, the reference only asks for it on three of them and\nit costs nothing elsewhere.\n\nExisting mmproj files must be converted again to carry the two keys.\n\nOn a long french text the port now lands within 2% of the reference:\n22.96s against 23.44s, with the same peak level and the same amount of\nsilence.\n\n* clip.gen.audio.model_variant\n\n* clean up code comments\n\n* nit: drop the dead flow_temp hparam, the pack table holds the default\n\n* update docs\n\n* address security problems\n\n* less invasive base.py\n\n* lint\n\n* add mtmd_gen_inp_default\n\n* add docs\n\n* rm gen_flow_temp\n\n---------\n\nCo-authored-by: Pascal ","tree":"0bc98945fb180d205070b9f2d96de7d1208c3e7d"}},{"name":"pockettts-seanet.cpp","mode":"0100644","size":7465,"lastCommit":{"hash":"6e62ba538478202094edc6c100c782719e310aa3","oid":"6e62ba538478202094edc6c100c782719e310aa3","parents":["8d274dd7c6233ed73c7509cc2a8be9960f7df7d5"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-08-11T14:18:30+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-11T14:18:30+02:00"},"message":"mtmd: support pocket-tts (#26871)\n\n* adapt the api\n\n* text model ok\n\n* working impl, need verify and clean up\n\n* mtmd: build the pocket-tts transposed convolutions as GEMM + col2im\n\nggml_conv_transpose_1d has no grouped mode, so the depthwise upsample\nwas built as one convolution and one concat per channel, which floods\nthe graph with small nodes and makes kernel launches dominate the\ndecoder.\n\nFold both cases into the column form the seanet decoder already needs:\nthe general case reshapes the kernel to [IC, K * OC] and matmuls it\nwith the input, the depthwise case batches a matmul over the channels\nso a step scales its own kernel. A single col2im_1d then scatter-adds\nthe columns back to the signal, with the same shape as before, so the\noverlap-add tail, the streaming state and the bias are untouched.\n\nGeneration time per frame drops by 80% on CUDA and by 50% on CPU. The\noutput matches the previous implementation sample for sample, with a\ncorrelation of 0.999994 and identical frame counts.\n\n* flow_temp + frames_after_eos\n\n* chunking\n\n* mtmd: carry the remaining pocket-tts per-pack settings\n\nThe language packs also tune the end-of-speech padding and the padding\nof short prompts, next to the temperature already carried in the\nmmproj: french_24l asks for 8 tail frames instead of the guessed 3,\nenglish_2026-01 asks for short prompts to be padded with spaces.\n\nWrite both in the mmproj as clip.gen.audio.frames_after_eos and\nclip.gen.audio.pad_short_text, keyed on the pack in the conversion\nscript like the temperature. The loader keeps them optional, so a\nmmproj without them behaves as before. Map semicolons to commas for\nevery pack instead, the reference only asks for it on three of them and\nit costs nothing elsewhere.\n\nExisting mmproj files must be converted again to carry the two keys.\n\nOn a long french text the port now lands within 2% of the reference:\n22.96s against 23.44s, with the same peak level and the same amount of\nsilence.\n\n* clip.gen.audio.model_variant\n\n* clean up code comments\n\n* nit: drop the dead flow_temp hparam, the pack table holds the default\n\n* update docs\n\n* address security problems\n\n* less invasive base.py\n\n* lint\n\n* add mtmd_gen_inp_default\n\n* add docs\n\n* rm gen_flow_temp\n\n---------\n\nCo-authored-by: Pascal ","tree":"0bc98945fb180d205070b9f2d96de7d1208c3e7d"}},{"name":"pockettts-spkenc.cpp","mode":"0100644","size":3281,"lastCommit":{"hash":"6e62ba538478202094edc6c100c782719e310aa3","oid":"6e62ba538478202094edc6c100c782719e310aa3","parents":["8d274dd7c6233ed73c7509cc2a8be9960f7df7d5"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-08-11T14:18:30+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-11T14:18:30+02:00"},"message":"mtmd: support pocket-tts (#26871)\n\n* adapt the api\n\n* text model ok\n\n* working impl, need verify and clean up\n\n* mtmd: build the pocket-tts transposed convolutions as GEMM + col2im\n\nggml_conv_transpose_1d has no grouped mode, so the depthwise upsample\nwas built as one convolution and one concat per channel, which floods\nthe graph with small nodes and makes kernel launches dominate the\ndecoder.\n\nFold both cases into the column form the seanet decoder already needs:\nthe general case reshapes the kernel to [IC, K * OC] and matmuls it\nwith the input, the depthwise case batches a matmul over the channels\nso a step scales its own kernel. A single col2im_1d then scatter-adds\nthe columns back to the signal, with the same shape as before, so the\noverlap-add tail, the streaming state and the bias are untouched.\n\nGeneration time per frame drops by 80% on CUDA and by 50% on CPU. The\noutput matches the previous implementation sample for sample, with a\ncorrelation of 0.999994 and identical frame counts.\n\n* flow_temp + frames_after_eos\n\n* chunking\n\n* mtmd: carry the remaining pocket-tts per-pack settings\n\nThe language packs also tune the end-of-speech padding and the padding\nof short prompts, next to the temperature already carried in the\nmmproj: french_24l asks for 8 tail frames instead of the guessed 3,\nenglish_2026-01 asks for short prompts to be padded with spaces.\n\nWrite both in the mmproj as clip.gen.audio.frames_after_eos and\nclip.gen.audio.pad_short_text, keyed on the pack in the conversion\nscript like the temperature. The loader keeps them optional, so a\nmmproj without them behaves as before. Map semicolons to commas for\nevery pack instead, the reference only asks for it on three of them and\nit costs nothing elsewhere.\n\nExisting mmproj files must be converted again to carry the two keys.\n\nOn a long french text the port now lands within 2% of the reference:\n22.96s against 23.44s, with the same peak level and the same amount of\nsilence.\n\n* clip.gen.audio.model_variant\n\n* clean up code comments\n\n* nit: drop the dead flow_temp hparam, the pack table holds the default\n\n* update docs\n\n* address security problems\n\n* less invasive base.py\n\n* lint\n\n* add mtmd_gen_inp_default\n\n* add docs\n\n* rm gen_flow_temp\n\n---------\n\nCo-authored-by: Pascal ","tree":"0bc98945fb180d205070b9f2d96de7d1208c3e7d"}},{"name":"qwen2vl.cpp","mode":"0100644","size":7828,"lastCommit":{"hash":"31e82494c0a3913c919c1027fa70500fbf4c07dd","oid":"31e82494c0a3913c919c1027fa70500fbf4c07dd","parents":["6b80c74f285390368b3c99c5e750f19e9b096e98"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-06-06T21:17:25+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-06T21:17:25+02:00"},"message":"mtmd: support \"frame merge\" for qwen-vl-based models (#21858)\n\n* feat: add video support for Qwen3.5\n\n* various clean up\n\n* revise the design\n\n* fix llava-uhd case\n\n* nits\n\n* nits 2\n\n---------\n\nCo-authored-by: andrewmd5 <1297077+andrewmd5@users.noreply.github.com>","tree":"f90c3f16f97574a7671eb31d42a31b1a927ccbc0"}},{"name":"qwen3a.cpp","mode":"0100644","size":3576,"lastCommit":{"hash":"72e60f500d739eb9dc26e51cda5fdfe062faf766","oid":"72e60f500d739eb9dc26e51cda5fdfe062faf766","parents":["8be1786707cdac011b90bf0d6ddb67c3eb6877f3"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-05-15T19:32:47+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-05-15T19:32:47+02:00"},"message":"mtmd: add chunks and fix preproc for qwen3a (#23073)\n\n* mtmd: add chunks and fix preproc for qwen3a\n\n* add attn_mask\n\n* limit mtmd_chunk size (avoid blow up memory)\n\n* correct audio tokens\n\n* re-order the set_input case\n\n* remove attn_mask","tree":"deefdb3a8d221c842300200f8672529da4c65421"}},{"name":"qwen3tts-gen.cpp","mode":"0100644","size":36306,"lastCommit":{"hash":"6e62ba538478202094edc6c100c782719e310aa3","oid":"6e62ba538478202094edc6c100c782719e310aa3","parents":["8d274dd7c6233ed73c7509cc2a8be9960f7df7d5"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-08-11T14:18:30+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-11T14:18:30+02:00"},"message":"mtmd: support pocket-tts (#26871)\n\n* adapt the api\n\n* text model ok\n\n* working impl, need verify and clean up\n\n* mtmd: build the pocket-tts transposed convolutions as GEMM + col2im\n\nggml_conv_transpose_1d has no grouped mode, so the depthwise upsample\nwas built as one convolution and one concat per channel, which floods\nthe graph with small nodes and makes kernel launches dominate the\ndecoder.\n\nFold both cases into the column form the seanet decoder already needs:\nthe general case reshapes the kernel to [IC, K * OC] and matmuls it\nwith the input, the depthwise case batches a matmul over the channels\nso a step scales its own kernel. A single col2im_1d then scatter-adds\nthe columns back to the signal, with the same shape as before, so the\noverlap-add tail, the streaming state and the bias are untouched.\n\nGeneration time per frame drops by 80% on CUDA and by 50% on CPU. The\noutput matches the previous implementation sample for sample, with a\ncorrelation of 0.999994 and identical frame counts.\n\n* flow_temp + frames_after_eos\n\n* chunking\n\n* mtmd: carry the remaining pocket-tts per-pack settings\n\nThe language packs also tune the end-of-speech padding and the padding\nof short prompts, next to the temperature already carried in the\nmmproj: french_24l asks for 8 tail frames instead of the guessed 3,\nenglish_2026-01 asks for short prompts to be padded with spaces.\n\nWrite both in the mmproj as clip.gen.audio.frames_after_eos and\nclip.gen.audio.pad_short_text, keyed on the pack in the conversion\nscript like the temperature. The loader keeps them optional, so a\nmmproj without them behaves as before. Map semicolons to commas for\nevery pack instead, the reference only asks for it on three of them and\nit costs nothing elsewhere.\n\nExisting mmproj files must be converted again to carry the two keys.\n\nOn a long french text the port now lands within 2% of the reference:\n22.96s against 23.44s, with the same peak level and the same amount of\nsilence.\n\n* clip.gen.audio.model_variant\n\n* clean up code comments\n\n* nit: drop the dead flow_temp hparam, the pack table holds the default\n\n* update docs\n\n* address security problems\n\n* less invasive base.py\n\n* lint\n\n* add mtmd_gen_inp_default\n\n* add docs\n\n* rm gen_flow_temp\n\n---------\n\nCo-authored-by: Pascal ","tree":"0bc98945fb180d205070b9f2d96de7d1208c3e7d"}},{"name":"qwen3tts-spkenc.cpp","mode":"0100644","size":8560,"lastCommit":{"hash":"07132750825a4f2d27a547cd9cdde1c6f6001885","oid":"07132750825a4f2d27a547cd9cdde1c6f6001885","parents":["1c3c9674de4d455f1e571bed808252af54932767"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-08-04T17:26:15+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-04T17:26:15+02:00"},"message":"mtmd: support Qwen3-TTS (note: breaking change to llama-tts binary) (#26254)\n\n* convert text model\n\n* main model load ok\n\n* convert encoder ok\n\n* speaker encoder loading ok\n\n* speaker enc graph\n\n* adapt vocab for backbone (with some tricks)\n\n* add suppress_tokens\n\n* poc new mtmd gen api\n\n* convert code_predictor to gguf\n\n* load gen_code model ok\n\n* add clip_encode\n\n* wire up\n\n* code gen cgraph init version\n\nCo-authored-by: Pascal \n\n* code2wav convert to gguf\n\n* code2wav graph ok\n\n* wire up in/out\n\n* (wip) subgraph\n\n* wire up\n\n* wip, correct code2wav\n\n* demo (to be removed)\n\n* code2wav preserve kv between calls\n\n* demo voice clone\n\n* llama: add llama_model_get_tok_embd\n\n* mtmd_helper_gen_audio API\n\n* fix clamp cold prefix\n\nCo-authored-by: Pascal \n\n* fuse snake op\n\nCo-authored-by: Pascal \n\n* demo: use proper sampling\n\n* update dev docs\n\n* polymorphism helper\n\n* revamp llama-tts binary\n\n* update docs\n\n* fix compile\n\n* fix lint\n\n* nits\n\n* add guide + docs\n\n* more timings info\n\n* clean up code comments\n\n* security fixes\n\n* update docs\n\n* use ggml_build_forward_select, clean up comments\n\n* fix ci\n\n* use ISO 639-1 language code\n\n* rename CODE2WAV --> GEN_WAV, update docs\n\n* clean up\n\n* clean up tts.cpp\n\n* add seq_id\n\n* add step_prompt()\n\n* mtmd_helper_model_can_chat\n\n* clean up comments\n\n---------\n\nCo-authored-by: Pascal ","tree":"46f77bf060878e2b3b9d7c43b4d8a3a566ba3384"}},{"name":"qwen3vl.cpp","mode":"0100644","size":6702,"lastCommit":{"hash":"b4aa7dd477acf065a3b9c6a8cf324c904da1a834","oid":"b4aa7dd477acf065a3b9c6a8cf324c904da1a834","parents":["71102a73f2042d9aacae2d6dc70874a8611b8361"],"author":{"name":"Gerben van V","email":"gerbenvv@gmail.com","when":"2026-07-21T23:58:34+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-07-21T23:58:34+02:00"},"message":"mtmd : use align_corners for qwen3vl vision position embedding interpolation (#25781)\n\nThe Qwen3-VL learned position embedding is interpolated to the runtime patch\ngrid with the default bilinear+antialias (align_corners=False) sampling, while\nthe transformers reference uses align_corners=True (torch.linspace(0, side-1, T)).\nThe mismatch scales grounding coordinates about the image center, growing with\nimage size and per-axis for non-square images (see #16880).","tree":"1a09ba0e06dbaf88676785b4a05b370ecf00e590"}},{"name":"siglip.cpp","mode":"0100644","size":3317,"lastCommit":{"hash":"a970515bdb0b1d09519106847660b0d0c84d2472","oid":"a970515bdb0b1d09519106847660b0d0c84d2472","parents":["056b50c319724ef4788da46c605673b94dc2374e"],"author":{"name":"Saba Fallah","email":"10401143+sfallah@users.noreply.github.com","when":"2026-03-25T19:57:40+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-03-25T19:57:40+01:00"},"message":"mtmd: Add DeepSeekOCR Support (#17400)\n\n* mtmd: llama.cpp DeepSeekOCR support\ninit commit\n\n* loading sam tensors\n\n* mtmd: fix vision model processing\n\n* deepseek-ocr clip-vit model impl\n\n* mtmd: add DeepSeek-OCR LM support with standard attention\n\n* mtmd: successfully runs DeepSeek-OCR LM in llama-cli\n\n* mtmd: Fix RoPE type for DeepSeek-OCR LM.\n\n* loading LM\ntesting Vision model loading\n\n* sam warmup working\n\n* sam erroneous return corrected\n\n* clip-vit: corrected cls_embd concat\n\n* clip-vit: model convert qkv_proj split\n\n* corrected combining of image encoders' results\n\n* fix: update callback for ffn_moe_weighted and add callback for attn_out in deepseek2 model\n\n* concat image_newline and image_seperator tokens\n\n* visual_model warmup (technically) works\n\n* window partitioning using standard ggml ops\n\n* sam implementation without using CPU only ops\n\n* clip: fixed warnings\n\n* Merge branch 'sf/deepseek-ocr' of github.com:sfallah/llama.cpp into sf/deepseek-ocr\n\n* mtmd: fix get_rel_pos\n\n* mtmd: fixed the wrong scaler for get_rel_pos\n\n* image encoding technically works but the output can't be checked singe image decoding fails\n\n* mtmd: minor changed\n\n* mtmd: add native resolution support\n\n* - image encoding debugged\n- issues fixed mainly related wrong config like n_patches etc.\n- configs need to be corrected in the converter\n\n* mtmd: correct token order\n\n* - dynamic resizing\n- changes are concerning PR https://github.com/sfallah/llama.cpp/pull/4\n\n* mtmd: quick fix token order\n\n* mtmd: fix danling pointer\n\n* mtmd: SAM numerically works\n\n* mtmd: debug CLIP-L (vit_pre_ln)\n\n* mtmd: debug CLIP-L & first working DeepSeek-OCR model\n\n* mtmd : add --dsocr-mode CLI argument for DeepSeek-OCR resolution control & all native resolution modes work\n\n* mtmd: simplify SAM patch embedding\n\n* mtmd: adapt Pillow image resizing function\n\n* mtmd: simplify DeepSeek-OCR dynamic resolution preprocessing\n\n* mtmd: remove --dsocr-mode argument\n\n* mtmd: refactor code & remove unused helper functions\n\n* mtmd: fix tensor names for image newlines and view separator\n\n* clean up\n\n* reverting automatically removed spaces\n\n* reverting automatically removed spaces\n\n* mtmd: fixed bad ocr check in Deepseek2 (LM)\n\n* mtmd: support combined QKV projection in buid_vit\n\n* using common build_attn in sam\n\n* corrected code-branch when flash-attn disabled\nenabling usage of --flash-attn option\n\n* mtmd: minor fix\n\n* minor formatting and style\n\n* fixed flake8 lint issues\n\n* minor editorconfig-check fixes\n\n* minor editorconfig-check fixes\n\n* mtmd: simplify get_rel_pos\n\n* mtmd: make sam hparams configurable\n\n* mtmd: add detailed comments for resize_bicubic_pillow\n\n* mtmd: fixed wrong input setting\n\n* mtmd: convert model in FP16\n\n* mtmd: minor fix\n\n* mtmd: remove tweak to llama-mtmd-cli & deepseek-ocr template\n\n* fix: test-1.jpg ORC issue with small (640) resolution\nsetting min-resolution base (1024) max large (1280) for dynamic-resolution\n\n* minor: editconfig-check fix\n\n* merge with changes from https://github.com/ggml-org/llama.cpp/pull/17909\nadded new opt to tests.sh to disable flash-attn\n\n* minor: editconfig-check fix\n\n* testing deepseek-ocr\nquick and dirty test script comparing results of Qwen2.5-VL vs DeepSeek-OCR\n\n* quick and (potential) dirty merge with https://github.com/ggml-org/llama.cpp/pull/17909\n\n* refactoring, one single builder function and static helpers\n\n* added deepseek-ocr test to tests.sh\n\n* minor formatting fixes\n\n* check with fixed expected resutls\n\n* minor formatting\n\n* editorconfig-check fix\n\n* merge with changes from https://github.com/ggml-org/llama.cpp/pull/18042\n\n* minor\n- added GLM-4.6V to big tests\n- added missing deps for python test\n\n* convert: minor fix\n\n* mtmd: format code\n\n* convert: quick fix\n\n* convert: quick fix\n\n* minor python formatting\n\n* fixed merge build issue\n\n* merge resolved\n- fixed issues in convert\n- tested several deepseek models\n\n* minor fix\n\n* minor\n\n* Update convert_hf_to_gguf.py\n\nCo-authored-by: Sigbjørn Skjæret \n\n* - removed clip_is_deepseekocr\n- removed redundant RESIZE_ALGO_BICUBIC_PILLOW resize-algo\n- simplified image-preprocessing\n- removed/simplified debug functions\n\n* - cleaning commented out code\n\n* fixing instabilities issues reintroducing resize_bicubic_pillow\n\n* - use f16 model for deepseek-ocr test\n- ignore llama-arch test for deepseek-ocr\n\n* rename fc_w --> mm_fc_w\n\n* add links to OCR discussion\n\n* cleaner loading code\n\n* add missing .weight to some tensors\n\n* add default jinja template (to be used by server)\n\n* move test model to ggml-org\n\n* rolling back upscale change\n\n* Update convert_hf_to_gguf.py\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: bluebread \nCo-authored-by: Sigbjørn Skjæret \nCo-authored-by: Xuan Son Nguyen \nCo-authored-by: Xuan-Son Nguyen ","tree":"bd791d6f890823d4bd6a2ef26e2c52898f1fc449"}},{"name":"step3vl.cpp","mode":"0100644","size":2708,"lastCommit":{"hash":"09343c01986667b26d414733950c9d38e91c3065","oid":"09343c01986667b26d414733950c9d38e91c3065","parents":["97508acb17ff933c67edb3a2f0997bc19a6abb98"],"author":{"name":"forforever73","email":"63285796+forforever73@users.noreply.github.com","when":"2026-04-08T15:51:31+08:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-04-08T09:51:31+02:00"},"message":"model : support step3-vl-10b (#21287)\n\n* feat: support step3-vl-10b\n\n* use fused QKV && mapping tensor in tensor_mapping.py\n\n* guard hardcoded params and drop crop metadata\n\n* get understand_projector_stride from global config\n\n* img_u8_resize_bilinear_to_f32 move in step3vl class\n\n* Apply suggestions from code review\n\nCo-authored-by: Sigbjørn Skjæret \n\n* fix the \\r\\n mess\n\n* add width and heads to MmprojModel.set_gguf_parameters\n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"c15390c410b5267d74bf5db61c542e8f763a3e34"}},{"name":"whisper-enc.cpp","mode":"0100644","size":4727,"lastCommit":{"hash":"f5c6ae18278b75712fc992587d9ea527d4ea2218","oid":"f5c6ae18278b75712fc992587d9ea527d4ea2218","parents":["5a69c974392020e514c3b2b2910bb92f847cb4c9"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-06-06T11:06:51+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-06T11:06:51+02:00"},"message":"mtmd, server: add \"placeholder bitmap\" for counting tokens , add */input_tokens API (#23913)\n\n* mtmd: add \"placeholder bitmap\" for counting tokens w/o preprocessing\n\n* fast path skip preproc for placeholder\n\n* fix build\n\n* correct the api\n\n* add server endpoint + tests\n\n* add object name\n\n* update docs\n\n* add proxy handling\n\n* fix build\n\n* fix audio input path\n\n* use is_placeholder in process_mtmd_prompt()\n\n* nits\n\n* nits (2)\n\n* docs: clarify chat/completions/input_tokens is not official\n\n* fix merge problem","tree":"f14db5652990fbaf3c81bc03a91f0bc27b98c8bd"}},{"name":"yasa2.cpp","mode":"0100644","size":8725,"lastCommit":{"hash":"98d2d2884e28ca8718774caad691ddd525f2f7b2","oid":"98d2d2884e28ca8718774caad691ddd525f2f7b2","parents":["84652b80cfd165c1b6f8c0541803e1967e5c4c34"],"author":{"name":"Kwa Jie Hao","email":"31984694+kwajiehao@users.noreply.github.com","when":"2026-04-22T02:02:49+08:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-04-21T20:02:49+02:00"},"message":"mtmd: Add support for Reka Edge 2603 (#21616)\n\n* feat: (vocab) fix stray text appended in llama_decode_text\n\nRemove accidental concatenation of the full `text` string when\nformatting UNK_BYTE hex escapes. Only the closing \"]\" should be appended.\n\n* feat(mtmd): add Yasa2 vision encoder support\n\nAdd a Yasa2 (ConvNeXtV2-based) vision encoder for reka-edge:\n- Register PROJECTOR_TYPE_YASA2 and tensor name definitions\n- Add yasa2_block/yasa2_stage model structs\n- Implement graph builder with ConvNeXt stages, GRN, adaptive pooling\n- Wire into clip.cpp switch statements and mtmd.cpp init_vision\n- Use mtmd_image_preprocessor_fixed_size for image preprocessing\n\n* feat(chat): add reka-edge template handler (tools, thinking)\n\n- Add chat-reka.cpp/h implementing PEG-based parser for reka-edge format\n- Add Reka-Edge.jinja chat template\n- Detect reka-edge template in try_specialized_template()\n- Add LLAMA_EXAMPLE_MTMD to chat-template-file arg\n\n* feat: add reka vlm to gguf conversion script\n\nConverts Reka Yasa2 hf checkpoints to GGUF format:\n- Text decoder: Llama-arch with tiktoken/BPE vocab\n- Mmproj (--mmproj): ConvNeXt vision backbone + language_projection\n- Generates 2D sincos positional embeddings for vision encoder\n\n* test: add Reka Edge chat template and parser tests\n\n- test-chat-template: oracle tests comparing Jinja engine output vs\n common_chat_templates_apply for text, tools, thinking, images, video\n- test-chat: PEG parser tests for Reka Edge format, round-trip tests\n for image/video content parts, common path integration tests\n\n* scripts: add Reka Edge mixed quantization helper\n\nQ4_0 base quantization with Q8_0 override for the last 8 transformer\nblocks (layers 24-31) via --tensor-type regex.\n\n* fix: adapt chat-reka and tests to upstream API\n\n- Use autoparser::generation_params (not templates_params)\n- Add p.prefix(generation_prompt) to PEG parser\n- Simplify reasoning parser to match LFM2 pattern\n- Remove image/video oracle tests (unsupported by oaicompat parser;\n no other multimodal models test this path)\n\n* fix: avoid duplicate tensor loading in yasa2 vision encoder\n\nTN_YASA_PATCH_W and TN_PATCH_EMBD both resolve to \"v.patch_embd.weight\",\ncausing the same tensor to be loaded twice into ctx_data and overflowing\nthe memory pool. Reuse the tensors already loaded by the common section.\n\n* chore: update image pre-processing settings\n\nThe reka-edge model depends on the following settings in an older\nfork of llama.cpp:\n1. Fixed square resize\n2. BICUBIC\n3. add_padding=false\n\nIn current llama.cpp, this means setting:\n- image_resize_algo = RESIZE_ALGO_BICUBIC\n- image_resize_pad = false\n\n* chore: remove reka gguf conversion script\n\n* chore: remove reka quantization script\n\n* chore: remove unnecessary changes from PR scope\n\nThis commit removes a couple of unnecessary changes for the PR scope:\n1. BPE decoder bug fix - this affects reka edge because there's a bug\nin our tokenization that doesn't represent tokens as special\ntokens. However this isn't meant to be a thinking model so when run\nwith --reasoning off the edge case does not affect us\n\n2. --chat-template-file support from llama-mtmd-cli - the focus is on\nllama-server and the reka edge gguf contains the necessary metadata\nto detect the chat template\n\n3. reka edge oracle test cases - no other model has similar test cases,\nso I removed it for standardization\n\n* chore: remove unnecessary ggml_cast\n\nThis commit removes unnecessary ggml_cast after updating the\nreka vlm -> gguf conversion script on hugging face.\n\n* chore: remove redundant code\n\n* chore: remove unnecessary ggml_cont calls\n\nThis commit removes all ggml_cont calls except the four that\nprecede ggml_reshape_3d/ggml_reshape_4d. Those are necessary\nbecause ggml_reshape recomputes strides assuming contiguous\nlayout and asserts ggml_is_contiguous.\n\nOther operations (ggml_mean, ggml_add, ggml_mul etc.) use\nstride-based indexing and handle non-contiguous inputs\ncorrectly and so we are ok to remove ggml_cont for those.\n\n* chore: remove unnecessary ggml_repeat calls\n\nThis commit removes unnecessary ggml_repeat calls because the underlying\nops already broadcast automatically.\n\nEvery ggml_repeat in yasa2.cpp was expanding a smaller tensor to match\na larger one's shape before passing both to an elementwise op (ggml_add,\nggml_sub, ggml_mul, or ggml_div). This is unnecessary because all four\nof these ops already support broadcasting internally.\n\n* chore: restore ggml_cont needed for cpu operations\n\n* refactor: locate reka chat template handler in chat.cpp\n\n* chore: remove unnecessary warmup tokens\n\n* chore: add code comments on image_resize_pad\n\n* chore: remove custom reka parsing code\n\n* chore: revert common/chat.cpp\n\n* Uncomment debug logging for PEG input parsing\n\n---------\n\nCo-authored-by: Piotr Wilkin (ilintar) ","tree":"ea39c0ae47ca7df9d71bad3287df2f9c1ea2b9de"}},{"name":"youtuvl.cpp","mode":"0100644","size":6827,"lastCommit":{"hash":"1e645345702154ba4813d3d9bbdbd97718de82c0","oid":"1e645345702154ba4813d3d9bbdbd97718de82c0","parents":["cd708db0ccd59e0770b5ff655db55c995b6fd174"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-03-19T13:11:39+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-03-19T13:11:39+01:00"},"message":"mtmd: add clip_graph::build_mm() (#20751)\n\n* clip: add build_mm()\n\n* apply to all models\n\n* add TODO for bias overload","tree":"199f6c4e93e7126cb52a6a91ed19e502fd794cfa"}}],"lastCommit":{"hash":"b21e4de74567f5eef213765c9476a843c2e43f0d","oid":"b21e4de74567f5eef213765c9476a843c2e43f0d","parents":["d9f918d2d06079b4336688e819eee821c8a9cd9e"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-08-22T16:33:47+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-22T16:33:47+02:00"},"message":"mtmd: use ggml_rope_set_offset (#27521)\n\n* mtmd: use ggml_rope_set_offset\n\n* add comment","tree":"612d00b0e6ee334c61cbdff8c05eabf45eb8a98a"}}