{"entries":[{"name":"afmoe.cpp","mode":"0100644","size":12187,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"apertus.cpp","mode":"0100644","size":7005,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"arcee.cpp","mode":"0100644","size":5443,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"arctic.cpp","mode":"0100644","size":6371,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"arwkv7.cpp","mode":"0100644","size":9322,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"baichuan.cpp","mode":"0100644","size":5263,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"bailingmoe.cpp","mode":"0100644","size":6903,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"bailingmoe2.cpp","mode":"0100644","size":9853,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"bailingmoe3.cpp","mode":"0100644","size":30014,"lastCommit":{"hash":"2115b73d8ebdbd659075cce66c609506863bc826","oid":"2115b73d8ebdbd659075cce66c609506863bc826","parents":["54ee5ee643f29abba6852903ddfdb688c2361b5b"],"author":{"name":"Tiwei Bie","email":"tiwei.btw@antgroup.com","when":"2026-08-22T17:19:48+08:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-22T12:19:48+03:00"},"message":"model : support DSpark for bailingmoe3 (#27508)","tree":"84f228da5692685fdc47f011bd467a1ff749d563"}},{"name":"bert.cpp","mode":"0100644","size":10160,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"bitnet.cpp","mode":"0100644","size":6483,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"bloom.cpp","mode":"0100644","size":5508,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"chameleon.cpp","mode":"0100644","size":7540,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"chatglm.cpp","mode":"0100644","size":5296,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"clip.cpp","mode":"0100644","size":636,"lastCommit":{"hash":"86c298fb8ac1ee9da24e3c4d68b71b1247dc151a","oid":"86c298fb8ac1ee9da24e3c4d68b71b1247dc151a","parents":["2e2d99cfd27d8bab08b6ee36836e20979ba48445"],"author":{"name":"Pedro Cuenca","email":"pedro@huggingface.co","when":"2026-08-10T11:58:32+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-10T11:58:32+02:00"},"message":"llama: Restore quantization of mmprojs (#26818)\n\n* Restore quantization of mmprojs\n\nThis was lost in the refactor undertaken in #22004.\n\n* add noreturn\n\n---------\n\nCo-authored-by: Xuan Son Nguyen ","tree":"3d5d50a6f36e045a85f7d20ca1b8c26a5141ad3f"}},{"name":"codeshell.cpp","mode":"0100644","size":5468,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"cogvlm.cpp","mode":"0100644","size":6341,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"cohere2.cpp","mode":"0100644","size":5815,"lastCommit":{"hash":"4988f6e866057afd130c1515ecef0c9bab9a15f8","oid":"4988f6e866057afd130c1515ecef0c9bab9a15f8","parents":["f05cf4676af46c2f017c0e6ba25b6e20204f700e"],"author":{"name":"Michael Wand","email":"michael@baybridgeaquarium.com","when":"2026-06-13T10:49:00-07:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-13T19:49:00+02:00"},"message":"Add arch support for cohere2-MoE (#24260)\n\n* Add arch support for cohere2-MoE\n\n* Removed redundant gating_func checks\n\n* Changed ffn lookup to prefer prefix_dense_intermediate_size\n\n* Renamed arch to cohere2moe\n\n* Removed redundant lmhead check and chat template changes\n\n* Removed lm_head.weight check from modify tensors, load output tensor not required, fallback to token_embd.weight\n\n* Changed to (routed+shared)*0.5 for shared expert combined avg\n\n* fixed sliding_window_pattern issue and pattern\n\n* Fixed transformers crash 'first_k_dense_replace' error\n\n* Remove comment\n\n* Removed cohere2-moe as a tokenizer type and kept as tiny_aya. Renamed North-Mini-Code-1.0.\n\n* Fixed MTP fail, changed to use iSWA\n\n* Fixed remaining todos: cohere2moe renamed, changed swa parsing to use get_key_or_arr, removed extra get_arr use\n\n* Force metadata usage\n\nCo-authored-by: Sigbjørn Skjæret \n\n* Remove Cohere2 checkpoint comment\n\nCo-authored-by: Sigbjørn Skjæret \n\n* Remove MTP comment\n\nCo-authored-by: Sigbjørn Skjæret \n\n* Regenerate cohere2moe tokenizer hash\n\n* Add cohere2moe to Llama Model Saver supported list\n\n* Check for zerobios tensors and add support for Command to use LayerNorm\n\n* Map expert_selection_fn to sigmoid in base.py instead of command.py\n\n* use bools for foundnorm/foundnormrms\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"97c854a834b63fb5cca6e76ad6a5248ce90ade19"}},{"name":"cohere2moe.cpp","mode":"0100644","size":19689,"lastCommit":{"hash":"82dbc4f017a7b005f993ac2e7af9c048ad686c04","oid":"82dbc4f017a7b005f993ac2e7af9c048ad686c04","parents":["6f3c0a790be2c8599663a98f03b95fb3eae767f8"],"author":{"name":"fairydreaming","email":"166155368+fairydreaming@users.noreply.github.com","when":"2026-07-31T14:57:02+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-07-31T14:57:02+02:00"},"message":"llama : load MTP tensors only if they are really used (#26296)\n\n* llama : load MTP tensors only if they are really used\n\n* llama : skip loading MTP (if not used) in remaining models that support MTP\n\n---------\n\nCo-authored-by: Stanisław Szymczyk ","tree":"e6a7f01c7e397557178a1bb21b1383659cd0095c"}},{"name":"command-r.cpp","mode":"0100644","size":5325,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"dbrx.cpp","mode":"0100644","size":5250,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"deci.cpp","mode":"0100644","size":7920,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"deepseek.cpp","mode":"0100644","size":7981,"lastCommit":{"hash":"42928bc14d33bde7b8fe8ea436ca44b40357737e","oid":"42928bc14d33bde7b8fe8ea436ca44b40357737e","parents":["59778f0196a82db32580bb649d5d839355d6d7bf"],"author":{"name":"ynankani","email":"ynankani@nvidia.com","when":"2026-05-16T09:09:27Z"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-05-16T11:09:27+02:00"},"message":"model : NvFP4 quantized LM head support (#23046)\n\n* NvFP4 quantized LM head support\n\nSigned-off-by: ynankani \n\n* Address review commnets\n\nSigned-off-by: ynankani \n\n* Add assert for NvFp4 lm head and tied embeddings\n\nSigned-off-by: ynankani \n\n* Address review commnets\n\nSigned-off-by: ynankani \n\n* Create output_s tensor only when LM head NvFp4\n\nSigned-off-by: ynankani \n\n---------\n\nSigned-off-by: ynankani ","tree":"50aa23949b519f164ae9f7332b3a1b9b018bb931"}},{"name":"deepseek2.cpp","mode":"0100644","size":32435,"lastCommit":{"hash":"873e5d8e39feb34a376e0efd01bf3f665dfffeb5","oid":"873e5d8e39feb34a376e0efd01bf3f665dfffeb5","parents":["d7fa69b7de2186b16c1024572273130d631038d7"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-08-21T18:54:29+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-21T18:54:29+02:00"},"message":"model: use ggml_rope_set_offset() (#27382)\n\n* model: use ggml_rope_set_offset()\n\n* partially apply to deepseek2","tree":"f508518e9602084c61100b527fa1fa579383f983"}},{"name":"deepseek2ocr.cpp","mode":"0100644","size":4150,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"deepseek32.cpp","mode":"0100644","size":35570,"lastCommit":{"hash":"2e92ecd0247d25f09797f8fdb044a166522fc05d","oid":"2e92ecd0247d25f09797f8fdb044a166522fc05d","parents":["645ca2834bc16e7eab112a91aeb282ebb913f935"],"author":{"name":"Sigbjørn Skjæret","email":"sigbjorn.skjaeret@huggingface.co","when":"2026-08-19T16:05:50+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-19T16:05:50+02:00"},"message":"models : remove duplicate metadata load (#27378)","tree":"8239b71f52571faf4f8971afe691600e4969feaa"}},{"name":"deepseek4.cpp","mode":"0100644","size":64287,"lastCommit":{"hash":"b0539c43ed13b16bf0d8a0840646faea65469702","oid":"b0539c43ed13b16bf0d8a0840646faea65469702","parents":["d3371929bb1b6982cf73f1e54156d3d426cd80a1"],"author":{"name":"Aman Gupta","email":"amangupta052@gmail.com","when":"2026-08-23T16:27:49+05:30"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-23T13:57:49+03:00"},"message":"DeepseekV4: fix rollback with multi-seq (#26756)\n\n* DeepseekV4: fix rollback with multi-seq\n\n* fix model loading\n\n* make pending rollback single use\n\n* only clear cache for seq_id for full load\n\n* add assert for compress ratio\n\n* make graph topology static\n\n* pass true instead of flags in clear_compressed\n\n* cont : clean-up + TODOs\n\n---------\n\nCo-authored-by: Georgi Gerganov ","tree":"322c1a6c7a2cb6f689e20906f4818e5d30900f3a"}},{"name":"delta-net-base.cpp","mode":"0100644","size":26208,"lastCommit":{"hash":"6d1917060d72cfffc6b54391ef8006b7f51361b9","oid":"6d1917060d72cfffc6b54391ef8006b7f51361b9","parents":["c0209b63a39312975f3cbff7670afda41f653cd9"],"author":{"name":"nekomimi.pet","email":"meowskulls@nekomimi.pet","when":"2026-10-13T19:13:57-04:00"},"committer":{"name":"nekomimi.pet","email":"meowskulls@nekomimi.pet","when":"2026-10-13T19:13:57-04:00"},"message":"bonsai: fused CUDA kernel for DeltaNet conv path (gdn-conv)\n\nOne CUDA kernel for the DeltaNet conv path (concat -> ssm_conv -> silu -> conv-state writes): the concat is never materialized; same arithmetic. Bit-identical; MTP decode +1.5%. GGML_CUDA_SSM_CONV_CONCAT_FUSION=0 turns it off.\n","tree":"b62a7b2173f152483ed21ab164952353e634072b"}},{"name":"dflash.cpp","mode":"0100644","size":66706,"lastCommit":{"hash":"279df66449f597f9b91d1976865486ba33e2b984","oid":"279df66449f597f9b91d1976865486ba33e2b984","parents":["9a5c96acb158e92e4c7910b0986c3927025777f0"],"author":{"name":"Chirag Patnaik","email":"chirag.patnaik@gmail.com","when":"2026-09-25T23:52:35+05:30"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-09-25T11:22:35-07:00"},"message":"spec : add DFlash2 support (local convolution + candidate selector) (#27342) (#27816) (#261)\n\n* spec : add DFlash2 support (local convolution + candidate selector) (#27342)\n\n* support DFlash2\n\n* Add p_min in DFlash2\n\nAssisted-by: Claude Opus 5\n\n* Revert unnecessary changes\n\nAssisted-by: Claude Opus 5\n\n* Revert draft sampling in rejection sampling\n\nAssisted-by: Claude Opus 5\n\n* Refactor code structure\n\nAssisted-by: Claude Opus 5\n\n* Delete embedding scaling\n\nAssisted-by: Claude Opus 5\n\n* Gate output transforms on DFlash2\n\nAssisted-by: Claude Opus 5\n\n* Optimize Dflash 2 cost\n\nAssisted-by: Claude Opus 5\n\n* Avoid using atoi\n\nAssisted-by: Claude Opus 5\n\n* Modify comments\n\nAssisted-by: Claude Opus 5\n\n* Move llama_model_dflash_selector_top_k to llama-ext.h\n\nAssisted-by: Claude Opus 5\n\n* Formatting\n\nAssisted-by: Claude Opus 5\n\n* Apply patch to fix the mrope bug\n\nAssisted-by: Claude Opus 5\n\n* fix ci\n\nAssisted-by: Claude Opus 5\n\n* Fix graph number calculation\n\nAssisted-by: Claude Opus 5\n\n* rename hid and unary\n\nAssisted-by: Claude Opus 5\n\n---------\n\n\n\n\n* revert top-k.cu changes\n\n---------\n\n\n\n(cherry picked from commit b10f9ca58c89ccfc3653ac01e979dd085d582b76)\n\nCo-authored-by: Xuan-Son Nguyen \nCo-authored-by: Zihan Zhang \nCo-authored-by: Jian Chen ","tree":"8ac2a17f083de363d9ffe06d35e10c1e1d4eff58"}},{"name":"dots1.cpp","mode":"0100644","size":8250,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"dots3note.cpp","mode":"0100644","size":24717,"lastCommit":{"hash":"5a32f7b66ef6cfb3e60deea26e3454cc6ad3438c","oid":"5a32f7b66ef6cfb3e60deea26e3454cc6ad3438c","parents":["873e5d8e39feb34a376e0efd01bf3f665dfffeb5"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-08-21T19:52:34+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-21T19:52:34+02:00"},"message":"model: add dots3-note (#27060)\n\n* text: conversion\n\n* init impl\n\n* address review comments\n\n* fix rope\n\n* move to a new llama_kv_cache_dsa_iswa","tree":"163ec692ca037c7897f5fb92a2fe6ce9c8083ea1"}},{"name":"dream.cpp","mode":"0100644","size":4951,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"dspark.cpp","mode":"0100644","size":32108,"lastCommit":{"hash":"fa8a8838582cec412a6f8ebc2239f57be752d5de","oid":"fa8a8838582cec412a6f8ebc2239f57be752d5de","parents":["6067dbea986ed1359db98c42ab06eb49426bbf16"],"author":{"name":"bri-prism","email":"288398250+bri-prism@users.noreply.github.com","when":"2026-09-17T21:46:43-07:00"},"committer":{"name":"bri-prism","email":"288398250+bri-prism@users.noreply.github.com","when":"2026-09-17T21:46:59-07:00"},"message":"dspark: clarify runtime terminology\n","tree":"1370643609f10c87bae8c3fde4b2892cab5d3478"}},{"name":"eagle3.cpp","mode":"0100644","size":13966,"lastCommit":{"hash":"f87067841bac583bc089a225382248d857791ca8","oid":"f87067841bac583bc089a225382248d857791ca8","parents":["c6292cfb8e964bd9f2146f3c603507627e4a22c3"],"author":{"name":"Ruixiang Wang","email":"wangruixiang07@outlook.com","when":"2026-07-28T08:58:16+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-07-28T09:58:16+03:00"},"message":"spec: add eagle3-v3 support for gpt-oss model (#25794)","tree":"2f9b35cb629916c84aa53b336299621434988ab5"}},{"name":"ernie4-5-moe.cpp","mode":"0100644","size":5091,"lastCommit":{"hash":"42928bc14d33bde7b8fe8ea436ca44b40357737e","oid":"42928bc14d33bde7b8fe8ea436ca44b40357737e","parents":["59778f0196a82db32580bb649d5d839355d6d7bf"],"author":{"name":"ynankani","email":"ynankani@nvidia.com","when":"2026-05-16T09:09:27Z"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-05-16T11:09:27+02:00"},"message":"model : NvFP4 quantized LM head support (#23046)\n\n* NvFP4 quantized LM head support\n\nSigned-off-by: ynankani \n\n* Address review commnets\n\nSigned-off-by: ynankani \n\n* Add assert for NvFp4 lm head and tied embeddings\n\nSigned-off-by: ynankani \n\n* Address review commnets\n\nSigned-off-by: ynankani \n\n* Create output_s tensor only when LM head NvFp4\n\nSigned-off-by: ynankani \n\n---------\n\nSigned-off-by: ynankani ","tree":"50aa23949b519f164ae9f7332b3a1b9b018bb931"}},{"name":"ernie4-5.cpp","mode":"0100644","size":6933,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"eurobert.cpp","mode":"0100644","size":4167,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"exaone-moe.cpp","mode":"0100644","size":11240,"lastCommit":{"hash":"59917d3922e976ae4d7a86eb976bd4c330fb5391","oid":"59917d3922e976ae4d7a86eb976bd4c330fb5391","parents":["7acb4e8cd2ce21f457d1298e75fad729520d263c"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:17:54+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:17:54+03:00"},"message":"minor : fix lint issues (#24165)","tree":"8b09e924bce0c6b4e7a9fe2d80e6ae3839d9f53b"}},{"name":"exaone.cpp","mode":"0100644","size":5123,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"exaone4.cpp","mode":"0100644","size":7895,"lastCommit":{"hash":"14e78ddef7a2061e7d5a31dce4eb7ee0bcdbc840","oid":"14e78ddef7a2061e7d5a31dce4eb7ee0bcdbc840","parents":["48d22e295e2b86b47366c16390794f3e05ba970a"],"author":{"name":"Junmo Kim","email":"me@junmo.kim","when":"2026-08-11T13:20:17+09:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-11T07:20:17+03:00"},"message":"model : fix SWA not being enabled for EXAONE 4.5 (#26848)\n\n* model : fix SWA not being enabled for EXAONE 4.5\n\nload_arch_hparams tests `hparams.n_layer() == 64` before\nLLM_KV_NEXTN_PREDICT_LAYERS has been read. n_layer() returns\nn_layer_all - n_layer_nextn and n_layer_nextn defaults to 0, so a GGUF\ncarrying the MTP head (block_count=65, nextn=1) evaluates to 65 and the\nwhole SWA block is skipped. The model type switch further down in the\nsame function reads 64, because by then the key has been loaded.\n\nn_swa is still filled in by the unconditional get_key below the block, so\nllama_model_n_swa() reports 4096 and the logs look correct while only\nswa_type stays LLAMA_SWA_TYPE_NONE.\n\nThis affects the official LGAI-EXAONE GGUF release as well. EXAONE 4.0 has\nno MTP head, so block_count is 64 there and the check matches.\n\n* model-loader : skip TENSOR_SKIP tensors in the metadata-only path\n\ncreate_tensor asserts on a null buffer type when building from metadata\nalone, but buft_for_tensor returns null by design for tensors marked\nTENSOR_SKIP, which is how architectures with nextn/MTP layers mark theirs.\nThose models cannot be constructed by llama_model_init_from_user at all.\n\nThe file-backed path below already returns nullptr for the same tensors, so\ncallers see the same thing either way.\n\n* tests : cover exaone4 hparams ordering\n\nBuilds a synthetic exaone4 model with the layout the shipped EXAONE 4.5\nGGUFs use (block_count 65 + nextn 1). The swa_type check is the one that\ncatches the ordering bug; the n_layer_nextn and n_layer() checks only tell\na broken fixture apart from a real regression.\n\nFails before the ordering fix with \"swa_type is not STANDARD\", passes after.\n\n* Revert \"tests : cover exaone4 hparams ordering\"\n\nThis reverts commit d2f3bafeee591ad691396b2708de4baef3aaf602.\n\n* Revert \"model-loader : skip TENSOR_SKIP tensors in the metadata-only path\"\n\nThis reverts commit aecb9bc0c7896b52afbc43921a1f572aa7b5e53c.","tree":"52df1dfa1d8c0e37bb748fdd3c0ff1c7fdf2aeb2"}},{"name":"falcon-h1.cpp","mode":"0100644","size":9017,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"falcon.cpp","mode":"0100644","size":5607,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"gemma-embedding.cpp","mode":"0100644","size":7484,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"gemma.cpp","mode":"0100644","size":4858,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"gemma2.cpp","mode":"0100644","size":6841,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"gemma3.cpp","mode":"0100644","size":8784,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"gemma3n.cpp","mode":"0100644","size":23665,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"gemma4-assistant.cpp","mode":"0100644","size":8758,"lastCommit":{"hash":"7d2b45b4f7b663cda74f23fbc3ce6dc3bd4f6545","oid":"7d2b45b4f7b663cda74f23fbc3ce6dc3bd4f6545","parents":["42a0afd59431815585ef8daa3277f9206b235531"],"author":{"name":"Max Krasnyansky","email":"maxk@qti.qualcomm.com","when":"2026-06-08T13:48:52-07:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-08T13:48:52-07:00"},"message":"mtp: support for gemma-4 E2B and E4B assistants (#24282)\n\n* models: update converter to support smaller assistants\n\n* models: add masked_embd tensors to gemma4-assist arch\n\n* gemma-4: remove temp debug for conversion\n\n* gemma-4-mtp: filter out masked_embedding tensors during conversion","tree":"eff64d75169aea06940e29b3d32f29eb9d5033ec"}},{"name":"gemma4.cpp","mode":"0100644","size":21905,"lastCommit":{"hash":"afeebe103bd99cda8f5dfaefcabadf890db7fda7","oid":"afeebe103bd99cda8f5dfaefcabadf890db7fda7","parents":["caa596ab3f0f8768ee326d6e3d5d39782194676c"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-07-29T18:02:30+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-07-29T18:02:30+02:00"},"message":"llama: move suppress_tokens handling to common/sampling (#26276)\n\n* llama: move suppress_tokens handling to common/sampling\n\n* address security issues\n\n* rm has_logit_bias","tree":"72039035a1efbc92c69d85859cafba91d1ece6ec"}},{"name":"glm-dsa.cpp","mode":"0100644","size":37666,"lastCommit":{"hash":"2e92ecd0247d25f09797f8fdb044a166522fc05d","oid":"2e92ecd0247d25f09797f8fdb044a166522fc05d","parents":["645ca2834bc16e7eab112a91aeb282ebb913f935"],"author":{"name":"Sigbjørn Skjæret","email":"sigbjorn.skjaeret@huggingface.co","when":"2026-08-19T16:05:50+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-19T16:05:50+02:00"},"message":"models : remove duplicate metadata load (#27378)","tree":"8239b71f52571faf4f8971afe691600e4969feaa"}},{"name":"glm4-moe.cpp","mode":"0100644","size":19297,"lastCommit":{"hash":"c060ca974c773c7c3d17fd1b66dc9d312bc292c0","oid":"c060ca974c773c7c3d17fd1b66dc9d312bc292c0","parents":["ccc8fd2baa64ab26c7210b02765d3604069ec7ee"],"author":{"name":"jacekpoplawski","email":"67507230+jacekpoplawski@users.noreply.github.com","when":"2026-08-23T20:20:44+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-23T21:20:44+03:00"},"message":"model : support MTP in GLM-4.5-Air (#26534)","tree":"1e6240a2b7065d828b817793a35f5f39a4f369fc"}},{"name":"glm4.cpp","mode":"0100644","size":8019,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"gpt2.cpp","mode":"0100644","size":5368,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"gptneox.cpp","mode":"0100644","size":8009,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"granite-hybrid.cpp","mode":"0100644","size":13612,"lastCommit":{"hash":"7221e24f579efbb5fd2a2ced1fffaef1aff2fe0d","oid":"7221e24f579efbb5fd2a2ced1fffaef1aff2fe0d","parents":["6cc504a2e90d5e41fb37e98b1da879f72d4d9f31"],"author":{"name":"Gabe Goodhart","email":"ghart@us.ibm.com","when":"2026-08-19T08:53:31-06:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-19T16:53:31+02:00"},"message":"model : GraniteSWAForCausalLM / GraniteMoeSWAForCausalLM (#25505)\n\n* feat(convert): Add conversion for GraniteSWAForCausalLM\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob, OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* feat(llama): Add granite_swa support\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob, OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* feat(conversion): Add conversion infra for rope_pattern array\n\nNOTE: There is other work also targeting this, so this may be\nremoved depending on merge order.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* fix(conversion): Fix SWA pattern logic and support for non-rope layers\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* feat(conversion): Add support for GraniteMoeSWA\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* feat: Add llama_hparams::has_rope and arch constants\n\nNOTE: This shadows the work done for Granite Speech\nhttps://github.com/ggml-org/llama.cpp/pull/25107\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* feat: Add support for per-layer rope determination\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* style: Fix failing flake8 for extra newlines\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* test: Write out SLIDING_WINDOW_PATTERN in llama-model-saver\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* fix(convert): Fix missing registration for GraniteMoeSWAForCausalLM\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Load MoE params as optional\n\nBranch: GraniteSWAForCausalLM\nAI-usage: draft (OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* feat: Handle MoE params in conversion\n\nbranch: GraniteSWAForCausalLM\nAI-usage: full (OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* style: Remove unnecessary newline\n\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Remove unnecessary tensor additions to GRANITE architecture\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Correctly handle naming for ffn gate inp\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Always default hparams.rope_pattern to 1s\n\nThis isn't strictly necessary, but it will allow other models to rely on\nhparams.has_rope(il) without needting to prepopulate.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* feat: Move to has_rope for all granite model architectures\n\nNow that we have a proper hparam for this, it's better to use it and not\nrequire a hacky fallback in the hparam method itself.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* feat: No hacky rope_finetuned fallback in has_rope\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Fully remove rope hparam filling in granitemoe\n\nThere are no granitemoe models that use NoPE (it's not actually used in the\nlayer building below), so this was just dead code.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Save out rope_pattern in model-saver\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Set hparams.rope_finetuned for round trip\n\nSince the value is _read_ from rope_finetuned, we need to persist it when\nthe model is saved with the saver.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Code review cleanup\n\nSigned-off-by: Gabe Goodhart \nCo-authored-by: Sigbjørn Skjæret \n\nCo-authored-by: Sigbjørn Skjæret \n\n* refactor: Keep gate/up fused for MoE path\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Claude + Sonnet 5)\nSigned-off-by: Gabe Goodhart \n\n* fix: Skip GRANITE_SWA in model saver\n\nhttps://github.com/ggml-org/llama.cpp/pull/25505#discussion_r3773175651\n\nKeeping is_swa_impl in the saver can break other models.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* add sliding window pattern for model in test\n\n* style: Fix indentation\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Fix \\r\\n\n\nThanks Claude!\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* feat: Keep shared expert fused\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Claude + Sonnet 5)\nSigned-off-by: Gabe Goodhart \n\n* style: More indentation fixes\n\nSigned-off-by: Gabe Goodhart \n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nSigned-off-by: Gabe Goodhart \nCo-authored-by: Sigbjørn Skjæret ","tree":"88310e187bddd2dabd71e7b4aabbf0567620d0a1"}},{"name":"granite-moe.cpp","mode":"0100644","size":4583,"lastCommit":{"hash":"7221e24f579efbb5fd2a2ced1fffaef1aff2fe0d","oid":"7221e24f579efbb5fd2a2ced1fffaef1aff2fe0d","parents":["6cc504a2e90d5e41fb37e98b1da879f72d4d9f31"],"author":{"name":"Gabe Goodhart","email":"ghart@us.ibm.com","when":"2026-08-19T08:53:31-06:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-19T16:53:31+02:00"},"message":"model : GraniteSWAForCausalLM / GraniteMoeSWAForCausalLM (#25505)\n\n* feat(convert): Add conversion for GraniteSWAForCausalLM\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob, OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* feat(llama): Add granite_swa support\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob, OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* feat(conversion): Add conversion infra for rope_pattern array\n\nNOTE: There is other work also targeting this, so this may be\nremoved depending on merge order.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* fix(conversion): Fix SWA pattern logic and support for non-rope layers\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* feat(conversion): Add support for GraniteMoeSWA\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* feat: Add llama_hparams::has_rope and arch constants\n\nNOTE: This shadows the work done for Granite Speech\nhttps://github.com/ggml-org/llama.cpp/pull/25107\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* feat: Add support for per-layer rope determination\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* style: Fix failing flake8 for extra newlines\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* test: Write out SLIDING_WINDOW_PATTERN in llama-model-saver\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* fix(convert): Fix missing registration for GraniteMoeSWAForCausalLM\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Load MoE params as optional\n\nBranch: GraniteSWAForCausalLM\nAI-usage: draft (OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* feat: Handle MoE params in conversion\n\nbranch: GraniteSWAForCausalLM\nAI-usage: full (OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* style: Remove unnecessary newline\n\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Remove unnecessary tensor additions to GRANITE architecture\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Correctly handle naming for ffn gate inp\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Always default hparams.rope_pattern to 1s\n\nThis isn't strictly necessary, but it will allow other models to rely on\nhparams.has_rope(il) without needting to prepopulate.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* feat: Move to has_rope for all granite model architectures\n\nNow that we have a proper hparam for this, it's better to use it and not\nrequire a hacky fallback in the hparam method itself.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* feat: No hacky rope_finetuned fallback in has_rope\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Fully remove rope hparam filling in granitemoe\n\nThere are no granitemoe models that use NoPE (it's not actually used in the\nlayer building below), so this was just dead code.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Save out rope_pattern in model-saver\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Set hparams.rope_finetuned for round trip\n\nSince the value is _read_ from rope_finetuned, we need to persist it when\nthe model is saved with the saver.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Code review cleanup\n\nSigned-off-by: Gabe Goodhart \nCo-authored-by: Sigbjørn Skjæret \n\nCo-authored-by: Sigbjørn Skjæret \n\n* refactor: Keep gate/up fused for MoE path\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Claude + Sonnet 5)\nSigned-off-by: Gabe Goodhart \n\n* fix: Skip GRANITE_SWA in model saver\n\nhttps://github.com/ggml-org/llama.cpp/pull/25505#discussion_r3773175651\n\nKeeping is_swa_impl in the saver can break other models.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* add sliding window pattern for model in test\n\n* style: Fix indentation\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Fix \\r\\n\n\nThanks Claude!\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* feat: Keep shared expert fused\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Claude + Sonnet 5)\nSigned-off-by: Gabe Goodhart \n\n* style: More indentation fixes\n\nSigned-off-by: Gabe Goodhart \n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nSigned-off-by: Gabe Goodhart \nCo-authored-by: Sigbjørn Skjæret ","tree":"88310e187bddd2dabd71e7b4aabbf0567620d0a1"}},{"name":"granite-swa.cpp","mode":"0100644","size":12528,"lastCommit":{"hash":"7221e24f579efbb5fd2a2ced1fffaef1aff2fe0d","oid":"7221e24f579efbb5fd2a2ced1fffaef1aff2fe0d","parents":["6cc504a2e90d5e41fb37e98b1da879f72d4d9f31"],"author":{"name":"Gabe Goodhart","email":"ghart@us.ibm.com","when":"2026-08-19T08:53:31-06:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-19T16:53:31+02:00"},"message":"model : GraniteSWAForCausalLM / GraniteMoeSWAForCausalLM (#25505)\n\n* feat(convert): Add conversion for GraniteSWAForCausalLM\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob, OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* feat(llama): Add granite_swa support\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob, OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* feat(conversion): Add conversion infra for rope_pattern array\n\nNOTE: There is other work also targeting this, so this may be\nremoved depending on merge order.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* fix(conversion): Fix SWA pattern logic and support for non-rope layers\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* feat(conversion): Add support for GraniteMoeSWA\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* feat: Add llama_hparams::has_rope and arch constants\n\nNOTE: This shadows the work done for Granite Speech\nhttps://github.com/ggml-org/llama.cpp/pull/25107\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* feat: Add support for per-layer rope determination\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* style: Fix failing flake8 for extra newlines\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* test: Write out SLIDING_WINDOW_PATTERN in llama-model-saver\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* fix(convert): Fix missing registration for GraniteMoeSWAForCausalLM\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Load MoE params as optional\n\nBranch: GraniteSWAForCausalLM\nAI-usage: draft (OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* feat: Handle MoE params in conversion\n\nbranch: GraniteSWAForCausalLM\nAI-usage: full (OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* style: Remove unnecessary newline\n\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Remove unnecessary tensor additions to GRANITE architecture\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Correctly handle naming for ffn gate inp\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Always default hparams.rope_pattern to 1s\n\nThis isn't strictly necessary, but it will allow other models to rely on\nhparams.has_rope(il) without needting to prepopulate.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* feat: Move to has_rope for all granite model architectures\n\nNow that we have a proper hparam for this, it's better to use it and not\nrequire a hacky fallback in the hparam method itself.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* feat: No hacky rope_finetuned fallback in has_rope\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Fully remove rope hparam filling in granitemoe\n\nThere are no granitemoe models that use NoPE (it's not actually used in the\nlayer building below), so this was just dead code.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Save out rope_pattern in model-saver\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Set hparams.rope_finetuned for round trip\n\nSince the value is _read_ from rope_finetuned, we need to persist it when\nthe model is saved with the saver.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Code review cleanup\n\nSigned-off-by: Gabe Goodhart \nCo-authored-by: Sigbjørn Skjæret \n\nCo-authored-by: Sigbjørn Skjæret \n\n* refactor: Keep gate/up fused for MoE path\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Claude + Sonnet 5)\nSigned-off-by: Gabe Goodhart \n\n* fix: Skip GRANITE_SWA in model saver\n\nhttps://github.com/ggml-org/llama.cpp/pull/25505#discussion_r3773175651\n\nKeeping is_swa_impl in the saver can break other models.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* add sliding window pattern for model in test\n\n* style: Fix indentation\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Fix \\r\\n\n\nThanks Claude!\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* feat: Keep shared expert fused\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Claude + Sonnet 5)\nSigned-off-by: Gabe Goodhart \n\n* style: More indentation fixes\n\nSigned-off-by: Gabe Goodhart \n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nSigned-off-by: Gabe Goodhart \nCo-authored-by: Sigbjørn Skjæret ","tree":"88310e187bddd2dabd71e7b4aabbf0567620d0a1"}},{"name":"granite-switch.cpp","mode":"0100644","size":18868,"lastCommit":{"hash":"7221e24f579efbb5fd2a2ced1fffaef1aff2fe0d","oid":"7221e24f579efbb5fd2a2ced1fffaef1aff2fe0d","parents":["6cc504a2e90d5e41fb37e98b1da879f72d4d9f31"],"author":{"name":"Gabe Goodhart","email":"ghart@us.ibm.com","when":"2026-08-19T08:53:31-06:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-19T16:53:31+02:00"},"message":"model : GraniteSWAForCausalLM / GraniteMoeSWAForCausalLM (#25505)\n\n* feat(convert): Add conversion for GraniteSWAForCausalLM\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob, OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* feat(llama): Add granite_swa support\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob, OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* feat(conversion): Add conversion infra for rope_pattern array\n\nNOTE: There is other work also targeting this, so this may be\nremoved depending on merge order.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* fix(conversion): Fix SWA pattern logic and support for non-rope layers\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* feat(conversion): Add support for GraniteMoeSWA\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* feat: Add llama_hparams::has_rope and arch constants\n\nNOTE: This shadows the work done for Granite Speech\nhttps://github.com/ggml-org/llama.cpp/pull/25107\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* feat: Add support for per-layer rope determination\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* style: Fix failing flake8 for extra newlines\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* test: Write out SLIDING_WINDOW_PATTERN in llama-model-saver\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* fix(convert): Fix missing registration for GraniteMoeSWAForCausalLM\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Load MoE params as optional\n\nBranch: GraniteSWAForCausalLM\nAI-usage: draft (OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* feat: Handle MoE params in conversion\n\nbranch: GraniteSWAForCausalLM\nAI-usage: full (OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* style: Remove unnecessary newline\n\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Remove unnecessary tensor additions to GRANITE architecture\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Correctly handle naming for ffn gate inp\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Always default hparams.rope_pattern to 1s\n\nThis isn't strictly necessary, but it will allow other models to rely on\nhparams.has_rope(il) without needting to prepopulate.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* feat: Move to has_rope for all granite model architectures\n\nNow that we have a proper hparam for this, it's better to use it and not\nrequire a hacky fallback in the hparam method itself.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* feat: No hacky rope_finetuned fallback in has_rope\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Fully remove rope hparam filling in granitemoe\n\nThere are no granitemoe models that use NoPE (it's not actually used in the\nlayer building below), so this was just dead code.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Save out rope_pattern in model-saver\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Set hparams.rope_finetuned for round trip\n\nSince the value is _read_ from rope_finetuned, we need to persist it when\nthe model is saved with the saver.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Code review cleanup\n\nSigned-off-by: Gabe Goodhart \nCo-authored-by: Sigbjørn Skjæret \n\nCo-authored-by: Sigbjørn Skjæret \n\n* refactor: Keep gate/up fused for MoE path\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Claude + Sonnet 5)\nSigned-off-by: Gabe Goodhart \n\n* fix: Skip GRANITE_SWA in model saver\n\nhttps://github.com/ggml-org/llama.cpp/pull/25505#discussion_r3773175651\n\nKeeping is_swa_impl in the saver can break other models.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* add sliding window pattern for model in test\n\n* style: Fix indentation\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Fix \\r\\n\n\nThanks Claude!\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* feat: Keep shared expert fused\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Claude + Sonnet 5)\nSigned-off-by: Gabe Goodhart \n\n* style: More indentation fixes\n\nSigned-off-by: Gabe Goodhart \n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nSigned-off-by: Gabe Goodhart \nCo-authored-by: Sigbjørn Skjæret ","tree":"88310e187bddd2dabd71e7b4aabbf0567620d0a1"}},{"name":"granite.cpp","mode":"0100644","size":12275,"lastCommit":{"hash":"7221e24f579efbb5fd2a2ced1fffaef1aff2fe0d","oid":"7221e24f579efbb5fd2a2ced1fffaef1aff2fe0d","parents":["6cc504a2e90d5e41fb37e98b1da879f72d4d9f31"],"author":{"name":"Gabe Goodhart","email":"ghart@us.ibm.com","when":"2026-08-19T08:53:31-06:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-19T16:53:31+02:00"},"message":"model : GraniteSWAForCausalLM / GraniteMoeSWAForCausalLM (#25505)\n\n* feat(convert): Add conversion for GraniteSWAForCausalLM\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob, OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* feat(llama): Add granite_swa support\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob, OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* feat(conversion): Add conversion infra for rope_pattern array\n\nNOTE: There is other work also targeting this, so this may be\nremoved depending on merge order.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* fix(conversion): Fix SWA pattern logic and support for non-rope layers\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* feat(conversion): Add support for GraniteMoeSWA\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* feat: Add llama_hparams::has_rope and arch constants\n\nNOTE: This shadows the work done for Granite Speech\nhttps://github.com/ggml-org/llama.cpp/pull/25107\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* feat: Add support for per-layer rope determination\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Bob)\nSigned-off-by: Gabe Goodhart \n\n* style: Fix failing flake8 for extra newlines\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* test: Write out SLIDING_WINDOW_PATTERN in llama-model-saver\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* fix(convert): Fix missing registration for GraniteMoeSWAForCausalLM\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Load MoE params as optional\n\nBranch: GraniteSWAForCausalLM\nAI-usage: draft (OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* feat: Handle MoE params in conversion\n\nbranch: GraniteSWAForCausalLM\nAI-usage: full (OpenCode + Qwen3.6-35b)\nSigned-off-by: Gabe Goodhart \n\n* style: Remove unnecessary newline\n\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Remove unnecessary tensor additions to GRANITE architecture\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Correctly handle naming for ffn gate inp\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Always default hparams.rope_pattern to 1s\n\nThis isn't strictly necessary, but it will allow other models to rely on\nhparams.has_rope(il) without needting to prepopulate.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* feat: Move to has_rope for all granite model architectures\n\nNow that we have a proper hparam for this, it's better to use it and not\nrequire a hacky fallback in the hparam method itself.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* feat: No hacky rope_finetuned fallback in has_rope\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Fully remove rope hparam filling in granitemoe\n\nThere are no granitemoe models that use NoPE (it's not actually used in the\nlayer building below), so this was just dead code.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Save out rope_pattern in model-saver\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Set hparams.rope_finetuned for round trip\n\nSince the value is _read_ from rope_finetuned, we need to persist it when\nthe model is saved with the saver.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Code review cleanup\n\nSigned-off-by: Gabe Goodhart \nCo-authored-by: Sigbjørn Skjæret \n\nCo-authored-by: Sigbjørn Skjæret \n\n* refactor: Keep gate/up fused for MoE path\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Claude + Sonnet 5)\nSigned-off-by: Gabe Goodhart \n\n* fix: Skip GRANITE_SWA in model saver\n\nhttps://github.com/ggml-org/llama.cpp/pull/25505#discussion_r3773175651\n\nKeeping is_swa_impl in the saver can break other models.\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* add sliding window pattern for model in test\n\n* style: Fix indentation\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* fix: Fix \\r\\n\n\nThanks Claude!\n\nBranch: GraniteSWAForCausalLM\nAI-usage: none\nSigned-off-by: Gabe Goodhart \n\n* feat: Keep shared expert fused\n\nBranch: GraniteSWAForCausalLM\nAI-usage: full (Claude + Sonnet 5)\nSigned-off-by: Gabe Goodhart \n\n* style: More indentation fixes\n\nSigned-off-by: Gabe Goodhart \n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nSigned-off-by: Gabe Goodhart \nCo-authored-by: Sigbjørn Skjæret ","tree":"88310e187bddd2dabd71e7b4aabbf0567620d0a1"}},{"name":"grok.cpp","mode":"0100644","size":8962,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"grovemoe.cpp","mode":"0100644","size":7954,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"hunyuan-dense.cpp","mode":"0100644","size":195,"lastCommit":{"hash":"994118a1831fdede28ee2d228e0570db9b728c45","oid":"994118a1831fdede28ee2d228e0570db9b728c45","parents":["c84e6d6db5c637ad8aacfdd41cf96fc1d2b85cec"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-05-04T12:36:59+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-05-04T12:36:59+02:00"},"message":"model: move `load_hparams` and `load_tensors` to per-model definition (#22004)\n\n* git-friendly migration\n\n* add build_graph\n\n* nits\n\n* exclude old code from build\n\n* wip\n\n* add llm_arch_model_i\n\n* prepare downstream functions\n\n* nits\n\n* nits\n\n* wip\n\n* wip\n\n* add back create_tensor_qkv\n\n* fix files missing include\n\n* enforce one llm_build per arch\n\n* cmake: use glob\n\n* missing model params\n\n* nits\n\n* wip\n\n* wip (2)\n\n* wip (3)\n\n* test-llama-archs is happy\n\n* improve switch case\n\n* move more stuff into llm_arch_model_i\n\n* fix downstream code\n\n* nits\n\n* nits (2)\n\n* fix order\n\n* llama_model_base\n\n* LLAMA_LOAD_LOCALS\n\n* small fix\n\n* fix build errors\n\n* auto\n\n* rm migration script and ifdef","tree":"e907b3bfe32589d24e2328448be8c5fa913f94ee"}},{"name":"hunyuan-moe.cpp","mode":"0100644","size":7274,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"hunyuan-vl.cpp","mode":"0100644","size":7306,"lastCommit":{"hash":"42928bc14d33bde7b8fe8ea436ca44b40357737e","oid":"42928bc14d33bde7b8fe8ea436ca44b40357737e","parents":["59778f0196a82db32580bb649d5d839355d6d7bf"],"author":{"name":"ynankani","email":"ynankani@nvidia.com","when":"2026-05-16T09:09:27Z"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-05-16T11:09:27+02:00"},"message":"model : NvFP4 quantized LM head support (#23046)\n\n* NvFP4 quantized LM head support\n\nSigned-off-by: ynankani \n\n* Address review commnets\n\nSigned-off-by: ynankani \n\n* Add assert for NvFp4 lm head and tied embeddings\n\nSigned-off-by: ynankani \n\n* Address review commnets\n\nSigned-off-by: ynankani \n\n* Create output_s tensor only when LM head NvFp4\n\nSigned-off-by: ynankani \n\n---------\n\nSigned-off-by: ynankani ","tree":"50aa23949b519f164ae9f7332b3a1b9b018bb931"}},{"name":"hy-v3.cpp","mode":"0100644","size":18551,"lastCommit":{"hash":"82dbc4f017a7b005f993ac2e7af9c048ad686c04","oid":"82dbc4f017a7b005f993ac2e7af9c048ad686c04","parents":["6f3c0a790be2c8599663a98f03b95fb3eae767f8"],"author":{"name":"fairydreaming","email":"166155368+fairydreaming@users.noreply.github.com","when":"2026-07-31T14:57:02+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-07-31T14:57:02+02:00"},"message":"llama : load MTP tensors only if they are really used (#26296)\n\n* llama : load MTP tensors only if they are really used\n\n* llama : skip loading MTP (if not used) in remaining models that support MTP\n\n---------\n\nCo-authored-by: Stanisław Szymczyk ","tree":"e6a7f01c7e397557178a1bb21b1383659cd0095c"}},{"name":"internlm2.cpp","mode":"0100644","size":4864,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"jais.cpp","mode":"0100644","size":5002,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"jais2.cpp","mode":"0100644","size":6300,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"jamba.cpp","mode":"0100644","size":7853,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"jina-bert-v2.cpp","mode":"0100644","size":3645,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"jina-bert-v3.cpp","mode":"0100644","size":3538,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"kimi-k3.cpp","mode":"0100644","size":30162,"lastCommit":{"hash":"ad1de39e0708e3ced9c71bb3c82d93a2c046a73f","oid":"ad1de39e0708e3ced9c71bb3c82d93a2c046a73f","parents":["22b8e310b921d568e013e4533002be5a8fe53f17"],"author":{"name":"Piotr Wilkin (ilintar)","email":"piotr.wilkin@syndatis.com","when":"2026-08-15T17:11:05+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-15T17:11:05+02:00"},"message":"model: add Kimi-K3 text model (#26185)\n\n* model: add Kimi-K3 text model\n\nHybrid KDA (linear) + MLA (full) attention as in Kimi-Linear-48B, plus five\nthings that architecture does not have:\n\n 1. cross-layer residual attention (attn_res_block_size)\n 2. latent MoE (routed experts run at n_expert_latent)\n 3. situ activation (replaces SwiGLU everywhere)\n 4. MLA output gate (sigmoid gate before o_proj)\n 5. full-rank KDA gate (single ssm_g instead of ssm_g_a/ssm_g_b)\n\nK3's text_config reports KimiLinearForCausalLM - the older 48B architecture -\nso get_model_architecture routes on the top-level name instead.\n\nThe KDA decay gate has two forms, selected by linear_attn_config's\ngate_lower_bound. It is not a clamp: when set it swaps the activation entirely\n(fla/ops/kda/gate.py), from -exp(A_log)*softplus(x) to\nlower_bound*sigmoid(exp(A_log)*x). K3 sets it to -5.0; kimi-linear leaves it\nunset, so that path is unchanged.\n\nCross-layer residuals reuse ggml_dsv4_hc_pre for the weighted sum. That op is\nCPU + CUDA only, so Metal/Vulkan will fall back per-node until those kernels\nexist.\n\nThe routed experts ship as compressed-tensors \"mxfp4-pack-quantized\". That is\nbit-compatible with ggml's MXFP4 - same E2M1 code assignment, same E8M0 scale\nbyte, only the nibble positions within a block differ - so they are repacked\nrather than dequantized, losslessly and without a ~5.5 TB bf16 round-trip.\nThe repack is built lazily because gguf_writer holds every added tensor until\nthe final write. DeepSeek-V4 was already doing the identical bit-shuffling, so\nit now shares the helper.\n\nVerified against Moonshot's own code path (transformers + fla's Triton KDA\nkernels) on a tiny model exercising every K3-specific feature. Final-position\nlogits vs the fp32 reference: 6.7e-05 rel / corr 1.00000000 for both the\nchunked and the recurrent delta-net path. MXFP4 blocks dequantize to the source\nweights with 0.0e+00 error.\n\nAssisted-By: Claude Opus 5 (1M context) \n\n* model: fix ty errors in the Kimi-K3 converter\n\n- `_res_parts` buffers (kind, tensor) pairs, not bare tensors\n- `get_tensors` must return an Iterator, matching ModelBase\n- LazyBase's `func` takes one argument, so pass the expert loaders through\n `args` instead of the closure\n- borrowing KimiLinearModel.set_vocab from an unrelated TextModel is\n deliberate and safe, but not expressible in the signature\n\nNo behaviour change: the MXFP4 repack still dequantizes to the source weights\nwith 0.0e+00 error and end-to-end logits are unchanged (8.386e-03 rel,\ncorr 0.99996630).\n\nAssisted-By: Claude Opus 5 (1M context) \n\n* Update conversion/kimi_k3.py\n\nCo-authored-by: Boris Dvorkin \n\n* Increase LLAMA_MAX_EXPERTS from 512 to 1024\n\n* tests : support for Kimi K3 in archs test\n\n* chat : add Kimi K3 chat format (reasoning, content, typed tool calls)\n\nK3's assistant output is an XTML-ish tagged format built by the template's\nopen_tag/close_tag macros. Two properties break generic parsing:\n\n1. The generation prompt ends with open_tag('think'), so the completion\n starts inside the think section with no opening marker in the output\n (thinking_forced_open).\n2. Only <|open|>/<|close|>/<|sep|>/<|end_of_msg|> are special tokens; tag\n names (\"think\", \"response\", \"message\") are ordinary text tokens.\n\nAdds common_chat_params_init_kimi_k3 (PEG_NATIVE) with detection on the\nmarker trio, reasoning extraction, response unwrapping, and tool-call\nparsing of the tools/call/argument tag structure with argument types\ntaken from the tool schema. Includes the K3 chat template fixture and 9\ntest-chat cases derived from real generations of the full 2.8T model.\n\nVerified end-to-end against Kimi-K3-Q2_K (GrEarl/Kimi-K3-GGUF) on 8x B200:\ncontent, reasoning_content, streaming deltas, and tool_calls all correct;\nfinish_reason stop/tool_calls as appropriate.\n\nCo-Authored-By: Claude Fable 5 \n\n* chat : add message_delimiters for Kimi K3\n\nPer-role message-start markers for token-level span splitting. User and\nassistant messages carry only the role attribute, so their full opener\n(through <|sep|>) is used; system and tool messages continue with more\nattributes (type=/tool=/index=), so those delimiters stop after the\nrole's closing quote. Verified against the K3 tiktoken vocabulary that\nthe closing quote is always a standalone token across all attribute\nvariants, so the token-level prefix match stays exact.\n\nCo-Authored-By: Claude Fable 5 \n\n* fix: apply nits from @ngxson and text fixes from @danielhanchen\n\n* tests : added missing hyperparameters and tensors for Kimi K3 in test-llama-archs\n\n* chore : move overly verbose header file comments to Kimi K3 source file\n\n* tests : re-enabled KIMI_K3 in test-llama-archs for WebGPU backend\n\n* model-saver : emit kda_gate_lower_bound for Kimi K3\n\nQuick fix. The Kimi K3 loader reads kda_gate_lower_bound and gates a graph branch on it (it scales the KDA gate when the bound is above -INFINITY), but the model\nsaver never wrote the key, so a save->load roundtrip silently dropped it back to the -INFINITY default and changed the model's output. The real K3 config sets gate_lower_bound = -5.0.\n\nI propose to emit it from the saver, and set it to -5.0 in the test-llama-archs K3 case so the roundtrip check exercises it (the roundtrip fails without the saver line).\n\n* Refactor conditional for model architecture check\n\n* tests : re-enabled (again) KIMI_K3 and MINIMAX_M3 in test-llama-archs for WebGPU backend\n\n* fix code comments\n\n* add template on conversion\n\n* move repack_mxfp4_blocks to model base\n\n* nits\n\n* add_value_length\n\n* optimize res_stack construction\n\n* nits\n\n---------\n\nCo-authored-by: Boris Dvorkin \nCo-authored-by: Stanisław Szymczyk \nCo-authored-by: Deepankar Singh \nCo-authored-by: Claude Fable 5 \nCo-authored-by: Caleb DeLeeuw \nCo-authored-by: Xuan Son Nguyen ","tree":"2a6544a5f78f459a7f7d480cc765755f46605fa9"}},{"name":"kimi-linear.cpp","mode":"0100644","size":30609,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"laguna.cpp","mode":"0100644","size":17226,"lastCommit":{"hash":"e9fa0781f1c25fc4fe8c86be1edc6970661ad6f0","oid":"e9fa0781f1c25fc4fe8c86be1edc6970661ad6f0","parents":["bc71c24c9da1e7ba6f5993da0e61476155720b06"],"author":{"name":"Guido Imperiale","email":"crusaderky@gmail.com","when":"2026-07-28T20:02:33+01:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-07-28T21:02:33+02:00"},"message":"model: Add Laguna-S-2.1 LLM_TYPE (#26233)","tree":"4163df59fb46557a444ba9b44d37e70cc421e9d2"}},{"name":"lfm2.cpp","mode":"0100644","size":14429,"lastCommit":{"hash":"07822bddf80d73f1168e592c52e69caaff820f9c","oid":"07822bddf80d73f1168e592c52e69caaff820f9c","parents":["78ec4c378031811671d1c76a067acbee4f4c56ce"],"author":{"name":"Tarek Dakhran","email":"tarek@liquid.ai","when":"2026-08-20T16:36:57+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-20T16:36:57+02:00"},"message":"model : support DSpark for LFM2 models (#27383)","tree":"7075371a5bf4ecc766e6171667ec2ed782ace4ac"}},{"name":"lfm2moe.cpp","mode":"0100644","size":4448,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"llada-moe.cpp","mode":"0100644","size":5948,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"llada.cpp","mode":"0100644","size":6191,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"llama-embed.cpp","mode":"0100644","size":199,"lastCommit":{"hash":"994118a1831fdede28ee2d228e0570db9b728c45","oid":"994118a1831fdede28ee2d228e0570db9b728c45","parents":["c84e6d6db5c637ad8aacfdd41cf96fc1d2b85cec"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-05-04T12:36:59+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-05-04T12:36:59+02:00"},"message":"model: move `load_hparams` and `load_tensors` to per-model definition (#22004)\n\n* git-friendly migration\n\n* add build_graph\n\n* nits\n\n* exclude old code from build\n\n* wip\n\n* add llm_arch_model_i\n\n* prepare downstream functions\n\n* nits\n\n* nits\n\n* wip\n\n* wip\n\n* add back create_tensor_qkv\n\n* fix files missing include\n\n* enforce one llm_build per arch\n\n* cmake: use glob\n\n* missing model params\n\n* nits\n\n* wip\n\n* wip (2)\n\n* wip (3)\n\n* test-llama-archs is happy\n\n* improve switch case\n\n* move more stuff into llm_arch_model_i\n\n* fix downstream code\n\n* nits\n\n* nits (2)\n\n* fix order\n\n* llama_model_base\n\n* LLAMA_LOAD_LOCALS\n\n* small fix\n\n* fix build errors\n\n* auto\n\n* rm migration script and ifdef","tree":"e907b3bfe32589d24e2328448be8c5fa913f94ee"}},{"name":"llama.cpp","mode":"0100644","size":10554,"lastCommit":{"hash":"88a39274ecf88ba11686acd357b59685b1cbf03d","oid":"88a39274ecf88ba11686acd357b59685b1cbf03d","parents":["85f99dca8b4539163e6858c9df7c9e8b0ad3632a"],"author":{"name":"Ruixiang Wang","email":"wangruixiang07@outlook.com","when":"2026-06-12T09:21:06+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-12T10:21:06+03:00"},"message":"spec: add EAGLE3 speculative decoding support (#18039)\n\n* llama : enable layer input extraction\n\n* spec: support eagle3\n\n* eagle3: fix params bug\n\n* eagle3: support Gemma4 eagle3 from RedHatAI\n\n* eagle3: set sync when get features from target\n\nCo-authored-by: tnhnyzc <115956684+tnhnyzc@users.noreply.github.com>\n\n* eagle3 : fix ubatch handling in embd_layer_inp extraction and encoder\n\nCo-authored-by: Doğaç Eldenk \n\n* eagle3: adapt to upstream changes\n\n* eagle3: fix rebase issues and adapt to upstream changes\n\n* eagle3:exclude the eagle3 arch from test-llama-archs\n\n* eagle3: fix editorconfig check failures\n\n* eagle3: fix multi-seq issue in d2t vocab mapping\n\n* cont : minor style / clean-up\n\n* spec : remove `common_speculative_setup_draft_model()`\n\n* llama : clean-up unused API\n\n* eagle3: set d2t vocab mapping in decode graph\n\n* cont : assert layer inputs are configured\n\n* hparams : use n_embd_inp instead of n_embd_target_features\n\n* eagle3: make output.weight optional and inherit from target model when needed\n\n* haparams : generic norm-before-residual param\n\n* llama-ext : consistent names\n\n* cont : fix\n\n* hparams : remove target_hidden_size\n\n* cparams : rename output_layer_inp -> embeddings_layer_inp\n\n* arch : reuse ATTN_NORM_2 instead of adding new hidden norm\n\n* llama : clean-up names\n\n* cont : add assert + comment\n\n* Update conversion/llama.py\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Georgi Gerganov \nCo-authored-by: tnhnyzc <115956684+tnhnyzc@users.noreply.github.com>\nCo-authored-by: Doğaç Eldenk \nCo-authored-by: Sigbjørn Skjæret ","tree":"350a48ee94f9fcab60faab12901bd57e79486604"}},{"name":"llama4.cpp","mode":"0100644","size":11327,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"maincoder.cpp","mode":"0100644","size":5427,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"mamba-base.cpp","mode":"0100644","size":15434,"lastCommit":{"hash":"a130532ae1c4c54daaae5527795f5b19c184f269","oid":"a130532ae1c4c54daaae5527795f5b19c184f269","parents":["bf0a29cc16066d313f6a8b52025620287a75cbc4"],"author":{"name":"Prabhsimran Singh","email":"pskrunner14@gmail.com","when":"2026-08-24T11:55:11+05:30"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-24T09:25:11+03:00"},"message":"mamba2 : Flatten in/out projections to dispatch GEMM instead of GEMV (#27513)\n\n* mamba2 : flatten mamba2 in/out projections to dispatch gemm instead of gemv\n\n* mamba2 : remove redundant output reshape","tree":"3105aa69f642a3b9e0b4e4db7e5f75434d50b7ab"}},{"name":"mamba.cpp","mode":"0100644","size":4918,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"mamba2.cpp","mode":"0100644","size":3517,"lastCommit":{"hash":"960d628f46d17ea9c10aa944b061b2341afeb3b5","oid":"960d628f46d17ea9c10aa944b061b2341afeb3b5","parents":["5c7c22c3e134cf7e2681fdddefd78b3dcada7154"],"author":{"name":"Arsen Arutunan","email":"58118221+limloop@users.noreply.github.com","when":"2026-06-26T08:50:54+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-26T08:50:54+03:00"},"message":"mamba2: remove hardcoded 2x expansion factor and invalid d_inner % d_state check (#23082)\n\n* mamba2: remove hardcoded 2x expansion factor, support any expand value\n\n* mamba2: remove invalid d_inner %% d_state check (unrelated parameters)\n\n* Update convert_hf_to_gguf.py: make expand optional with default 2\n\nCo-authored-by: Sigbjørn Skjæret \n\n* mamba2: apply expand fix to refactored conversion/mamba.py\n\n* also check for mamba_expand\n\n---------\n\nCo-authored-by: Sigbjørn Skjæret \nCo-authored-by: Sigbjørn Skjæret <1629204+CISC@users.noreply.github.com>","tree":"18ba1c58547e78226a1dcde16a366a88e5d1c14b"}},{"name":"mellum.cpp","mode":"0100644","size":8442,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"mimo2.cpp","mode":"0100644","size":17202,"lastCommit":{"hash":"75587a05b39e5d5866a2e120178c75c12dfb75cb","oid":"75587a05b39e5d5866a2e120178c75c12dfb75cb","parents":["7a2db1a0cf38e5f4eb278a58c3903d295258d0ea"],"author":{"name":"Sigbjørn Skjæret","email":"sigbjorn.skjaeret@scala.com","when":"2026-08-02T09:03:05+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-02T09:03:05+02:00"},"message":"model : load MiMo V2 MTP tensors only if used (#26412)","tree":"782caae6772ddbab4437afaf1ee92deb299af604"}},{"name":"minicpm.cpp","mode":"0100644","size":4737,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"minicpm3.cpp","mode":"0100644","size":10832,"lastCommit":{"hash":"873e5d8e39feb34a376e0efd01bf3f665dfffeb5","oid":"873e5d8e39feb34a376e0efd01bf3f665dfffeb5","parents":["d7fa69b7de2186b16c1024572273130d631038d7"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-08-21T18:54:29+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-21T18:54:29+02:00"},"message":"model: use ggml_rope_set_offset() (#27382)\n\n* model: use ggml_rope_set_offset()\n\n* partially apply to deepseek2","tree":"f508518e9602084c61100b527fa1fa579383f983"}},{"name":"minimax-01.cpp","mode":"0100644","size":21121,"lastCommit":{"hash":"16d222fc5ead59d20039501a37251c9ed457a454","oid":"16d222fc5ead59d20039501a37251c9ed457a454","parents":["6fed9f6ff7a603b124cb8c5864fca6ea879f9f99"],"author":{"name":"fairydreaming","email":"166155368+fairydreaming@users.noreply.github.com","when":"2026-08-15T00:02:38+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-15T00:02:38+02:00"},"message":"model : add support for MiniMaxText01ForCausalLM and MiniMaxM1ForCausalLM (#27018)\n\n* llama : support for MiniMax-Text-01 model\n\n* chore : renames to match the other MiniMax models\n\n* model : add logits mask as MiniMax-Text-01 embeddings tensor has zero-valued embeddings for tokens >= 200032 that produce zero logits disrupting the token sampling process\n\n* llama : replace hardcoded conditions with hparams.is_recr()\n\n* model : used build_rs() for recurrent state management\n\n* chore : code cleanup\n\n* model : optimized MiniMax-Text-01 by removing the state tranpose operations\n\n* chore : removed unnecessary ggml_cont() in MiniMax-Text-01 implementation\n\n* llama : add generic logits mask graph input\n\n* model : permuted diag_decay dimensions to avoid doing it inside MiniMax-Text-01 graph\n\n* chore : code cleanup\n\n* chore : code cleanup\n\n* model : use token positions when calculating MiniMax-Text-01 decay tensors\n\n* convert : add support for MiniMaxM1ForCausalLM as it seems to be the same as MiniMaxText01ForCausalLM\n\n* chat : add jinja template for MiniMax-M1\n\nCo-authored-by: QscQ \n\n* chore : code cleanup\n\n* tests : MINIMAX_01-related fixes\n\n* chore : silence Python lint errors\n\n* vocab : remove unnecessary vocab type\n\n* convert : update MiniMaxText01Model conversion to use yield when modifying tensors\n\n* convert : suppress tokens with zero-valued embeddings during MiniMax-Text-01 conversion\n\n* llama : removed logits mask - no longer necessary as token suppression is used instead\n\n* model : use common functions to make MiniMax-Text-01 implementation more concise\n\nCo-authored-by: Sigbjørn Skjæret \n\n* model : use common functions to make MiniMax-Text-01 implementation more concise\n\nCo-authored-by: Sigbjørn Skjæret \n\n* convert : override non-working built-in chat template during conversion\n\n* tests : skip arch MINIMAX_01 tests for WebGPU backend (it breaks again)\n\n---------\n\nCo-authored-by: Stanisław Szymczyk \nCo-authored-by: QscQ \nCo-authored-by: Sigbjørn Skjæret ","tree":"935703438a4fa57137a858e4eed8ad1dec4e8a46"}},{"name":"minimax-m2.cpp","mode":"0100644","size":6369,"lastCommit":{"hash":"259ae1df8b5277a57a5092636bfcbcbb2f753219","oid":"259ae1df8b5277a57a5092636bfcbcbb2f753219","parents":["4193ea697f0406e2ccbc16f84251ca5d76a83914"],"author":{"name":"Adrian","email":"40185566+adrianisk@users.noreply.github.com","when":"2026-07-13T06:22:37-07:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-07-13T15:22:37+02:00"},"message":"spec: add Minimax2 eagle3 support\n\n* Fix nullptr in minimax2 EAGLE3\n\n* minor : add newline\n\n---------\n\nCo-authored-by: Georgi Gerganov ","tree":"052cc283d233847f3651d57b9434299b2d0e57d0"}},{"name":"minimax-m3.cpp","mode":"0100644","size":31554,"lastCommit":{"hash":"6fed9f6ff7a603b124cb8c5864fca6ea879f9f99","oid":"6fed9f6ff7a603b124cb8c5864fca6ea879f9f99","parents":["9e40df63ba151d771d8b247ac4011cf203337e99"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-08-14T23:34:56+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-14T23:34:56+02:00"},"message":"mtmd, common: various fixes (#27071)\n\n* apply fixes\n\n* cont\n\n* revert gguf fix","tree":"6e34817cdb942ecf65a677396fd28297d1188213"}},{"name":"mistral3.cpp","mode":"0100644","size":9878,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"mistral4.cpp","mode":"0100644","size":190,"lastCommit":{"hash":"994118a1831fdede28ee2d228e0570db9b728c45","oid":"994118a1831fdede28ee2d228e0570db9b728c45","parents":["c84e6d6db5c637ad8aacfdd41cf96fc1d2b85cec"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-05-04T12:36:59+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-05-04T12:36:59+02:00"},"message":"model: move `load_hparams` and `load_tensors` to per-model definition (#22004)\n\n* git-friendly migration\n\n* add build_graph\n\n* nits\n\n* exclude old code from build\n\n* wip\n\n* add llm_arch_model_i\n\n* prepare downstream functions\n\n* nits\n\n* nits\n\n* wip\n\n* wip\n\n* add back create_tensor_qkv\n\n* fix files missing include\n\n* enforce one llm_build per arch\n\n* cmake: use glob\n\n* missing model params\n\n* nits\n\n* wip\n\n* wip (2)\n\n* wip (3)\n\n* test-llama-archs is happy\n\n* improve switch case\n\n* move more stuff into llm_arch_model_i\n\n* fix downstream code\n\n* nits\n\n* nits (2)\n\n* fix order\n\n* llama_model_base\n\n* LLAMA_LOAD_LOCALS\n\n* small fix\n\n* fix build errors\n\n* auto\n\n* rm migration script and ifdef","tree":"e907b3bfe32589d24e2328448be8c5fa913f94ee"}},{"name":"models.h","mode":"0100644","size":98628,"lastCommit":{"hash":"6d1917060d72cfffc6b54391ef8006b7f51361b9","oid":"6d1917060d72cfffc6b54391ef8006b7f51361b9","parents":["c0209b63a39312975f3cbff7670afda41f653cd9"],"author":{"name":"nekomimi.pet","email":"meowskulls@nekomimi.pet","when":"2026-10-13T19:13:57-04:00"},"committer":{"name":"nekomimi.pet","email":"meowskulls@nekomimi.pet","when":"2026-10-13T19:13:57-04:00"},"message":"bonsai: fused CUDA kernel for DeltaNet conv path (gdn-conv)\n\nOne CUDA kernel for the DeltaNet conv path (concat -> ssm_conv -> silu -> conv-state writes): the concat is never materialized; same arithmetic. Bit-identical; MTP decode +1.5%. GGML_CUDA_SSM_CONV_CONCAT_FUSION=0 turns it off.\n","tree":"b62a7b2173f152483ed21ab164952353e634072b"}},{"name":"modern-bert.cpp","mode":"0100644","size":6483,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"mpt.cpp","mode":"0100644","size":6991,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"muse-glimmer.cpp","mode":"0100644","size":8799,"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":"nanbeige.cpp","mode":"0100644","size":7018,"lastCommit":{"hash":"b77d646751d01c0962bc203b6809e9d94f7d50b7","oid":"b77d646751d01c0962bc203b6809e9d94f7d50b7","parents":["0324696b8e5fe340dc94b64714e4c9aab03084a2"],"author":{"name":"zql","email":"37731799+zqlcode@users.noreply.github.com","when":"2026-07-27T23:04:18+08:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-07-27T17:04:18+02:00"},"message":"model: Add support for Nanbeige4.2 (#25994)\n\n* support nanbeige4.2 model\n\n* fix\n\n* fix flake8 Lint check\n\n* fix loop bound check and drop redundant head_dim\n\n---------\n\nCo-authored-by: root ","tree":"b0ea961c811201ee10089d1d8e4a6fa70cce3000"}},{"name":"nemotron-h-moe.cpp","mode":"0100644","size":6093,"lastCommit":{"hash":"7a20b417f4526cae073bd997af5020cea3e7ccbe","oid":"7a20b417f4526cae073bd997af5020cea3e7ccbe","parents":["e23e9440eb0c625c30d6c40266e9335071a4debc"],"author":{"name":"Ruixiang Wang","email":"wangruixiang07@outlook.com","when":"2026-08-10T10:25:24+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-10T11:25:24+03:00"},"message":"model: add MTP support for Nemotron model (#26725)\n\n* model: add MTP support for Nemotron Nano model\n\n* model: add mtp_flags for nemotron model\n\n* address review comments","tree":"1c3b8b7626fbf7bb86e6cad65094eb16a2d211f0"}},{"name":"nemotron-h.cpp","mode":"0100644","size":15821,"lastCommit":{"hash":"cc078b45b635b3a59aa9adc1b888150ab67798a8","oid":"cc078b45b635b3a59aa9adc1b888150ab67798a8","parents":["6e62ba538478202094edc6c100c782719e310aa3"],"author":{"name":"lnigam","email":"lnigam@nvidia.com","when":"2026-08-11T18:46:26+05:30"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-11T18:46:26+05:30"},"message":"Dflash support for nemotron-3.5 (#26905)\n\n* conversion: skip untrained DFlash embeddings\n\n* Add Nemotron DFlash support\n\n* Add DFlash NVFP4 support\n\n* Address review comments\n\n* add missing output_s for nvfp4\n\n* Include change for keeping residual for last layer also if requested in future dflash models\n\n* Update conversion/qwen.py\r\n\r\nDefensive check, not needed\n\nCo-authored-by: Sigbjørn Skjæret \n\n* Fixing bug introduced by merge conflict\n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"9b0b59a9b919f645324ca24fc7b4d5d0a661f3a1"}},{"name":"nemotron.cpp","mode":"0100644","size":5406,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"neo-bert.cpp","mode":"0100644","size":4730,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"nomic-bert-moe.cpp","mode":"0100644","size":3763,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"nomic-bert.cpp","mode":"0100644","size":3751,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"olmo.cpp","mode":"0100644","size":4728,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"olmo2.cpp","mode":"0100644","size":7933,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"olmoe.cpp","mode":"0100644","size":6205,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"openai-moe.cpp","mode":"0100644","size":6929,"lastCommit":{"hash":"f87067841bac583bc089a225382248d857791ca8","oid":"f87067841bac583bc089a225382248d857791ca8","parents":["c6292cfb8e964bd9f2146f3c603507627e4a22c3"],"author":{"name":"Ruixiang Wang","email":"wangruixiang07@outlook.com","when":"2026-07-28T08:58:16+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-07-28T09:58:16+03:00"},"message":"spec: add eagle3-v3 support for gpt-oss model (#25794)","tree":"2f9b35cb629916c84aa53b336299621434988ab5"}},{"name":"openelm.cpp","mode":"0100644","size":6334,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"orion.cpp","mode":"0100644","size":4986,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"paddleocr.cpp","mode":"0100644","size":3495,"lastCommit":{"hash":"42928bc14d33bde7b8fe8ea436ca44b40357737e","oid":"42928bc14d33bde7b8fe8ea436ca44b40357737e","parents":["59778f0196a82db32580bb649d5d839355d6d7bf"],"author":{"name":"ynankani","email":"ynankani@nvidia.com","when":"2026-05-16T09:09:27Z"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-05-16T11:09:27+02:00"},"message":"model : NvFP4 quantized LM head support (#23046)\n\n* NvFP4 quantized LM head support\n\nSigned-off-by: ynankani \n\n* Address review commnets\n\nSigned-off-by: ynankani \n\n* Add assert for NvFp4 lm head and tied embeddings\n\nSigned-off-by: ynankani \n\n* Address review commnets\n\nSigned-off-by: ynankani \n\n* Create output_s tensor only when LM head NvFp4\n\nSigned-off-by: ynankani \n\n---------\n\nSigned-off-by: ynankani ","tree":"50aa23949b519f164ae9f7332b3a1b9b018bb931"}},{"name":"pangu-embed.cpp","mode":"0100644","size":5938,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"phi2.cpp","mode":"0100644","size":5461,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"phi3.cpp","mode":"0100644","size":7368,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"phimoe.cpp","mode":"0100644","size":2764,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"plamo.cpp","mode":"0100644","size":4560,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"plamo2.cpp","mode":"0100644","size":19847,"lastCommit":{"hash":"1692f9e50bb20fd96b963af38a282daf78feea64","oid":"1692f9e50bb20fd96b963af38a282daf78feea64","parents":["4c1a0af40d88c7fbb3b15c85bf2e8016d1d5b64c"],"author":{"name":"lnigam","email":"lnigam@nvidia.com","when":"2026-08-14T19:50:40+05:30"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-14T17:20:40+03:00"},"message":"ggml : recurrent state rollback for ggml_ssm_scan (#26623)\n\n* Initial changes for Recurrent state rollback for nemotron for cpu and cuda\n\n* Removing CPU RS rollback. Will enable it in subsequent PRs\n\n* addition of test case\n\n* Removing assert and calling runtime API to check if op is supported\n\n* removing extra API and updating the call sites for K\n\n* replace static cuda detection to runtime fused_op api\n\n* address review comments and fallback when SSM rollback not supprted\n\n* Adding changes for supporting RS-rollback in CPU. Also added test-backend-ops for cpu and cuda\n\n* removing memory manipulation as rs rollback is now supported in CPU\n\n* removing the static probe which is not needed now\n\n* correcting the format\n\n* address review comments\n\n* enabling test for all the backends, unsupported backends will fallback to CPU\n\n* Apply suggestions from code review\n\nCo-authored-by: Georgi Gerganov \n\n* choose different graph based on the result of fused_ssm_op is supported or not and also handled memory->n_rs_seq >1 case incase of op is not supported\n\n* Support K > 1 in ssm_scan for all backends\n\n* Fix CI Issues\n\n---------\n\nCo-authored-by: Georgi Gerganov \nCo-authored-by: Gaurav Garg ","tree":"849b53f780d78a5308e177885994e6d3aadbb904"}},{"name":"plamo3.cpp","mode":"0100644","size":7986,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"plm.cpp","mode":"0100644","size":8631,"lastCommit":{"hash":"873e5d8e39feb34a376e0efd01bf3f665dfffeb5","oid":"873e5d8e39feb34a376e0efd01bf3f665dfffeb5","parents":["d7fa69b7de2186b16c1024572273130d631038d7"],"author":{"name":"Xuan-Son Nguyen","email":"son@huggingface.co","when":"2026-08-21T18:54:29+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-08-21T18:54:29+02:00"},"message":"model: use ggml_rope_set_offset() (#27382)\n\n* model: use ggml_rope_set_offset()\n\n* partially apply to deepseek2","tree":"f508518e9602084c61100b527fa1fa579383f983"}},{"name":"pockettts.cpp","mode":"0100644","size":5177,"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":"qwen.cpp","mode":"0100644","size":4831,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"qwen2.cpp","mode":"0100644","size":5535,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"qwen2moe.cpp","mode":"0100644","size":7300,"lastCommit":{"hash":"7acb4e8cd2ce21f457d1298e75fad729520d263c","oid":"7acb4e8cd2ce21f457d1298e75fad729520d263c","parents":["3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985"],"author":{"name":"Georgi Gerganov","email":"ggerganov@gmail.com","when":"2026-06-05T11:09:36+03:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-05T11:09:36+03:00"},"message":"hparams : refactor `hparams.n_layer` (#24060)\n\n* hparams : refactor hparams.n_layer\n\n* cont : remove `n_layer_kv()`, use n_layer_all instead\n\n* cont : type consistency\n\n* pi : update SYSTEM.md\n\n* models : fix Step3.5 MTP\n\n* cont : remove duplicate switch cases\n\n* cont : explicitly set `false` to extra layers for `is_swa` and `is_recr`\n\n* cont : fix nextn layer count handling\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Sigbjørn Skjæret ","tree":"5a77024ea13397365f3be1610bee5573b0c42f1f"}},{"name":"qwen2vl.cpp","mode":"0100644","size":5062,"lastCommit":{"hash":"42928bc14d33bde7b8fe8ea436ca44b40357737e","oid":"42928bc14d33bde7b8fe8ea436ca44b40357737e","parents":["59778f0196a82db32580bb649d5d839355d6d7bf"],"author":{"name":"ynankani","email":"ynankani@nvidia.com","when":"2026-05-16T09:09:27Z"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-05-16T11:09:27+02:00"},"message":"model : NvFP4 quantized LM head support (#23046)\n\n* NvFP4 quantized LM head support\n\nSigned-off-by: ynankani \n\n* Address review commnets\n\nSigned-off-by: ynankani \n\n* Add assert for NvFp4 lm head and tied embeddings\n\nSigned-off-by: ynankani \n\n* Address review commnets\n\nSigned-off-by: ynankani \n\n* Create output_s tensor only when LM head NvFp4\n\nSigned-off-by: ynankani \n\n---------\n\nSigned-off-by: ynankani ","tree":"50aa23949b519f164ae9f7332b3a1b9b018bb931"}},{"name":"qwen3.cpp","mode":"0100644","size":5872,"lastCommit":{"hash":"88a39274ecf88ba11686acd357b59685b1cbf03d","oid":"88a39274ecf88ba11686acd357b59685b1cbf03d","parents":["85f99dca8b4539163e6858c9df7c9e8b0ad3632a"],"author":{"name":"Ruixiang Wang","email":"wangruixiang07@outlook.com","when":"2026-06-12T09:21:06+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-12T10:21:06+03:00"},"message":"spec: add EAGLE3 speculative decoding support (#18039)\n\n* llama : enable layer input extraction\n\n* spec: support eagle3\n\n* eagle3: fix params bug\n\n* eagle3: support Gemma4 eagle3 from RedHatAI\n\n* eagle3: set sync when get features from target\n\nCo-authored-by: tnhnyzc <115956684+tnhnyzc@users.noreply.github.com>\n\n* eagle3 : fix ubatch handling in embd_layer_inp extraction and encoder\n\nCo-authored-by: Doğaç Eldenk \n\n* eagle3: adapt to upstream changes\n\n* eagle3: fix rebase issues and adapt to upstream changes\n\n* eagle3:exclude the eagle3 arch from test-llama-archs\n\n* eagle3: fix editorconfig check failures\n\n* eagle3: fix multi-seq issue in d2t vocab mapping\n\n* cont : minor style / clean-up\n\n* spec : remove `common_speculative_setup_draft_model()`\n\n* llama : clean-up unused API\n\n* eagle3: set d2t vocab mapping in decode graph\n\n* cont : assert layer inputs are configured\n\n* hparams : use n_embd_inp instead of n_embd_target_features\n\n* eagle3: make output.weight optional and inherit from target model when needed\n\n* haparams : generic norm-before-residual param\n\n* llama-ext : consistent names\n\n* cont : fix\n\n* hparams : remove target_hidden_size\n\n* cparams : rename output_layer_inp -> embeddings_layer_inp\n\n* arch : reuse ATTN_NORM_2 instead of adding new hidden norm\n\n* llama : clean-up names\n\n* cont : add assert + comment\n\n* Update conversion/llama.py\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Georgi Gerganov \nCo-authored-by: tnhnyzc <115956684+tnhnyzc@users.noreply.github.com>\nCo-authored-by: Doğaç Eldenk \nCo-authored-by: Sigbjørn Skjæret ","tree":"350a48ee94f9fcab60faab12901bd57e79486604"}},{"name":"qwen35.cpp","mode":"0100644","size":35603,"lastCommit":{"hash":"6d1917060d72cfffc6b54391ef8006b7f51361b9","oid":"6d1917060d72cfffc6b54391ef8006b7f51361b9","parents":["c0209b63a39312975f3cbff7670afda41f653cd9"],"author":{"name":"nekomimi.pet","email":"meowskulls@nekomimi.pet","when":"2026-10-13T19:13:57-04:00"},"committer":{"name":"nekomimi.pet","email":"meowskulls@nekomimi.pet","when":"2026-10-13T19:13:57-04:00"},"message":"bonsai: fused CUDA kernel for DeltaNet conv path (gdn-conv)\n\nOne CUDA kernel for the DeltaNet conv path (concat -> ssm_conv -> silu -> conv-state writes): the concat is never materialized; same arithmetic. Bit-identical; MTP decode +1.5%. GGML_CUDA_SSM_CONV_CONCAT_FUSION=0 turns it off.\n","tree":"b62a7b2173f152483ed21ab164952353e634072b"}},{"name":"qwen35moe.cpp","mode":"0100644","size":33091,"lastCommit":{"hash":"00e8187618590ca52e6737f97d38173087458ebf","oid":"00e8187618590ca52e6737f97d38173087458ebf","parents":["1bc46a8a70bdbaf3e2e458a30b219800bb4e6f2b"],"author":{"name":"bri-prism","email":"288398250+bri-prism@users.noreply.github.com","when":"2026-09-08T15:47:22-07:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-09-08T15:47:22-07:00"},"message":"graph: six fusions for the hybrid-attention decode graph (-25% ops/token; Metal 2B +7.7%, 27B +2.6% decode; CUDA 9B +2.5%) (#165)\n\n* metal: fuse the Hadamard sign flip into the FWHT kernel\n\nEvery folded weight's activation went through three encoded ops: an elementwise\nMUL by the +-1 sign vector, a RESHAPE and the FWHT-hint MUL_MAT. The sign factors\nare exact, so the transform now applies them while loading its row and the MUL\nis elided, matching what the CUDA backend already does. The encode list drops\nempty ops, so the MUL / RESHAPE / MUL_MAT pattern is matched on the graph's own\nindices and checked with ggml_can_fuse_subgraph_ext; the fused kernel writes the\nmatmul's output, so its ranges are checked for concurrency explicitly.\n\nMeasured on an M5 Pro, largest band PQ2_0 folded (block 1024), r=3 x2 interleaved:\nencoded nodes per decoded token 2359 -> 2103, tg128 27.05 -> 27.45 (+1.5%),\npp512 390.5 -> 397.2 (+1.7%). Greedy output byte-identical over 3 prompts;\ntest-backend-ops MUL_MAT_HADAMARD 24/24.\n\n* ggml, metal: fold the GDN gate activations into the fused op (raw gates)\n\nThe hybrid-attention layers spent four elementwise launches per layer per\ntoken on the gate scalars: sigmoid(beta), and softplus(alpha + dt_bias) * A\nfor the decay. ggml_gated_delta_net_set_raw_gates() marks a gated_delta_net\nop as taking beta and alpha pre-activation, with dt_bias and A as two extra\nsources, and the kernel applies the same formulas as the unary kernels while\nreading the scalars it already read. The qwen35 builder hands the fused op the\nraw projections (GGML_GDN_RAW_GATES_DISABLE=1 restores the separate ops); the\nnon-fused chunked path keeps the activated tensors. CPU reference and Metal\nkernel implemented here; CUDA follows in the next commit.\n\ntest-backend-ops -o GATED_DELTA_NET on MTL0: 39/39 incl. 4 raw-gate cases\n(decode, 2 seqs, prefill 64, rows mode). Encoded nodes per decoded token on the\nlargest band folded PQ2_0: 2103 -> 1910. Greedy output byte-identical to base.\nM5 Pro, r=3 x2 interleaved, both fusions vs base: 2B tg128 206.4 -> 214.8\n(+4.1%), pp512 +1.0%; largest band tg128 27.08 -> 27.55 (+1.7%), pp512 +1.3%.\n\n* cuda: raw gates for the fused gated_delta_net kernel\n\nSame folding as the Metal and CPU implementations: with the op's raw-gate flag\nset, beta goes through sigmoid and the decay is raw_a[h] * softplus(g +\nraw_dt_bias[h]) inside the kernel, using the formulas of the unary kernels.\nThe template gains a RAW parameter (scalar gate only; KDA asserts it off) and\nthe two tables are passed as extra pointers.\n\nCompile-checked only by inspection here; needs a CUDA build plus\ntest-backend-ops -o GATED_DELTA_NET before this branch is pushed.\n\n* ggml: decline raw-gate gated_delta_net on backends without it; gate the builder on device backend\n\nSYCL, OpenCL, WebGPU and Hexagon compute gated_delta_net but do not implement the\nraw-gate flag, so their supports_op now declines such ops (the CPU fallback is\ncorrect). The qwen35 builder only emits raw gates when every device is CPU,\nMetal, CUDA, ROCm or MUSA, so those backends keep their native op instead of a\nCPU fallback for the whole recurrence.\n\n* qwen35: gated norm as one swiglu_split op\n\nsilu(z) * rms_norm(o) was a unary and a mul; ggml_swiglu_split computes the same\nexpression in one kernel. Encoded nodes per decoded token on the largest folded band\nPQ2_0: 1910 -> 1862. Greedy output byte-identical. M5 Pro decode with the three\ngraph changes so far vs base: 2B 205.3 -> 219.9 tok/s (+7.1%), largest band 27.12 ->\n27.72 (+2.2%).\n\n* qwen35: one l2_norm over the adjacent q and k head groups\n\nq and k occupy adjacent, equally shaped head groups in the conv output, so a\nsingle l2_norm over the joint view replaces two launches per recurrent layer;\nq and k become views into its result. Encoded nodes per decoded token on the\nlargest band: 1862 -> 1814. Greedy output byte-identical.\n\n* metal: fuse the trailing silu into the ssm_conv kernels\n\nWhen the conv output is consumed only by a silu, the conv kernels apply it on\nstore (function constant) and write the silu's tensor directly, matching what\nthe CUDA backend already does. Encoded nodes per decoded token on the largest band:\n1814 -> 1766. test-backend-ops -o SSM_CONV 45/45; greedy output byte-identical.\nCumulative decode vs base on an M5 Pro: 2B 207.0 -> 223.0 tok/s (+7.7%),\nlargest band 27.15 -> 27.85 (+2.6%).\n\n* qwen35: scan the devices once per graph build for the GDN path choices\n\nThe fused-GDN state-rows check and the raw-gates check each looped over the\nmodel's devices on every recurrent layer, asking the backend for the device\ntype each time. Both flags are now computed once in the graph constructor.\n\n* tests: register the raw-gate gated_delta_net cases in the eval list\n\nThey had been appended to the perf-mode list, so test mode never ran them.","tree":"588bbdc561eeacfee8279a74ac0a8d8b4dd9641b"}},{"name":"qwen3moe.cpp","mode":"0100644","size":6603,"lastCommit":{"hash":"88a39274ecf88ba11686acd357b59685b1cbf03d","oid":"88a39274ecf88ba11686acd357b59685b1cbf03d","parents":["85f99dca8b4539163e6858c9df7c9e8b0ad3632a"],"author":{"name":"Ruixiang Wang","email":"wangruixiang07@outlook.com","when":"2026-06-12T09:21:06+02:00"},"committer":{"name":"GitHub","email":"noreply@github.com","when":"2026-06-12T10:21:06+03:00"},"message":"spec: add EAGLE3 speculative decoding support (#18039)\n\n* llama : enable layer input extraction\n\n* spec: support eagle3\n\n* eagle3: fix params bug\n\n* eagle3: support Gemma4 eagle3 from RedHatAI\n\n* eagle3: set sync when get features from target\n\nCo-authored-by: tnhnyzc <115956684+tnhnyzc@users.noreply.github.com>\n\n* eagle3 : fix ubatch handling in embd_layer_inp extraction and encoder\n\nCo-authored-by: Doğaç Eldenk \n\n* eagle3: adapt to upstream changes\n\n* eagle3: fix rebase issues and adapt to upstream changes\n\n* eagle3:exclude the eagle3 arch from test-llama-archs\n\n* eagle3: fix editorconfig check failures\n\n* eagle3: fix multi-seq issue in d2t vocab mapping\n\n* cont : minor style / clean-up\n\n* spec : remove `common_speculative_setup_draft_model()`\n\n* llama : clean-up unused API\n\n* eagle3: set d2t vocab mapping in decode graph\n\n* cont : assert layer inputs are configured\n\n* hparams : use n_embd_inp instead of n_embd_target_features\n\n* eagle3: make output.weight optional and inherit from target model when needed\n\n* haparams : generic norm-before-residual param\n\n* llama-ext : consistent names\n\n* cont : fix\n\n* hparams : remove target_hidden_size\n\n* cparams : rename output_layer_inp -> embeddings_layer_inp\n\n* arch : reuse ATTN_NORM_2 instead of adding new hidden norm\n\n* llama : clean-up names\n\n* cont : add assert + comment\n\n* Update conversion/llama.py\n\nCo-authored-by: Sigbjørn Skjæret \n\n---------\n\nCo-authored-by: Georgi Gerganov