src · llama-cpp-bonsai · Tangled

llama-cpp-bonsai

llama.cpp fork for Ternary Bonsai 2 27B: MTP speculation, suffix cache reuse (SCR), q8 KV, two-slot fairness, fused DeltaNet kernels

llama-cpp-bonsai src

1 folder · 64 files

nekomimi.pet · bonsai: fused CUDA kernel for DeltaNet conv path (gdn-conv) · 2w ago

bonsai: fused CUDA kernel for DeltaNet conv path (gdn-conv)

model: add dots3-note (#27060)

llama-adapter.cpp

llama : check LoRA tensor data is within file bounds (#27056)

llama-adapter.h

llama : re-enable manual LoRA adapter free (#19983)

spec : add DFlash2 support (local convolution + candidate selector) (#27342) (#27816) (#261)

spec : add DFlash2 support (local convolution + candidate selector) (#27342) (#27816) (#261)

llama-batch.cpp

bonsai: suffix cache reuse (SCR) for linear attention hybrids (scr)

bonsai: suffix cache reuse (SCR) for linear attention hybrids (scr)

chat : add Granite 4.1 chat template (#23518)

chat : add Granite 4.1 chat template (#23518)

llama-context.cpp

bonsai: suffix cache reuse (SCR) for linear attention hybrids (scr)

llama-context.h

bonsai: suffix cache reuse (SCR) for linear attention hybrids (scr)

llama-cparams.cpp

cparams : rename LLAMA_MAX_PARALLEL_SEQUENCES to LLAMA_MAX_SEQ (#14188)

llama-cparams.h

dspark: add corrected draft runtime and shared packed target head

bonsai: suffix cache reuse (SCR) for linear attention hybrids (scr)

llama-grammar.cpp

grammar : degrade max repetition >= 2000 to unbounded (#26613)

llama-grammar.h

common/grammar : replace problematic backtracking regex `[\s\S]*` (#18342)

llama-graph.cpp

bonsai: in-place GDN recurrent state reading from cache rows (gdn-rows)

bonsai: in-place GDN recurrent state reading from cache rows (gdn-rows)

llama-hparams.cpp

model : GraniteSWAForCausalLM / GraniteMoeSWAForCausalLM (#25505)

llama-hparams.h

spec : add DFlash2 support (local convolution + candidate selector) (#27342) (#27816) (#261)

llama : correct platform-independent loading of BOOL metadata (#21428)

llama: refactor fused ops (#24646)

server : avoid checkpoint data host copies (#22558)

llama : add option to save memory in device buffers (#22679)

llama-kv-cache-dsa-iswa.cpp

model: add dots3-note (#27060)

llama-kv-cache-dsa-iswa.h

model: add dots3-note (#27060)

llama-kv-cache-dsa.cpp

llama : allocate indexer cache only in "full" indexer layers (#26474)

llama-kv-cache-dsa.h

llama : allocate indexer cache only in "full" indexer layers (#26474)

llama-kv-cache-dsv4.cpp

Deepseek 4: `-sm tensor` (#26490)

llama-kv-cache-dsv4.h

DeepseekV4 MTP + DSpark (#25784)

llama-kv-cache-iswa.cpp

llama-batch: add n_keep_tail in split_equal for recurrent models (#25278)

llama-kv-cache-iswa.h

DeepSeek V4 (#24162)

llama-kv-cache-msa.cpp

model: M3: Move MSA into a new memory implementation (#26338)

llama-kv-cache-msa.h

model: M3: Move MSA into a new memory implementation (#26338)

llama-kv-cache.cpp

bonsai: suffix cache reuse (SCR) for linear attention hybrids (scr)

llama-kv-cache.h

bonsai: suffix cache reuse (SCR) for linear attention hybrids (scr)

llama-kv-cells.h

bonsai: suffix cache reuse (SCR) for linear attention hybrids (scr)

llama-memory-hybrid-iswa.cpp

llama-batch: add n_keep_tail in split_equal for recurrent models (#25278)

llama-memory-hybrid-iswa.h

llama + spec: MTP Support (#22673)

llama-memory-hybrid.cpp

llama-batch: add n_keep_tail in split_equal for recurrent models (#25278)

llama-memory-hybrid.h

llama + spec: MTP Support (#22673)

llama-memory-recurrent.cpp

bonsai: in-place GDN recurrent state reading from cache rows (gdn-rows)

llama-memory-recurrent.h

bonsai: in-place GDN recurrent state reading from cache rows (gdn-rows)

llama-memory.cpp

memory : correctly handle failure in apply() (#14438)

llama : add Gemma4 MTP (#23398)

Update llama-mmap to use ftello/fseeko (#22497)

llama: fix llama-model-saver (#20503)

llama-model-loader.cpp

ggml: add PTQ1_0, ternary at group 128 (1.75 bpw, lossless vs PQ2_0) (#148)

llama-model-loader.h

quant : Optimise memory usage by evicting weights after processing each layer (#22877)

llama-model-saver.cpp

Deepseek 4: `-sm tensor` (#26490)

llama-model-saver.h

llama: fix llama-model-saver (#20503)

llama-model.cpp

spec : add DFlash2 support (local convolution + candidate selector) (#27342) (#27816) (#261)

spec : add DFlash2 support (local convolution + candidate selector) (#27342) (#27816) (#261)

llama-quant.cpp

ggml: add PTQ1_0, ternary at group 128 (1.75 bpw, lossless vs PQ2_0) (#148)

llama : refactor `src/llama.cpp` (#10902)

llama-sampler.cpp

llama : support multi-output backend sampling (#25532)

llama-sampler.h

llama : support multi-output backend sampling (#25532)

llama-vocab.cpp

vocab : support integer tokenizer scores (#27260)

Add support for Laguna XS.2 & M.1 (#25165)

llama : allow virtual igpu devices (#26953)

unicode-data.cpp

server : better security control for public deployments (#9776)

llama : reduce compile time and binary size (#9712)

unicode : include '~' in collapsed symbol class (#26972)

vocab: fix Gemma4 tokenizer (#21343)