Lumpiasty c49d19cc64 llama: free backend compute scratch on cold; guard memory_breakdown ...

On device release with KV eviction, also call ggml_backend_free_scratch()
on each backend so a cold model drops its Vulkan compute preallocations
(reallocated lazily on the next compute via restore).

Also guard llama_context::memory_breakdown() against a freed scheduler:
evict_kv release resets sched to null, so a cold model that is then torn
down (e.g. terminated by the process manager) hit GGML_ASSERT(sched) in
ggml_backend_sched_get_buffer_type. Skip the compute-buffer accounting
when sched is null.

Assisted-by: Claude

2026-09-09 00:38:39 +02:00

..

models

ggml : update ggml_prec specification (#26675)

2026-09-08 09:06:24 +03:00

CMakeLists.txt

llama: fix whole source code rebuilt on each new commit (#28278)

2026-09-03 23:53:04 +02:00

llama-adapter.cpp

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

2026-08-16 09:38:01 +03:00

llama-adapter.h

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

2026-03-18 12:03:26 +02:00

llama-arch.cpp

model : support Kimi-K3 recurrent-state rollback (#28466)

2026-09-08 11:31:48 +08:00

llama-arch.h

[Model] Support for Spark2_5ForCausalLM implementation (#27868)

2026-09-06 17:43:58 +02:00

llama-batch.cpp

llama-batch: fix allowed decreasing pos in a seq (#25449)

2026-07-08 19:24:34 +03:00

llama-batch.h

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

2026-07-08 15:55:19 +08:00

llama-chat.cpp

chat : add Granite 4.1 chat template (#23518)

2026-05-28 13:13:33 +02:00

llama-chat.h

chat : add Granite 4.1 chat template (#23518)

2026-05-28 13:13:33 +02:00

llama-context.cpp

llama: free backend compute scratch on cold; guard memory_breakdown

2026-09-09 00:38:39 +02:00

llama-context.h

server: optionally evict the KV cache too (Phase 2 of VRAM sharing)

2026-09-09 00:36:40 +02:00

llama-cparams.cpp

cparams : rename LLAMA_MAX_PARALLEL_SEQUENCES to LLAMA_MAX_SEQ (#14188)

2025-06-15 10:08:58 +03:00

llama-cparams.h

context : autoscale n_ctx_train when yarn scaling specified (#28030)

2026-09-01 19:59:54 +03:00

llama-ext.h

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

2026-08-27 19:17:07 +02:00

llama-grammar.cpp

grammar : fix max repetition threshold (#28469)

2026-09-06 11:59:10 +03:00

llama-grammar.h

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

2026-01-03 16:02:43 -06:00

llama-graph.cpp

ggml : update ggml_prec specification (#26675)

2026-09-08 09:06:24 +03:00

llama-graph.h

llama : add missing headers (#28566)

2026-09-08 12:59:53 +02:00

llama-hparams.cpp

src : add n_expert_used_max function (#28323)

2026-09-04 06:36:51 +02:00

llama-hparams.h

model : add Tencent Hy 4 (hy_v4) preview architecture support (#28127)

2026-09-04 14:31:36 +02:00

llama-impl.cpp

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

2026-04-06 01:40:38 +02:00

llama-impl.h

llama: refactor fused ops (#24646)

2026-07-08 18:18:09 +08:00

llama-io.cpp

server : avoid checkpoint data host copies (#22558)

2026-05-02 18:03:25 +03:00

llama-io.h

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

2026-05-05 06:35:07 +03:00

llama-kv-cache-dsa-iswa.cpp

model: add dots3-note (#27060)

2026-08-21 19:52:34 +02:00

llama-kv-cache-dsa-iswa.h

model: add dots3-note (#27060)

2026-08-21 19:52:34 +02:00

llama-kv-cache-dsa.cpp

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

2026-08-03 14:56:30 +02:00

llama-kv-cache-dsa.h

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

2026-08-03 14:56:30 +02:00

llama-kv-cache-dsv4.cpp

Deepseek 4: -sm tensor (#26490)

2026-08-24 09:20:25 +03:00

llama-kv-cache-dsv4.h

DeepseekV4 MTP + DSpark (#25784)

2026-08-02 20:55:34 +08:00

llama-kv-cache-iswa.cpp

server: optionally evict the KV cache too (Phase 2 of VRAM sharing)

2026-09-09 00:36:40 +02:00

llama-kv-cache-iswa.h

server: optionally evict the KV cache too (Phase 2 of VRAM sharing)

2026-09-09 00:36:40 +02:00

llama-kv-cache-msa.cpp

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

2026-08-03 16:30:08 +03:00

llama-kv-cache-msa.h

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

2026-08-03 16:30:08 +03:00

llama-kv-cache.cpp

llama: guard buffer iteration against released device buffers

2026-09-09 00:36:40 +02:00

llama-kv-cache.h

server: optionally evict the KV cache too (Phase 2 of VRAM sharing)

2026-09-09 00:36:40 +02:00

llama-kv-cells.h

kv-cells: look up the n-gram history in the sequence position index (#28040)

2026-09-01 20:16:07 +02:00

llama-memory-hybrid-idx.cpp

qwen4exp: fix seq_cp, block position keying, mtmd input, cuda abort, add tests (#27941)

2026-09-01 13:22:04 +03:00

llama-memory-hybrid-idx.h

qwen4exp: fix seq_cp, block position keying, mtmd input, cuda abort, add tests (#27941)

2026-09-01 13:22:04 +03:00

llama-memory-hybrid-iswa.cpp

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

2026-07-08 15:55:19 +08:00

llama-memory-hybrid-iswa.h

llama + spec: MTP Support (#22673)

2026-05-16 20:06:23 +08:00

llama-memory-hybrid.cpp

llama: evict recurrent/SSM state on device release

2026-09-09 00:36:41 +02:00

llama-memory-hybrid.h

server: optionally evict the KV cache too (Phase 2 of VRAM sharing)

2026-09-09 00:36:40 +02:00

llama-memory-recurrent.cpp

llama: evict recurrent/SSM state on device release

2026-09-09 00:36:41 +02:00

llama-memory-recurrent.h

llama: evict recurrent/SSM state on device release

2026-09-09 00:36:41 +02:00

llama-memory.cpp

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

2025-06-30 18:03:03 +03:00

llama-memory.h

server: optionally evict the KV cache too (Phase 2 of VRAM sharing)

2026-09-09 00:36:40 +02:00

llama-mmap.cpp

llama : add missing headers (#28566)

2026-09-08 12:59:53 +02:00

llama-mmap.h

llama: model_loader: add TENSOR_READ_LAZY (#27794)

2026-08-27 15:14:34 +02:00

llama-model-loader.cpp

src : add n_expert_used_max function (#28323)

2026-09-04 06:36:51 +02:00

llama-model-loader.h

llama: improve TENSOR_READ_LAZY handling (#27837)

2026-08-30 16:59:48 +02:00

llama-model-saver.cpp

[Model] Support for Spark2_5ForCausalLM implementation (#27868)

2026-09-06 17:43:58 +02:00

llama-model-saver.h

model: add Qwen3.8-Flash-Next (qwen4exp) (#27742)

2026-08-27 21:32:31 +02:00

llama-model.cpp

llama: guard buffer iteration against released device buffers

2026-09-09 00:36:40 +02:00

llama-model.h

server: on-demand VRAM sharing to time-share one GPU between models

2026-09-09 00:36:34 +02:00

llama-quant.cpp

quantize: row-slab stream to avoid thread starvation (#27830)

2026-09-01 11:18:54 +02:00

llama-quant.h

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

2025-01-03 10:18:53 +02:00

llama-sampler.cpp

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

2026-08-10 16:58:56 +03:00

llama-sampler.h

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

2026-08-10 16:58:56 +03:00

llama-version.h.in

llama: fix whole source code rebuilt on each new commit (#28278)

2026-09-03 23:53:04 +02:00

llama-vocab.cpp

llama : add missing headers (#28566)

2026-09-08 12:59:53 +02:00

llama-vocab.h

[Model] Support for Spark2_5ForCausalLM implementation (#27868)

2026-09-06 17:43:58 +02:00

llama.cpp

llama: fix whole source code rebuilt on each new commit (#28278)

2026-09-03 23:53:04 +02:00

unicode-data.cpp

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

2024-10-08 13:27:04 +02:00

unicode-data.h

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

2024-10-02 15:49:55 +02:00

unicode.cpp

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

2026-08-18 15:15:22 +02:00

unicode.h

vocab: fix Gemma4 tokenizer (#21343)

2026-04-03 10:33:03 +02:00