The destroy() function in server_context_impl only cleaned up the main
model and context (via llama_init.reset()) but did not free the speculative
decoder (spec), draft context (ctx_dft), or draft model (model_dft).
For MTP (Multi-Token Prediction) models, ctx_dft holds GPU-allocated
resources (KV cache, compute buffers) that are not freed when entering
the sleeping state. On each sleep/resume cycle, new resources are
allocated without the old ones being freed, leading to a VRAM leak
that eventually crashes the server with out-of-memory errors.
Fix by explicitly resetting spec, ctx_dft, and model_dft in destroy()
before resetting llama_init, ensuring proper cleanup order to avoid
use-after-free.
ref: https://github.com/ggml-org/llama.cpp/issues/23395
Assisted-by: llama.cpp:local pi
* spec: support MTP
* fix batch size
* rename files
* cont : simplify (#7)
* MTP: clean-up (#9)
* MTP: clean-up
* review: use llama_context_type instead of llama_graph_type
* review: remove llama_model_has_mtp
* review: fix convert issues
* convert: fix pycheck
* review: formatting
* use `mtp-` for identifying mtp models
* convert: fix mtp conversion
* mtp -> draft-mtp
* remove unused llama_arch
* add need_embd in speculative
* llama: allow partial seq_rm for GDN models for speculative decoding
Currently speculative checkpoint needs to restart from a checkpoint
after some draft tokens are not accepted, this leads to some wastage in
running the target again. This PR adds the ability to rollback upto
`draft_max` by storing the GDN intermediates.
* fix pending state
* vulkan: add GDN partial rollback
* meta: extend check to axis 1
* metal: add GDN partial rollback
Extend the gated delta net kernel to store intermediate states for
partial rollback support on the Metal backend.
- Add K (snapshot slot count) as a function constant
- Read input state from slot 0 of the 3D state tensor
- Write intermediate states to different slots during token loop
- For K=1, maintain backward-compatible single-slot behavior
Ref: https://github.com/ggml-org/llama.cpp/commit/8c05923630110223669f069af2000e9cf10c02bc
Assisted-by: llama.cpp:local pi
* delta_net_base: use ggml_pad instead of new_tensor
* review: add need_rs_seq
* review: rename part_bounded to n_rs
* review: deslop comments
* review: rename, add asserts
* server : adjust checkpoint logic (#11)
* server : adjust checkpoint logic
* cont : rm asserts
* server-context: fix early exit
* spec : fix compatibility with n-gram and add TODOs (#13)
* metal : cleanup
* llama : fix faulty bitwise check in recurrent memory
* server : disable RS-based MTP in combination with other spec types
* spec : add TODOs
* cont : fix comment
* cont : update comment
* common : fix logic for ngram + mtp compat
* llama-memory: enable checkpointing with partial rollback
* cont: add test-case for loading into a dirty ctx
* llama-memory-recurrent: clear rs_idx in clear
* download: fix mtp path
* llama-arch: fix enorm op
* docs: update docs
* conversion: fix type annotations
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* ggml: add graph_reused
* use versioning instead of reuse flag
* increment version with atomic
* use top bits for split numbering
* add assert
* move counter to ggml.c
* set uid in split_graph only
* fix windows
* address further review comments
* get next_uid rather than doing bit manipulation
* rename + add comment about uid
* ggml-cuda: add mem check for fusion
* Replace NaNs with -FLT_MAX
* fix typo
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* CUDA: use shared mem for ssm_conv
* fuse silu + ssm_conv
* fuse unary + mul
* enable for fp16
* formatting
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* ggml-cpu: FA add GEMM microkernel
* add guard for sizeless vector types
* fix case where DV % GGML_F32_EPR !=0
* move memset out of the loop
* move another memset out of the loop
* use RM=4 for arm
* simd_gemm: convert everything to int
* convert everything to size_t to avoid warnings
* fixup
* add pragma for ignoring aggressive loop optimizations
* CUDA: use mmvq for mul-mat-id for small batch sizes
* add mmvq too
* Fix perf issue on ampere. Use mmvf mm-id only for non-nvidia GPUs
* templatize multi_token_path
* ggml-cpu: Use tiled FA for prompt-processing
the FA performance is gimped on CPU on long contexts because it essentially uses a vector kernel. This PR adds a tiled FA for PP. Perf tuning for tile sizes done on a AMD EPYC single-socket 64-c machine.
* fix out of bounds for mask
* skip rows where there are all masks
* skip tile if mask is inf
* store mask in worksize
* check inf tile earlier
* ggml-cuda: fix blackwell native builds
Replace 12x in native architectures by 12xa
* replace for GGML_NATIVE=OFF too
* only replace for native
* remove 120f-virtual for default compilation
---------
Co-authored-by: Aman Gupta <aman>
* llama-server: friendlier error msg when ctx < input
This PR adds formatted strings to the server's send_error function
* llama-server: use string_format inline
* fix test
* CUDA: add fused rope
* move k forward_expand up
* create helper function instead of re-using params
* make assert statement more in line with comment
* rope_norm: coalesced writes to global mem
* CUDA: Fix bug in topk-moe for gpt-oss
When using ggml_can_fuse_subgraph, the output nodes which are passed are wrong. This causes `test-backend-ops` to still fuse ndoes (because the nodes are not used elsewhere in the graph),
but it actually doesn't fuse in the actual gpt-oss
* fix for qwen3 too
* change ifndef to ifdef
* ggml: add ggml_can_fuse_subgraph
* ggml-cuda: use ggml_can_fuse_subgraph for topk-moe
* format
* 1. remove inputs from signature as they are transient nodes
2. add check for views: view_src should be part of the subgraph
* - combine check into one loop
- check all view_src parents
- other minor review comments
* remove redudant if test
* - rename and other minor review comments
* add assert about count < 32
* CUDA: mul_mat_id for mmf for bs <= 64 for f16 and bs <= 32 for f32
This commit adds mul_mat_id support for ncols_dst >= 16. It does this by
packing ncols_dst tiles into the blockDim.y.
My tests on a RTX 3090 show that this is faster than the cuBLAS fallback
for f16 till bs=64, and for f32 till bs=32
* Review: refactor if statement
* CUDA: add a fused top-K MoE kernel
This kernel does the following:
1. softmax over the logits per token [n_experts, n_tokens]
2. argmax reduce over the top-k (n_experts_used) logits
3. write weights + ids to global memory
It is intended as fusion of softmax->top-k->get_rows pipeline for MoE models
* Refactor into ggml_cuda_should_use_topk_moe
* Review: Use better coalescing pattern, use WARP_SIZE, store logits into registers before
* Review: format + micro-optimizations
* Fix bug: fix tie breakers
* Add optional norm + clean-up code
* Use smem for final write
* Add bounds check
* Use better memory pattern for writeback
* CUDA: Add mul_mat_id support the mmf
Add support for mul_mat_id for bs < 16
* Review: use warp_size, fix should_use_mmf condition
* Launch one block per expert, stride along n_expert_used
* templatize mul_mat_id
* Pad shmem to 16 bytes, add helper function mul_mat_f_switch_ids
* Reduce compile times by dividing mmf into f16, bf16 and f32 variants
* Divide mmf by ncols_dst
* Add missing files
* Fix MUSA/HIP builds
* Add support for Llada-8b: diffusion model
* Add README
* Fix README and convert_hf_to_gguf
* convert_hf_to_gguf.py: address review comments
* Make everything in a single example
* Remove model-specific sampling
* Remove unused argmax
* Remove braced initializers, improve README.md a bit
* Add diffusion specific gguf params in set_vocab, remove setting rope_theta and rms_norm_eps
* Remove adding the mask token
* Move add_add_bos_token to set_vocab
* use add_bool in gguf_writer.py
