Skip to content

Pull requests: ikawrakow/ik_llama.cpp

New pull request New
34 Open 1,421 Closed
34 Open 1,421 Closed
Author
Filter by author
Loading
Label
Filter by label
Loading
Use alt + click/return to exclude labels
or ⇧ + click/return for logical OR
Projects
Filter by project
Loading
Milestones
Filter by milestone
Loading
Reviews
Filter by reviews
No reviews Review required Approved review Changes requested
Assignee
Filter by who’s assigned
Assigned to nobody Loading
Sort
Sort by
Newest Oldest Most commented Least commented Recently updated Least recently updated Best match
Most reactions
👍 👎 😄 🎉 😕 ❤️ 🚀 👀

Pull requests list

Fix DSV4 tool calls and reasoning
#2242 opened Aug 2, 2026 by sayap Contributor Loading…
2 of 4 tasks
1
Another mul_mat_iq4_nl_r4
#2241 opened Aug 2, 2026 by cora4 Contributor • Draft
deepseek4: add --dsv4-cache-cpu to keep compressed-attention K caches in host memory
#2239 opened Aug 2, 2026 by nmccrory Loading…
Apply helper functions to the MTP graph
#2238 opened Aug 2, 2026 by SamuelOliveirads Collaborator Loading…
Chores : tidy up more typos project wide (ggml directory excluded), new -ptcall alias
#2237 opened Aug 2, 2026 by Nexesenex Contributor Loading…
2 tasks done
Allow concatenating quantized tensors
#2232 opened Aug 2, 2026 by ikawrakow Owner Loading…
1
Another minor indexer optimization on the CPU
#2231 opened Aug 2, 2026 by ikawrakow Owner Loading…
deepseek4: fix multi-stream (parallel sequence) decoding
#2229 opened Aug 2, 2026 by nmccrory Loading…
5
Bucket top_k (CPU): ~3% better TG at 128k context
#2225 opened Aug 1, 2026 by ikawrakow Owner Loading…
window-sized SWA ring KV cache (--swa-compress to opt in)
#2171 opened Jul 23, 2026 by joaomdsg Loading…
35
CUDA sm_60: bound the fp16 flash-attention error on P100 prefill instead of disabling the fast path
#2151 opened Jul 18, 2026 by mb8565 Contributor Loading…
6
ggml : fix build and MoE inference on CPUs without AVX2 (AVX1-only)
#2138 opened Jul 15, 2026 by neomindryan Loading…
2 of 4 tasks
6
llama: disable fused up-gate when an offload backend lacks GGML_OP_FUSED_UP_GATE (fixes silent 3.5x Metal slowdown)
#2133 opened Jul 15, 2026 by hchengit Contributor Loading…
4
Scope split-mode graph mmap to the tensor being loaded. Prevents memory leak
#2126 opened Jul 13, 2026 by Farmadupe Contributor • Draft
2 of 4 tasks
Reduce host memory pressure when performing a sm graph load.
#2125 opened Jul 13, 2026 by Farmadupe Contributor • Draft
1
CUDA: update captured graph when a source tensor's shape changes
#2096 opened Jul 8, 2026 by fibrahimov Loading…
3
Be able to force synchronization when copying graph inputs
#2092 opened Jul 6, 2026 by ikawrakow Owner Loading…
1
add --split-output-tensor / -sot CLI parameter
#2056 opened Jun 29, 2026 by Nexesenex Contributor • Draft
2 of 4 tasks
2
implement perplexity in llama-server
#2011 opened Jun 22, 2026 by magikRUKKOLA Contributor • Draft
2
fix: checkpoint restore for hybrid/recurrent models
#1976 opened Jun 16, 2026 by razlani Loading…
1
5
Fix misc. expiring logit/sparam bias bugs
#1914 opened Jun 2, 2026 by dungquixote42 Contributor Loading…
2 of 4 tasks
9
Qwen3.5 MTP: extract selected tokens earlier
#1892 opened May 28, 2026 by ikawrakow Owner Loading…
server: enable checkpoint reuse for recurrent/hybrid models (qwen3next, Mamba)
#1888 opened May 26, 2026 by localweights Loading…
1
3
Fix prompt cache viability
#1877 opened May 25, 2026 by zeljkokalezic Loading…
2 of 4 tasks
5
docs: Complete rewrite of build.md – CMake-only, focused on supported backends
#1853 opened May 21, 2026 by maddes8cht Loading…
2 of 4 tasks
76
ProTip! Follow long discussions with comments:>50.