Repository navigation

Issues

New issue

changelog : libllama API
#9289 · ggerganov opened on Sep 3, 2024
15
changelog : llama-server REST API
#9291 · ggerganov opened on Sep 3, 2024
21
tutorials : list for llama.cpp
#13523 · ggerganov opened on May 14, 2025
24

Feature request: make mixed-vendor (e.g. Intel Vulkan target + NVIDIA CUDA draft) speculative decoding actually worth using
ggml-org/llama.cpp#31830 · opened on Nov 6, 2026
Vulkan output corruption on Intel Arc B-series persists under long-context/RAG prompts despite GGML_VK_DISABLE_COOPMAT workaround
ggml-org/llama.cpp#31828 · opened on Nov 6, 2026
Feature Request: improve large MoE model loading time when using offloading
ggml-org/llama.cpp#31825 · opened on Nov 6, 2026
Misc. bug: allOf with non-object types generates an empty object grammar
ggml-org/llama.cpp#31805 · opened on Nov 6, 2026
1
Eval bug: draft dflash on other device
ggml-org/llama.cpp#31800 · opened on Nov 6, 2026
Qwen2.5-Omni audio: clips are padded to 30 s and the encoder does not use 2 s windows
ggml-org/llama.cpp#31796 · opened on Nov 6, 2026
Feature Request: Video with audio support
ggml-org/llama.cpp#31792 · opened on Nov 6, 2026
Eval bug: Nemotron Labs Dory model not supported with llama.cpp
ggml-org/llama.cpp#31780 · opened on Nov 6, 2026
Eval bug: [Vulkan] Windows VIDEO_TDR_FAILURE (0x116) during Intel Arc multi-GPU inference; reduced micro-batch size avoids crash
ggml-org/llama.cpp#31777 · opened on Nov 6, 2026
Feature Request: Implement OpenAI's TTS /v1/audio/speech API
ggml-org/llama.cpp#31754 · opened on Nov 5, 2026
1
Eval bug: new Better UD 3.1 quants support on vulkan backend in AMD strix halo gfx1151
ggml-org/llama.cpp#31739 · opened on Nov 5, 2026
1
Eval bug: CPU F16 dot product accumulates in F16 on ARM (FP16 arithmetic, no SVE), overflows to inf
ggml-org/llama.cpp#31734 · opened on Nov 5, 2026
4
rpc-server: document or release cached device memory after the last client disconnects?
ggml-org/llama.cpp#31723 · opened on Nov 5, 2026
[ggml-cpu] AVX-512BW _1x8/_1x4/_1x3/_1x2 multi-column kernels, BMI2+SWAR IQ2_XXS sign unpacker & SwiGLU-H2 quantization for 284B DeepSeek-V4-Flash in 57.61 GiB RAM (+77.1% pp)
ggml-org/llama.cpp#31713 · opened on Nov 5, 2026
Misc. bug: CUDA decode fragmentation with non-unified KV and 32 variable-length requests
ggml-org/llama.cpp#31711 · opened on Nov 5, 2026
3
Feature Request: Either GPU alone produces perfect output. The corruption appears only when layers are split across both.
ggml-org/llama.cpp#31710 · opened on Nov 5, 2026
Eval bug: cancelled long embedding is never released — server then serves nothing while /health reports ok
ggml-org/llama.cpp#31704 · opened on Nov 5, 2026
1
Misc. bug: cache_reuse disabled for multimodal models (Qwen-VL)
ggml-org/llama.cpp#31700 · opened on Nov 4, 2026
Misc. bug: server: realtime session.update asking for 24 kHz is accepted, the audio is still read as 16 kHz
ggml-org/llama.cpp#31696 · opened on Nov 4, 2026
server: /v1/realtime audio output ignores output language/voice (non-English speech gets English accent; llama-tts with --tts-lang works)
ggml-org/llama.cpp#31689 · opened on Nov 4, 2026
Eval bug: Chatterbox Multilingual drops letters with diacritics (ą ś ż …) because the original NFKD/lowercase preprocessing is missing
ggml-org/llama.cpp#31686 · opened on Nov 4, 2026
SYCL AOT (GGML_SYCL_DEVICE_ARCH=arl-h) produces garbage output (GGGG...) for models > 4 GB, fixed by buffer size flag
ggml-org/llama.cpp#31685 · opened on Nov 4, 2026
Eval bug: POWER10: llamafile sgemm (MMA) produces corrupted output for Flint (MoE); correct with GGML_LLAMAFILE=OFF
ggml-org/llama.cpp#31680 · opened on Nov 4, 2026
1
Feature Request: SYCL FP16 build for Windows
ggml-org/llama.cpp#31678 · opened on Nov 4, 2026
2
Compile bug: cannot build for spacemit soc with upstream binutils
ggml-org/llama.cpp#31676 · opened on Nov 4, 2026