Issues

iOS: `releaseContext` does not free Metal memory — `delete ctx` runs on a pool worker with no `@autoreleasepool`

#398

· juliusbangert opened

on Sep 17, 2026

`nextToken()` decodes with `llama_batch_get_one`: wrong positions and a KV-cell leak on iSWA models after prefix reuse past the sliding window

#396

· juliusbangert opened

on Sep 17, 2026

Android: SIGABRT in lm_ggml_backend_sched_split_graph on OpenCL (Q4_K_M) and HTP (q4_0 KV) — v0.13.0-rc.3

#389

· bmb-li opened

on Sep 9, 2026

OpenCL: incorrect R4 attention decode on Adreno 730 / E031.38

#386

· zraisan opened

on Sep 5, 2026

qwen2 (DeepSeek-R1-Distill) produces gibberish on iOS CPU path; same llama.cpp generation clean on macOS CPU

#374

· KA-devl opened

on Jul 30, 2026

Vulkan support for Android

#364

· roshansoni-io opened

on Jul 11, 2026

[Feature Request] Support draftless n-gram speculative decoding (ngram-mod / --spec-default)

#359

· abhinavk454 opened

on Jun 28, 2026

mutlimodel issues

#356

· animesh65432 opened

on Jun 18, 2026

The inference fail on Qualcomm NPU

#353

· ZhangWei125521 opened

on May 28, 2026

SIGSEGV in `lm_ggml_compute_forward_mul_mat+22XX → __memcpy_aarch64_simd` on aarch64 / Samsung Exynos under sustained completion load (Qwen3-0.6B-Q4_K_M, llama.rn 0.12.3 / llama.cpp b9254) — refines #336

#350

· eckonode opened

on May 24, 2026

GLM-OCR fails to load — "Failed to load model" on all configurations

#337

· tutak opened

on May 6, 2026

First `ctx.completion()` throws `Exception in HostFunction: <unknown>` on Samsung Exynos (qwen3-0.6b / TinyLlama / gemma4-e2b-lite, both with and without `no_extra_bufts: true`)

#336