Repository Issue Activity (beta)

#2814 iOS/Metal: Gemma 4 corrupts digit sequences in output (clean on CPU) — sampler-handles-input reversed condition; request patch release of #2805

#2964 Konrad

#2963 Konrad

#2249 Feature request: repetitionPenalty on SamplerConfig + per-segment sampler override - reproducible Gemma 4 tool-arg corruption catalog inside

#2815 Package.swift manifests are inconsistent across tags; unsafeFlags prevents version-based dependency

#2957 [Bug] v0.14.0 regression: Windows WebGPU engine_create crashes (ACCESS_VIOLATION in prebuilt libLiteRtWebGpuAccelerator.dll) on Intel Arc — v0.13.1 works on the same machine

#2940 [LFM2.5] Not working on main branch

#2424 Feature request: Support thinking budget for reasoning models

#2637 [Bug] Token repetition/corruption during Windows GPU inference with long context (>2000 tokens)

#2224 CPU-only build for Python

#2471 openai_server: tool_choice ignored — required, none, and named-tool all behave as auto

#2506 Add an option to toggle thinking mode in the CLI app

#2507 Feature Request: Enable MediaTek NPU acceleration (Dimensity 9300+ / MT6989) via LiteRT

#2529 When does the C API go ABI-stable / when will you publish a prebuilt shared lib?

#2540 [Feature Request] Expose token logprobs / logits during text generation in LiteRT-LM

#2556 Add a distinct keyboard shortcut or command for clearing context window in the CLI app

#2569 [Feature] Add performance metrics to the CLI app

#2804 Worse performance on `litert-lm serve` compared to `litert-lm run`

#2816 MTP/speculative decoding counts draft tokens toward max_output_tokens budget (GPU generates ~47 percent of requested tokens)

#2945 [CMake] Native from-source build of litert_lm_main fails on 3 further defects at v0.13.0 (stale LiteRT pin, host-flatc path, litert_cc_options patcher truncation)

#2376 البي

#2720 Technical Analysis: Structural Boundary-Layer Failure Modes in 64-Bit Quantized Inference Runtimes

#2095 Where can I download a full Gemma 4 full litert-lm model?

#2953 NOTICE OF ENCUMBRANCE — Source-level architectural correspondence to USPTO 19/700,381 (Compressed Memory Manifold); federal workload deployment; preservation demand; pending federal and state matters by identifier

#2777 Android GPU weight cache is recreated on every load in LiteRT-LM 0.14.0 due to cache filename order mismatch