Repository Issue Activity (beta)

#3508 [Qualcomm NPU] Only the first prefill chunk survives: prompts longer than the prefill signature (128) lose everything after token 128

#3444 Conversation fails mid-flight with "Chosen prefill work group size exceeds available state entries" when remaining context < smallest prefill signature; EngineConfig accepts maxNumTokens above what the bundle serves, and an app has no way to query the effective ceiling

#3418 README: the top gemma-4 command pairs the E4B file with the E2B repo, so it fails with 404 as published

#3334 litertlm-android 0.16.1 crashes at runtime: NoSuchMethodError SendChannel.close$default (no public kotlinx-coroutines release carries that synthetic)

#3364 npm @litert-lm/[email protected] package contains only package.json

#3507 Android GPU: rearranged weights are copied to the GPU instead of mmap'd (iOS keeps mmap) — +780 MB resident vs iOS at the same 8192 window

#3135 GPU samplers not loading on Android and Windows due to missing symbols in LiteRT-LM v0.15.0 prebuilt (libLiteRtTopKOpenClSampler.so, libLiteRtTopKWebGpuSampler.dll)

#2149 CPU decode crash on Gemma 4 (engine_cpu) — segfault on E4B, silent hang on E2B post-RunPrefillAsync

#3443 Qwen3-0.6B bundle: second message on any conversation fails with INTERNAL "rendered template string does not start with the previous" (the bundle's embedded HF template renders the trailing assistant turn differently once the next message is appended)

#1748 Custom INT4 .litertlm conversion report for Phi-4-mini and Qwen2.5-1.5B

#3194 Request: binary-specific SBOM and third-party notices for CLiteRTLM.xcframework v0.15.0

#3495 README: Gemma4-E4B command points to the Gemma4-E2B Hugging Face repository

#2424 Feature request: Support thinking budget for reasoning models

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

#2779 [Bug] OpenAI server treats `model_id,gpu` as full model ID in LiteRT-LM 0.14.0

#2812 litertlm-android 0.14.0: NoSuchMethodError SendChannel.close$default in Conversation.onDone — AAR built against kotlinx-coroutines 1.11.0 but POM declares 1.9.0

#3477 macOS arm64 prebuilt libLiteRtTopKWebGpuSampler.dylib exports an older ABI than the engine expects (_UpdateConfig missing)

#3165 Cross-conversation token dedup reuses recurrent state on running-state models (observation from shipping Mamba2 / gated-delta hybrids on v0.15.0)

#3348 fast_vlm: prompt text containing added-token specials (e.g. <|im_start|>) appears not to encode them to their token ids

#3486 # License compliance defect distributed binary carries no valid Apache-2.0 Section 4 grant, and executes a format layout of record predating its release — reproducible verification and preservation request for COMP-26-005779 referral of GAO to the U. S. Department of Justice (DOJ), Office of the Inspector General (OIG )

#3417 FastVLM-0.5B: the bundle's jinja template renders history turns only for role 'model' — a conversation created with role 'assistant' (the runtime's own convention) silently loses those turns from the prompt

#3439 Speculative decoding: turn-final stop token is never committed, so multi-turn transcripts lose <|im_end|> after every assistant turn

#3445 Structured think-prefix history: re-rendered assistant turns keep the pre-opened `<think>` but never close it, so every past answer reads as an unfinished thought — long conversations quietly lose coherence

#3446 ~1 MB of phys_footprint retained per conversation create/destroy cycle (197 cycles over 30 min → +198 MB, macOS/Metal, v0.16.0)

#3463 `ThinkingConfig(enableThinking: true, budget: N)` makes every `ResponseFormat` turn fail (llguidance `ff_tokens` assertion at token 0)