Issues

Windows Intel GPU (D3D12): prompts prefilled in prefill_8 chunks give empty or wrong replies, and the result depends on earlier sessions (0.19.1, nightly 0.20.0.dev20261106)

google-ai-edge/LiteRT-LM#4460 · opened

on Nov 8, 2026

1

Omni GPU backend crashes on Exynos 2600 / Xclipse: SIGSEGV in vendor OpenCL during AHWB weight import (Qwen3-ASR)

google-ai-edge/LiteRT-LM#4423 · opened

on Nov 5, 2026

1

Feature Request: Add Qwen3-ForcedAligner support for word-level timestamps

google-ai-edge/LiteRT-LM#4420 · opened

on Nov 5, 2026

2

Feature request: run a session on caller-computed input embeddings and read the hidden states (speech output for LFM3-Audio without a second copy of the decoder)

google-ai-edge/LiteRT-LM#4408 · opened

on Nov 4, 2026

2

GPU: a clone's first send that repeats its parent's message gets a wrong or empty reply after another conversation generated

google-ai-edge/LiteRT-LM#4394 · opened

on Nov 4, 2026

1

Heuristic KV capacity reads the head count for [batch, seq, heads, dim] caches: litert-community Qwen2.5-0.5B-Instruct refuses every prompt (max tokens = 2)

google-ai-edge/LiteRT-LM#4384 · opened

on Nov 4, 2026

NPU executor loads the uncompiled TF_LITE_AUX and TF_LITE_EMBEDDER sections with kNpu, so the NPU compiler plugin runs on them on every load

google-ai-edge/LiteRT-LM#4362 · opened

on Nov 3, 2026

2

Qwen3-ASR (LiteRT Omni): no way to pass a context / system prompt; patched prefill restores accuracy

google-ai-edge/LiteRT-LM#4360 · opened

on Nov 3, 2026

3

[NPU] Several conversations on one engine share state: CloneContext does not save KV cache buffers the prefill signature does not take

google-ai-edge/LiteRT-LM#4355 · opened

on Nov 3, 2026

1

Feature request: TTS in litert-lm serve

google-ai-edge/LiteRT-LM#4353 · opened

on Nov 2, 2026

2

Android GPU: `Engine.initialize()` freezes the app's screen, even on a background thread

google-ai-edge/LiteRT-LM#4315 · opened

on Oct 30, 2026

1

[NPU][MediaTek MT6989] Second message in an existing Conversation returns an empty reply (0.18.1, no tokens on turn 2)

google-ai-edge/LiteRT-LM#4256 · opened

on Oct 28, 2026

1

Feature request: cache encoded image/video tokens so prompts on the same image don't re-run the vision encoder

google-ai-edge/LiteRT-LM#4211 · opened

on Oct 27, 2026

1

Rewind leaves `last_prefill_token_id` stale, so the first decode/scoring step after it feeds the wrong token

google-ai-edge/LiteRT-LM#4195 · opened

on Oct 26, 2026

2

Gemma 4 12B: could the per-channel int4 checkpoint behind gemma-4-12B-it.litertlm be published? Re-quantizing the Q4_0 QAT weights to per-channel int4 breaks audio

google-ai-edge/LiteRT-LM#4185 · opened

on Oct 26, 2026

1

NPU executor: LlmMetadata.max_num_tokens larger than the compiled KV cache becomes the token limit (no clamp to the cache length)

google-ai-edge/LiteRT-LM#4184 · opened

on Oct 26, 2026

1

Gemma 4 E2B NPU bundle: CPU audio takes ~47 s per request for a 1.9 s clip on Snapdragon 8 Gen 3 (about 1.4 s with the CPU build)

google-ai-edge/LiteRT-LM#4183 · opened

on Oct 26, 2026

1

Expose max_input_length in CLI embedding-server configuration

google-ai-edge/LiteRT-LM#4179 · opened

on Oct 25, 2026

2

iOS Metal, v0.18.0: Qwen3.5 GPU programs fail to load from the cache (`DATA_LOSS`), so every engine creation compiles them again (about a minute)

google-ai-edge/LiteRT-LM#4178 · opened

on Oct 25, 2026

2

Docs: delivering a .litertlm model to Android users with Play AI packs

google-ai-edge/LiteRT-LM#4177 · opened

on Oct 25, 2026

3

Title: UTF-8 / French accent corruption with Gemma 4 E2B on Raspberry Pi 5 using LiteRT-LM

google-ai-edge/LiteRT-LM#4175 · opened

on Oct 24, 2026

5

[GPU] Static KV cache allocates full max_num_tokens for all layers, ignoring sliding-window layers (Gemma 4 E4B: ~76 KB/token, ~5x over-allocation)

google-ai-edge/LiteRT-LM#4096 · opened

on Oct 21, 2026

1

Gemma 4 12B on Windows Intel GPU (D3D12): no text when the prompt is under 512 tokens (prefill_128-only path)

google-ai-edge/LiteRT-LM#4083 · opened

on Oct 20, 2026

2

litert_lm_builder: let callers keep uuid and creation_timestamp so that a rebuilt .litertlm can be byte-identical

google-ai-edge/LiteRT-LM#4064 · opened

on Oct 19, 2026

[FEATURE REQUEST] C API for LIteRT-LM Omni

google-ai-edge/LiteRT-LM#4058 · opened

on Oct 19, 2026

1

Navigation Menu

Repository navigation

Issues

Windows Intel GPU (D3D12): prompts prefilled in prefill_8 chunks give empty or wrong replies, and the result depends on earlier sessions (0.19.1, nightly 0.20.0.dev20261106)

Omni GPU backend crashes on Exynos 2600 / Xclipse: SIGSEGV in vendor OpenCL during AHWB weight import (Qwen3-ASR)

Feature Request: Add Qwen3-ForcedAligner support for word-level timestamps

Feature request: run a session on caller-computed input embeddings and read the hidden states (speech output for LFM3-Audio without a second copy of the decoder)

GPU: a clone's first send that repeats its parent's message gets a wrong or empty reply after another conversation generated

Heuristic KV capacity reads the head count for [batch, seq, heads, dim] caches: litert-community Qwen2.5-0.5B-Instruct refuses every prompt (max tokens = 2)

NPU executor loads the uncompiled TF_LITE_AUX and TF_LITE_EMBEDDER sections with kNpu, so the NPU compiler plugin runs on them on every load

Qwen3-ASR (LiteRT Omni): no way to pass a context / system prompt; patched prefill restores accuracy

[NPU] Several conversations on one engine share state: CloneContext does not save KV cache buffers the prefill signature does not take

Feature request: TTS in litert-lm serve

Android GPU: `Engine.initialize()` freezes the app's screen, even on a background thread

[NPU][MediaTek MT6989] Second message in an existing Conversation returns an empty reply (0.18.1, no tokens on turn 2)

Feature request: cache encoded image/video tokens so prompts on the same image don't re-run the vision encoder

Rewind leaves `last_prefill_token_id` stale, so the first decode/scoring step after it feeds the wrong token

Gemma 4 12B: could the per-channel int4 checkpoint behind gemma-4-12B-it.litertlm be published? Re-quantizing the Q4_0 QAT weights to per-channel int4 breaks audio

NPU executor: LlmMetadata.max_num_tokens larger than the compiled KV cache becomes the token limit (no clamp to the cache length)

Gemma 4 E2B NPU bundle: CPU audio takes ~47 s per request for a 1.9 s clip on Snapdragon 8 Gen 3 (about 1.4 s with the CPU build)

Expose max_input_length in CLI embedding-server configuration

iOS Metal, v0.18.0: Qwen3.5 GPU programs fail to load from the cache (`DATA_LOSS`), so every engine creation compiles them again (about a minute)

Docs: delivering a .litertlm model to Android users with Play AI packs

Title: UTF-8 / French accent corruption with Gemma 4 E2B on Raspberry Pi 5 using LiteRT-LM

[GPU] Static KV cache allocates full max_num_tokens for all layers, ignoring sliding-window layers (Gemma 4 E4B: ~76 KB/token, ~5x over-allocation)

Gemma 4 12B on Windows Intel GPU (D3D12): no text when the prompt is under 512 tokens (prefill_128-only path)

litert_lm_builder: let callers keep uuid and creation_timestamp so that a rebuilt .litertlm can be byte-identical

[FEATURE REQUEST] C API for LIteRT-LM Omni