Issues

Real-time person matting on an Intel iGPU in Edge: how to reduce frame-time and stutter with LiteRT.js WebGPU?

google-ai-edge/LiteRT#11612 · opened

on Nov 5, 2026

2

litert_quantizer: a quantized STABLEHLO_COMPOSITE and its decomposition subgraph can get different output scales, so the result is off by their ratio

google-ai-edge/LiteRT#11478 · opened

on Nov 3, 2026

2

MediaTek MT6993 / LiteRT 2.3.0: GetSupportedOperations preflight stops without operation or NeuronAdapter error diagnostic

google-ai-edge/LiteRT#11428 · opened

on Nov 1, 2026

1

[WebGPU] Depthwise shader leaves channels unwritten on Intel D3D12 with FXC; DXC works

google-ai-edge/LiteRT#11420 · opened

on Oct 31, 2026

1

MediaTek NPU (MT6993) can't fully run Ultralytics YOLO depth and OBB models: EXP, SIN, COS, FLOOR, ROUND, CAST and SELECT are not offloaded

google-ai-edge/LiteRT#11413 · opened

on Oct 31, 2026

4

algorithm time cost fluctuates greatly：google pixel 10

google-ai-edge/LiteRT#11390 · opened

on Oct 30, 2026

2

[TFLite] Quantized int16 EQUAL/NOT_EQUAL/LESS/LESS_EQUAL give wrong results; unquantized int16 NOT_EQUAL writes no output

google-ai-edge/LiteRT#11363 · opened

on Oct 29, 2026

2

[GPU/OpenCL] FP16 SpaceToDepth/DepthToSpace performance regression on Adreno 730 in 2.2.0 and 2.3.2

google-ai-edge/LiteRT#11301 · opened

on Oct 27, 2026

1

[TFLite] VerifyAndBuildFromBuffer accepts a model with no subgraphs; InterpreterBuilder dereferences null

google-ai-edge/LiteRT#11291 · opened

on Oct 27, 2026

1

Qualcomm dispatch: models with several QNN contexts in different HTP PDs fail to run, because every tensor buffer is registered with the most recently created context

google-ai-edge/LiteRT#11252 · opened

on Oct 26, 2026

2

[Google Tensor] On-device (JIT) compilation is refused for third-party apps ("not in the EdgeTPU allowed list", error code 16) and the model silently runs on CPU

google-ai-edge/LiteRT#11251 · opened

on Oct 26, 2026

7

An app that ships both the Google Tensor and Qualcomm NPU runtimes can't use the Pixel NPU: LiteRT always loads the first libLiteRtDispatch_*.so it finds

google-ai-edge/LiteRT#11250 · opened

on Oct 26, 2026

2

odml.sdpa_transposed on the Metal accelerator returns wrong values when query heads are grouped over 2+ key/value heads

google-ai-edge/LiteRT#11179 · opened

on Oct 23, 2026

2

Memory Leak in MetalAccelerator?

google-ai-edge/LiteRT#11150 · opened

on Oct 22, 2026

2

TranslateGemma 4B prefill AOT compile exhausts 125 GiB RAM and 157 GiB swap (G3 Target)

google-ai-edge/LiteRT#11129 · opened

on Oct 22, 2026

4

Control-flow tensor copying should reject out-of-range tensor indices

google-ai-edge/LiteRT#11093 · opened

on Oct 21, 2026

2

FULLY_CONNECTED VerifySparsity misses 1x4 block-sparse d1 index scaling and CSR array_segments monotonicity

google-ai-edge/LiteRT#11084 · opened

on Oct 21, 2026

3

Galaxy S26 (SM8850) NPU: the `encode` signature of the parakeet SM8850 AOT models returns NaN on every input I tried (LiteRT 2.2.0)

google-ai-edge/LiteRT#11038 · opened

on Oct 20, 2026

1

Python GIL held by CompiledModel API?

google-ai-edge/LiteRT#10994 · opened on Oct 18, 2026

3

GPU accelerator fails on FULLY_CONNECTED with int8 weights + DEQUANTIZE ("Shape mismatch: {4096, 1, 1, 1024} vs {1, 1, 4096, 1024}")

google-ai-edge/LiteRT#10934 · opened

on Oct 15, 2026

3

[Web] Support multithreading with standard SIMD (without Relaxed SIMD)

google-ai-edge/LiteRT#10903 · opened

on Oct 15, 2026

1

UINT8 output remains unchanged after LiteRT OpenCL inference on Android

google-ai-edge/LiteRT#10800 · opened

on Oct 13, 2026

3

Quantized int8 DIV saturates small quotients to the int8 endpoints

google-ai-edge/LiteRT#10745 · opened

on Oct 12, 2026

2

[Google Tensor SDK Beta][Tensor G6] AOT compiler returns INTERNAL for the Qwen3-0.6B decoder of Qwen3-ASR-0.6B and Fun-ASR-Nano-2512 (audio encoders compile fully)

google-ai-edge/LiteRT#10479 · opened

on Oct 2, 2026

6

[Feature] Support standard shape broadcasting in MLDrift GPU delegate (or automate 4D Reshape insertion for lower/mixed-rank inputs)

Feature

google-ai-edge/LiteRT#10445 · opened

on Oct 1, 2026

2

Navigation Menu

Repository navigation

Issues

Real-time person matting on an Intel iGPU in Edge: how to reduce frame-time and stutter with LiteRT.js WebGPU?

litert_quantizer: a quantized STABLEHLO_COMPOSITE and its decomposition subgraph can get different output scales, so the result is off by their ratio

MediaTek MT6993 / LiteRT 2.3.0: GetSupportedOperations preflight stops without operation or NeuronAdapter error diagnostic

[WebGPU] Depthwise shader leaves channels unwritten on Intel D3D12 with FXC; DXC works

MediaTek NPU (MT6993) can't fully run Ultralytics YOLO depth and OBB models: EXP, SIN, COS, FLOOR, ROUND, CAST and SELECT are not offloaded

algorithm time cost fluctuates greatly：google pixel 10

[TFLite] Quantized int16 EQUAL/NOT_EQUAL/LESS/LESS_EQUAL give wrong results; unquantized int16 NOT_EQUAL writes no output

[GPU/OpenCL] FP16 SpaceToDepth/DepthToSpace performance regression on Adreno 730 in 2.2.0 and 2.3.2

[TFLite] VerifyAndBuildFromBuffer accepts a model with no subgraphs; InterpreterBuilder dereferences null

Qualcomm dispatch: models with several QNN contexts in different HTP PDs fail to run, because every tensor buffer is registered with the most recently created context

[Google Tensor] On-device (JIT) compilation is refused for third-party apps ("not in the EdgeTPU allowed list", error code 16) and the model silently runs on CPU

An app that ships both the Google Tensor and Qualcomm NPU runtimes can't use the Pixel NPU: LiteRT always loads the first libLiteRtDispatch_*.so it finds

odml.sdpa_transposed on the Metal accelerator returns wrong values when query heads are grouped over 2+ key/value heads

Memory Leak in MetalAccelerator?

TranslateGemma 4B prefill AOT compile exhausts 125 GiB RAM and 157 GiB swap (G3 Target)

Control-flow tensor copying should reject out-of-range tensor indices

FULLY_CONNECTED VerifySparsity misses 1x4 block-sparse d1 index scaling and CSR array_segments monotonicity

Galaxy S26 (SM8850) NPU: the `encode` signature of the parakeet SM8850 AOT models returns NaN on every input I tried (LiteRT 2.2.0)

Python GIL held by CompiledModel API?

GPU accelerator fails on FULLY_CONNECTED with int8 weights + DEQUANTIZE ("Shape mismatch: {4096, 1, 1, 1024} vs {1, 1, 4096, 1024}")

[Web] Support multithreading with standard SIMD (without Relaxed SIMD)

UINT8 output remains unchanged after LiteRT OpenCL inference on Android

Quantized int8 DIV saturates small quotients to the int8 endpoints

[Google Tensor SDK Beta][Tensor G6] AOT compiler returns INTERNAL for the Qwen3-0.6B decoder of Qwen3-ASR-0.6B and Fun-ASR-Nano-2512 (audio encoders compile fully)

[Feature] Support standard shape broadcasting in MLDrift GPU delegate (or automate 4D Reshape insertion for lower/mixed-rank inputs)