[Bug]: GPU Inference Failure (Status 13) - Gemma3-1B-IT - LiteRT Executor

[synapstorm]

The application fails during the inference phase when using the GPU delegate on a Qualcomm chipset. The model compiles successfully, but the invocation of the compiled graph triggers a Status 13 error in the LiteRT runtime.

Error Signature:

Error: Status Code: 13. Message: ERROR: [third_party/odml/litert_lm/runtime/executor/llm_litert_compiled_model_executor.cc:836] Failed to invoke the compiled model

Model: Gemma3-1B-IT
Chipset: Qualcomm Adreno GPU
Runtime: LiteRT

[synapstorm]

Note: The model runs on CPU, but fails consistently on GPU with Status 13.

[dpknag]

Hello @synapstorm,

Thank you for reporting this issue. To help us investigate the "Status 13" GPU inference failure, could you please provide some additional information?

Device: What is the exact model and chipset of the device you are using (e.g., Samsung S24 Ultra, Snapdragon 8 Gen 3)?
Version: Which version of application are you running? (v.1.0.8, v.1.0.9..)
Reproduction Steps: Detailed steps to reproduce this error.
Other Models: Have you tried running other models on the GPU? If so, did they work correctly?
Logs: Could you Please attach the full adb logcat output capturing the crash.

We appreciate your help. Thanks!

[synapstorm]

Hello!

Thanks for the follow-up. To provide more context: I was running the Google AI Edge Gallery on a device with a Snapdragon 685 chipset.

My goal was to evaluate whether the Gemma3-1B-IT model could execute on the GPU of a lower-resource phone. I am specifically testing lower-end hardware to validate real-world compatibility and report whether users can expect stable GPU execution or not, helping to ensure a better experience for the community.

Here is the requested information:

• Device: Xiaomi Redmi Note 13, Snapdragon 685.
• Version: v1.0.9.
• Reproduction Steps: I selected the Gemma3-1B-IT model and attempted to initialize/run inference using the GPU delegate. The "Status 13" error appears during initialization.
• Other Models: Yes, I tested FastVLM-0.5B https://huggingface.co/litert-community/FastVLM-0.5B. This model worked correctly on both GPU and CPU on this same device.

It seems the issue might be specific to the 1B model's requirements on this chipset's GPU.

Thanks for reading my comment!.

[dpknag]

Hi @synapstorm,

Thank you for the detailed information.

Given that FastVLM-0.5B works but Gemma3-1B fails, this is likely a resource limitation. We will share these findings with the team to investigate further.

Thanks!