mediapipe
mediapipe copied to clipboard
Getting gibberish output using phi-2 when building Mediapipe from source
Have I written custom code (as opposed to using a stock example script provided in MediaPipe)
None
OS Platform and Distribution
Android 14
Mobile device if the issue happens on mobile device
Pixel 8 pro
Browser and version if the issue happens on browser
N/A
Programming Language and version
C++/Java
MediaPipe version
v0.10.14
Bazel version
6.1.1
Solution
MediaPipe
Android Studio, NDK, SDK versions (if issue is related to building in Android environment)
NDK r25c
Xcode & Tulsi version (if issue is related to building for iOS)
N/A
Describe the actual behavior
Running phi-2 after conversion (with my build from source) is producing gibberish output.
Describe the expected behaviour
Running phi-2 after conversion (with my build from source) should produce non-gibberish output.
Standalone code/steps you may have used to try to get what you need
I have used this colab notebook to convert phi-2 model to mediapipe format:
https://colab.research.google.com/github/googlesamples/mediapipe/blob/main/examples/llm_inference/conversion/llm_conversion.ipynb
Then I have built mediapipe + XNNPACK from soruce using v0.10.14 commit, I have used the following command:
$ bazel build -c opt \
--config=android_arm64 \
--strip=ALWAYS \
--legacy_whole_archive=0 \
--features=-legacy_whole_archive \
--copt=-fvisibility=hidden \
--copt=-ffunction-sections \
--copt=-fdata-sections \
--copt=-fstack-protector \
--copt=-Oz \
--copt=-fomit-frame-pointer \
--copt=-DABSL_MIN_LOG_LEVEL=2 \
--linkopt=-Wl,--gc-sections,--strip-all \
//mediapipe/tasks/java/com/google/mediapipe/tasks/genai:tasks_genai.aar
Did the same for tasks_core.aar
Then, I have used Mediapipe LLM sample app to test this: https://github.com/google-ai-edge/mediapipe-samples/tree/main/examples/llm_inference/android
When running phi-2 with my build I am getting gibberish output.
It is worth noting that Gemma model works fine after conversion with my build, I am facing this issue with the other models (e.g. phi-2). For some reason, when I use the released AARs from Maven, all models works fine, the issue happens when I build MediaPipe+XNNPACK from source.
### Other info / Complete Logs
```shell
The current output I am getting with my build from source:
Ġ=ĊTheĠlatestĠnewsĠandĠresearchĠonĠtheĠglobalĠeconomyĊĊInĠthisĠarticle,ĠweĠwillĠdiscussĠtheĠimpactĠofĠtheĠCOVID-19ĠpandemicĠonĠtheĠglobalĠeconomy.ĊTheĠglobalĠeconomyĠisĠexpectedĠtoĠshrinkĠbyĠ4.32%ĠinĠ2020,ĠandĠitĠisĠprojectedĠtoĠrecoverĠtoĠpre-pandemicĠlevelsĠbyĠ2023.ĊTheĠglobalĠeconomyĠisĠtheĠsumĠofĠallĠtheĠeconomicĠactivitiesĠthatĠtakeĠplaceĠwithinĠaĠcountry'sĠborders.ĠItĠisĠaĠcountry'sĠGrossĠDomesticĠProductĠ(GDP)ĠandĠisĠaĠmeasureĠofĠtheĠvalueĠofĠallĠtheĠ