400 Client Error on Phenom-Beta Inference API endpoint

[l-haus]

Description
I am encountering a 400 Bad Request error when attempting to use the API's inference endpoint. However, uploading and deleting assets are working correctly without any issues. This problem only arises during the inference operation.

Steps to Reproduce

1. Upload an asset (asset uploads successfully).
2. Attempt to run inference on the uploaded asset.
3. Receive a 400 Bad Request error.

Expected Behavior
The inference operation should process the asset and return a successful response with the inference results.

Actual Behavior
The API returns a 400 Bad Request error, and no inference results are received.

Additional Information
Endpoint: https://api.nvcf.nvidia.com/v2/nvcf/exec/functions/7db32b36-ec04-43a6-a78f-1d8296accd8d/versions/3d73b252-008d-4469-b4c3-b25b9cbec654
Request Method: POST
Payload Used: { "asset_id": "f2af3e59-aea2-49b8-a567-98ab4e226410" }
Error Message: Bad Request

Attempts to Resolve
Verified that the asset exists and is accessible (accessible via delete endpoint).
Ensured that the request payload is structured correctly according to the documentation.
Checked API documentation for any recent changes (https://developer.nvidia.com/docs/bionemo-service/phenom-beta.html).

Request
Could you please investigate this issue? The request appears to be structured correctly, and other endpoints are functioning as expected. I would greatly appreciate any guidance or corrections to my request.

[kian-kd]

Hi @aluettringhaus , thanks for reaching out! Currently the public endpoint is very particular and requires that the input TIFFs be a rectangle with a shape that's exactly divisible by 256x256. We did not want to include logic that applies clipping the edges of the images behind the scenes, and instead leave that to the user. The error may not be surfacing correctly, currently. So if you don't mind, please try re-uploading your images after trimming them, ensuring that the Height and Width of your input images are of a dimensionality that is evenly divisible by 256x256.

[l-haus]

Hi @kian-kd , Thanks for getting back to me! I have double-checked the input images. They are multi-channel images with a shape of (4, 1536, 2048), so the height and width dimensions are exactly divisible 256. Additionally, I have tried the inference endpoint with LZW-compressed and uncompressed 8-bit TIFF images. Are there any other requirements for the TIFFs?