Question 1

Why are the ENCODED_VIDEO_LOCATION and THUMB_LOCATION environment variables not taking effect in Immich?

Accepted Answer

These environment variables are not read by the Immich server code. They are only intended for variable substitution in the docker-compose.yml to simplify volume mount paths. To actually use custom storage locations, update the volume mounts for the relevant containers to map your desired host directories to the expected internal paths (e.g., mount /your/custom/thumbs to /usr/src/app/upload/thumbs). Refer to the official custom-locations guide for examples.

Question 2

Why does Immich "Download all" save an album as a file named `archive` with no .zip extension?

Accepted Answer

This happens in Chromium-based browsers when the album name contains an RFC 8187 reserved character, especially an apostrophe, e.g. `Sam's birthday`. Immich sends `filename*=UTF-8''Sam's%20birthday.zip`; the unescaped apostrophe acts as the header delimiter, so Chromium rejects the filename and falls back to the URL endpoint name `archive`.

The fix is to escape `'`, `(`, `)`, and `*` after `encodeURIComponent` when building the `Content-Disposition` header:

```ts
filename*=UTF-8''${
  encodeURIComponent(dto.archiveName)
    .replace(/['()*]/g, c => `%${c.charCodeAt(0).toString(16).toUpperCase()}`)
}.zip
```

Apply PR #32297 or upgrade to a release containing it. Temporary workarounds: rename the album to remove apostrophes/parentheses/asterisks, use Firefox or Safari, or manually rename the downloaded `archive` file to `album-name.zip`.

Question 3

Why does Immich OCR fail with TensorRT "Incompatible dimensions" when maximum resolution is set to 1080?

Accepted Answer

The OCR text detector requires both canvas dimensions to be divisible by 32. In Immich v3.3.1 with ML model revision v2, `ocr_canvases()` rounds only one dimension and uses the configured maximum resolution as-is for the other, so values such as `1080` cause TensorRT/MIGraphX model preparation to fail.

Workaround: go to **Machine Learning → OCR → Maximum resolution** and set it to a multiple of 32. Use `1088` instead of `1080`; other known working values include `960` and `1056`. Save the setting, restart `immich-machine-learning` if necessary, then rerun the OCR job. Avoid switching OCR models just to work around this, because switching models removes existing OCR results.

A permanent fix rounds the configured maximum resolution up to a multiple of 32; upgrade to a release containing PR "fix(ml): Round OCR max resolution up to a multiple of 32" when available.

Question 4

How to fix Immich Machine Learning ROCm error “Cannot read TensileLibrary.dat: Illegal seek for GPU arch : gfx900” after upgrading to 2.4.0?

Accepted Answer

Immich ML 2.4.0’s newer ROCm image no longer ships `TensileLibrary_lazy_gfx900.dat`, which is required when forcing an unsupported Vega/Renoir GPU to gfx900. Copy the file from the last pre-2.4.0 ROCm image and bind-mount it into the new container.

1. Extract the missing library from an older Immich ML image, using the last version that worked for you:
```sh
OLD_TAG=v2.3.0-rocm
cid=$(docker create ghcr.io/immich-app/immich-machine-learning:$OLD_TAG)
docker cp "$cid":/opt/rocm/lib/rocblas/library/TensileLibrary_lazy_gfx900.dat ./TensileLibrary_lazy_gfx900.dat
docker rm "$cid"
```

2. Add the file as a bind mount in `docker-compose.yml`:
```yaml
volumes:
  - ./TensileLibrary_lazy_gfx900.dat:/opt/rocm/lib/rocblas/library/TensileLibrary_lazy_gfx900.dat:ro
```

3. Keep the gfx900 override present:
```yaml
environment:
  - HSA_OVERRIDE_GFX_VERSION=9.0.0
  - HSA_USE_SVM=0
```

4. Recreate the container:
```sh
docker compose up -d immich-machine-learning
```

If the log later reports other missing gfx900 `.dat` files, extract those from the same older image using the same method. This is a workaround because gfx900 is not officially supported by the newer ROCm stack; alternatives are pinning Immich ML to a 2.3.x ROCm image or falling back to CPU execution.

Question 5

How do I fix the rocBLAS "TensileLibrary.dat" error for GPU arch gfx90c in Immich Docker?

Accepted Answer

For AMD Vega-based gfx90c APUs, force ROCm to use the compatible gfx900 architecture by adding `HSA_OVERRIDE_GFX_VERSION=9.0.0` to the `immich-machine-learning` service. In `docker-compose.yml`, add:

```yaml
environment:
  - HSA_OVERRIDE_GFX_VERSION=9.0.0
```

Alternatively, put `HSA_OVERRIDE_GFX_VERSION=9.0.0` in the env file loaded by the service (`stack.env` in the reported setup). Ensure `/dev/dri` and `/dev/kfd` are passed through and the container is in the `video` group, then run:

```bash
docker compose up -d --force-recreate immich-machine-learning
```

This resolves `rocBLAS error: Cannot read ... TensileLibrary.dat ... for GPU arch : gfx90c` by reporting the APU as gfx900. Note: newer Immich ROCm images may use a ROCm version that no longer supports Vega APUs; if the override stops working, pin the `immich-machine-learning:<version>-rocm` image to an older release whose ROCm still includes gfx900 support.

Immich

Core Features

What It Can't Do

Use Cases

Tags

Getting Started

Similar Projects