Skip to main content

Enable Vision Preprocessing

This guide explains how to enable vision preprocessing on a running PaletteAI Inference Launchpad appliance. After you finish, a text-only model such as GLM 5.2 can answer questions about screenshots and other images. Coding assistants keep calling the text model they already use. For what the feature does and what clients receive, refer to Vision Preprocessing.

Prerequisites​

  • A running PaletteAI Inference Launchpad appliance, with the admin console reachable and operator access.

  • A text-only model in the appliance catalog, ready to deploy. GLM 5.2 is the validated text model for this path. To place a model in the catalog, refer to Upload a Model. To deploy it, refer to Deploy a Model.

  • A vision model in the appliance catalog. Qwen 3.5 9B multimodal is the validated vision model for this path.

  • Hardware that can run both models on the same node. Spectro Cloud has validated GLM 5.2 with Qwen 3.5 9B multimodal on 8 x MI325X, 8 x B200, and 8 x H200. Both models run on the same physical GPUs; no manual GPU reservation is needed. For certified text models by GPU, refer to Certified Models by Hardware.

  • The catalog entries glm-5.2-shared-with-vision (text half) and qwen-3.5-9B-vision (vision half). Vision preprocessing needs a vision-mode text model, which is a text model whose catalog entry ends in -shared-with-vision. It runs with the -vision model of its group. You deploy these two entries in Deploy Both Models and do not edit their settings by hand. For the per-GPU-family variants each entry carries, refer to Certified Models by Hardware.

  • An appliance that runs no text-mode model. An appliance runs its models in one mode, so the deploy picker offers the vision-mode entries only when no text-mode model is deployed. For how a model's catalog entry decides its mode, refer to One Mode per Appliance.

Deploy Both Models​

Vision preprocessing needs two models serving at the same time: the text model that answers users, and the vision model that converts images to text. Deploy the text model first, then the vision model. On the validated pairing, choose glm-5.2-shared-with-vision as the text half and qwen-3.5-9B-vision as the vision half. Those are the catalog entries whose per-hardware variants carry the tuned deploy configuration.

  1. From the left main menu, select Cluster.

  2. If the text model is not already serving, select Deploy Model, choose the text model, and complete the deploy flow with GPU sharing turned on. Refer to Share a GPU Between Models.

  3. Wait until the text model health reads N/N healthy and its state reaches ready or serving. A large text model such as GLM 5.2 can take 5 to 10 minutes on the first-ever cold start while it loads weights, compiles kernels, and captures CUDA graphs. Do not cancel the deploy during this window. Subsequent restarts reuse the compiled kernels from the on-appliance cache and start faster.

  4. Select Deploy Model again, choose the vision model, and complete the deploy flow with GPU sharing turned on.

  5. Wait until the vision model also reaches ready or serving. The vision model is small and starts in under two minutes on validated hardware.

info

The Multimodal preprocessing card on Settings > Configurations appears after a second model is available, or when vision preprocessing is already on. If the card does not appear, confirm both models are serving, then open Settings again.

If the deploy panel reports that no node has enough free GPUs, the vision model cannot be placed next to the text model on this appliance. Free capacity, choose a smaller vision model, or contact Spectro Cloud to confirm a supported pairing for your hardware.

Enable Multimodal Preprocessing​

  1. From the left main menu, select Settings.

  2. Select the Configurations tab.

  3. In the Multimodal preprocessing card, select Enable vision preprocessing.

  4. Open the Vision engine drop-down menu and select the engine that serves the vision model. The list is filled from the models that are currently serving.

  5. Open the Vision model drop-down menu and select the vision model. The field is marked optional. Set it anyway so the appliance sends extracts to a named model.

  6. Leave Max images per request at 8, or enter a different per-turn cap. If a turn includes more images than this value, the appliance processes images up to the cap and skips the rest with a notice. It does not drop extra images silently.

  7. Select Turn on vision preprocessing.

The setting applies to new requests. Requests already in progress are not changed. Text-only requests never use the vision model, whether the setting is on or off. On this vision-mode appliance, image requests that send auto are routed by the semantic router too, and the Overview page shows a Vision preprocessing card.

Verify Image Requests​

Confirm that an image sent to the text model comes back with a correct answer. You can use a coding assistant that is already pointed at the appliance.

  1. Connect Claude Code, or another supported coding assistant, to the appliance. Refer to Use Claude Code.

  2. Start a session that uses the text model, not the vision model.

  3. Paste a simple image, such as a solid-color PNG, or attach a screenshot, and ask a direct question about it. For example, ask what color the image is, or what a labeled value in the screenshot reads.

  4. Confirm the reply matches the image. A solid red image should be answered as red. A screenshot question should quote or describe what is visible.

  5. Send a follow-up that contains only text and no image. Confirm you still get a normal reply. That turn must not depend on the vision model.

tip

If a short test question comes back with no visible answer, raise the output token limit and try again. Reasoning models use some of that budget before they write the visible reply, so a one-word answer can still need more room than it appears.

Turn Off Vision Preprocessing​

  1. From the left main menu, select Settings.

  2. Select the Configurations tab.

  3. In the Multimodal preprocessing card, clear the Enable vision preprocessing checkbox.

  4. Select Turn off vision preprocessing.

Image requests then fail until you turn the feature on again, or until you send the request without images. Text-only requests continue to work.

Operational Caveats​

  • Do not use the vision model as a chat model. It exists to convert images to text. Do not point coding-assistant routing at it; keep clients on the text model. Refer to Manage a Client's Model Access.
  • Keep both models serving. If the vision model stops serving, image requests fail instead of continuing without the images. Text-only requests are unaffected.
  • Stay within the image cap. The default cap is 8 images per turn. Extra images in the same turn are skipped with a notice. Ask about them in a follow-up message.
  • Expect the vision model to occupy GPU memory alongside the text model. Qwen 3.5 9B is sized to run next to GLM 5.2 rather than instead of it, and together the two reserve most of the VRAM on each shared GPU. Do not treat the remainder as spare capacity for another large model. For the per-GPU memory split, refer to Certified Models by Hardware.
  • This path covers vision preprocessing only. It does not configure arbitrary multi-model placement or a separate vision API.

Next Steps​