[skyrl][tinker] Multi-modal Tinker Sampling#1484
Open
nithinvc wants to merge 6 commits intoNovaSky-AI:mainfrom
Open
[skyrl][tinker] Multi-modal Tinker Sampling#1484nithinvc wants to merge 6 commits intoNovaSky-AI:mainfrom
nithinvc wants to merge 6 commits intoNovaSky-AI:mainfrom
Conversation
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
pcmoritz
pushed a commit
that referenced
this pull request
Apr 17, 2026
## Summary Integrates the VLLMRenderer (landed in #1464) into the SkyRL train backend so that VLM training batches include image placeholder tokens and decoded vision tensors (`pixel_values`, `image_grid_thw`). - When using new inference (`_SKYRL_USE_NEW_INFERENCE`), `_to_training_batch` lazily creates a `VLLMRenderer` and renders all `ModelInput`s through it. - Extracts `pixel_values` and `image_grid_thw` from rendered outputs and adds them to the `TrainingInputBatch` as `TensorList` entries (one tensor per batch element, since patch counts vary per image). - Extends `_pad_batch` to handle `TensorList` fields by cycling and cloning entries, matching the existing padding strategy for regular tensors. - Reorders `forward_backward` and `forward` to call `_to_training_batch` before `_sleep_inference_engines`, since the renderer needs the inference servers need to be initialized. Note that this does not wake the KV cache or model GPU memory since that is explicitly done in `save_weights_for_sampler` via the dispatcher. ## E2E Tinker VLM Classifier Curves With #1484 , we can now run tinker vision language recipes against the SkyRL. Merging both closes #1200 Example: ```bash _SKYRL_USE_NEW_INFERENCE=1 uv run --extra tinker --extra fsdp -m skyrl.tinker.api \ --base-model "Qwen/Qwen3-VL-8B-Instruct" \ --backend fsdp \ --backend-config '{"trainer.placement.policy_num_gpus_per_node": 8, "generator.inference_engine.num_engines": 8, "trainer.placement.colocate_all": true, "trainer.use_sample_packing": false}' ``` Cookbook ```bash TINKER_API_KEY=tml-dummy uv run --with tinker --with datasets --with torch python -m \ tinker_cookbook.recipes.vlm_classifier.train \ base_url=http://localhost:8000 \ model_name="Qwen/Qwen3-VL-4B-Instruct" \ dataset=caltech101 ``` Train nll: <img width="1200" height="675" alt="train_nll" src="https://github.com/user-attachments/assets/82e36767-edee-43b7-ab4a-7fbf496c8cbb" /> Val nll: <img width="1200" height="675" alt="val_nll" src="https://github.com/user-attachments/assets/1dc6e96b-7e1b-4ead-bf0e-71e42eab0491" /> Val accuracy: <img width="1200" height="675" alt="accuracy" src="https://github.com/user-attachments/assets/ec6f92b8-a544-42d9-9a00-4c06292e7ae3" /> <!-- devin-review-badge-begin --> --- <a href="https://app.devin.ai/review/novasky-ai/skyrl/pull/1496" target="_blank"> <picture> <source media="(prefers-color-scheme: dark)" srcset="https://static.devin.ai/assets/gh-open-in-devin-review-dark.svg?v=1"> <img src="https://static.devin.ai/assets/gh-open-in-devin-review-light.svg?v=1" alt="Open with Devin"> </picture> </a> <!-- devin-review-badge-end -->
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Adds VLM sampling support to
RemoteInferenceClientsample endpoint. Finalizes inference side changes for #1200._render_for_sampleto handle both text-only and image-containing prompts. For text-only prompts, it flattens chunk tokens directly. When images are present, it calls/v1/chat/completions/renderto process images, then splices placeholder tokens into the pre-tokenized text stream with adjusted offsets.sample()to pass multi-modal features through to the generate payload when present.Test plan
None)test_vlm_inference_generation.py