fix: align visual-embedding feature selection with LLaVA semantics

[kaiitunnz]

Purpose

The visual-embedding path built LLaVA image embeddings by indexing the vision tower's hidden states directly, diverging from transformers' own LlavaModel.get_image_features: it never dropped the leading CLS token under the default vision-feature-select strategy, and it could not handle a list-valued vision_feature_layer. This aligns the selection with LLaVA semantics so the emitted embeddings match the model's canonical image features.

This change builds on #112 (merged), which relocated vision_tower / multi_modal_projector under self._model.model; the divergence is only reachable once that relocation is in place.

Changes

• src/worker/executors/transformers_executor.py — add _select_vision_features, mirroring get_image_features (CLS crop on the default strategy; per-layer gather + last-dim concat for a list-valued vision_feature_layer), and route the visual-embedding block through it.
• tests/worker/test_transformers_visual_embedding_features.py — cover int/list feature layer against default/full select strategy.

Design

Replicates only the feature-selection logic rather than calling base_model.get_image_features, which returns a BaseModelOutputWithPooling with image_sizes-driven splitting — a different shape and artifact contract than the current tensor path. The grouping/split and saved visual_embeddings.pt format are unchanged.

Test Plan

uv run pre-commit run --all-files
uv run pytest tests/ --ignore=tests/worker/test_mp_executor_cleanup_gpu.py

An isolated local-stack E2E builds current-branch server and GPU-worker images, brings up the stack with a single GPU worker, and submits a LLaVA visual-embedding workflow (llava-hf/llava-1.5-7b-hf) over one image. It waits for the task to reach a terminal state, downloads the resulting visual_embeddings.pt artifact, and inspects the embedding tensor to confirm the leading CLS token was dropped under the default strategy.

Test Result

• uv run pre-commit run --all-files all passed.
• uv run pytest tests/ --ignore=tests/worker/test_mp_executor_cleanup_gpu.py passed: 1416 tests, including the 4 new selection tests.
• The local-stack E2E passed all 5 assertions: the GPU worker registered, the visual-embedding task was submitted and reached DONE, the visual_embeddings.pt artifact was downloadable, and the first embedding tensor had shape (1, 576, 4096) — sequence length 576, confirming the CLS token was cropped (for llava-1.5 the CLIP tower emits 576 patches plus one CLS token, so an un-cropped run would report 577).

---

Pre-submission Checklist

• I have read the contribution guidelines.
• I have run pre-commit run --all-files and fixed any issues.
• I have added or updated tests covering my changes (if applicable).
• I have verified that uv run pytest tests/ passes locally.
• If I changed shared schemas or proto definitions, I have checked downstream compatibility across Server and Worker.
• If I changed the SDK or CLI, I have verified the affected packages work (uv sync --all-packages --group ci --frozen).
• If this is a breaking change, I have prefixed the PR title with [BREAKING] and described migration steps above.
• I have updated documentation or config examples if user-facing behavior changed.

[timzsu]

LGTM.