[Bug]: Qwen3-VL-Embedding multimodal embedding model crashes on oMLX v0.3.8

[Footree]

Summary

Qwen3-VL-Embedding (multimodal vision-language embedding model) crashes during embedding generation on oMLX v0.3.8. The server application stays alive, but the model crashes internally, returning empty responses.

Environment

• oMLX version: v0.3.8
• Hardware: Mac Studio (Apple Silicon, M4 Max)
• Model: Qwen3-VL-Embedding (multimodal, supports text + image embeddings)
• Reference: https://huggingface.co/Qwen/Qwen3-VL-Embedding

Steps to Reproduce

1. Download Qwen3-VL-Embedding via oMLX
2. Send embedding request via OpenAI-compatible /v1/embeddings endpoint
3. Model crashes during embedding generation

Expected Behavior

Qwen3-VL-Embedding should generate embeddings (same as Qwen3-Embedding, which works perfectly).

Actual Behavior

curl: (18) transfer closed with outstanding read data remaining

• Response is empty
• Model crashes internally during embedding generation
• oMLX application process stays alive (health endpoint still responds)
• Only the embedding API endpoint returns errors

Comparison

Model	Type	Status
Qwen3-Embedding-8B-mxfp8	Text-only embedding	✅ Works perfectly
Qwen3-VL-Embedding	Multimodal VL embedding	❌ Crash on embed

Impact

Cannot use multimodal embeddings for applications that need to embed both text and images (e.g., visual memory search, image-text cross-modal retrieval). This blocks the upgrade path to VL-capable embedding models on oMLX.

Logs / Debug Info

• Swagger UI: transfer closed error
• curl: transfer closed with outstanding read data remaining
• No visible error in health endpoint — application remains healthy
• Suggests model-level crash inside the embedding pipeline, not a server-level crash