Add Qwen3.5 model support (VLM + hybrid GatedDeltaNet text model)

[Copilot]

Adds OpenVINO export and inference support for Qwen/Qwen3.5-0.8B — a visual language model with a hybrid text backbone combining GatedDeltaNet (linear attention) and standard full attention layers. Follows the patching patterns established in PR huggingface#1523 (Qwen3-Next CausalConv1D/GatedDeltaNet) and PR huggingface#1551 (Qwen3VL VLM structure).

Export pipeline (optimum/exporters/openvino/)

• _ov_ops.py (new): convert_recurrent_attention_cell — OpenVINO conversion rule that lowers the GatedDeltaNet recurrent loop into an OV Loop op via ModuleExtension
• model_configs.py: Qwen3_5TextOpenVINOConfig (registered for qwen3_5_text) with hybrid cache layout (conv + recurrent + KV), Qwen3_5OpenVINOConfig (registered for qwen3_5) splitting the VLM into vision patch embed / pos embed / merger / language components
• model_patcher.py: Qwen3_5ModelPatcher patches CausalConv1D via ov_causal_conv1d and GatedDeltaNet via RecurrentAttentionCell + ModuleExtension; handles both standalone text and VLM language-model contexts. Qwen3_5VisionEmbMergerPatcher replaces cu_seqlens with attention_mask for traceable vision blocks
• utils.py: qwen3_5_text → SSM_MODELS, qwen3_5 → MULTI_MODAL_TEXT_GENERATION_MODELS

Inference pipeline (optimum/intel/openvino/)

• modeling_visual_language.py: _OVQwen3_5ForCausalLM with vision embedding interpolation, rot_pos_emb, and rope index computation using mm_token_type_ids (Qwen3.5's token-type-based modality dispatch differs from Qwen2VL's token-ID-based approach)
• modeling_decoder.py: qwen3_5_text added to full-context attention mask list (same requirement as LFM2/GraniteMoeHybrid)

Key architectural differences from Qwen3VL

• No deepstack — vision merger output is a single tensor, not (hidden_states, deepstack_features)
• Text model uses Qwen3_5DynamicCache (hybrid conv/recurrent/KV) instead of standard DynamicCache
• get_rope_index requires mm_token_type_ids parameter (constructed from image_token_id/video_token_id)
• Text embeddings accessed via model.model.language_model.embed_tokens (not model.model.embed_tokens)

Warning

Firewall rules blocked me from connecting to one or more addresses (expand for details)

I tried to connect to the following addresses, but was blocked by firewall rules:

• huggingface.co
  • Triggering command: /home/REDACTED/work/_temp/ghcca-node/node/bin/node /home/REDACTED/work/_temp/ghcca-node/node/bin/node --enable-source-maps /home/REDACTED/work/_temp/copilot-developer-action-main/dist/index.js (dns block)

If you need me to access, download, or install something from one of these locations, you can either:

• Configure Actions setup steps to set up my environment, which run before the firewall is enabled
• Add the appropriate URLs or hosts to the custom allowlist in this repository's Copilot coding agent settings (admins only)

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.