Add Qwen3-VL export support for multimodal text-to-text pipeline

[seyeong-han]

Overview

Enables exporting Qwen3-VL vision-language models through the multimodal-text-to-text task. Qwen3-VL uses M-RoPE (Multi-dimensional Rotary Position Embeddings) and a Conv3d-based vision encoder, both of which require special handling during torch.export.

Changes

Three changes to the export pipeline:

1. M-RoPE vision encoder positions — The Qwen3-VL visual encoder computes position embeddings via data-dependent ops (torch.linspace, repeat_interleave on image_grid_thw) that torch.export cannot trace. VisionExportableModule now pre-computes pos_embeds, rotary_pos_emb, and cu_seqlens eagerly and stores them as buffers so they become constants in the exported graph.

2. M-RoPE text decoder hook — During text decoder export only inputs_embeds and cache_position are provided (no input_ids). M-RoPE models call get_rope_index which requires input_ids and crashes. A forward pre-hook injects position_ids derived from cache_position so the model skips that code path.

3. Model loading / modality detection — AutoModelForPreTraining doesn't resolve Qwen3-VL, so we fall back to AutoModelForImageTextToText. Modality detection now handles models that report more than two modalities (Qwen3-VL reports ("image", "video", "text")) by picking the first supported non-text modality.

optimum-cli export executorch \
  --model "Qwen/Qwen3-VL-2B-Instruct" \
  --task "multimodal-text-to-text" \
  --recipe "xnnpack" \
  --use_custom_sdpa \
  --use_custom_kv_cache \
  --qlinear "8da4w" \
  --qlinear_group_size 32 \
  --qlinear_encoder "8da4w,8da8w" \
  --qlinear_encoder_group_size 32 \
  --qembedding "8w" \
  --qembedding_encoder "8w" \
  --dtype "float32" \
  --output_dir="qwen3/Qwen3-VL-2B-Instruct-xnnpack"

Quantized model is ~1.4 GB (down from ~4.4 GB bf16).
Text decoder runs at ~25-29 tok/s on Apple Silicon via XNNPACK.