models/qwen3_vl_compat.py

"""
Helpers for Qwen3-VL (transformers): vision lives on model.model.visual,
and ViT+merger returns BaseModelOutputWithDeepstackFeatures.pooler_output.
"""

from __future__ import annotations

import torch


def get_visual_module(model: torch.nn.Module):
    inner = getattr(model, "model", None)
    if inner is not None and hasattr(inner, "visual"):
        return inner.visual
    return model.visual


def set_visual_module(model: torch.nn.Module, value) -> None:
    inner = getattr(model, "model", None)
    if inner is not None and hasattr(inner, "visual"):
        inner.visual = value
        return
    model.visual = value


def merged_spatial_grid(image_grid_thw: torch.Tensor, spatial_merge_size: int) -> torch.Tensor:
    """Processor grid (T,H,W) -> LLM-side grid after spatial merge on H,W."""
    g = image_grid_thw.clone()
    g[:, 1] = g[:, 1] // spatial_merge_size
    g[:, 2] = g[:, 2] // spatial_merge_size
    return g


def pack_image_features(model, pixel_values: torch.Tensor, image_grid_thw: torch.Tensor):
    """
    Returns flat merger embeddings [total_tokens, D] and LLM grid [num_imgs, 3]
    consistent with OpticalCompressor token counts.
    """
    inner = model.model if hasattr(model, "model") else model
    out = inner.get_image_features(pixel_values, image_grid_thw)
    parts = out.pooler_output
    if isinstance(parts, (list, tuple)):
        packed = torch.cat(parts, dim=0)
    else:
        packed = parts
    vis = get_visual_module(model)
    grid_llm = merged_spatial_grid(image_grid_thw, vis.spatial_merge_size)
    return packed, grid_llm