Summary

• Adds safety bounds to dynamic_preprocess to prevent out-of-memory (OOM) errors when processing images with extreme aspect ratios or when max_num is set to very high values.
• Caps max_num at MAX_PATCHES_LIMIT (24) to prevent runaway patch counts that lead to excessive memory allocation.
• Filters out target patch grid ratios with extreme aspect ratios (beyond MAX_ASPECT_RATIO_THRESHOLD of 200) to avoid degenerate patch grids like (12, 1) that produce large intermediate images without useful visual information.
• Adds a safety fallback to (1, 1) grid if all candidate ratios are filtered out.
• Applies the fix consistently across all three copies of the function in internvl_chat, internvl_chat_gpt_oss, and streamlit_demo.

Closes #1221

Details

When dynamic_preprocess receives an image with extreme proportions (e.g., a very wide panorama or a very tall screenshot), the generated target ratios can include highly elongated grids. Combined with large max_num values that can be configured via training metadata (max_dynamic_patch), this causes:

1. Large intermediate resize buffers (e.g., 5376x448 pixels for a (12, 1) grid with image_size=448)
2. Many patch tensors being created simultaneously
3. OOM crashes, especially on GPU-constrained environments

The fix introduces two configurable module-level constants:

• MAX_PATCHES_LIMIT = 24: Hard upper bound on patch count regardless of max_num parameter
• MAX_ASPECT_RATIO_THRESHOLD = 200: Maximum allowed ratio between grid dimensions

These values are intentionally generous to avoid affecting normal usage while preventing pathological cases.

Test plan

• Verify normal images (standard aspect ratios like 4:3, 16:9) produce identical results before and after the change
• Test with extreme aspect ratio images (e.g., 10000x100, 100x10000) to confirm no OOM
• Test with max_num values > 24 to confirm capping works correctly
• Verify the fallback to (1, 1) works when all ratios are filtered
• Run existing eval scripts (e.g., MME, VQA) to confirm no regression