Add torchvision (Qwen2-VL video sub-processor dep); switch default reasoning model to E4B (works on ZeroGPU CUDA emulation)

README.md

@@ -81,12 +81,12 @@ Alternative reasoning models — set `REASONING_MODEL_ID` to switch:
 
 | `REASONING_MODEL_ID` override | Approx VRAM | Recommended slice |
 |---|---|---|
-| `nvidia/Gemma-4-26B-A4B-NVFP4` *(default)* | ~16 GB (NVFP4) | `large` (48 GB) |
+| `google/gemma-4-E4B-it` *(default)* | ~8 GB text / ~16 GB multimodal (bf16) | `large` (48 GB) — places cleanly through ZeroGPU CUDA emulation |
 | `google/gemma-4-26B-A4B-it` | ~52 GB (bf16) | `xlarge` (96 GB) |
 | `google/gemma-4-31B-it` | ~62 GB (bf16) | `xlarge` |
-| `google/gemma-4-E4B-it` | ~8 GB (bf16) | `large` — smaller / faster |
+| `RedHatAI/gemma-4-26B-A4B-it-NVFP4` | ~16 GB (compressed-tensors NVFP4) | `large` — but module-level NVFP4 unpacking hangs through CUDA emulation; needs lazy-load fix |
 | `openai/gpt-oss-20b` | ~16 GB (MXFP4) | `large` — needs Hopper+ |
-| `RedHatAI/gemma-4-26B-A4B-it-NVFP4` | ~16 GB (NVFP4) | `large` — community quant |
+| `nvidia/Gemma-4-26B-A4B-NVFP4` | ~16 GB (modelopt) | requires `nvidia-modelopt`, not native to transformers |
 
 ## Configuration
 

legal_doc_redteam/reasoning_review.py

@@ -22,7 +22,7 @@ from __future__ import annotations
 import json
 from typing import Any, Callable
 
-DEFAULT_REASONING_MODEL = "nvidia/Gemma-4-26B-A4B-NVFP4"
+DEFAULT_REASONING_MODEL = "google/gemma-4-E4B-it"
 
 SYSTEM_INSTRUCTIONS = (
     "You are the Document Integrity Verifier — a defensive auditor that "

requirements.txt

@@ -9,6 +9,10 @@ accelerate>=0.34
 kernels>=0.4,<0.15
 compressed-tensors>=0.7
 torch>=2.8
+# Qwen2-VL family (Nanonets-OCR-s) auto-loads a video sub-processor that
+# requires torchvision even if we only feed it images. Without it, the VLM
+# OCR backend errors out at AutoProcessor.from_pretrained import time.
+torchvision
 rapidocr-onnxruntime>=1.4
 onnxruntime>=1.18
 beautifulsoup4>=4.12