Document LightOnOCR-2 vLLM nightly regression

- Revert to using nightly wheels (required for LightOnOCR-2 support)
- Add known issue note about image processor loading failure
- Update CLAUDE.md with detailed status and test commands
- vLLM nightly broke between v0.15.0rc2.dev73 and dev81

The script was working before the nightly update. Error:
"OSError: Can't load image processor for lightonai/LightOnOCR-2-1B"

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

CLAUDE.md

@@ -0,0 +1,145 @@
+# OCR Scripts - Development Notes
+
+## Active Scripts
+
+### DeepSeek-OCR v1 (`deepseek-ocr-vllm.py`)
+✅ **Production Ready**
+- Fully supported by vLLM
+- Fast batch processing
+- Tested and working on HF Jobs
+
+### LightOnOCR-2-1B (`lighton-ocr2.py`)
+⚠️ **Temporarily Broken** (2026-01-29)
+
+**Status:** vLLM nightly regression - image processor loading fails
+
+**What happened:**
+- Script was working with vLLM nightly `v0.15.0rc2.dev73`
+- Nightly updated to `v0.15.0rc2.dev81` and broke
+- Error: `OSError: Can't load image processor for 'lightonai/LightOnOCR-2-1B'`
+- Both nightly and stable vLLM 0.14.x have this issue now
+
+**Initial test results (before breakage):**
+- 8/10 samples had good OCR output
+- 2/10 samples showed repetition loops (likely due to max_tokens=6144)
+- Changed max_tokens default from 6144 → 4096 (per model card recommendation)
+
+**Fixes applied:**
+- `max_tokens`: 6144 → 4096 (model card recommends 4096 for arXiv papers)
+- Fixed pyarrow compatibility (>=17.0.0,<18.0.0)
+- Replaced deprecated `huggingface-hub[hf_transfer]` with `hf-xet`
+
+**To verify when vLLM is fixed:**
+```bash
+hf jobs uv run --flavor a100-large \
+    -s HF_TOKEN \
+    https://huggingface.co/datasets/uv-scripts/ocr/raw/main/lighton-ocr2.py \
+    davanstrien/ufo-ColPali davanstrien/lighton-ocr2-test-v3 \
+    --max-samples 10 --shuffle --seed 42
+```
+
+**Model Info:**
+- Model: `lightonai/LightOnOCR-2-1B`
+- Architecture: Pixtral ViT encoder + Qwen3 LLM
+- Training: RLVR (Reinforcement Learning with Verifiable Rewards)
+- Performance: 83.2% on OlmOCR-Bench, 42.8 pages/sec on H100
+
+## Pending Development
+
+### DeepSeek-OCR-2 (Visual Causal Flow Architecture)
+
+**Status:** ⏳ Waiting for vLLM upstream support
+
+**Context:**
+DeepSeek-OCR-2 is the next generation OCR model (3B parameters) with Visual Causal Flow architecture offering improved quality. We attempted to create a UV script (`deepseek-ocr2-vllm.py`) but encountered a blocker.
+
+**Blocker:**
+vLLM does not yet support `DeepseekOCR2ForCausalLM` architecture in the official release.
+
+**PR to Watch:**
+🔗 https://github.com/vllm-project/vllm/pull/33165
+
+This PR adds DeepSeek-OCR-2 support but is currently:
+- ⚠️ **Open** (not merged)
+- Has unresolved review comments
+- Pre-commit checks failing
+- Issues: hardcoded parameters, device mismatch bugs, missing error handling
+
+**What's Needed:**
+1. PR #33165 needs to be reviewed, fixed, and merged
+2. vLLM needs to release a version including the merge
+3. Then we can add these dependencies to our script:
+   ```python
+   # dependencies = [
+   #     "datasets>=4.0.0",
+   #     "huggingface-hub",
+   #     "pillow",
+   #     "vllm",
+   #     "tqdm",
+   #     "toolz",
+   #     "torch",
+   #     "addict",
+   #     "matplotlib",
+   # ]
+   ```
+
+**Implementation Progress:**
+- ✅ Created `deepseek-ocr2-vllm.py` script
+- ✅ Fixed dependency issues (pyarrow, datasets>=4.0.0)
+- ✅ Tested script structure on HF Jobs
+- ❌ Blocked: vLLM doesn't recognize architecture
+
+**Partial Implementation:**
+The file `deepseek-ocr2-vllm.py` exists in this repo but is **not functional** until vLLM support lands. Consider it a draft.
+
+**Testing Evidence:**
+When we ran on HF Jobs, we got:
+```
+ValidationError: Model architectures ['DeepseekOCR2ForCausalLM'] are not supported for now.
+Supported architectures: [...'DeepseekOCRForCausalLM'...]
+```
+
+**Next Steps (when PR merges):**
+1. Update `deepseek-ocr2-vllm.py` dependencies to include `addict` and `matplotlib`
+2. Test on HF Jobs with small dataset (10 samples)
+3. Verify output quality
+4. Update README.md with DeepSeek-OCR-2 section
+5. Document v1 vs v2 differences
+
+**Alternative Approaches (if urgent):**
+- Create transformers-based script (slower, no vLLM batching)
+- Use DeepSeek's official repo setup (complex, not UV-script compatible)
+
+**Model Information:**
+- Model ID: `deepseek-ai/DeepSeek-OCR-2`
+- Model Card: https://huggingface.co/deepseek-ai/DeepSeek-OCR-2
+- GitHub: https://github.com/deepseek-ai/DeepSeek-OCR-2
+- Parameters: 3B
+- Resolution: (0-6)×768×768 + 1×1024×1024 patches
+- Key improvement: Visual Causal Flow architecture
+
+**Resolution Modes (for v2):**
+```python
+RESOLUTION_MODES = {
+    "tiny": {"base_size": 512, "image_size": 512, "crop_mode": False},
+    "small": {"base_size": 640, "image_size": 640, "crop_mode": False},
+    "base": {"base_size": 1024, "image_size": 768, "crop_mode": False},  # v2 optimized
+    "large": {"base_size": 1280, "image_size": 1024, "crop_mode": False},
+    "gundam": {"base_size": 1024, "image_size": 768, "crop_mode": True},  # v2 optimized
+}
+```
+
+## Other OCR Scripts
+
+### Nanonets OCR (`nanonets-ocr.py`, `nanonets-ocr2.py`)
+✅ Both versions working
+
+### PaddleOCR-VL (`paddleocr-vl.py`)
+✅ Working
+
+---
+
+**Last Updated:** 2026-01-29
+**Watch PRs:**
+- DeepSeek-OCR-2: https://github.com/vllm-project/vllm/pull/33165
+- LightOnOCR-2 regression: Check https://github.com/vllm-project/vllm/issues?q=LightOnOCR

lighton-ocr2.py

@@ -6,11 +6,18 @@
 #     "huggingface-hub",
 #     "hf-xet",
 #     "pillow",
-#     "vllm>=0.11.1",
+#     "vllm",
 #     "tqdm",
 #     "toolz",
 #     "torch",
+#     "triton-kernels @ git+https://github.com/triton-lang/triton.git@v3.5.0#subdirectory=python/triton_kernels",
 # ]
+#
+# [[tool.uv.index]]
+# url = "https://wheels.vllm.ai/nightly"
+#
+# [tool.uv]
+# prerelease = "allow"
 # ///
 
 """
@@ -23,6 +30,10 @@ Uses Reinforcement Learning with Verifiable Rewards (RLVR) for improved quality.
 NOTE: Requires vLLM nightly wheels for LightOnOCR-2 support. First run may take
 a few minutes to download and install dependencies.
 
+KNOWN ISSUE (2026-01-29): vLLM nightly may have a regression causing
+"Can't load image processor" errors. If this occurs, try again later
+or check vLLM issues for updates.
+
 Features:
 - ⚡ Fastest: 42.8 pages/sec on H100 GPU (7× faster than v1)
 - 🎯 High accuracy: 83.2 ± 0.9% on OlmOCR-Bench (+7.1% vs v1)