--- title: MiniCPM-OCR-Jev-BF16-iluvatar-FlagOS canonical_url: "https://www.modelscope.cn/models/FlagRelease/MiniCPM-OCR-Jev-BF16-iluvatar-FlagOS" md_url: "https://www.modelscope.cn/models/FlagRelease/MiniCPM-OCR-Jev-BF16-iluvatar-FlagOS.md" repository: FlagRelease/MiniCPM-OCR-Jev-BF16-iluvatar-FlagOS last_updated: 2026-10-28 license: apache-2.0 model_type: - qwen3_5 architectures: - Qwen3_5ForConditionalGeneration library_name: - pytorch - transformer - safetensors language: - zh - en downloads: 1 stars: 0 --- # MiniCPM-OCR-Jev-BF16-iluvatar-FlagOS > MiniCPM-OCR-Jev-BF16-iluvatar-FlagOS - FlagRelease 在 ModelScope 开源的模型。Introduction ModelBest, together with OpenBMB, has open-sourced the MiniCPM-OCR series for document parsing and text recognition, comprising two models: MiniCPM-OCR 1.0 (1.3B) and… FlagRelease/MiniCPM-OCR-Jev-BF16-iluvatar-FlagOS 是 ModelScope 魔搭社区上的机器学习模型，采用 apache-2.0 许可。 - **Repository**: FlagRelease/MiniCPM-OCR-Jev-BF16-iluvatar-FlagOS - **License**: apache-2.0 - **Downloads**: 1 - **Stars**: 0 - **Last updated**: 2026-10-28 Source: https://www.modelscope.cn/models/FlagRelease/MiniCPM-OCR-Jev-BF16-iluvatar-FlagOS --- # Introduction ModelBest, together with OpenBMB, has open-sourced the MiniCPM-OCR series for document parsing and text recognition, comprising two models: MiniCPM-OCR 1.0 (1.3B) and MiniCPM-OCR Jev (2B). Built on a unified multi-chip software stack, the FlagOS community provides Day-0 cross-chip adaptation for both models, covering 12 platforms in total: T-Head, NVIDIA, Huawei Ascend, Hygon, KUNLUNXIN, MetaX, Moore Threads, Tsingmicro, Iluvatar CoreX, Xiwang, Mac (ARM), and CIX. In this round of evaluation, the FlagOS build of Jev reached 98.7% of NVIDIA native compute throughput, while on the same platform OCR 1.0 delivered up to 1.69× native P50 throughput per unit of peak compute. On ARM edge devices, Jev completes a page-block assessment in an average of 1.20 seconds using the M5 CPU alone. Backed by FlagOS support for both multi-chip servers and ARM terminals, industry users can pick a document parsing and quality-inspection deployment that matches their business requirements and existing compute resources. MiniCPM-OCR 1.0 targets real-world document scenarios — scans, photos, curved pages, and re-photographed screens — with support for text, formula, and complex table parsing as well as key information extraction. In official evaluations, this lightweight 1.3B-parameter model scored 93.68 on Real5-OmniDocBench, and with DSpark decoding acceleration its end-to-end processing efficiency improved by 55%. To address the problem that parsed output still needs manual verification, the 2B-parameter MiniCPM-OCR Jev acts as an "AI quality inspector" for documents: it scores the reliability of parsing results and can be plugged into different OCR models to help flag content that requires human review. Its overall AUROC on JevOCRBench reaches 0.936, close to the 0.938 of Qwen3.5-397B-A17B. ### Integrated Deployment - Out-of-the-box inference scripts with pre-configured hardware and software parameters - Released **FlagOS-Iluvatar** container image supporting deployment within minutes ### Consistency Validation - Rigorously evaluated through benchmark testing: Performance and results from the FlagOS software stack are compared against native stacks on multiple public. # Evaluation Results ## Benchmark Result | Metrics | MiniCPM-OCR-Jev-Nvidia-Origin | MiniCPM-OCR-Jev-Iluvatar-FlagOS | |--------------|--------------------------------|--------------------------------------| | ocr_rm_mini(12 entries) avg_prob_1 | 0.4817 | 0.482352 | | omnidocbench-render-compare(60 entries) avg_prob_1 | 0.2397 | 0.241341 | # User Guide Environment Setup | Item | Version | |------------------|----------------------| | Docker Version | Docker version 28.1.1 | | Operating System | 22.04.4 LTS | | Python | 3.12 | | torch | 2.10.0+cpu | | torch_fl | 2.10.0+corex4.5.0 | | FlagTree | 0.6.2a3+iluvatar3.6 | | FlagGems | 5.4.0 | | FlagCX | not installed (not needed for a single card) | | transformers | 5.17.0 | | tokenizers | 0.23.1 | | corex | 4.5.0 | ## Operation Steps ### Download FlagOS Image ```bash docker pull harbor.baai.ac.cn/flagrelease-public/torch-fl-iluvatar:jev-ocr-rm-2b-20261023 ``` ### Download Open-source Model Weights ```bash pip install modelscope modelscope download --model FlagRelease/MiniCPM-OCR-Jev-BF16-iluvatar-FlagOS --local_dir /data2/models/Jev-OCR-RM-2B ``` ### Start the Container ```bash H=/usr/local/corex-4.5.0.20260509 docker run -d --name jev-ilu-tfl --privileged --ipc=host --net=host --shm-size 64g \ -v /dev:/dev -v /lib/modules:/lib/modules -v /sys:/sys -v /data2:/data2 \ -v $H:/opt/host-corex45:ro \ -v $H/lib64/libcuda.so.1:/usr/local/corex/lib64/libcuda.so.1:ro \ -v $H/lib64/libixml.so:/usr/local/corex/lib64/libixml.so:ro \ -v $H/bin/ixsmi:/usr/local/corex/bin/ixsmi:ro \ -e LD_LIBRARY_PATH=/opt/host-corex45/lib64:/usr/local/corex/lib64:/usr/local/openmpi/lib:/usr/local/lib \ --entrypoint /bin/bash \ harbor.baai.ac.cn/flagrelease-public/torch-fl-iluvatar:jev-ocr-rm-2b-20261023 \ -c 'sleep infinity' ``` ### Start the Server ```bash No need to start server. ``` ## Service Invocation ### Invocation Script ```bash docker exec -i jev-ilu-tfl bash -lc \ 'source /opt/torch-fl-iluvatar-venv/bin/activate && \ export LD_LIBRARY_PATH=/opt/host-corex45/lib64:/usr/local/corex/lib64:/usr/local/openmpi/lib:/usr/local/lib && \ DEVICE=flagos:0 python3 - s02' <<'PY' """Jev (OCR-RM-2B) @ Iluvatar BI-V150 -- single-sample scoring on Torch-FL (the `flagos` device), bypassing vendor torch. Key points: 1. **`import torch_fl` MUST come before `import torch`** (MetaX docs: CUDA / cu-bridge ABI compatibility) 2. The device string is **`flagos:0`**, not `cuda:0` 3. Build the model on CPU first, then `.to("flagos:0")`; transformers' `device_map=` does not accept the `flagos` device name Usage (`python3 -` reads this block from stdin; the arguments after `-` are sample names): python3 - s02 # run the default single sample s02 python3 - s01 s03 s12 # run the given samples """ import json import os import sys _SKIP_TFL = os.environ.get("SKIP_TFL") == "1" # 1 = plain-transformers baseline (do not import torch_fl) if not _SKIP_TFL: import torch_fl # noqa: F401 WARNING: must be first (MetaX docs: cu-bridge ABI) import torch from PIL import Image from transformers import AutoProcessor, Qwen3_5ForSequenceClassification CKPT = os.environ.get("CKPT", "/data2/models/Jev-OCR-RM-2B/qwen35_2b_l2_triple_value_checkpoint-3000") DEMO = os.environ.get("DEMO", "/data2/models/Jev-OCR-RM-2B/rm_hf_infer_demo_v1_20261015") DEV = os.environ.get("DEVICE", "flagos:0") # set DEVICE=cpu for a CPU baseline # NOTE: the prompt below is intentionally kept verbatim in Chinese -- it is the exact # text used to produce every number reported in this README, so translating it would # break reproducibility. PROMPT = ( "<|im_start|>user\n" "<|vision_start|><|image_pad|><|vision_end|>OCR文本：\n{ocr}\n" "<|vision_start|><|image_pad|><|vision_end|>" "第一张图是原图，第二张图是上述 OCR 文本渲染得到的图片。\n" "请判断该 OCR 文本及其对应渲染图是否与原图内容一致。\n" "这是一个二分类任务：一致标为 1，不一致标为 0。<|im_end|>\n" ) # --------------------------------------------------------------------------- # Torch-FL compatibility patch (**scoped to the get_rope_index call only**) # # Problem: `torch.tensor([<0-d tensor living on flagos>])` raises # TypeError: len() of a 0-d tensor # Minimal reproduction (3 lines): # import torch_fl, torch # torch.tensor([torch.tensor(3, device="flagos:0")], device="flagos:0") # FAIL # torch.tensor([torch.tensor(3)]) # OK (CPU 0-d is fine) # transformers' Qwen3.5-VL `get_rope_index()` happens to use exactly that pattern: # mrope_position_deltas.append(llm_positions.max() + 1 - len(current_input_ids)) # mrope_position_deltas = torch.tensor(mrope_position_deltas, device=input_ids.device) # # WARNING: **never patch torch.tensor globally**. Measured: a global patch drags the # same model on CPU from 0.9952 down to 0.49 (all values garbage) and is very hard to # debug. Here the patch is installed only while get_rope_index is executing. # --------------------------------------------------------------------------- _orig_tensor = torch.tensor def _tensor_compat(data=None, *args, **kwargs): if isinstance(data, (list, tuple)) and any(torch.is_tensor(x) for x in data): try: return _orig_tensor(data, *args, **kwargs) except TypeError as e: if "0-d tensor" not in str(e): raise data = [x.item() if torch.is_tensor(x) else x for x in data] return _orig_tensor(data, *args, **kwargs) def _install_rope_index_compat() -> list[str]: """Wrap every qwen3_5 class that defines get_rope_index; the compat shim is active only while that method runs.""" patched = [] try: import transformers.models.qwen3_5.modeling_qwen3_5 as q35 except Exception: return patched for cls in list(vars(q35).values()): if not isinstance(cls, type) or "get_rope_index" not in cls.__dict__: continue orig = cls.__dict__["get_rope_index"] def wrapper(self, *a, _orig=orig, **kw): prev = torch.tensor torch.tensor = _tensor_compat try: return _orig(self, *a, **kw) finally: torch.tensor = prev setattr(cls, "get_rope_index", wrapper) patched.append(cls.__name__) return patched DTYPE = getattr(torch, os.environ.get("DTYPE", "bfloat16")) def describe_backend() -> None: """Print hard evidence that we run on the device provided by Torch-FL and not on vendor torch.""" print("=" * 70) if _SKIP_TFL: print("[SKIP_TFL=1] plain-transformers baseline, torch_fl not imported") print("torch.__version__ :", torch.__version__) print("torch.version.cuda:", torch.version.cuda) print("torch.__file__ :", torch.__file__) print("=" * 70) return import torch_fl as tfl print("torch.__version__ :", torch.__version__) print("torch.version.cuda:", torch.version.cuda, " <- should be None under the Torch-FL CPU core") print("torch.__file__ :", torch.__file__) print("torch_fl.__file__ :", tfl.__file__) print("torch_fl version :", getattr(tfl, "__version__", "")) print("flagos available :", torch.flagos.is_available()) print("flagos devices :", torch.flagos.device_count()) try: props = torch.flagos.get_device_properties(0) print("flagos dev0 :", props.name, f"{props.total_memory / 1024**3:.1f} GiB") except Exception as e: # noqa: BLE001 print("flagos dev0 props : ", repr(e)[:160]) print("=" * 70) assert "corex" not in torch.__version__ and "ixml" not in torch.__version__, ( f"torch is still the vendor CoreX build ({torch.__version__}), not the Torch-FL CPU core" ) assert getattr(torch, "flagos", None) is not None, "torch has no flagos namespace" assert torch.flagos.is_available(), "flagos device is not available" def main() -> None: describe_backend() names = sys.argv[1:] or ["s02"] manifest = json.load(open(f"{DEMO}/samples/manifest.json"))["samples"] by_name = {s["name"]: s for s in manifest} proc = AutoProcessor.from_pretrained(CKPT) model = Qwen3_5ForSequenceClassification.from_pretrained( CKPT, dtype=DTYPE, attn_implementation="sdpa", device_map=None ) model = model.to(DEV).eval() # Known issue with Iluvatar + Torch-FL: without explicit synchronization the results # are **unstable** (the same input gives prob_1 = 0.994 on one run and 0.844 on # the next, occasionally NaN). Calling torch.flagos.synchronize() after every # layer makes it fully stable. Enable with SYNC_LAYERS=1 (on by default, because # it is a prerequisite for correctness). if os.environ.get("SYNC_LAYERS", "1") == "1" and hasattr(torch, "flagos"): def _sync_hook(mod, inp, out): torch.flagos.synchronize() for _n, _mod in model.named_modules(): if _n.rsplit(".", 1)[-1] in ("layers", "blocks"): for _idx, _child in _mod.named_children(): _child.register_forward_hook(_sync_hook) print("[info] SYNC_LAYERS=1: torch.flagos.synchronize() after every layer", flush=True) if os.environ.get("ROPE_PATCH", "1") == "1": print("[info] rope compat patched:", _install_rope_index_compat(), flush=True) print(f"[info] model on {DEV} dtype={DTYPE}", flush=True) for name in names: s = by_name[name] ocr = open(f"{DEMO}/{s['ocr_text_file']}", encoding="utf-8").read().rstrip("\r\n") images = [ Image.open(f"{DEMO}/{s['orig_image']}").convert("RGB"), Image.open(f"{DEMO}/{s['render_image']}").convert("RGB"), ] inputs = proc(text=[PROMPT.format(ocr=ocr)], images=images, return_tensors="pt") inputs = {k: v.to(DEV) for k, v in inputs.items() if hasattr(v, "to")} with torch.inference_mode(): logits = model(**inputs).logits.float() prob_1 = torch.softmax(logits, dim=-1)[0, 1].item() ref = s["ref_prob_1"] print( f"{name}: prob_1={prob_1:.8f} ref={ref:.8f} delta={prob_1 - ref:+.2e} " f"n_tokens={inputs['input_ids'].shape[1]}", flush=True, ) if __name__ == "__main__": main() PY ``` ### AnythingLLM Integration Guide #### 1. Download & Install - Visit the official site: https://anythingllm.com/ - Choose the appropriate version for your OS (Windows/macOS/Linux) - Follow the installation wizard to complete the setup #### 2. Configuration - Launch AnythingLLM - Open settings (bottom left, fourth tab) - Configure core LLM parameters - Click "Save Settings" to apply changes #### 3. Model Interaction - After model loading is complete: - Click **"New Conversation"** - Enter your question (e.g., "Explain the basics of quantum computing") - Click the send button to get a response # Technical Overview **FlagOS** is a fully open-source system software stack designed to unify the "model–system–chip" layers and foster an open, collaborative ecosystem. It enables a "develop once, run anywhere" workflow across diverse AI accelerators, unlocking hardware performance, eliminating fragmentation among vendor-specific software stacks, and substantially lowering the cost of porting and maintaining AI workloads. With core technologies such as the **FlagScale**, together with vllm-plugin-fl, distributed training/inference framework, **FlagGems** universal operator library, **FlagCX** communication library, and **FlagTree** unified compiler, the **FlagRelease** platform leverages the **FlagOS** stack to automatically produce and release various combinations of . This enables efficient and automated model migration across diverse chips, opening a new chapter for large model deployment and application. ## FlagGems FlagGems is a high-performance, generic operator libraryimplemented in [Triton](https://github.com/openai/triton) language. It is built on a collection of backend-neutralkernels that aims to accelerate LLM (Large-Language Models) training and inference across diverse hardware platforms. ## FlagTree FlagTree is an open source, unified compiler for multipleAI chips project dedicated to developing a diverse ecosystem of AI chip compilers and related tooling platforms, thereby fostering and strengthening the upstream and downstream Triton ecosystem. Currently in its initial phase, the project aims to maintain compatibility with existing adaptation solutions while unifying the codebase to rapidly implement single-repository multi-backend support. Forupstream model users, it provides unified compilation capabilities across multiple backends; for downstream chip manufacturers, it offers examples of Triton ecosystem integration. ## FlagScale and vllm-plugin-fl Flagscale is a comprehensive toolkit designed to supportthe entire lifecycle of large models. It builds on the strengths of several prominent open-source projects, including [Megatron-LM](https://github.com/NVIDIA/Megatron-LM) and [vLLM](https://github.com/vllm-project/vllm), to provide a robust, end-to-end solution for managing and scaling large models. vllm-plugin-fl is a vLLM plugin built on the FlagOS unified multi-chip backend, to help flagscale support multi-chip on vllm framework. ## **FlagCX** FlagCX is a scalable and adaptive cross-chip communication library. It serves as a platform where developers, researchers, and AI engineers can collaborate on various projects, contribute to the development of cutting-edge AI solutions, and share their work with the global community. ## **FlagEval Evaluation Framework** FlagEval is a comprehensive evaluation system and open platform for large models launched in 2023. It aims to establish scientific, fair, and open benchmarks, methodologies, and tools to help researchers assess model and training algorithm performance. It features: - **Multi-dimensional Evaluation**: Supports 800+ modelevaluations across NLP, CV, Audio, and Multimodal fields,covering 20+ downstream tasks including language understanding and image-text generation. - **Industry-Grade Use Cases**: Has completed horizonta1 evaluations of mainstream large models, providing authoritative benchmarks for chip-model performance validation. # Contributing We warmly welcome global developers to join us: 1. Submit Issues to report problems 2. Create Pull Requests to contribute code 3. Improve technical documentation 4. Expand hardware adaptation support # License The model weights are derived from OpenBMB/MiniCPM-OCR-Jev and are open‑sourced under the Apache License 2.0: https://www.apache.org/licenses/LICENSE-2.0.txt