feat: add llm_transformers_manual_vision_deepseek_ocr and deploy deepseek_ocr vision bot with fastapi_webrtc_vision_bot_serve

[weedge]

• DeepEncoder ( ImageEncoderViT (sam) + a 16× token compressor + VitModel(clip) ) (有个想法是否可以使用这部分用于RAG的embedding模型，将文本记忆体的信息以图片encoder进行压缩(损失在可接受范围，本身长记忆体也存在信息损失的情况))
• DeepSeek-3B-MoE decoder (DeepseekV2Decoder MOE-A570M)

---

AI Podcast:

• DeepSeek-OCR：开启长上下文光学压缩新纪元: https://podcast-997.pages.dev/podcast/a39bf3b1332e490ab4e619ea6424eb2f

---

feat:

• add add deepseek-OCR transformers/vllm cases on modal

modal run src/download_models.py --repo-ids "deepseek-ai/DeepSeek-OCR" --revision "refs/pr/23"

# HF transformers
IMAGE_GPU=L4 modal run src/llm/transformers/vlm/ocr_deepseek.py --task dump_model
IMAGE_GPU=L4 modal run src/llm/transformers/vlm/ocr_deepseek.py --task infer
IMAGE_GPU=L4 modal run src/llm/transformers/vlm/ocr_deepseek.py --task infer_filter
BACKEND=achatbot IMAGE_GPU=L4 modal run src/llm/transformers/vlm/ocr_deepseek.py --task achatbot_infer

# vllm
IMAGE_GPU=L40s modal run src/llm/vllm/vlm/ocr_deepseek.py --task stream_infer

• add llm_transformers_manual_vision_deepseek_ocr
• deploy deepseek_ocr vision ocr bot with fastapi_webrtc_vision_bot_serve

# deepseek single room bot
modal run src/download_models.py --repo-ids "FunAudioLLM/SenseVoiceSmall"
modal run src/download_models.py --repo-ids "deepseek-ai/DeepSeek-OCR" --revision "refs/pr/23"
modal volume put config ./config/bots/daily_ocr_vision_bot.json /bots/ -f
EXTRA_INDEX_URL=https://pypi.org/simple/ \
    SERVE_TYPE=room_bot \
    CONFIG_FILE=/root/.achatbot/config/bots/daily_ocr_vision_bot.json \
    ACHATBOT_VERSION=0.0.28 \
    IMAGE_NAME=deepseek_ocr IMAGE_CONCURRENT_CN=1 IMAGE_GPU=L4 \
    modal serve src/fastapi_webrtc_vision_bot_serve.py

# run DailyOCRVisionBot with config
curl -XPOST "https://weedge--fastapi-webrtc-vision-deepseek-ocr-bot-srv-app-dev.modal.run/bot_join/chat-room/DailyOCRVisionBot"

daily_ocr_vision_bot.json

{
  "chat_bot_name": "DailyOCRVisionBot",
  "room_name": "chat-room",
  "room_url": "",
  "token": "",
  "services": {},
  "config": {
    "vad": {
      "tag": "silero_vad_analyzer",
      "args": {
        "start_secs": 0.0,
        "stop_secs": 0.0,
        "confidence": 0.7,
        "min_volume": 0.6,
        "onnx": true
      }
    },
    "asr": {
      "tag": "sense_voice_asr",
      "args": {
        "language": "zn",
        "model_name_or_path": "/root/.achatbot/models/FunAudioLLM/SenseVoiceSmall"
      }
    },
    "vision_ocr": {
      "tag": "llm_transformers_manual_vision_deepseek_ocr",
      "args": {
        "lm_model_name_or_path": "/root/.achatbot/models/deepseek-ai/DeepSeek-OCR",
        "lm_device": "cuda",
        "ocr_base_size":1024,
        "ocr_image_size":640,
        "ocr_crop_mode":true
      }
    },
    "tts": {
      "tag": "tts_edge",
      "args": {
        "voice_name": "zh-CN-YunjianNeural",
        "language": "zh",
        "gender": "Male"
      }
    }
  },
  "config_list": []
}

---

reference

• ⭐️⭐️ 2021.03 Learning Transferable Visual Models From Natural Language Supervision | paper code (Embedding)

• VITDET: Exploring Plain Vision Transformer Backbones for Object Detection

• ⭐️⭐️ 2023.04 Segment Anything | paper code (Tokenizer)

• https://www.youtube.com/watch?v=AhkMKIQsBrw (非常好高效讲解)

• 2024.09 General OCR Theory: Towards OCR-2.0 via a Unified End-to-end Model | paper code

• 2024.12 DeepSeek-VL2: Mixture-of-Experts Vision-Language Models for Advanced Multimodal Understanding | paper code

• 2025.10 DeepSeek-OCR: Contexts Optical Compression | paper code

• 2024.12 OmniDocBench: Benchmarking Diverse PDF Document Parsing with Comprehensive Annotations | paper code

[gemini-code-assist]

Summary of Changes

Hello @weedge, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request significantly enhances the platform's optical character recognition capabilities by integrating the DeepSeek-OCR model. It provides both Hugging Face Transformers and vLLM backends for the model and deploys a new vision bot that utilizes this advanced OCR technology. The changes also include necessary dependency updates and a refactoring of the core OCR interface to streamline functionality.

Highlights

• DeepSeek-OCR Integration: Added comprehensive support for the DeepSeek-OCR model, including both Hugging Face Transformers and vLLM implementations, enabling advanced Optical Character Recognition capabilities.
• New Vision Bot Deployment: Introduced a new vision bot that leverages DeepSeek-OCR, deployed via fastapi_webrtc_vision_bot_serve, expanding the platform's multimodal bot offerings.
• Dependency and Version Updates: Updated the achatbot project version to 0.0.28 and refined various dependencies, including accelerate and onnxruntime, to support the new OCR functionalities.
• Refactored OCR Interface: The IVisionOCR interface was updated to consolidate streaming and generation logic into a single generate method, improving consistency and simplifying future implementations.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request introduces support for the DeepSeek-OCR model, adding implementations for both standard transformers and vLLM, along with Modal deployment configurations. The changes are substantial, including new Modal applications, a new OCR processor, and updates to dependencies and configurations. My review has identified a critical security vulnerability related to the use of eval() on model outputs, which must be addressed. I've also highlighted several areas for improvement in correctness, robustness, and adherence to best practices, such as unsafe path manipulations, potential runtime errors, and insecure use of subprocess. Addressing this feedback will significantly enhance the security and maintainability of the new functionality.

[weedge]

DeepSeek-OCR 3336.10624 M parameters

DeepseekOCRForCausalLM(
  (model): DeepseekOCRModel(
    (embed_tokens): Embedding(129280, 1280)
    (layers): ModuleList(
      (0): DeepseekV2DecoderLayer(
        (self_attn): LlamaFlashAttention2(
          (q_proj): Linear(in_features=1280, out_features=1280, bias=False)
          (k_proj): Linear(in_features=1280, out_features=1280, bias=False)
          (v_proj): Linear(in_features=1280, out_features=1280, bias=False)
          (o_proj): Linear(in_features=1280, out_features=1280, bias=False)
          (rotary_emb): LlamaRotaryEmbedding()
        )
        (mlp): DeepseekV2MLP(
          (gate_proj): Linear(in_features=1280, out_features=6848, bias=False)
          (up_proj): Linear(in_features=1280, out_features=6848, bias=False)
          (down_proj): Linear(in_features=6848, out_features=1280, bias=False)
          (act_fn): SiLU()
        )
        (input_layernorm): DeepseekV2RMSNorm()
        (post_attention_layernorm): DeepseekV2RMSNorm()
      )
      (1-11): 11 x DeepseekV2DecoderLayer(
        (self_attn): LlamaFlashAttention2(
          (q_proj): Linear(in_features=1280, out_features=1280, bias=False)
          (k_proj): Linear(in_features=1280, out_features=1280, bias=False)
          (v_proj): Linear(in_features=1280, out_features=1280, bias=False)
          (o_proj): Linear(in_features=1280, out_features=1280, bias=False)
          (rotary_emb): LlamaRotaryEmbedding()
        )
        (mlp): DeepseekV2MoE(
          (experts): ModuleList(
            (0-63): 64 x DeepseekV2MLP(
              (gate_proj): Linear(in_features=1280, out_features=896, bias=False)
              (up_proj): Linear(in_features=1280, out_features=896, bias=False)
              (down_proj): Linear(in_features=896, out_features=1280, bias=False)
              (act_fn): SiLU()
            )
          )
          (gate): MoEGate()
          (shared_experts): DeepseekV2MLP(
            (gate_proj): Linear(in_features=1280, out_features=1792, bias=False)
            (up_proj): Linear(in_features=1280, out_features=1792, bias=False)
            (down_proj): Linear(in_features=1792, out_features=1280, bias=False)
            (act_fn): SiLU()
          )
        )
        (input_layernorm): DeepseekV2RMSNorm()
        (post_attention_layernorm): DeepseekV2RMSNorm()
      )
    )
    (norm): DeepseekV2RMSNorm()
    (sam_model): ImageEncoderViT(
      (patch_embed): PatchEmbed(
        (proj): Conv2d(3, 768, kernel_size=(16, 16), stride=(16, 16))
      )
      (blocks): ModuleList(
        (0-11): 12 x Block(
          (norm1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
          (attn): Attention(
            (qkv): Linear(in_features=768, out_features=2304, bias=True)
            (proj): Linear(in_features=768, out_features=768, bias=True)
          )
          (norm2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
          (mlp): MLPBlock(
            (lin1): Linear(in_features=768, out_features=3072, bias=True)
            (lin2): Linear(in_features=3072, out_features=768, bias=True)
            (act): GELU(approximate='none')
          )
        )
      )
      (neck): Sequential(
        (0): Conv2d(768, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (1): LayerNorm2d()
        (2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (3): LayerNorm2d()
      )
      (net_2): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
      (net_3): Conv2d(512, 1024, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
    )
    (vision_model): VitModel(
      (embeddings): CLIPVisionEmbeddings(
        (patch_embedding): Conv2d(3, 1024, kernel_size=(14, 14), stride=(14, 14), bias=False)
        (position_embedding): Embedding(257, 1024)
      )
      (transformer): NoTPTransformer(
        (layers): ModuleList(
          (0-23): 24 x NoTPTransformerBlock(
            (self_attn): NoTPAttention(
              (qkv_proj): Linear(in_features=1024, out_features=3072, bias=True)
              (out_proj): Linear(in_features=1024, out_features=1024, bias=True)
            )
            (mlp): NoTPFeedForward(
              (fc1): Linear(in_features=1024, out_features=4096, bias=True)
              (fc2): Linear(in_features=4096, out_features=1024, bias=True)
            )
            (layer_norm1): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
            (layer_norm2): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
          )
        )
      )
      (pre_layrnorm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
    )
    (projector): MlpProjector(
      (layers): Linear(in_features=2048, out_features=1280, bias=True)
    )
  )
  (lm_head): Linear(in_features=1280, out_features=129280, bias=False)
)