GLM-OCR README.md and concurrency modification (#549)

Co-authored-by: AlexKer <AlexKer@users.noreply.github.com>

Author: AlexKer

glm-ocr/README.md

@@ -0,0 +1,75 @@
+# GLM-OCR Truss Model
+
+This is a Truss deployment of the [GLM-OCR](https://huggingface.co/zai-org/GLM-OCR) model for optical character recognition using vLLM engine on Baseten served on an L4 GPU. With only 0.9B parameters, GLM-OCR delivers strong OCR performance while being lightweight enough for high-concurrency and edge deployments.
+
+GLM-OCR integrates the CogViT visual encoder, a lightweight cross-modal connector with efficient token downsampling, and a GLM-0.5B language decoder. It supports a two-stage pipeline (layout analysis + parallel recognition) for processing complex documents.
+
+## Quick Start
+
+### 1. Deploy to Baseten
+
+```bash
+# Clone this repo and cd into this folder
+git clone https://github.com/basetenlabs/truss-examples.git
+cd truss-examples/glm-ocr
+
+# Deploy the model
+truss push --publish
+# This assumes you have truss installed, if not follow the instructions here:
+# https://docs.baseten.co/development/model/build-your-first-model
+```
+
+### 2. Test with the OpenAI Client
+
+Replace the `api_key` and `base_url` in `test.py` with your specific deployment credentials and URL.
+
+```bash
+pip install openai
+python test.py
+```
+
+### 3. Project Structure
+
+```
+glm-ocr/
+├── config.yaml              # Truss configuration
+├── test.py                  # Test script (OpenAI client)
+└── README.md                # Documentation
+```
+
+## Model Information
+
+- **Model**: [zai-org/GLM-OCR](https://huggingface.co/zai-org/GLM-OCR)
+- **Parameters**: 0.9B
+- **Framework**: vLLM (OpenAI-compatible API)
+- **GPU**: L4 (24GB)
+- **API**: OpenAI Chat Completions (`/v1/chat/completions`)
+
+## Usage
+
+### Using the OpenAI Client
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-XXXX.api.baseten.co/deployment/YYYY/sync/v1"
+)
+
+response = client.chat.completions.create(
+    model="zai-org/GLM-OCR",
+    messages=[{
+        "role": "user",
+        "content": [
+            {"type": "image_url", "image_url": {"url": "https://example.com/document.png"}},
+            {"type": "text", "text": "Text Recognition:"}
+        ]
+    }],
+    max_tokens=4096
+)
+
+print(response.choices[0].message.content)
+```
+
+The model accepts images via URL or base64-encoded data URIs, and returns recognized text in markdown format.

glm-ocr/config.yaml

@@ -28,6 +28,6 @@ resources:
   accelerator: L4
   use_gpu: true
 runtime:
-  predict_concurrency: 128
+  predict_concurrency: 32
 secrets:
   hf_access_token: null