sie_id: zai-org/GLM-OCR hf_id: zai-org/GLM-OCR hf_revision: ca5d8b3e287e52589e37c28385d9655ee4372f9d inputs: text: false image: true audio: false video: false tasks: encode: null score: null extract: {} max_sequence_length: 16384 profiles: transformers: max_batch_tokens: 16384 compute_precision: bfloat16 adapter_path: sie_server.adapters.glm_ocr:GlmOcrAdapter adapter_options: loadtime: {} runtime: max_new_tokens: 8192 num_beams: 1 default: max_batch_tokens: 16384 compute_precision: bfloat16 adapter_path: sie_server.adapters.sglang_vision_extract.adapter:SGLangVisionExtractAdapter adapter_options: loadtime: # 64 pages in flight, each up to 16384 tokens of page image and output, # need far more KV than the automatic one-sequence pool of 16384. mem_fraction_static: 0.85 served_model_name: zai-org/GLM-OCR max_concurrent_requests: 64 # SGLang's default fast (torch) image processor. The slow PIL processor runs # single-threaded in SGLang's tokenizer process and capped throughput on a # 4-vCPU lane. default_prompt: "Text Recognition:" runtime: max_new_tokens: 8192 num_beams: 1