sie_id: Qwen/Qwen3-Reranker-4B hf_id: Qwen/Qwen3-Reranker-4B hf_revision: 22e683669bc0f0bd69640a1354a6d0aebcfeede5 inputs: text: true image: false audio: false video: false tasks: encode: null score: {} extract: null max_sequence_length: 32768 profiles: default: max_batch_tokens: 16384 compute_precision: bfloat16 adapter_path: sie_server.adapters.qwen2_flash_cross_encoder.adapter:Qwen2FlashCrossEncoderAdapter adapter_options: loadtime: yes_token: "yes" no_token: "no" input_format: qwen3 score_mode: log_softmax default_instruction: "Given a web search query, retrieve relevant passages that answer the query" # The layers' elementwise work as fused Triton kernels: the same scores, # bit for bit, in less GPU time. fused_kernels: true runtime: {} # The default profile plus two serving options that move scores by # floating-point rounding (server README, "Qwen3 reranker serving options"): # the template, instruction and query are encoded once for the rows and the # requests that share them, and a one-row request over a cached prefix replays # as a CUDA graph. An extending profile restates the load-time options. fast: extends: default adapter_options: loadtime: yes_token: "yes" no_token: "no" input_format: qwen3 score_mode: log_softmax default_instruction: "Given a web search query, retrieve relevant passages that answer the query" fused_kernels: true prefix_reuse: true prefix_cache_mb: 512 cuda_graphs: bucketed cuda_graph_tokens: 512