diff --git a/packages/terratorch/models/prithvi/benchmark_instances/vllm_performance/space.yaml b/packages/terratorch/models/prithvi/benchmark_instances/vllm_performance/space.yaml new file mode 100644 index 0000000..1bb12e2 --- /dev/null +++ b/packages/terratorch/models/prithvi/benchmark_instances/vllm_performance/space.yaml @@ -0,0 +1,96 @@ + +entitySpace: + - identifier: "model" + propertyDomain: + variableType: "CATEGORICAL_VARIABLE_TYPE" + values: ["ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11"] + + - identifier: "request_rate" + propertyDomain: + variableType: "DISCRETE_VARIABLE_TYPE" + values: [1, 16, 23, 32, 64] + + # Docker images: one with threadpool support (v0.20.1), one without (v0.18.0) + - identifier: "image" + metadata: + description: "Docker image with vLLM + terratorch" + propertyDomain: + variableType: "CATEGORICAL_VARIABLE_TYPE" + values: + - "quay.io/mgazzetti/vllm-performance/vllm:vllm.v0.20.1-tt.v1.2.7" + + # Optional: Number of GPUs + - identifier: "n_gpus" + propertyDomain: + variableType: "DISCRETE_VARIABLE_TYPE" + values: [1] + + # Optional: GPU type + - identifier: "gpu_type" + propertyDomain: + variableType: "CATEGORICAL_VARIABLE_TYPE" + values: ["NVIDIA-A100-80GB-PCIe"] + + - identifier: "max_num_seq" + propertyDomain: + variableType: "DISCRETE_VARIABLE_TYPE" + values: [256] + + - identifier: "max_batch_tokens" + propertyDomain: + variableType: "DISCRETE_VARIABLE_TYPE" + values: [16384] + + # Optional: Number of CPUs + - identifier: "n_cpus" + propertyDomain: + variableType: "DISCRETE_VARIABLE_TYPE" + values: [16] + + - identifier: "memory" + propertyDomain: + variableType: "CATEGORICAL_VARIABLE_TYPE" + values: ['64Gi'] + + - identifier: "gpu_memory_utilization" + propertyDomain: + variableType: "DISCRETE_VARIABLE_TYPE" + values: [0.9] + + - identifier: "dtype" + propertyDomain: + variableType: "CATEGORICAL_VARIABLE_TYPE" + values: ["float16"] + + - identifier: "num_prompts" + propertyDomain: + variableType: "DISCRETE_VARIABLE_TYPE" + values: [500] + + - identifier: "dataset" + propertyDomain: + variableType: "CATEGORICAL_VARIABLE_TYPE" + values: ["india_url_in_b64_out"] + + # Geospatial-specific: Skip tokenizer initialization + - identifier: "skip_tokenizer_init" + propertyDomain: + variableType: "DISCRETE_VARIABLE_TYPE" + values: [1] + + # Geospatial-specific: Enforce eager mode + - identifier: "enforce_eager" + propertyDomain: + variableType: "DISCRETE_VARIABLE_TYPE" + values: [1] + + # Geospatial-specific: IO processor plugin + - identifier: "io_processor_plugin" + propertyDomain: + variableType: "CATEGORICAL_VARIABLE_TYPE" + values: ["terratorch_segmentation"] + + +experiments: + - actuatorIdentifier: "vllm_performance" + experimentIdentifier: "test-geospatial-deployment-custom-dataset-v1" diff --git a/packages/terratorch/nexus.yaml b/packages/terratorch/nexus.yaml index adeaede..e68258b 100644 --- a/packages/terratorch/nexus.yaml +++ b/packages/terratorch/nexus.yaml @@ -1,2 +1,8 @@ package: name: terratorch + + benchmark_packages: + # Local package stored in this repository + - requirement_specifier: "ado-vllm-performance==1.8.0" + experiments: + - "test-geospatial-deployment-custom-dataset-v1"