PAI-Blade optimization report structure and parameters - Platform For AI - Alibaba Cloud - Platform For AI

After you call blade.optimize to optimize a model with PAI-Blade, the system generates an optimization report in JSON format. This reference describes the report structure and each field.

Report structure

The following example shows a complete optimization report:

{
  "software_context": [
    {
      "software": "tensorflow",
      "version": "1.15.0"
    },
    {
      "software": "cuda",
      "version": "9.0.176"
    }
  ],
  "hardware_context": {
    "device_type": "gpu",
    "microarchitecture": "T4"
  },
  "user_config": "",
  "diagnosis": {
    "model": "tmp_graph.pbtxt",
    "test_data_source": "user provided",
    "shape_variation": "dynamic",
    "message": "",
    "test_data_info": "input_ids_a_1:0 shape: (1, 9240) data type: int32"
  },
  "optimizations": [
    {
      "name": "TfStripUnusedNodes",
      "status": "effective",
      "speedup": "na",
      "pre_run": "na",
      "post_run": "na"
    },
    {
      "name": "TfAutoMixedPrecisionGpu",
      "status": "effective",
      "speedup": "1.42",
      "pre_run": "9.35 ms",
      "post_run": "6.59 ms"
    }
  ],
  "overall": {
    "baseline": "10.00 ms",
    "optimized": "4.38 ms",
    "speedup": "2.28"
  },
  "model_info": {
    "input_format": "frozen_pb"
  },
  "compatibility_list": [
    {
      "device_type": "gpu",
      "microarchitecture": "T4"
    }
  ],
  "model_sdk": {}
}

The two most important fields to check first are overall.speedup (the end-to-end speedup ratio) and optimizations[].status (whether each optimization took effect). Use the remaining fields to understand the environment, diagnose issues, and determine deployment constraints.

Parameter reference

Primary fields

Parameter	Description
`software_context`	Software environment, including the ML framework and Compute Unified Device Architecture (CUDA) version.
`hardware_context`	Hardware environment, including the device type and GPU microarchitecture.
`user_config`	User configuration. This field is blank in PAI-Blade optimization reports.
`diagnosis`	Diagnostic information about the model and test data. See Diagnosis fields.
`optimizations`	List of optimizations that PAI-Blade attempted. See Optimizations fields.
`overall`	End-to-end optimization results. See Overall fields.
`model_info`	Model format information.
`compatibility_list`	Hardware requirements for the optimized model. Deploy the model only on devices listed here to ensure correct behavior and expected performance.
`model_sdk`	Additional SDK information required for model deployment. This field is blank.

Diagnosis fields

Parameter	Description
`model`	Name of the model file.
`test_data_source`	Source of the test data. `user provided` means the data came from you; `blade deduced` means PAI-Blade inferred it automatically.
`shape_variation`	How input shapes vary. Valid values: `static`, `multiple_static_shapes`, `dynamic`.
`message`	Error message returned when PAI-Blade fails to infer test data.
`test_data_info`	Shape and data type of the test data.

Optimizations fields

Each entry in the optimizations array describes one optimization that PAI-Blade attempted.

Parameter	Description
`name`	Name of the optimization.
`status`	Whether the optimization took effect. `effective` means it was applied; `ineffective` means it did not take effect.
`speedup`	Speedup ratio for this optimization, calculated as `pre_run / post_run`.
`pre_run`	Model latency before this optimization.
`post_run`	Model latency after this optimization.

Overall fields

The overall object reports the end-to-end result across all optimizations applied.

Parameter	Description
`baseline`	Model latency before any optimization.
`optimized`	Model latency after all optimizations.
`speedup`	Overall speedup ratio, calculated as `baseline / optimized`. For example, a value of `2.28` means the optimized model runs 2.28× faster than the original.

Model info fields

Parameter	Description
`model_info.input_format`	Format of the original model.