Skip to content

Latest commit

 

History

History
1115 lines (850 loc) · 37.4 KB

config.md

File metadata and controls

1115 lines (850 loc) · 37.4 KB
Raw

Table of Contents


Configuring Model Analyzer

Model Analyzer can be configured using either a YAML config file, the `command line interface* (CLI), or a combination of both.

  • Every flag supported by Model Analyzer can be configured using a YAML config file
  • Only a subset of flags can be configured using the CLI

The placeholders listed below are used throughout the configuration:

  • <boolean>: a boolean that can take true or false as value
  • <string>: a regular string
  • <comma-delimited-list>: a list of comma separated items
  • <int>: a regular integer value
  • <list>: a list of values
  • <range>: An object containing start and stop key with an optional step value
    • If step is not defined, 1 is the default step value
    • Types that support <range> can be described by a list or by using the example structure below, which declares the value of batch_sizes to be an array [2, 4, 6]
batch_sizes:
  start: 2
  stop: 6
  step: 2
  • <dict>: a set of key-value pairs
triton_server_flags:
  log_verbose: True
  exit_timeout_secs: 120

Config options for profile

CLI and YAML Config Options

A list of all the configuration options supported by both the CLI and YAML config file are shown below.

  • Brackets indicate that a parameter is optional.
  • For non-list and non-object parameters the value is set to the specified default.
  • The CLI flags corresponding to each of the options below are obtained by converting the snake_case options to --kebab-case.
    For example, profile_models in the YAML would be --profile-models in the CLI.
# Path to the Triton Model Repository
# https://github.com/triton-inference-server/server/blob/main/docs/user_guide/model_repository.md
model_repository: <string>

# List of the model names to be profiled
profile_models: <comma-delimited-string-list>

# List of composing models for BLS models
bls_composing_models: <comma-delimited-string-list>

# List of composing models that should be profiled using CPU instances only
cpu_only_composing_models: <comma-delimited-string-list>

# Full path to directory to which to read and write checkpoints and profile data
[ checkpoint_directory: <string> | default: './checkpoints' ]

# The directory to which the model analyzer will save model config variants
[ output_model_repository_path: <string> | default: 'output_model_repository' ]

# Allow model analyzer to overwrite contents of the output model repository
[ override_output_model_repository: <boolean> | default: false ]

# Export path to be used
[ export_path: <string> | default: '.' ]

# Concurrency values to be used
[ concurrency: <comma-delimited-string|list|range> ]

# Batch size values to be used
[ batch_sizes: <comma-delimited-string|list|range> | default: 1 ]

# Request rate values to be used
[ request_rate: <comma-delimited-string|list|range> ]

# Specifies the maximum number of retries for any retry attempt
[ client_max_retries: <int> | default: 50 ]

# Specifies how long (seconds) to gather server-only metrics
[ duration_seconds: <int> | default: 3 ]

# Duration of waiting time between each metric measurement in seconds
[ monitoring_interval: <float> | default: 1 ]

# Specifies which metric(s) are to be collected.
[ collect_cpu_metrics: <bool> | default: false ]

# The protocol used to communicate with the Triton Inference Server. Only 'http' and 'grpc' are allowed for the values
[ client_protocol: <string> | default: grpc ]

# The full path to the perf_analyzer binary executable
[ perf_analyzer_path: <string> | default: perf_analyzer ]

# Perf analyzer timeout value in seconds
[ perf_analyzer_timeout: <int> | default: 600]

# Maximum CPU utilization value allowed for the perf_analyzer
[ perf_analyzer_cpu_util: <float> | default: 80.0 ]

# Enables writing the output from the perf_analyzer to a file or stdout
[ perf_output: <bool> | default: false ]

# If specified, setting --perf-output will write the perf_analyzer output to the file at # this location
[ perf_output_path: <str> ]

# Maximum number of times perf_analyzer is launched with auto adjusted parameters in an attempt to profile a model
[ perf_analyzer_max_auto_adjusts: <int> | default: 10 ]

# Disables model loading and unloading in remote mode
[ reload_model_disable: <bool> | default: false]

# Triton Docker image tag used when launching using Docker mode
[ triton_docker_image: <string> | default: nvcr.io/nvidia/tritonserver:24.10-py3 ]

# Triton Server HTTP endpoint url used by Model Analyzer client"
[ triton_http_endpoint: <string> | default: localhost:8000 ]

# The full path to the parent directory of 'lib/libtritonserver.so. Only required when using triton_launch_mode=c_api
[ triton_install_path: <string> | default: /opt/tritonserver ]

# Triton Server GRPC endpoint url used by Model Analyzer client
[ triton_grpc_endpoint: <string> | default: localhost:8001 ]

# Triton Server metrics endpoint url used by Model Analyzer client
[ triton_metrics_url: <string> | default: http://localhost:8002/metrics ]

# The full path to the tritonserver binary executable
[ triton_server_path: <string> | default: tritonserver ]

# The full path to a file to write the Triton Server output log
[ triton_output_path: <string> ]

# List of strings containing the paths to the volumes to be mounted into the tritonserver docker
# containers launched by model-analyzer. Will be ignored in other launch modes
[ triton_docker_mounts: <list of strings> ]

# The size of /dev/shm for the triton docker container
[ triton_docker_shm_size: <string>]

# How Model Analyzer will launch triton. It should
# be either "docker", "local", "remote" or "c_api".
# See docs/launch_modes.md for more information
[ triton_launch_mode: <string> | default: 'local' ]

# List of GPU UUIDs to be used for the profiling. Use 'all' to profile all the GPUs visible by CUDA
[ gpus: <string|comma-delimited-list-string> | default: 'all' ]

# Search mode. Options are "brute", "quick", and "optuna"
[ run_config_search_mode: <string> | default: brute]

# Minimum concurrency used for the automatic/quick/optuna config search
[ run_config_search_min_concurrency: <int> | default: 1 ]

# Maximum concurrency used for the automatic/quick/optuna config search
[ run_config_search_max_concurrency: <int> | default: 1024 ]

# Minimum max_batch_size used for the automatic/quick/optuna config search
[ run_config_search_min_model_batch_size: <int> | default: 1 ]

# Maximum max_batch_size used for the automatic/quick/optuna config search
[ run_config_search_max_model_batch_size: <int> | default: 128 ]

# Minimum instance group count used for the automatic/quick/optuna config search
[ run_config_search_min_instance_count: <int> | default: 1 ]

# Maximum instance group count used for the automatic/quick/optuna config search
[ run_config_search_max_instance_count: <int> | default: 5 ]

# Minimum request rate used for the automatic/quick/optuna config search
[ run_config_search_min_request_rate: <int> | default: 16 ]

# Maximum request rate used for the automatic/quick/optuna config search
[ run_config_search_max_request_rate: <int> | default: 8092 ]

# Maximum number of steps taken during a binary search
[ run_config_search_max_binary_search_steps: <int> | default: 5 ]

# Disables automatic config search
[ run_config_search_disable: <bool> | default: false ]

# Enables the profiling of all supplied models concurrently
[ run_config_profile_models_concurrently_enable: <bool> | default: false]

# Enables the searching of request rate (instead of concurrency)
[ request_rate_search_enable: <bool> | default: false]

# Minimum percentage of the search space to profile when using Optuna
[ min_percentage_of_search_space: <int> | default: 5]

# Maximum percentage of the search space to profile when using Optuna
[ max_percentage_of_search_space: <int> | default: 10]

# Minimum number of trials to profile when using Optuna
[ optuna_min_trials: <int> | default: None]

# Maximum number of trials to profile when using Optuna
[ optuna_max_trials: <int> | default: None]

# Number of trials without improvement before triggering early exit when using Optuna
[ optuna_early_exit_threshold: <int> | default: 10]

# Use the concurrency formula instead of searching the concurrency space in Optuna search mode
[ use_concurrency_formula: <bool> | default: false]

# Disables the sweeping of concurrencies for the top-N models after quick/optuna search completion
[ concurrency_sweep_disable: <bool> | default: false]

# Always report GPU metrics, even if the model(s) is cpu_only
[ always_report_gpu_metrics: <bool> | default: false]

# Skips the generation of summary reports and tables
[ skip_summary_reports: <bool> | default: false]

# Skips the generation of detailed reports and tables
[ skip_detailed_reports: <bool> | default: false]

# Type of model being profiled: generic or LLM
[ model_type: <string> | default: generic]

# Number of top configs to show in summary plots
[ num_configs_per_model: <int> | default: 3]

# Number of top model configs to save across ALL models, none saved by default
[ num_top_model_configs: <int> | default: 0 ]

# File name to be used for the model inference results
[ filename_model_inference: <string> | default: metrics-model-inference.csv ]

# File name to be used for the GPU metrics results
[ filename_model_gpu: <string> | default: metrics-model-gpu.csv ]

# File name to be used for storing the server only metrics
[ filename_server_only: <string> | default: metrics-server-only.csv ]

# Specifies columns keys for model inference metrics table
[ inference_output_fields: <comma-delimited-string-list> | default: See [Config Defaults](#config-defaults) section]

# Specifies columns keys for model gpu metrics table
[ gpu_output_fields: <comma-delimited-string-list> | default: See [Config Defaults](#config-defaults) section]

# Specifies columns keys for server only metrics table
[ server_output_fields: <comma-delimited-string-list> | default: See [Config Defaults](#config-defaults) section]

# Shorthand that allows a user to specify a max latency constraint in ms
[ latency_budget: <int>]

# Shorthand that allows a user to specify a min throughput constraint
[ min_throughput: <int>]

# Specify path to config YAML file
[ config_file: <string> ]

YAML Only Options

The following config options are supported only by the YAML config file.

# YAML config section for each model to be profiled
profile_models: <comma-delimited-string-list|list|profile_model>

# List of composing models for BLS models
bls_composing_models: <comma-delimited-string-list>

# List of composing models that should be profiled using CPU instances only
cpu_only_composing_models: <comma-delimited-string-list>

# List of constraints placed on the config search results
[ constraints: <constraint> ]

# List of objectives that user wants to sort the results by it
[ objectives: <objective|list> ]

# Weighting used to bias the model's objectives (against the other models) in concurrent multi-model mode
[ weighting: <int>]

# Specify flags to pass to the Triton instances launched by model analyzer
[ triton_server_flags: <dict> ]

# Allows custom configuration of perf analyzer instances used by model analyzer
[ perf_analyzer_flags: <dict> ]

# Allows custom configuration of GenAI-perf instances used by model analyzer
[ genai_perf_flags: <dict> ]

# Allows custom configuration of the environment variables for tritonserver instances
# launched by model analyzer
[ triton_server_environment: <dict> ]

# Dict of name=value pairs containing metadata for the tritonserver docker container
# launched in docker launch mode
[ triton_docker_labels: <dict> ]

# Dict of name=value pairs containing arguments (applied to the run command) that will be
# applied to the tritonserver docker container launched in docker launch mode
[ triton_docker_args: <dict> ]

Config Options for report

CLI and YAML options

A list of all the configuration options supported by both the CLI and YAML config file are shown below.

  • Brackets indicate that a parameter is optional.
  • For non-list and non-object parameters the value is set to the specified default.
# Comma-delimited list of the model names for which to generate detailed reports
report_model_configs: <comma-delimited-string-list>

# Full path to directory to which to read and write checkpoints and profile data
[ checkpoint_directory: <string> | default: '.' ]

# Export path to be used
[ export_path: <string> | default: '.' ]

# Specify path to config YAML file
[ config_file: <string> ]

# Always report GPU metrics
[ always_report_gpu_metrics: <bool> | default: false]

YAML only options

The following config options are support by the YAML config file only.

# YAML config section for each model config for which to generate detailed reports
report_model_configs: <comma-delimited-string-list|list|report_model_config>

# YAML sections to configure the plots that should be shown in the detailed report
[ plots: <dict-plot-configs> | default: See [Config Defaults](#config-defaults) section ]

Field Descriptions

Before proceeding, it will be helpful to see the documentation on Model Analyzer Metrics regarding what metric tags are and how to use them.

<constraint>

A constraint, specifies the bounds that determine a successful run. The table below shows examples of the types of constraints allowed:

Option Name Units Constraint Description
perf_throughput inf / sec min Specify minimum desired throughput.
perf_latency_p99 ms max Specify maximum tolerable latency or latency budget.
output_token_throughput tok / sec min Specify minimum desired output token throughput.
inter_token_latency_p99 ms max Specify maximum tolerable inter token latency.
time_to_first_token_p99 ms max Specify maximum tolerable time to first token latency.
gpu_used_memory MB max Specify maximum GPU memory used by model.

Examples

To filter out the results when perf_throughput is less than 5 infer/sec:

perf_throughput:
  min: 5

To filter out the results when perf_latency_p99 is larger than 100 milliseconds:

perf_latency_p99:
  max: 100

To filter out the results when gpu_used_memory is larger than 200 MBs:

gpu_used_memory:
  max: 200

Keys can be combined for more complex constraints:

gpu_used_memory:
  max: 200
perf_latency_p99:
  max: 100

This will filter out the results when gpu_used_memory is larger than 200 MBs and their latency is larger than 100 milliseconds.

The values described above can be specified both globally and on a per model basis.

The global example looks like below:

model_repository: /path/to/model-repository
profile_models:
  - model_1
  - model_2

constraints:
  gpu_used_memory:
    max: 200

In the global mode, the constraint specified will be enforced on every model. To have different constraints for each model, version below can be used:

model_repository: /path/to/model-repository
profile_models:
  model_1:
    constraints:
      gpu_used_memory:
        max: 200
  model_2:
    constraints:
      perf_latency_p99:
        max: 50

<objective>

Objectives specify the sorting criteria for the final results. The fields below are supported under this object type:

Option Name Description
perf_throughput Use throughput as the objective.
perf_latency_p99 Use latency as the objective.
gpu_used_memory Use GPU memory used by the model as the objective.
gpu_free_memory Use GPU memory not used by the model as the objective.
gpu_utilization Use the GPU utilization as the objective.
cpu_used_ram Use RAM used by the model as the objective.
cpu_free_ram Use RAM not used by the model as the objective.
output_token_throughput Use output token throughput as the objective.
inter_token_latency_p99 Use inter token latency as the objective.
time_to_first_token_p99 Use time to first token latency as the objective.

An example objectives that will sort the results by throughput looks like below:

objectives:
  - perf_throughput

To sort the results by latency, objectives should look like:

objectives:
  - perf_latency_p99

Weighted Objectives

In addition to the mode discussed above, multiple values can be provided in the objectives key in order to provide more generalized control over how model analyzer sorts results. For example:

objectives:
  - perf_latency_p99
  - perf_throughput

The above config is telling model analyzer to compare two measurements by finding relative gain from one measurement to the other, and computing the weighted average of this gain across all listed metrics. In the above example, the relative weights for each metric is equal by default. So if we have two measurements of latency and throughput, model analyzer employs the following logic:

measurement_A = (latency_A, throughput_A)
measurement_B = (latency_B, throughput_B)

gain_A_B = (latency_A - latency_B, throughput_A - throughput_B)

weighted_average_gain = 0.5*(latency_A - latency_B) + 0.5*(throughput_A - throughput_B)

If weighted_average_gain exceeds a threshold then measurement_A is declared to be "better" than measurement_B. Model Analyzer will automatically account for metrics in which less is better and those which more is better.

An extension of the above objectives is explicitly specifying the weights. For example:

objectives:
  perf_latency_p99: 2
  perf_throughput: 3

The score for each measurement will be a weighted average using the weights specified here. The above config will tell Model Analyzer to multiply the throughput gain by 0.4 and latency gain by by 0.6.

The objectives section can be specified both globally and on a per model basis.


Test Configuration <parameter>

A user can specify a range of test configurations that Model Analyzer will profile over. The possible configuration parameters are concurrency and batch_sizes. One or more parameters are specified per model only. Parameters cannot be specified globally.

Options available under this parameter are described in table below:

Option Name Description Supporting Types
concurrency Request concurrency used for generating the input load. <range>, <comma-delimited-list>, or a <list>
request_rate Request rate used for generating the input load. <range>, <comma-delimited-list>, or a <list>
batch_sizes Static batch size used for generating requests. <range>, <comma-delimited-list>, or a <list>

An example <parameter> looks like below:

parameters:
  concurrency:
    start: 2
    stop: 64
    step: 8
  batch_sizes: 1,2,3

A complete config example looks like below:

model_repository: /path/to/model-repository
profile_models:
  model_1:
    parameters:
      concurrency:
        start: 2
        stop: 64
        step: 8
      batch_sizes: 1,2,3

These parameters will result in testing the concurrency configurations of 2, 10, 18, 26, 34, 42, 50, 58, and 64, for each of different batch sizes of 1, 2 and 3. This will result in 27 individual test runs of the model.

<weighting>

This field is used to bias a model's objective when performing a multi-model search.

See Multi-Model Search - Model Weighting for details and an example YAML configuration.

<model-config-parameters>

This field represents the values that can be changed or swept through using Model Analyzer. All the values supported in the Triton Config can be specified or swept through here. <model-config-parameters> should be specified on a per model basis and cannot be specified globally (like <parameter>).

Table below presents the list of common parameters that can be used for manual sweeping:

Option Description
dynamic_batching Dynamic batching is a feature of Triton that allows inference requests to be combined by the server, so that a batch is created dynamically.
max_batch_size The max_batch_size property indicates the maximum batch size that the model supports for the types of batching that can be exploited by Triton.
instance_group Triton can provide multiple instances of a model so that multiple inference requests for that model can be handled simultaneously. The model configuration ModelInstanceGroup property is used to specify the number of execution instances that should be made available and what compute resource should be used for those instances.

An example <model-config-parameters> look like below:

model_config_parameters:
  max_batch_size: [6, 8]
  dynamic_batching:
    max_queue_delay_microseconds: [200, 300]
  instance_group:
    - kind: KIND_GPU
      count: [1, 2]

Note that for values that accept a list by default the user needs to specify an additional list if want to sweep through it. Otherwise, it will only change the original model config to the value specified and it will not sweep through it. To read more about the automatic config search, check out Config Search docs.

A complete profile YAML config looks like below:

model_repository: /path/to/model/repository/

run_config_search_disable: True
profile_models:
  model_1:
    model_config_parameters:
      max_batch_size: 2
      dynamic_batching:
        max_queue_delay_microseconds: 200
      instance_group:
        - kind: KIND_GPU
          count: 1
        - kind: KIND_CPU
          count: 1

Note that in the above configuration, it will not sweep through any of the parameters. The reason is that instance_group accepts a list by default. Sweeping thorough different parameters can be achieved using the configuration below:

model_repository: /path/to/model/repository/

run_config_search_disable: True
profile_models:
  model_1:
    model_config_parameters:
      max_batch_size: 2
      dynamic_batching:
        max_queue_delay_microseconds: [200, 400, 600]
      instance_group:
        - - kind: KIND_GPU
            count: 1
        - - kind: KIND_CPU
            count: 1

This will lead to 6 different configurations (3 different max queue delays and two instance group count combinations). If both model_config_parameters and parameters keys are specified, the list of sweep configurations will be the cartesian product of both of the lists.

<cpu_only>

This flag tells the model analyzer that, whether performing a search during profiling or generating reports, this model should use CPU instances only. In order to run a model on CPU only you must provide a value of true for this flag.

model_repository: /path/to/model/repository/
profile_models:
  model_1:
    cpu_only: true
  model_2:
    perf_analyzer_flags:
      percentile: 95

The above config tells model analyzer to profile model_1 on CPU only, but profile model_2 using GPU.

<perf-analyzer-flags>

This field allows the user to pass perf_analyzer any CLI options it needs to execute properly. Refer to the perf_analyzer docs for more information on these options.

Global options to apply to all instances of Perf Analyzer

The perf_analyzer_flags section can be specified globally to affect perf_analyzer instances across all models in the following way:

model_repository: /path/to/model/repository/
profile_models:
  model_1:
    parameters:
      batch_sizes: 4
perf_analyzer_flags:
  percentile: 95

Model-specific options for Perf Analyzer

In order to set flags only for a specific model, you can specify the flags in the following way:

model_repository: /path/to/model/repository/
profile_models:
  model_1:
    perf_analyzer_flags:
      percentile: 95

Shape, Input-Data, and Streaming

The input-data, shape, and streaming perf_analyzer options are additive and can take either a single string (non-list) or a list of strings. Below is an example for shape argument:

model_repository: /path/to/model/repository/
profile_models:
  model_1:
    parameters:
      batch_sizes: 4
perf_analyzer_flags:
  shape:
    - INPUT0:1024,1024
    - INPUT1:1024,1024

Variable-sized dimensions

If a model configuration has variable-sized dimensions in the inputs section, then the shape option of the perf_analyzer_flags option must be specified. More information about this can be found in the Perf Analyzer documentation.

SSL Support:

Perf Analyzer supports SSL via GRPC and HTTP. It can be enabled via Model Analyzer configuration file updates.

GRPC example:

model_repository: /path/to/model/repository/
profile_models:
  model_1:
    perf_analyzer_flags:
      ssl-grpc-root-certifications-file: /path/to/PEM/encoded/server/root/cert

HTTPS example:

model_repository: /path/to/model/repository/
profile_models:
  model_1:
    perf_analyzer_flags:
      ssl-https-ca-certificates-file: /path/to/cert/authority/cert/file
      ssl-https-client-certificate-file: /path/to/client/cert/file
      ssl-https-private-key-file: /path/to/private/key/file

More information about this can be found in the Perf Analyzer documentation.

Important Notes:

  • Only a subset of flags can be specified on the command line. Use model-analyzer profile --help to see the list of flags that can be specified on the command line. If a flag isn't listed there, it can be specified via the YAML config file.
  • When providing arguments under perf_analyzer_flags, you must use - instead of _. This casing is important and Model Analyzer will not recognize snake_cased arguments.
  • Model Analyzer also provides certain arguments to the perf_analyzer instances it launches. They are the following:
    • concurrency-range
    • request-rate-range
    • batch-size
    • model-name
    • measurement-mode
    • service-kind
    • triton-server-directory
    • model-repository
    • protocol
    • url If provided under the perf_analyzer_flags section, their values will be overridden. Caution should therefore be exercised when overriding these.

<triton-server-flags>

This section of the config allows fine-grained control over the flags passed to the Triton instances launched by Model Analyzer when it is running in the docker or local Triton launch modes. Any argument to the server can be specified here.

model_repository: /path/to/model/repository/
profile_models:
  model_1:
    parameters:
      batch_sizes:
        start: 4
        stop: 9
      concurrency:
        - 2
        - 4
        - 8
triton_server_flags:
  strict_model_config: False
  log_verbose: True

Since Model Analyzer relaunches Triton Server each time a model config is loaded, you can also specify triton_server_flags on a per model basis. For example:

model_repository: /path/to/model/repository/
profile_models:
  model_1:
    triton_server_flags:
        strict_model_config: False
        log_verbose: True
  model_1:
    triton_server_flags:
        exit_timeout_secs: 120

Note: Triton Server supports SSL via GRPC. It can be enabled via Model Analyzer configuration file updates.

GRPC example:

model_repository: /path/to/model/repository/
profile_models:
  model_1:
    parameters:
      batch_sizes:
        start: 4
        stop: 9
      concurrency:
        - 2
        - 4
        - 8
triton_server_flags:
  grpc-use-ssl: 1
  grpc-server-cert: /path/to/grpc/server/cert
  grpc-server-key: /path/to/grpc/server/keyfile
  grpc-root-cert: /path/to/grpc/root/cert

More information about this can be found in the Triton Server documentation.

Important Notes:

  • The Model Analyzer also provides certain arguments to the tritonserver instances it launches. These cannot be overridden by providing those arguments in this section. An example of this is http-port, which is an argument to Model Analyzer itself.

<triton-server-environment>

This section enables setting environment variables for the tritonserver instances launched by Model Analyzer. For example, when a custom operation is required by a model, Triton requires the LD_PRELOAD and LD_LIBRARY_PATH environment variables to be set. See this link for details. The value for this section is a dictionary where the keys are the environment variable names and their values are the values to be set.

model_repository: /path/to/model/repository/
profile_models:
  - model_1
triton_server_environment:
  LD_PRELOAD: /path/to/custom/op/.so
  LD_LIBRARY_PATH: /path/to/shared/libararies

Since Model Analyzer relaunches Triton Server each time a model config is loaded, you can also specify triton_server_environment on a per model basis. For example:

model_repository: /path/to/model/repository/
profile_models:
  model_1:
    triton_server_environment:
      LD_PRELOAD: /path/to/custom/op

Important Notes:

  • The Model Analyzer also provides certain environment variables to the tritonserver instances it launches. These cannot be overridden by providing those arguments in this section. An example of this is CUDA_VISIBLE_DEVICES.

<plots>

This section is used to specify the kind of plots that will be displayed in the detailed report. The section is structured as a list of <plot> objects as follows:

plots:
  plot_name_1:
    title: Title
    x_axis: perf_latency_p99
    y_axis: perf_throughput
    monotonic: True
  plot_name_2: .
    .
    .

Each <plot> object consists of the plot_name which is also the name of the file to which the plot will be saved as a .png. Each plot object also requires specifying each of the following:

  • title : The title of the plot
  • x_axis : The metric tag for the metric that should appear in the x-axis, e.g. perf_latency_p99. The plotted points are also sorted by the values of this metric.
  • y_axis : The metric tag for the metric that should appear in the y-axis.
  • monotonic : Some plots may require consecutive points to be strictly increasing. A boolean value of true can be specified here to require this.

<profile-model>

The --profile-models argument can be provided as a list of strings (names of models) from the CLI interface, or as a more complex <profile-model> object but only through the YAML configuration file. The model object can contain <model-config-parameters>, <parameter>, <perf-analyzer-flags>,<triton-server-flags>,<triton-server-environment> and a flag cpu_only.

A profile model object puts together all the different parameters specified above. An example will look like:

model_repository: /path/to/model/repository/

run_config_search_disable: True
profile_models:
  model_1:
    perf_analyzer_flags:
      percentile: 95
    model_config_parameters:
      max_batch_size: 2
      dynamic_batching:
        max_queue_delay_microseconds: 200
      instance_group:
        - - kind: KIND_GPU
            count: 1
        - - kind: KIND_CPU
            count: 1
    parameters:
      concurrency:
        start: 2
        stop: 64
        step: 8
      batch_sizes: 1,2,3

Multiple models can be specified under the profile_models key.

Example:

model_repository: /path/to/model-repository
run_config_search_disable: true
triton_launch_mode: docker
profile_models:
  vgg_19_graphdef:
    parameters:
      batch_sizes: 1, 2, 3
    model_config_parameters:
      dynamic_batching:
        max_queue_delay_microseconds: 200
      instance_group:
        - - kind: KIND_CPU
            count: 1
    vgg_16_graphdef:
      parameters:
        concurrency:
          start: 2
          stop: 64
          step: 8
      model_config_parameters:
        dynamic_batching:
          max_queue_delay_microseconds: 200
        instance_group:
          - - kind: KIND_GPU
              count: 1
batch_sizes:
  start: 4
  stop: 9
concurrency:
  - 2
  - 4
  - 8

If this file is saved to the config.yml, Model Analyzer can be started using the -f, or --config-file flag.

model-analyzer -f /path/to/config.yml

It will run the model vgg_19_graphdef over combinations of batch sizes [1,2,3], concurrency [2,4,8] (taken from the global concurrency section), with dynamic batching enabled and a single CPU instance.

It will also run the model vgg_16_graphdef over combinations of batch sizes [4,5,6,7,8,9](taken from the global batch_sizes section), concurrency [2,10,18,26,34,42,50,58], with dynamic batching enabled and a single GPU instance.

<report-model-config>

The --report-model-configs argument can be provided as a list of strings (names of models) from the CLI interface, or as a more complex <report-model-config> object but only through the YAML configuration file. The model object can contain <plots> objects. An example looks like:

report_model_configs:
  model_config_0:
    plots:
      throughput_v_latency:
        title: Title
        x_axis: perf_latency_p99
        y_axis: perf_throughput
        monotonic: True

Multiple models can be specified under the report_model_configs key as well.

report_model_configs:
  model_config_default:
    plots:
      throughput_v_latency:
        title: Title
        x_axis: perf_latency_p99
        y_axis: perf_throughput
        monotonic: True
  model_config_0:
    plots:
      gpu_mem_v_latency:
        title: Title
        x_axis: perf_latency_p99
        y_axis: gpu_used_memory
        monotonic: False

Config Defaults

Up to date config defaults can be found in config_defaults.py