GitHub tag (latest by date) PyPI - Module Version PyPI - License PyPI - Downloads

A CLI tool and library for communicating with Thoth backend.

Using Thamos as a CLI tool

Thamos is released on PyPI. See installation instructions bellow to setup Thoth/Thamos for your repository:

# Install Thamos CLI tool:
$ pip3 install thamos  # keep in mind: requires Python 3.6+!!
# Go to repository that should be managed by Thoth which already has Pipfile present:
$ cd ~/git/repo/
# Setup Thamos configuration:
$ thamos config
# Add packages.
$ thamos add tensorflow
# Ask Thoth for software stack recommendations:
$ thamos advise
# Install packages:
$ thamos install  # can be merged using `thamos advise --install`
# Retrieve logs of the last analysis:
$ thamos log

As Thamos notes analysis ids for better UX of thamos log, it’s recommended to add .thoth_last_analysis_id file to .gitignore. Adding also .venv might be useful if virtual environment management is turned on.

See thoth-station/cli-examples repository with examples.

Recommendation types

See this document for a detailed info on the recommendation types Thoth provides.

Adjusting configuration based on environment variables

You can adjust content of configuration file each time Thamos CLI or Thamos library loads it by expanding entries with environment variables. This can be handy if you would like to parameterize some of the options at runtime (e.g. in deployment).

This behaviour is (due to security reasons) explicitly turned off by default. However you can turn it on by setting THAMOS_CONFIG_EXPAND_ENV environment variable to 1 (0 explicitly turns this behaviour off, default value): THAMOS_CONFIG_EXPAND_ENV=1 thamos advise
2019-03-13 11:22:59,562 [18639] INFO     thamos.config: Expanding configuration file based on environment variables

Entries which should be expanded have environment variables in curly braces like the following example:

host: {THOTH_HOST}

Note the expansion is done by replacing these values directly with values of environment variable, this means types need to be taken into account (environment variable with value "true" is put into configuration file as true).

Using custom configuration file template

You can use your own custom configuration file as a template. This is especially useful if you want to have some configuration entries constant and let expand only some of the configuration options. In other words, you can parametrize configuration file.

An example of configuration file template can be:

tls_verify: true
requirements_format: {requirements_format}

  - name: '{runtime_environment_name}'
      name: {os_name}
      version: '{os_version}'
      foo: bar
      key: value
      cpu_family: {cpu_family}
      cpu_model: {cpu_model}
      gpu_model: {gpu_model}
    python_version: '{python_version}'
    cuda_version: {cuda_version}
    recommendation_type: stable
    platform: '{platform}'

Then, you need to supply this configuration file to the following command:

thamos config --template template.yaml

Listing of automatically expanded configuration options which are supplied the config sub-command (these options are optional and will be expanded based on HW or SW discovery):

Configuration option




name of operating system



name of operating system



version of operating system



CPU family identifier



CPU model identifier



Python version (major.minor)



CUDA version (major.minor)



Platform used.



Requirements format.



Thoth base image used.

Platform corresponds to sysconfig.get_platform() call.

These configuration options are optional and can be mixed with adjustment based on environment variables (see THOTH_SERVICE_HOST example above). Note the environment variables are not expanded on thamos config call but rather on other sub-commands issued (e.g. thamos advise or others).

The output format coming out of recommendations can be compatible with Pipenv, raw pip or similar to the one provided by pip-tools (actually same as for pip as these formats are interchangeable). The format is configured using requirements_format configuration option, available options are:

  • requirements_format: pipenv for Pipenv compatible output

  • requirements_format: pip or requirements_format: pip-tools for pip or pip-tools compatible output


It is possible to label requests for user-specific needs. In such a case, resolver will include pipeline units that match labels with the ones provided on the request.

An example can be a CI system that is asking for an advise and labels the request with requester=ci_foo;team=thoth. In such a case, the resolution engine includes pipeline units that are specific to the CI system and the team specified (besides the ones that are added by default). Labels can be specified in the .thoth.yaml configuration file or using CLI (labels passed via CLI take precedence):

thamos advise --labels requester=ci_foo;team=thoth

See the following demo for more information.

Support for multiple runtime environments

Thoth performs recommendations based on your hardware and software environment, so called runtime environments. You can specify more than just one runtime environment that should be targetted during recommendations. This might be suitable if you would like to tweak some runtime environment specific configuration options. An example could be a deployment of a machine learning model to the cluster that uses CUDA, but you do not run CUDA locally (fast iterative development locally, subsequently training a model in the cluster on a large dataset). In such cases, you can specify two configuration entries in .thoth.yaml file:

tls_verify: true
requirements_format: pipenv

  - name: 'cuda'  # <<<
      name: fedora
      version: '32'
      cpu_family: 6
      cpu_model: 94
      gpu_model: 'GeForce GTX 680'
    python_version: '3.8'
    # <<< HERE
    cuda_version: '10.1'  # <<<
    # <<< HERE
    recommendation_type: stable
    platform: 'linux-x86_64'
    openblas_version: '0.3.13'
    openmpi_version: '4.1'
    cudnn_version: '8'
    mkl_version: '2021.1.1'
    base_image: ''

  - name: 'no_cuda'  # <<<
      name: fedora
      version: '32'
      cpu_family: 6
      cpu_model: 94
      gpu_model: null
    python_version: '3.8'
    # <<< HERE
    cuda_version: null  # <<<
    # <<< HERE
    recommendation_type: stable
    platform: 'linux-x86_64'
    openblas_version: '0.3.13'
    openmpi_version: '4.1'
    cudnn_version: null
    mkl_version: '2021.1.1'
    base_image: ''

The two runtime environments stated in the .thoth.yaml differ in cuda_version configuration and their names.

To trigger advises for runtime environment named cuda, issue:

thamos advise --runtime-environment cuda

To target the latter runtime environment named no_coda, you can issue:

thamos advise --runtime-environment no_cuda

This option can be also supplied via environment variable using THAMOS_RUNTIME_ENVIRONMENT=no_cuda.

If the runtime environment is not provided explictly, Thamos will take the first runtime environment entry stated in the runtime_environment listing. For the example showed above it will default to cuda environment:

# defaults to the first one - "cuda"
thamos advise

Multiple runtime environments can be used in conjunction with the automatically expanded configuration options and configuration file templating naturally.

By default, all the files produced during advises are stored in the project root directory. To maintain multiple lock files specific for runtime environments, it is possible to configure “overlays” directory in Thamos configuration file.

Listing available environments and container images

To list available environments for which the resolver can resolve dependencies, issue:

thamos environments

Each entry states configuration of operating system, its version and Python interpreter version that can be configured in each runtime environment section in .thoth.yaml.

If you wish to list available container images ready to be used:

thamos images

Each entry stated can be set as a base_image in .thoth.yaml in the respective runtime environment section and used as a base for running the Python applications.

Overlays directory

Multiple directories carrying requirement files can be configured using overlays_dir configuration option in .thoth.yaml file. This configuration is configured on a global scope and all the runtime environments inherit path from it.

An example configuration file states overlays_dir:

tls_verify: true
requirements_format: pipenv
overlays_dir: overlays

  - name: 'fedora-33'
      name: fedora
      version: '33'
    python_version: '3.8'

  - name: 'ubi-8'
      name: rhel
      version: '8'
    python_version: '3.8'

In such case, the directrory structure respecting the configuration supplied should be:

├── overlays
│   ├── fedora-33
│   │   ├── Pipfile
│   │   ├── Pipfile.lock
│   │   ├── .env
│   │   └── constraints.txt
│   └── ubi-8
│   │   ├── Pipfile
│   │   ├── Pipfile.lock
│   │   ├── .env
│   │   └── constraints.txt
└── .thoth.yaml

Each directory in the overlays directory should respect the runtime environment name stated in .thoth.yaml file and carries files specific for the given runtime environment.

Similarly as for Pipenv files, requirement files respecting pip-tools can be used ( and requirements.txt).

Constraints files (constraints.txt) are optional.

Optionally, users can provide .env file that can state environment variables that should be passed to the process when thamos run is executed. The .env file states each environment variable on a single line in a form of ENV_NAME=VALUE. Optionally, lines can be commented out with hash (#). An example of the file content:

# This is an example .env file.

Each .env file can be specified per overlay. If no overlay directories are used, .env file can be placed in the top level project directory (the directory where .thoth.yaml is present).

Installing requirements

Once a lock file is resolved after calling thamos advise, the application stack can be installed by using thamos install command. If you wish to pass additional options that should be used by pip, you can do so by passing them after --.

An example could be installing packages in a corporate network where packages should be installed through a proxy tunnel:

thamos install -- --proxy socks5h:// --trusted-host

Advise report structure

The structure of the advise report in JSON format produced with the thamos advise --json command is available in the Thoth API advise endpoint to retrieve advise results and in the corresponding schema specification for this endpoint.

Using Thoth and thamos in OpenShift’s s2i

Using configuration templates is especially useful for OpenShift builds where you can specify your template in an s2i repository (omit Pipfile.lock to enable call to thamos advise as shown in this repository).

Then, you need to provide following environment variables:

  • THAMOS_CONFIG_TEMPLATE - holds path to template - use /tmp/src prefix to point to root of s2i repository (e.g. /tmp/src/template.yaml if template.yaml is the configuration template and is stored in root of your Git repository).

  • THAMOS_NO_INTERACTIVE - set to 1 if you don’t want to omit interactive thamos (suitable for automated s2i builds happening in the cluster).

  • THAMOS_NO_PROGRESSBAR - set to 1 to disable progressbar while waiting for response from Thoth backend - it can cause annoying too verbose output printed to OpenShift console during the build.

  • THAMOS_CONFIG_EXPAND_ENV - set to 1 to enable expansion based on environment variables when generating .thoth.yaml file - this needs to be explicitly turned on due to possible security implications.

  • THAMOS_FORCE - set to 1 not use cached results, always force analysis on Thoth’s side (note this option can be ignored by a Thoth operator based on deployment configuration).

  • THAMOS_VERBOSE - set to 1 to run thamos in verbose mode to show what’s going on (verbosity on client side).

  • THAMOS_DEBUG - set to 1 to run analyzes (adviser, provenance checker, …) on Thoth’s backend side in debug mode, you can obtain logs by running thamos logs or directly on Thoth’s user API; the analysis id gets printed into the console during the build process in OpenShift (verbosity on server side).

  • THAMOS_DEV - set to 1 to consider also development dependencies, this flag defaults to 0 - by enabling development dependencies, adviser will need to browse larger space of software stacks possibly ending with a worse software stack advised (development dependencies are usually not used during application deployment)

  • THAMOS_DISABLE_CUDA - set to 1 to disable CUDA detection

  • THAMOS_NO_EMOJI - set to 1 to disable UTF-8 emojis (useful for dummy terminals)

  • THAMOS_NO_USER_STACK - set to 1 to disable sending lock file present in the directory - this lock file is used as a base when searching a better lock file for user needs

  • THAMOS_RETRY_ON_ERROR_COUNT - number of retries performed if the API server is responding with an error HTTP status (defaults to 3), this option is not usually needed to be adjusted

  • THAMOS_RETRY_ON_ERROR_SLEEP - sleep time when an error on the API server is spotted (see THAMOS_RETRY_ON_ERROR_COUNT), defaults to 3 seconds

  • THAMOS_NO_PROGRESSBAR - disable progress bar visualization, useful for dummy terminals

  • THAMOS_TIMEOUT - timeout period in seconds after which Thamos stops trying to fetch results

  • THAMOS_DISABLE_LAST_ANALYSIS_ID_FILE - set to 1 if you do not want to create a file that states last analysis id (used not to memorize the last analysis id across commands)

  • THAMOS_REQUIREMENTS_FORMAT - style of requirements used for managing dependencies - one of pip, pip-tools, pipenv, defaults to pipenv if not specified

  • THAMOS_TOKEN - token used for authenticated requests to the backend

See OpenShift s2i documentation on how to pin build to a specific node in the cluster. This is needed if you would like to perform automatic hardware discovery to get optimized stacks on your hardware.

Using Thamos as a library

from thamos.lib import image_analysis
from thamos.config import config

# Set global context.
# Host to Thoth's User API. API discovery will be done
# transparently and the most appropriate API version will be used.
config.explicit_host = ""
# TLS verification when communicating with Thoth API.
config.tls_verify = True

  # TLS verification when communicating with registry.

Autogenerated client from OpenAPI

Most parts of Thamos consist of automatic generated code. You can update Thamos by running the following command:

$ ./

The command above will download and run automatic code generation tool against the most recent OpenAPI specification of User API. Results of the tool are automatically placed into this repository in thamos/swagger_client/ and Documentation/. They consist of automatically generated code as well as documentation on how to use the code. Thamos itself provides routines built on top of this automated generated code to simplify usage in thamos/lib.

Thamos Python client


This documentation corresponds to implementation in version 1.28.1, documentation was generated on Oct 17, 2022.