Skip to content

Latest commit

 

History

History
47 lines (30 loc) · 2.89 KB

FAQ.md

File metadata and controls

47 lines (30 loc) · 2.89 KB
Raw

FAQ & Troubleshooting Guide

Resetting an accelerator board

If you encounter a situation where a Tenstorrent chip appears to be unresponsive or is exhibiting unusual behavior, a software reset of the board might be a viable solution.

For a software reset on a single chip use : tt-smi -r 0 .

For more information on performing reset on multiple chips or other specifics visit TT-SMI Resets

If you need additional assistance, you can access a detailed explanation of all available command options by appending the help flag to the command like so: tt-smi --help or tt-smi -h.

For comprehensive insights and detailed instructions on utilizing the command line GUI, we invite you to explore the Tenstorrent System Management Interface (TT-SMI) repository on GitHub at TT-SMI homepage. TT-SMI serves as a versatile command-line utility tailored to streamline interaction with all Tenstorrent devices on host.

If the software reset fails to resolve the issue, the next step would be to power cycle the board. This typically involves rebooting the host machine that the Tenstorrent board is connected to. Please note that any unsaved work may be lost during this process, so ensure all important data is saved before proceeding

PermissionError on /tmp/*.lock files

If multiple users are running on a system with a shared Tenstorrent device, you may encounter a PermissionError: [Errno 13] Permission denied: '/tmp/*.lock'.

You would need to remove these files between user sessions.

Python pip install dependency resolution issues

Your Python environment's pip version, if different than required version, may have dependency resolution issues and not correctly install pybuda (tt-buda) or its dependencies. To avoid dependency resolution issues, check that you are using the supported and tested version of pip for your pybuda release. For example, pip==24.0 is required for the most recent releases. To avoid changing the system installation of pip we generally recommend using the venv installation method (see guide in https://github.com/tenstorrent/tt-buda-demos/blob/main/first_5_steps/1_install_tt_buda.md#python-environment-installation).

# check your pip version
pip --version
# install a specific version of pip, e.g. 24.0 required for pybuda >= v0.10.5
pip install pip==24.0

For reference pip install error messages for dependency resolution issues with an unsupported pip version can look like:

ERROR: Invalid requirement: 'torch@ https://download.pytorch.org/whl/cpu-cxx11-abi/torch-2.1.0%2Bcpu.cxx11.abi-cp38-cp38-linux_x86_64.whl; python_version == "3.8"'

or:

ERROR: Package 'networkx' requires a different Python: 3.8.10 not in '>=3.9'