Skip to main content

Memory Visualizer

Memory visualizer (a.k.a L1 visualizer), Tenstorrent Visualizer, is a comprehensive tool for visualizing and analyzing model execution, offering interactive graphs, memory plots, tensor details, buffer overviews, operation flow graphs, and multi-instance support with file or SSH-based report loading.


Installation guide

Prerequisite​

You need a Linux environment.

Installation​

  1. Go to your virtual env in the tt-metal folder

  2. pip install ttnn-visualizer

  3. (optional) You should add your SSH agent, ensuring it is configured for the target folder you plan to use.

     eval $(ssh-agent)
     ssh-add ~/.ssh/id_rsa

  4. Run $ ttnn-visualizer and go to 127.0.0.0:8000

How to Open Project



Before fetching l1 visualizer data from remote server, run your model with these environment variables.

    export TTNN_CONFIG_OVERRIDES='{
        "enable_fast_runtime_mode": false,
        "enable_logging": true,
        "enable_graph_report": false,
        "report_name": "<name of the run in the visualizer>",
        "enable_detailed_buffer_report": true,
        "enable_detailed_tensor_report": false,
        "enable_comparison_mode": false
    }'

Then, directory "/to/your/tt-metal/generated/ttnn/reports" is created. You're going to fetch this.

Remote Access​

  1. First, click add_new_connection button.
  2. Fill up the pop-up.
  3. After establishing the connection, click fetch folder list.
  4. Folders are fetched, but the data aren't transferred. Click recycle button next to your folder name. That will fetch data inside the folder.
  5. See your reports.
  6. Your data is stored in ./virtual_env/lib/python3.10/site-packages/ttnn_visualizer/data/remote.
  7. Those reports can be shown by clicking Select local report button.

Download the report folder from remote server, and run in localhost​

  1. $ ttnn-visualizer --profiler-path [DOWNLOADED GENERATED FOLDER]
  2. go to 127.0.0.0:8000 and then just use.

How to Use

In the menu at the top, you can observe the memory area through [Operations, Tensors, Buffers, Graphs], and you can usually check the L1 area for each operation by pressing [Memory details] after each Op in the [Operations] menu.