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) I think you've already had id_rsa in .ssh directory. ssh-add to your agent. You should add your ssh agent to notify which folder you're going 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



  1. Before fetching l1 visualizer data from remote server, you should make some data. Run your model with this 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.

if you want remote access​

  1. First, click add_new_connection button.
  2. Fill up the pop-up.
  3. After making connections,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.

OR, download 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.