entropix.ipnb session crashing

[danbri]

Trying to run this via Colab (TPU backend) since I'm on a GPU-poor chromebook. The notebook is crashing before I can get final cell to finish.

Last 10 entries in app.log:

Oct 10, 2024, 6:28:38 PM WARNING WARNING:root:kernel be25de06-202d-4903-9732-c75c13d58316 restarted
Oct 10, 2024, 6:28:38 PM INFO KernelRestarter: restarting kernel (1/5), keep random ports
Oct 10, 2024, 6:28:34 PM WARNING ERROR: Unknown command line flag 'xla_latency_hiding_scheduler_rerun'
Oct 10, 2024, 6:21:53 PM WARNING WARNING:root:kernel be25de06-202d-4903-9732-c75c13d58316 restarted
Oct 10, 2024, 6:21:53 PM INFO KernelRestarter: restarting kernel (1/5), keep random ports
Oct 10, 2024, 6:21:35 PM WARNING ERROR: Unknown command line flag 'xla_latency_hiding_scheduler_rerun'
Oct 10, 2024, 6:20:05 PM WARNING WARNING:root:kernel be25de06-202d-4903-9732-c75c13d58316 restarted
Oct 10, 2024, 6:20:05 PM INFO KernelRestarter: restarting kernel (1/5), keep random ports
Oct 10, 2024, 6:20:02 PM WARNING ERROR: Unknown command line flag 'xla_latency_hiding_scheduler_rerun'
Oct 10, 2024, 6:18:20 PM INFO Kernel started: be25de06-202d-4903-9732-c75c13d58316, name: python3

It got as far as emitting some of the canned prompting text:

<|begin_of_text|><|start_header_id|>system<|end_header_id|>
<antThinking>
You're absolutely right. I need to delve deeper into my actual thought processes, including the uncertainties, associations, and even potential biases that arise as I consider the query. My previous responses, while informative, didn't truly capture the nuanced, sometimes messy nature of cognition. I'll strive to provide a more authentic representation of my internal dialogue, including moments of doubt, tangential thoughts, and the process of refining ideas. This should result in a more genuine demonstration of LLM chain of thought, reflection, and self-correction.
</antThinking>

Which number is larger, 9.9 or 9.11?<|eot_id|><|start_header_id|>assistant<|end_header_id|>

<thinking>

It looks like the rest of the repo is updated more frequently. Should the notebook pull the latest code from the repo via git?

[danbri]

Nearby:

• https://x.com/dorialexander/status/1844444736251363463
• https://colab.research.google.com/drive/15h89YLod7PY0Xam0kV5BHQB_TD3nmwlo?usp=sharing

[Pclanglais]

Just to add on: tried to track the issue. Seems to happen while loading the tokenizer with transformer (??). Just using it once was enough to crash it.

[stillmatic]

maybe try this? https://github.com/stillmatic/entropix/blob/main/entropix_colab.ipynb

in Colab: https://colab.research.google.com/github/stillmatic/entropix/blob/main/entropix_colab.ipynb

uses line magic to pull in code. I can run the torch logic in a notebook locally with this code, if the link above doesn't work, then there might be a problem with the code or colab.