Tom Tseng
Tom Tseng
### What happened + What you expected to happen I am able to reproduce on a macOS instance in CircleCI that `import ray` gives the following error when I install...
Here are some adjustments to the interface that would make this more useable for other researchers who want to run or benchmark this. (In the meantime, hopefully this helps anyone...
# Issue When evaluating an input using the TensorRT backend in `match` or `selfplay`, I sometimes hit a `Got nonfinite for policy sum` error that I never hit when using...
## Bug description Suppose we are running `match` with 3 bots called bot0, bot1, and bot2. bot0 and bot1 use the same model file. Due to model file deduping in...
Adds script for generating heat map visualizations of adversaries' cyclic group shapes. Maybe belongs in KataGoVisualizer, except we need to use `go_attack.go.Game` to simulate each game and find the cyclic...
Our bots are only trained with Tromp-Taylor rules, but KGS doesn't have Tromp-Taylor rules. If a bot plays with non-Tromp-Taylor rules, then because the rules are features of the NN...
KataGo-raw is still on v1.12, which is older than KataGo-custom. If we run GTP evals we should update KataGo-raw either to the latest version (v1.14.1) or to match KataGo-custom (v1.13.0)
* Increase game threads 128->256: I had found [several months back](https://www.notion.so/chaiberkeley/debugging-victimplay-throughput-being-lower-than-expected-bc376db69a5a4343bd6c609dcec1ca4b?pvs=4#82543a2589c348d3ae3940584fd72abf) that 256 game threads had a little more throughput than 128 threads * Increase server threads to match the...
Add lightvector's cyclic position evaluation script + games, with the following modifications to the script: * Linting * Change some hardcoded values into CLI arguments (config, model, katago executable, number...
## Question I'm seeing very large gradients when I use flash attention 2 + bfloat16 on Pythia-14m. I'm loading my models with `transformers`. Steps: * Load `EleutherAI/pythia-14m` * Run it...