HN user
GPU is good at doing parallel task that in nature is the same. So it fit perfectly in CFR context. CFR is basically doing a lot of similar work across nodes. So if you implememnt a cuda kernel for regret matching, it will accelerate the solving process a lot.
I made this project to combine advantages of alpaca-lora and fastchat.
I opensourced everything, including the share GPT dataset I used for training, model ,and training script.
I hope this project would allow people with lower-end GPU have a chance to train chatGPT-like LLM.
It's opensourced at https://github.com/bupticybee/TexasSolver , supports Windows,MacOs and Linux.
I also made a online demo: https://colab.research.google.com/github/bupticybee/TexasSol...
And an youtube introduction for it:
Really fun stuff, perhaps you should use a computer screen instead of producing one
After two weeks of hardwork, TexasHoldemSolverJava 0.2.0 is complete. A gui is created. Now it's a fully functional solver. It supports texas holdem and shortdeck. Most important, It's totally free and opensourced (piosolver costs hunderds of dollars). I really hope this project can bring some small change to the texas holdem community.
To be simple, python is not fast enought for solving, java is not fast enough for coding. So my solution is build prototype with python and transfer the code to java.
Wow. Thank you for your support
For now I still struggle to make it work in two player's preflop (take a ton of time and computation resources). In the future if a lot of people ask for multiplayer I think I will write the code~
My idea is the free version has some drawbacks: (1) people don't want to pay will stick with this free version (2) it's like a sample pack for solver, you may download the free version and say "it's not so good after all". Bad for business (3) it increase the probability of been hacked
The reason is simple, I'm not exactly a full time java engineer, as a matter of fact, I write python 80% of my time, all of this project's prototype is written in python in another project. I'm slowly transfering some of the code to java, for now the tree construction part is still in python.
The author of piosolver stop providing the free version for obvious reasons.
Yeah I am aware of this project, and I am current trying write an alpha-star like AI for another holdem simulator https://www.pettingzoo.ml/classic, I already made some progress and I somehow believe the code for a alpha-star like texas-holdem AI will be released in sometime of the future.
Piosolver is so expensive (pro version is about $1000) so I decide to write a opensource,free version of it. I already check it's preformance and result with piosolver and I believe it's a good competitor.