Flow and Diffusion models Lab 3 (Mac GPU and backup w pickle)
Diffusion Lab3, conditional generation
Posts by Tag
LLM
Flow and Diffusion models Lab 1 (and Jupyter installation)
Reviewed the diffusion courses and take notes for Lab1
Flow and Diffusion models Lab 2-2 - Gaussian Probability Path w Score matching
I separat the score matching part from the Lab 2 and all put into this blog. So we can see more clearly how it works
Deterministic in LLM
Deterministic in LLM is a much harder problem than I initially thought. I was working with a major customer SN for a couple of months for this issues and eve...
Flow and Diffusion models Lab 2-1 - Gaussian Probability Path w Flow matching
1 Gaussian Conditional Probability Path Gaussian conditional probability path is given by \(p_t(x|z) = N(x;\alpha_t z,\beta_t^2 I_d),\quad\quad\quad p_{\text...
Flow and Diffusion models Part 3 - Langevin and Matching
Lecture 3. Finished Lab 1, implemented flow and diffusion models and implemented Langevin Dynamics
Mixture of Recursions
EZ’s talk about Mixture of Recursions
DPO - Direct Preference Optimization
Continue with another blog from Cameron and greate explanation and super easy to digest math details
Reward Model
Read some latest publish from Cameron’s blog and got some knowledge refresh for Reward model, DPO (will be my next blog) and review on PPO (like always)
Mixin and super()
When working with Huggingface, there are lots of classes with suffix Mixin. I never thought about it’s actually meaning, but turn out to be an important feat...
VLM
Flow and Diffusion models Part 2 - Fokker-Planck
Lecture 2 This is the most mathmatical chanllenging part of the lecture, but also covers all the missing knowledge I would like to learn, like Langevin and F...
Flow and Diffusion models Part 1 - Flow and Diffusion
Will start to review Diffusion series from 苏神 blog. But then I realized that SDE, Langevin and Flow models, are things taught in statiscal mechanics which Im...
AniSora
Got back from a short vacation and jump back into animation generation
Kijia and HiDream, VACE
I am really into ComfyUI these days and try out couple of more models and pluggins
ComfyUI
It’s time for a bit of fun after learning and coding with VLM for a while. ComfyUI and SB WebUI both has been very matured and quite some new models in the G...
Video support in Nemontron Nano VL
The PR for Nemotron Nano VL is still on going. There are more modifications as below (Finally merged after I just finished this blog) 1 InternVL’s video exte...
Pytest Debug and vLLM Processing
Debug Pytest for vLLM and get a closer look at the processing workflow. So working from the test case is a much clearly way to understand the workflow. 1 Pyt...
VLM Nemotron-Nano-VL-8B support
I started working on this a month ago since Nemotron-Nano-VL-8B-V1 released. I thought it would be a good choice for me to add a model from scratch, and get ...
Ray/Anyscale
Ray continue on Two H200x8 nodes
Issues and fixes GPU limitations Setting export CUDA_VISIBLE_DEVICES=0,1,... at the top of _start_ray.sh does NOT work Instead, add -e CU...
KubeRay Setup and Trouble shooting
Trying to get a 2 node 8 GPU (4 GPU from each node) Ray cluster running with KubeRay
Multi-node vLLM on Ray cluster
It’s kind of funny that I barely started a Ray cluster without Anyscale. It actually a comlicated process but most current Enterprise are not aware of. Mainl...
Python
import, uv, and module/package in Python
A really good video talks about morden Python project structures and tools like uv/poetry. Also explained what exactly import does
Decorators and some little things
Learn a little bit of Python from time to time.
Misc
Quantum Computing
I knew that Quantum Computing is NOT about parallelism but never truely understand how it works. Thanks for 3Blue1Brown’s great explanation video here, I fin...
VAE and MLE
Now Im sure that all my VAE and ELBO related learning were NOT recorded in this blog. So it’s time to review some courses from Hung-yi.
Diffusion
DDIM and Classifer-free Guidance
I looked up in my blog, and there is no record of VAE, ELBO, or flow models. These are the contents I spent quite some time on, but it’s right before I start...
Finance
Something about IRA
I’m starting a new series, of Finance. Looked into IRA, backdoor, and mega backdoor. These CPA domain knowledge shouldn’t be too hard for STEM background ppl...
Math
Huhugui
Read a middle school math problem, which is called 胡不归。The purpose is to find shortest travel time between points with different velocity on different segmen...