there are a lot of benchmarks that suggest 5.6 sol is the best model in the world right now, but the most reliable way to tell is that elon is obsessed with me again
Don't expect motivation to carry you. Motivation comes and goes. Systems stick around. Design your environment so the right choice is the easiest one. The less you have to rely on self-control, the more consistent you'll be.
Day 4/90 of Inference Engineering
I built a CUDA kernel for matrix addition. The objective was to write correct CUDA while instilling the right intuition.
Mainly putting the reps in with allocating memory on the host and device and writing the actual kernel through the lens of writing a thread instead of a sequential program.
I learned that the idea of blocks and grids are just a software abstraction. However at launch time, grids are already partitioned into blocks. This allows the global block scheduler to schedule blocks and pin them to a SM for a lifetime. Without this abstraction we would have to schedule each thread individually for execution.
I also learned that the block abstractions are what allows threads to cooperate cheaply within a block. Since threads within the same block share the same shared memory, this gives threads visibility into each other's data, and __syncthreads__() synchronizes completion for every thread within the block.
Very interesting ideas. Love to hear a Kernel Engineer shed some light on this topic~
On the hardware side, I previously thought that the L1 Cache and Shared Memory were interchangeable terms. Turns out, they are two unique hardware components. I also forgot to include polymorph, TCP, Raster Engine within my old diagram. I attached an image below of the updated diagram that I drew.
I also went over some CUDA documentation to understand the syntax of cudamalloc(). Today was a great day, on to the next!
Leetcode interviews for ML engineers/researchers doesn't make any sense. Oftentimes, it just tests how much you memorized Blind 75.
At the same time, a lot of proclaimed "better" ML interviews are really not much better...
Finding 3 bugs in this transformer, implementing a tokenizer, implementing backprop, etc. are also just testing how much you memorized Karpathy tutorials.
IMO, my ideal interview would be something like:
- Given a wandb run where training became unstable
- You have to look through logged metrics and figure out why
- Make a hypothesis on why training collapsed
- Explain/implement a solution
Especially with coding agents, the most important skill is identifying what the right problem/solution to even work on is, rather than knowing rote syntax for pytorch operations.
No amount of admiring an activity will make you good at it. Whether it's math, music, coding, or anything else: you have to do the exercises the same way athletes have to train, not just watch the game.
@Sabj00@pakistanwalli Russia is done
China is very passive.
The only Strategic hurdle US is facing rn is 'strait of Hormuz'.
There is no other immediate challenge to US for now.
@pakistanwalli@JDVance And you think Powerful Zoinists Investors will back his campaign.
Your Analysis does not sit right to me from last couple of weeks.
It's more more like you are sharing your personal thoughts.