📅 September 26, 2026
*Why I started it, what it does today, and where I’d love some help. · 5-minute read*
<figure style="margin: 1.5rem 0; display: flex; flex-direction...
cat 2026-09-26_megabake_launch.md → Read
~/blog
$ ls -la blog-posts/
megabake_launch.md
September 26, 2026
torch-compile-end-to-end.md
September 2, 2026
torch-compile-in-the-wild.md
August 12, 2026
vllm-torch-compile-and-ir.md
April 16, 2026
torchdynamo-python-to-fx-graph.md
March 20, 2026
vllm-and-torch-compile.md
January 16, 2026
vllm-pytorch-meetup.md
December 20, 2025
vllm-architecture-deep-dive.md
October 8, 2025
$ cat blog-posts/*
📅 September 2, 2026
<!-- Source presentation: torch.compile e2e.pdf. Date is the blog publication date. -->
I wanted to connect the parts of `torch.compile` that are of...
cat 2026-09-02_torch-compile-end-to-end.md → Read
📅 August 12, 2026
<!-- Source presentation: torch.compile in the wild.pdf. Date is the blog publication date. -->
The interesting question about `torch.compile` is of...
cat 2026-08-12_torch-compile-in-the-wild.md → Read
📅 April 16, 2026
<!-- Source presentation: vLLM, torch.compile and IR.pdf. Date is the blog publication date. -->
vLLM already uses `torch.compile`, so why add anoth...
cat 2026-04-16_vllm-torch-compile-and-ir.md → Read
📅 March 20, 2026
<!-- Source presentation: March meetup talk.pdf. Date is the blog publication date. -->
Calling `torch.compile` takes one line. Understanding how it...
cat 2026-03-20_torchdynamo-python-to-fx-graph.md → Read
📅 January 16, 2026
<!-- Source presentation: vLLM & torch.compile.pdf. Date is the blog publication date. -->
After looking at vLLM's scheduler and KV cache, I wanted ...
cat 2026-01-16_vllm-and-torch-compile.md → Read
📅 December 20, 2025
<!-- Source presentation: vLLM presentation - Pytorch meetup.pdf. Date is the blog publication date. -->
For this PyTorch meetup talk, I started wit...
cat 2025-12-20_vllm-pytorch-meetup.md → Read
📅 October 8, 2025
<!-- Source presentation: vLLM presentation.pdf. Date is the blog publication date. -->
Running one prompt through an LLM is fairly easy. Keeping a ...
cat 2025-10-08_vllm-architecture-deep-dive.md → Read