Herdora recently launched!

‍

^{"Profile your inference pipeline in < 60 seconds with one line of code"}

^{‍‍
TL;DR Herdora is reverse engineering GPUs to give ML engineers the profiling tools they actually need. Cut inference latency by 50%+ with one line of code.}

‍

https://www.youtube.com/watch?v=KEZHky0Xexk

‍

Founded by Steven Arellano & Emilio Andere ‍

‍

Today, Herdora is releasing Keys & Caches. They aim to solve one of the most frustrating problems in ML infrastructure: you can't see why your model is slow.

‍

🔥 The Problem

If you're running any ML models in production, you know the pain:

• Your inference is inexplicably slow but existing profilers give you walls of incomprehensible data
• You're burning through GPU budget without knowing why
• You miss SLAs because you can't find the actual bottlenecks
• torch.profiler either overwhelms you with noise or misses the real issues entirely

‍

⚡ Their Solution

The team is reverse engineering NVIDIA GPUs to understand how they execute ML workloads. With Keys & Caches:

1. Add one decorator to your code
2. Get clear, actionable traces showing exactly where time and memory go
3. Drill down from Python to CUDA to PTX - see every layer of the stack
4. Find and fix bottlenecks in minutes, not days

Here's what it looks like in action.

They have already helped a team optimize their Llama deployment and cut latency by 67% by identifying a single overlooked kernel that was eating 40% of runtime. Read the full case study.

‍

Learn More

^‍

^{🌐 Visit www.herdora.com to learn more.}

^‍

^{📣 If you're scaling inference-heavy workloads: Try Keys & Caches free - Get 10 hours of profiling credits for FREE 💸💸💸, no credit card required!}

^‍

^{👉 Book a 20-min demo if you want to see it on your actual workload.}

^‍

^{⚙️ For the GPU-curious: Check out their deep dives on GPU internals and optimization techniques}

^‍

^{🤝 Reach out directly to the founders here.}

^‍

^{👣 Follow Herdora on LinkedIn & X.}

‍