Research
34 min read
A Tinyblog about Tinygrad
A deep dive through Tinygrad's compiler and runtime — rangeify, memory planning, beam search — and where its GPU-shaped assumptions start to bind.
Pavle Padjin
Blog
Research
6 min read
Day 0 to Day 1: DeepSeek V4 Flash from 340 to 2,910 tok/s/GPU on MI355X
The weights dropped Friday. By Sunday, our harness had beaten weeks of hand-tuning, dead ends included.
Momcilo Mrkaic, Vladimir Zeljkovic
Notes from the frontier of inference.
- Product
1 min read
Hoid now targets Huawei Ascend 950DT
Fifth hardware world, same one-line command.
Vladimir Zeljkovic
- Research
1 min read
The optimization space is 10¹⁷ wide. Here's how we search it.
Tiling, layout, fusion, scheduling and precision choices multiply fast.
Vladimir Zeljkovic
- Research
1 min read
Stopping an agent from shipping a fast, wrong kernel
An agent optimizing for tok/s will happily break numerics to get there.
Vladimir Zeljkovic
- Company
1 min read
Why we started Hoid
Optimization expertise is scarce, non-transferable and perishable. It shouldn't be.
Momcilo Mrkaic, Pavle Padjin, Vladimir Zeljkovic
- Research
1 min read
Attention is 8% of the FLOPs and 41% of the wall clock
A layer-by-layer roofline of five frontier MoE models across four accelerators.
Pavle Padjin, Vladimir Zeljkovic
- Research
1 min read
Hoid Runtime 0.4: continuous batching, disaggregated prefill, and a scheduler you can read
The runtime is 11k lines with no vendor SDK in the hot path.
Pavle Padjin, Vladimir Zeljkovic
- Research
1 min read
We let the harness redesign our MoE dispatch. It removed the all-to-all.
Decode latency cut 34% at EP=32. We didn't suggest it, and for two days we didn't believe it.
Vladimir Zeljkovic
- Product
1 min read
Performance regression CI for inference: catch the 6% you'd otherwise ship
Kernel changes, driver bumps and framework upgrades quietly cost throughput.
Momcilo Mrkaic
- Research
1 min read
FP4 is not one format: MXFP4, NVFP4, and the accuracy cliff nobody benchmarks
We quantized GLM-5 six ways and measured both the Pareto curve and the eval damage.
Momcilo Mrkaic, Pavle Padjin
- Company
1 min read
Hoid raises to make every accelerator worth buying
We're backed by a16z speedrun. What we're spending it on and what we're deliberately not building.
Momcilo Mrkaic