TechHuggingFace2 days ago
Profiling in PyTorch: From nn.Linear to Fused MLP
1 min read
The authors replaced a hand-written matmul-add pair with nn.Linear(bias=True) and stacked three layers with activations to form an MLP block. They used an NVIDIA A100-SXM4-80GB GPU to run the scripts. The post builds on Part 1's profiler trace analysis, covering CPU dispatch and torch.compile internals.
Level
Hype check
Tap to vote and see what everyone thinks.
