skyline
skyline copied to clipboard
skylineprof
🏙 Interactive in-editor performance profiling, visualization, and debugging for PyTorch neural networks.
Right now the `skyline_*_provider()` functions are quite complicated. We adopted these providers to help with our profiling needs. But with some creativity, we may be able to simplify these interfaces...
Right now the server disconnects the plugin if they cannot agree on a protocol version to use. The plugin just shows a generic "connection lost" message. A better user experience...
Profiling large models can be slow (~10 - 20 seconds). We should determine which parts of the profiling pass are slow and then look for ways to speed it up.
We currently use milliseconds for time measurements and store the measurements in floating point values. However, when we perform arithmetic on these values, they are susceptible to numerical error. Our...
For convenience, the analysis code runs in the same process as the Skyline server. However, the analysis code executes user code as well. Running user code in the main Skyline...
We need to update our tests for the server code: - Iteration profiler - Message handler - Connection message de/serialization logic
We use 32-bit unsigned integers to represent a message's sequence number. We assume that the sequence number increases monotonically. However when we reach 2^32-1, the next sequence number will be...
We should add tests for important plugin code: - `file_tracker.js` - `io/connection.js` - Iteration profiler in the CLI (move this to a separate issue) - [ ] Add initial tests...
When we update protobuf files, we need to ensure that the developer does not forget to check in the updated generated code. One way to accomplish this could be to...
Metadata
Owner
Metadata
🏙 Interactive in-editor performance profiling, visualization, and debugging for PyTorch neural networks.