DeepSpeed icon indicating copy to clipboard operation
DeepSpeed copied to clipboard

Prevent hangs in CI during parallel run compilation

Open mrwyattii opened this issue 2 years ago • 0 comments

We are seeing random hangs, which we believe are caused by multiple pytest processes trying to compile the same code at once creating a deadlock. This PR sets a seperate TORCH_EXTENSIONS_DIR for each pytest process to avoid this.

@tjruwase

mrwyattii avatar Feb 16 '23 18:02 mrwyattii