Add fp16 support #38
Add fp16 support #38
Conversation
| assert isinstance(output, tuple) | ||
| for i, v in enumerate(op.outputs): | ||
| value_map[v] = output[i] | ||
| if torch.any(torch.isnan(output[i])): |
There was a problem hiding this comment.
I'd put these under a debug flag to avoid slowing down executions
| args.dram_bandwidth = simulation_parameters["dram_bandwidth"] | ||
| args.kernel_launch_overhead = simulation_parameters["kernel_launch_overhead"] | ||
| args.device_throughput = 1.0 / simulation_parameters["device_parameters"][0] | ||
| args.dram_bandwidth = 1.0 / simulation_parameters["device_parameters"][1] |
There was a problem hiding this comment.
Won't this become infinity if one of the regression coefficients is 0?
There was a problem hiding this comment.
Yea but I think that's ok for now since we don't see much utility from the dram bandwidth at sufficiently large data sizes right?
There was a problem hiding this comment.
When I ran it before Tue's meeting, it threw a RuntimeError for dividing by zero (not sure if it was from this line though), so maybe safer to store the parameters if you run into that error again.
| "--use_gpu", | ||
| action="store_true", | ||
| default=torch.cuda.is_available(), | ||
| default=False, |
There was a problem hiding this comment.
Why did you change the default?
There was a problem hiding this comment.
There's no way to only use CPU otherwise, but maybe we should just make this --use_cpu instead
No description provided.