Default MIOPEN_FIND_MODE=FAST for amd instead of disabling cudnn#10448
Default MIOPEN_FIND_MODE=FAST for amd instead of disabling cudnn#10448alexheretic wants to merge 1 commit intoComfy-Org:masterfrom
Conversation
|
This doesn't really improve performance so much as it reduces initial latency. Running I think setting fast by default if it's unset might still be a good idea though, but you might want to give a hint to people about trying hybrid as well if they want to optimize performance. |
There is already a log line about setting this, I think we could add more info there. Perhaps a link to the miopen docs. But maybe it's enough just mentioning the env var, users can search for it easily enough. The crux of this PR is that cudnn is currently disabled because of the impact on initial runs. I don't have this issue myself on rocm 6.4 and disabling cudnn significantly reduces ImageUpscaleWithModel performance (#10447) for me. This PR is an attempt at a better workaround. |
|
I tried this, it doesn't change anything on my setups. |
|
Ok fair enough. Do you have a rocm 6.4 test case? Perhaps we can enable cudnn for that version? |
Defaulting
MIOPEN_FIND_MODE=FAST, when this env var is not already set, may fix warmup slowness on AMD without the downside of disabling cudnn fully which has negative perf effects on rocm 6.4.This is also easy for the user to opt out of, since they can simply set
MIOPEN_FIND_MODEthemselves.@comfyanonymous can you check if this works as well as disabling cudnn for your test scenarios?
Resolves #10447