Optimize flash bert path for hpu device #509
Conversation
Signed-off-by: kaixuanliu <[email protected]>
Signed-off-by: kaixuanliu <[email protected]>
Signed-off-by: kaixuanliu <[email protected]>
|
This PR optimized the performance of flashBert path for HPU device, with this optimization, the mean latency drops from 6.4 ms to 4.32 ms, which finally aligns with the perf of tei-gaudi. |
Signed-off-by: kaixuanliu <[email protected]>
|
Seems you changed modeling which cover other devices, do you validated GPU, CPU, XPU? what's the performance? |
|
For CPU and XPU device, I just passed 2 extra args to calc attention, and these 2 args are only used in hpu_attn calculation. The other changes is just replace torch.addmm to F.linear, which I suppose there should be no perf difference. I validated the output correctness of CPU. I will double check the perf of both CPU and XPU and output of XPU. |
|
Have double checked the output correctness of XPU devices and perf of both CPU/XPU, no change compared with original implementation. |
| @@ -15,6 +15,7 @@ | |||
| __all__ = ["Model"] | |||
|
|
|||
| TRUST_REMOTE_CODE = os.getenv("TRUST_REMOTE_CODE", "false").lower() in ["true", "1"] | |||
|
|
|||
There was a problem hiding this comment.
Let's remove this new blank line as it's the only change in the file
There was a problem hiding this comment.
Oops, have fixed it.
Signed-off-by: kaixuanliu <[email protected]>
No description provided.