Fp8 hipstream fix #3127
Fp8 hipstream fix #3127
Conversation
❌ Deploy Preview for pytorch-fbgemm-docs failed.
|
acoskunses-AMD
force-pushed
the
fp8_hipstream_fix
branch
2 times, most recently
from
0c9fc22
to
863f596
Compare
|
We use hipify script for this, is the change here needed? |
|
Rebase to resolve conflicts: also as Xiaodong mentioned, internally we don't need this (automatically hipify). Is this needed on OSS workflow? I guess hipify torch is not enabled? |
acoskunses-AMD
force-pushed
the
fp8_hipstream_fix
branch
from
c948fe8
to
e94d9e9
Compare
| @@ -12,10 +12,13 @@ | |||
| #include <numeric> | |||
|
|
|||
| #include <ATen/ATen.h> | |||
| #if !defined(USE_ROCM) | |||
There was a problem hiding this comment.
hipify script should hipify .h right?
There was a problem hiding this comment.
in any case, I think maybe we can just remove this block just to be clean.
| @@ -12,11 +12,14 @@ | |||
| #include <numeric> | |||
|
|
|||
| #include <ATen/ATen.h> | |||
| #include <c10/hip/HIPStream.h> | |||
| #if !defined(USE_ROCM) | |||
There was a problem hiding this comment.
is this block even needed? If USE_ROCM is not defined, it'll basically be an empty file. So we can just remove this block here.
There was a problem hiding this comment.
Let's just remove this include and move it down to the USE_ROCM. No need to include cuda header here.
There was a problem hiding this comment.
changed as required. Please review.
| @@ -12,10 +12,13 @@ | |||
| #include <numeric> | |||
|
|
|||
| #include <ATen/ATen.h> | |||
| #if !defined(USE_ROCM) | |||
There was a problem hiding this comment.
in any case, I think maybe we can just remove this block just to be clean.
acoskunses-AMD
force-pushed
the
fp8_hipstream_fix
branch
3 times, most recently
from
ee380a4
to
05e9b76
Compare
|
@xw285cornell has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
| @@ -201,4 +199,3 @@ at::Tensor f8f8bf16_rowwise_impl( | |||
| return Y; | |||
| } | |||
|
|
|||
There was a problem hiding this comment.
@acoskunses-AMD can you remove the blank lines at the end of each file? Otherwise we cannot land it internally due to some lint check
acoskunses-AMD
force-pushed
the
fp8_hipstream_fix
branch
from
05e9b76
to
1e4e129
Compare
|
@xw285cornell has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@xw285cornell merged this pull request in 8e7beba. |
Pick hip stream where cuda is not defined