-
Notifications
You must be signed in to change notification settings - Fork 94
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Develop upstream sync 230731 #2170
Conversation
…in global or local view. If the attribute is set on a CallOp, then verification logic converts the programs arguments and results from local view to global view to verify that local view shape + sharding is equivalent to the expected global view shape. PiperOrigin-RevId: 551222813
… wheel. PiperOrigin-RevId: 551224665
Updates LLVM usage to match [365d6eb1f7d8](llvm/llvm-project@365d6eb1f7d8) PiperOrigin-RevId: 551229328
…ion/configuration out of experimental. PiperOrigin-RevId: 551235514
PiperOrigin-RevId: 551238120
Also fix typo in SetAllowBufferHandleOutput comment: false->true. Also fix #include order to match style guide. PiperOrigin-RevId: 551247708
PiperOrigin-RevId: 551261650
PiperOrigin-RevId: 551275563
PiperOrigin-RevId: 551276374
`TF_STATUS_ASSIGN_OR_RETURN` and `TF_STATUS_RETURN_IF_ERROR` PiperOrigin-RevId: 551278625
PiperOrigin-RevId: 551290442
PiperOrigin-RevId: 551295403
…_heuristics PiperOrigin-RevId: 551297374
This CL will add patterns to fold Transpose and FC to covert into a BMM, like below- FC(lhs, Transpose(rhs)) -> BMM(lha, rhs, false, false) The right thing to do in this pattern will be to apply the pattern only if keep_num_dims==True. Because, if the output rank is less-than the input rank, it means `keep_num_dims` has reduced the output. But checking for rank will improve the coverage. This pattern will now work PiperOrigin-RevId: 551297769
PiperOrigin-RevId: 551313932
To improve debuggability, we want the shape refinement to make as few changes as possible to the module. In this change we remove one use of inlining. PiperOrigin-RevId: 551325242
PiperOrigin-RevId: 551347216
PiperOrigin-RevId: 551353292
…ac compiler error Apparently ssize_t is only a long sometimes (at least 32-bit), instead of long long (at least 64-bit). I don't have a mac so I can't repro the failing build, but hopefully this fixes it based on the error message. PiperOrigin-RevId: 551376003
PiperOrigin-RevId: 551401683
PiperOrigin-RevId: 551408554
PiperOrigin-RevId: 551410772
PiperOrigin-RevId: 551413305
The BFS algorithm didn't have a visited set, therefore had a complexity of O(N*E). PiperOrigin-RevId: 551414282
…sion pattern. This change implements a conversion pattern that converts stablehlo.convolution to tfl.conv_2d. This is a minimal version that converts quantized `stablehlo.convolution` with certain assumptions like that the filter has the format of `[0, 1, i, o]`. PiperOrigin-RevId: 551419638
…ll on SplitShardingDimension. PiperOrigin-RevId: 551438695
This is in preparation for another change improving the state of copies in while loops. PiperOrigin-RevId: 551451818
PiperOrigin-RevId: 551452964
…end-Recv sequence. This is to prevent the latency hiding scheduler to interleave two Send-Recv sequences. PiperOrigin-RevId: 552621536
PiperOrigin-RevId: 552621643
InitializeCreateGcsFileSystemFnPtr is a temporary fix and it is no longer needed. PiperOrigin-RevId: 552624923
…nalysis PiperOrigin-RevId: 552625083
PiperOrigin-RevId: 552626000
PiperOrigin-RevId: 552631765
This removes some unnecessary `cuDeviceGetCount()` calls when custom ops are used. PiperOrigin-RevId: 552634342
…c in tf.constant according to auto dtype conversion semantics. WeakTensor is created if it satisfies both of the following conditions: 1. tf.constant is called with no dtype arg specified. 2. Input is a nested Python type. PiperOrigin-RevId: 552634845
…mul. PiperOrigin-RevId: 552636662
…nd 0 of a gather, assume that the sharding of that operand does not matter. PiperOrigin-RevId: 552637713
Since I did the cover here https://github.com/openxla/xla/pull/4603/files#diff-fc02eb6aea06ad0d72011e9a64da28e8c6fae000a9e0cab9b15b40b4e914af4aR956-R958 could you try to enable triton-softmax flag to have a go, please? we need to see are any other unit tests affected on that flag. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Since gpu.graph PR is merged, you should enable xla_graph_level = 1
turn it on in the latest commit. |
turn it on in the latest commit. |
e1a2099
to
5ec1e3e
Compare
5ec1e3e
to
3131c3a
Compare
Jenkins: retest Ubuntu-GPU-single please |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This is added in tensorflow@56f261b we need a ticket to track this test
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Sure.
No description provided.