experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Disable functionalization of v1 graphs during import.

Browse Source 
Running functionalization during import would run it on the entire graph (instead of only the TPU partition) is not semantics preserving. Disabling it until functionalization is fixed to be able to filter the nodes and process only the TPU nodes in the graph.

PiperOrigin-RevId: 303406934
Change-Id: I35a5e728ac3f54a851103439721beefb09878ad3
This commit is contained in:
Prakalp Srivastava 2020-03-27 14:26:20 -07:00 committed by TensorFlower Gardener
parent 6d732cee5b
commit 0a4a6893a8
1 changed files with 3 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc
View File

@ -197,7 +197,9 @@ Status MlirV1CompatGraphOptimizationPass::Run(
RegisterDialects();
mlir::MLIRContext context;
GraphImportConfig import_config;
import_config.upgrade_legacy = true;
// TODO(b/150959075): Running functionalization before TPU cluster formation
// is not semantics preserving and should be disabled for now.
import_config.upgrade_legacy = false;
TF_ASSIGN_OR_RETURN(
auto module_ref,
ConvertGraphToMlir(**options.graph, debug_info, *options.flib_def,