OSSForks/tensorflow
1
0
Fork 0
mirror of https://github.com/zebrajr/tensorflow.git synced 2026-01-15 12:15:41 +00:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

Disable capturing of dot RHS operands

Browse Source 
This is proving to be unreliable.

PiperOrigin-RevId: 826395008
This commit is contained in:
A. Unique TensorFlower
2025-10-31 02:22:48 -07:00
committed by TensorFlower Gardener
parent d6d4e02248
commit 8734ec41d5
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
third_party/xla/xla/service/cpu/thunk_emitter.cc vendored
View File

@@ -1553,9 +1553,9 @@ absl::StatusOr<ThunkSequence> ThunkEmitter::EmitYnnFusionThunk(
absl::Span<const int64_t> captured_arguments_ids;
if (instruction->opcode() == HloOpcode::kDot) {
const HloDotInstruction* dot = Cast<HloDotInstruction>(instruction);
// TODO(ashaposhnikov): Revisit this if we ever get a reliable way
// to determine that RHS is constant.
bool capture_rhs = HloPredicateIsOp<HloOpcode::kParameter>(dot->operand(1));
// TODO(b/455903737): If we know the RHS is a constant, we should capture it
// here.
bool capture_rhs = false;
// Construct YNNPACK subgraph builder from the dot instruction.
TF_ASSIGN_OR_RETURN(builder, EmitYnnDotBuilder(dot, capture_rhs));
static constexpr int64_t kCapturedIds[1] = {1};