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

Allow more sharding propagation from broadcast operands

Browse Source 
MaybeImproveInstructionSharding() already checks that it needs to be compatible with existing tiled sharding (if any).

PiperOrigin-RevId: 356375771
Change-Id: Ic5e65e00cea0a25cec2cf93027632c4c7783a20e
This commit is contained in:
Yuanzhong Xu 2021-02-08 16:08:37 -08:00 committed by TensorFlower Gardener
parent 324ab95fb0
commit 18f2dd4262
2 changed files with 31 additions and 5 deletions

5
tensorflow/compiler/xla/service/sharding_propagation.cc
View File

@ -739,11 +739,6 @@ bool InferShardingFromOperands(HloInstruction* instruction,
if (aggressiveness < 3) { if (aggressiveness < 3) {
return false; return false;
} }
// Do not override existing tile sharding. This is likely from users.
if (IsSpatiallyPartitioned(instruction) &&
!instruction->sharding().IsTileMaximal()) {
return false;
}
const HloInstruction* op = instruction->operand(0); const HloInstruction* op = instruction->operand(0);
if (!IsSpatiallyPartitioned(op) || op->sharding().IsReplicated()) { if (!IsSpatiallyPartitioned(op) || op->sharding().IsReplicated()) {
return false; return false;

31
tensorflow/compiler/xla/service/sharding_propagation_test.cc
View File

@ -392,6 +392,37 @@ ENTRY %broadcast {
} }
} }
TEST_P(ParameterizedMetadataTest, BroadcastMerge) {
const char* const hlo_string = R"(
HloModule module
ENTRY %broadcast {
%param0 = f32[3,2048]parameter(0),
sharding={devices=[1,2,2]0,1,2,3 last_tile_dim_replicate metadata={op_name="a"}}
%broadcast = f32[3,2048,3] broadcast(%param0), dimensions={0,1}
ROOT %copy = f32[3,2048,3] copy(%broadcast),
sharding={devices=[1,1,2,2]0,2,1,3 last_tile_dim_replicate metadata={op_name="b"}}
})";
TF_ASSERT_OK_AND_ASSIGN(auto module,
ParseAndReturnVerifiedModule(hlo_string));
if (GetParam().clear_metadata) {
ClearMetadata(module.get());
}
TF_ASSERT_OK_AND_ASSIGN(
bool changed,
ShardingPropagation(/*is_spmd=*/true, GetParam().propagate_metadata)
.Run(module.get()));
EXPECT_TRUE(changed);
auto* instruction = FindInstruction(module.get(), "broadcast");
ASSERT_NE(instruction, nullptr);
EXPECT_THAT(instruction, op::Sharding("{devices=[1,2,2]0,1,2,3}"));
if (GetParam().propagate_metadata && !GetParam().clear_metadata) {
EXPECT_THAT(instruction->sharding(),
ShardingMetadata({CreateMetadata("a"), CreateMetadata("b")}));
} else {
EXPECT_THAT(instruction->sharding(), ShardingMetadata({}));
}
}
TEST_P(ParameterizedMetadataTest, BroadcastUser) { TEST_P(ParameterizedMetadataTest, BroadcastUser) {
const char* const hlo_string = R"( const char* const hlo_string = R"(
HloModule module HloModule module