[XLA:SPMD] Add HLO annotation to disable collective matmul in SPMD.

PiperOrigin-RevId: 698271808
openxla · Nov 20, 2024 · 74da9de · 74da9de
1 parent c738435
commit 74da9de
Show file tree

Hide file tree

Showing 5 changed files with 35 additions and 2 deletions.
diff --git a/xla/service/spmd/BUILD b/xla/service/spmd/BUILD
@@ -40,6 +40,7 @@ cc_library(
         "//xla:literal_util",
         "//xla:protobuf_util",
         "//xla:shape_util",
+        "//xla:side_effect_util",
         "//xla:status_macros",
         "//xla:types",
         "//xla:util",

diff --git a/xla/service/spmd/dot_handler.cc b/xla/service/spmd/dot_handler.cc
@@ -52,6 +52,7 @@ limitations under the License.
 #include "xla/service/spmd/spmd_partitioner_util.h"
 #include "xla/shape.h"
 #include "xla/shape_util.h"
+#include "xla/side_effect_util.h"
 #include "xla/status_macros.h"
 #include "xla/util.h"
 #include "xla/window_util.h"
@@ -1905,8 +1906,12 @@ absl::StatusOr<HloInstruction*> PartitionBaseCase(
            hlo.hlo()->opcode() == HloOpcode::kBitcast ||
            hlo.hlo()->opcode() == HloOpcode::kTranspose;
   };
-  bool should_skip_windowed_einsum = false;
-  if (options.disable_ag_rewrite_for_multiple_consumers) {
+  const auto& attrs = original_hlo->frontend_attributes().map();
+  bool should_skip_windowed_einsum =
+      attrs.contains(kXlaCollectiveMatmulAttr) &&
+      attrs.at(kXlaCollectiveMatmulAttr) == kXlaCollectiveMatmulNone;
+  if (!should_skip_windowed_einsum &&
+      options.disable_ag_rewrite_for_multiple_consumers) {
     auto lhs_operand =
         has_reshape_operand(lhs) ? lhs.hlo()->operand(0) : lhs.hlo();
     auto rhs_operand =

diff --git a/xla/service/spmd/spmd_partitioner_test.cc b/xla/service/spmd/spmd_partitioner_test.cc
@@ -4975,6 +4975,30 @@ ENTRY entry {
                           op::Shape("f32[16,256,1024]")));
 }
 
+TEST_P(SpmdPartitioningTest, DisableWindowedEinsumWithUserAnnotation) {
+  absl::string_view hlo_string = R"(
+HloModule module
+
+ENTRY entry {
+  %p0 = f32[2048,2,3264]{2,1,0} parameter(0), sharding={devices=[1,1,2]0,1}
+  %p1 = f32[2,3264,2176]{2,1,0} parameter(1), sharding={devices=[2,1,1]0,1}
+  ROOT %dot.224 = f32[2048,2176]{1,0} dot(f32[2048,2,3264]{2,1,0} %p0, f32[2,3264,2176]{2,1,0} %p1), lhs_contracting_dims={1,2}, rhs_contracting_dims={0,1}, sharding={devices=[1,2]0,1}, frontend_attributes={_xla_collective_matmul="none"}
+})";
+
+  TF_ASSERT_OK_AND_ASSIGN(
+      auto module,
+      PartitionComputation(hlo_string, /*num_devices=*/2,
+                           /*conv_halo_exchange_always_on_lhs=*/true,
+                           /*choose_faster_windowed_einsum=*/false,
+                           /*unroll_windowed_einsum=*/false,
+                           /*bidirectional_windowed_einsum=*/false,
+                           /*threshold_for_windowed_einsum_mib=*/0));
+  ASSERT_FALSE(absl::c_any_of(module->entry_computation()->instructions(),
+                              [](const HloInstruction* inst) {
+                                return inst->opcode() == HloOpcode::kWhile;
+                              }));
+}
+
 TEST_P(SpmdPartitioningTest, EinsumBatchPartitioned) {
   absl::string_view hlo_string = R"(
 HloModule module

diff --git a/xla/side_effect_util.cc b/xla/side_effect_util.cc
@@ -69,6 +69,8 @@ const char kXlaCollectiveMatmulRhsAg[] = "rhs_ag";
 
 const char kXlaCollectiveMatmulRs[] = "rs";
 
+const char kXlaCollectiveMatmulNone[] = "none";
+
 const char kXlaMultiRecvCountAttr[] = "_xla_multi_recv_count";
 
 }  // namespace xla
diff --git a/xla/side_effect_util.h b/xla/side_effect_util.h
@@ -77,6 +77,7 @@ extern const char kXlaCollectiveMatmulAttr[];
 extern const char kXlaCollectiveMatmulLhsAg[];
 extern const char kXlaCollectiveMatmulRhsAg[];
 extern const char kXlaCollectiveMatmulRs[];
+extern const char kXlaCollectiveMatmulNone[];
 
 // XLA frontend attribute for specifying the number of sends this recv should
 // match.