Skip to content

Commit 627fe8e

Browse files
loisloGoogle-ML-Automation
authored andcommitted
[XLA:GPU] Enable int4 matmul rewriting with Triton MLIR rewriter by default.
PiperOrigin-RevId: 715341409
1 parent c18b0ea commit 627fe8e

File tree

2 files changed

+1
-8
lines changed

2 files changed

+1
-8
lines changed

xla/debug_options_flags.cc

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -232,7 +232,7 @@ DebugOptions DefaultDebugOptionsIgnoringFlags() {
232232
opts.set_xla_gpu_exhaustive_tiling_search(false);
233233

234234
opts.set_xla_gpu_experimental_enable_triton_heroless_priority_fusion(false);
235-
opts.set_xla_gpu_experimental_enable_triton_i4_rewrites(false);
235+
opts.set_xla_gpu_experimental_enable_triton_i4_rewrites(true);
236236

237237
opts.set_xla_gpu_auto_spmd_partitioning_memory_budget_gb(0);
238238
opts.set_xla_gpu_auto_spmd_partitioning_memory_budget_ratio(1.1);

xla/service/gpu/fusions/triton/triton_fusion_emitter_int4_device_test.cc

Lines changed: 0 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -377,13 +377,6 @@ TEST_F(TritonTest, NonstandardLayoutInt4) {
377377
)";
378378

379379
TF_ASSERT_OK_AND_ASSIGN(auto module, GetOptimizedModule(kHloText));
380-
EXPECT_TRUE(*RunFileCheck(module->ToString(), R"(
381-
CHECK: %[[param_0:.*]] = s4[64,128]{0,1:E(4)} parameter(0)
382-
CHECK: %[[bitcast:.*]] = s4[128,64]{1,0:E(4)} bitcast(s4[64,128]{0,1:E(4)} %[[param_0]])
383-
CHECK: %[[convert:.*]] = bf16[128,64]{1,0} convert(s4[128,64]{1,0:E(4)} %[[bitcast]])
384-
CHECK: %[[param_1:.*]] = bf16[256,64]{1,0} parameter(1)
385-
CHECK: ROOT %dot.1 = bf16[128,256]{1,0} dot(bf16[128,64]{1,0} %[[convert]], bf16[256,64]{1,0} %[[param_1]]), lhs_contracting_dims={1}, rhs_contracting_dims={1}
386-
)"));
387380
EXPECT_TRUE(RunAndCompare(kHloText, ErrorSpec{/*aabs=*/1e-3, /*arel=*/1e-3}));
388381
}
389382

0 commit comments

Comments
 (0)