Remove LTS workaround (triton-lang#1595)

The workarounds were added in intel/intel-xpu-backend-for-triton#1275 and intel/intel-xpu-backend-for-triton#1337. All huggingface training float32 models pass with the LTS workaround removed: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/9865658421 Signed-off-by: Whitney Tsang <[email protected]>
ZzEeKkAa · Jul 10, 2024 · 03efe15 · 03efe15
1 parent c6394b4
commit 03efe15
Show file tree

Hide file tree

Showing 2 changed files with 1 addition and 7 deletions.
diff --git a/third_party/intel/lib/TritonIntelGPUToLLVM/TargetInfo.cpp b/third_party/intel/lib/TritonIntelGPUToLLVM/TargetInfo.cpp
@@ -88,10 +88,6 @@ bool TargetInfo::warpReduce(RewriterBase &rewriter, Location loc,
                             SmallVector<Value> &acc, triton::ReduceOp op,
                             unsigned numLaneToReduce,
                             unsigned interleave) const {
-  const bool isLTS =
-      op->getParentOfType<ModuleOp>()->hasAttr("triton_gpu.is_lts");
-  if (isLTS)
-    return false;
   // No horizontal reduce required.
   if (numLaneToReduce == 1)
     return false;

diff --git a/third_party/intel/lib/TritonIntelGPUTransforms/RemoveLayoutConversions.cpp b/third_party/intel/lib/TritonIntelGPUTransforms/RemoveLayoutConversions.cpp
@@ -978,13 +978,11 @@ void LayoutRematerialization::rewriteSlice(SetVector<Value> &slice,
   SetVector<Operation *> opsToRewrite;
   // Keep track of yield operands that need to be duplicated.
   DenseMap<Operation *, SmallVector<int>> yieldOperandsMap;
-  bool isLTS =
-      convertOp->getParentOfType<ModuleOp>()->hasAttr("triton_gpu.is_lts");
   for (Value v : slice) {
     auto layoutIt = layout.find(v);
     assert(layoutIt != layout.end());
     // If we already have a remat value for this value, use it.
-    if (!isLTS && hasRematValue(v, layoutIt->second)) {
+    if (hasRematValue(v, layoutIt->second)) {
       mapping.map(v, getRematValue(v, layoutIt->second));
       continue;
     }