Remove last mentions of IREE.

include/torch-mlir/Conversion/Passes.td

@@ -93,7 +93,7 @@ def ConvertTorchToLinalg : Pass<"convert-torch-to-linalg", "FuncOp"> {
     4. All this code operates on ranked tensors, for which using individual
        SSA values for sizes (rather than a "shape type") seems to
        work really well at this level of abstraction based on prior experience
-       in IREE. (unranked code tends to benefit from having a discrete
+       in other projects. (unranked code tends to benefit from having a discrete
        "shape type" to model shapes).
 
     We will see if we end up needing something like `shape.assuming`, but for

python/torch_mlir_e2e_test/linalg_on_tensors_backends/refbackend.py

@@ -88,7 +88,7 @@ LOWERING_PIPELINE = ",".join([
 
 
 class RefBackendLinalgOnTensorsBackend(LinalgOnTensorsBackend):
-    """Main entry-point for the backend."""
+    """Main entry-point for the reference backend."""
     def __init__(self):
         super().__init__()
 
@@ -102,10 +102,8 @@ class RefBackendLinalgOnTensorsBackend(LinalgOnTensorsBackend):
           imported_module: The MLIR module consisting of funcs in the torch
             dialect.
         Returns:
-          An opaque, backend specific module object that can be passed to load.
-          The object may actually be something more specific to the backend (i.e.
-          for IREE, it is a serialized VM flatbuffer) but the contract is that
-          it is operated on by methods on this class.
+          An opaque, backend specific compiled artifact object that can be
+          passed to `load`.
         """
         with imported_module.context:
             pm = PassManager.parse(LOWERING_PIPELINE)