Merge pull request #14 from cms-ml/feature/passthrough_xla_flags

Forward XLA_FLAGS and TF_XLA_FLAGS in AOT compilation.

Author: valsdav

cmsml/scripts/compile_tf_graph.py

@@ -20,6 +20,8 @@ def compile_tf_graph(
     output_serving_key: str | None = None,
     compile_prefix: str | None = None,
     compile_class: str | None = None,
+    xla_flags: list[str] | None = None,
+    tf_xla_flags: list[str] | None = None,
 ) -> None:
     """
     For AOT compilation a static memory layout at runtime is required. This function prepares the given input SavedModel
@@ -35,6 +37,8 @@ def compile_tf_graph(
 
     An optional AOT compilation is initiated if *compile_class* and *compile_prefix* are given. In this case
     *compile_prefix* is the file prefix, while *compile_class* is the name of the AOT class within the generated files.
+
+    *xla_flags* and *tf_xla_flags* are forwarded to :py:func:`aot_compile`.
     """
     tf = import_tf()[0]
 
@@ -96,6 +100,8 @@ def compile_tf_graph(
             compile_class,
             batch_sizes=batch_sizes,
             serving_key=output_serving_key,
+            xla_flags=xla_flags,
+            tf_xla_flags=tf_xla_flags,
         )
 
 
@@ -106,13 +112,18 @@ def aot_compile(
     class_name: str,
     batch_sizes: tuple[int] = (1,),
     serving_key: str = r"serving_default_bs{}",
+    xla_flags: list[str] | None = None,
+    tf_xla_flags: list[str] | None = None,
 ) -> None:
     """
     Loads the graph from the SavedModel located at *model_path*, extracts the static graph specified by *serving_key*
     from it, AOT compiles it.
 
     This process generates header and object files at *output_path*. The *class_name* is used as class name within the
     header access the AOT-compiled network.
+
+    When *xla_flags* and *tf_xla_flags* are given, they are forwarded as comma-separated values to the *XLA_FLAGS*
+    and *TF_XLA_FLAGS* environment variables, respectively.
     """
     # prepare model path
     model_path = os.path.abspath(os.path.expandvars(os.path.expanduser(str(model_path))))
@@ -131,6 +142,19 @@ def aot_compile(
     # get the compilation executable
     exe = _which_saved_model_cli()
 
+    # ammend the env when xla flags were passed
+    env = os.environ.copy()
+    if xla_flags:
+        xla_flags_orig = env.get("XLA_FLAGS", "")
+        if xla_flags_orig:
+            xla_flags = [xla_flags_orig.rstrip(",")] + xla_flags
+        env["XLA_FLAGS"] = ",".join(map(str, xla_flags))
+    if tf_xla_flags:
+        tf_xla_flags_orig = env.get("TF_XLA_FLAGS", "")
+        if tf_xla_flags_orig:
+            tf_xla_flags = [tf_xla_flags_orig.rstrip(",")] + tf_xla_flags
+        env["TF_XLA_FLAGS"] = ",".join(map(str, tf_xla_flags))
+
     # compile for each batch size
     for bs in sorted(set(map(int, batch_sizes))):
         cmd = (
@@ -143,7 +167,7 @@ def aot_compile(
         )
 
         print(f"compiling for batch size {colored(bs, 'magenta')}")
-        code = interruptable_popen(cmd, executable="/bin/bash", shell=True, cwd=output_path)[0]
+        code = interruptable_popen(cmd, executable="/bin/bash", shell=True, cwd=output_path, env=env)[0]
         if code != 0:
             raise Exception(f"aot compilation using {exe} failed with exit code {code}")