[5597780] Add support for FP16-only custom ops (#460)

Signed-off-by: gcunhase <[email protected]>

Author: gcunhase

modelopt/onnx/autocast/__main__.py

@@ -162,6 +162,19 @@ def get_parser() -> argparse.ArgumentParser:
             "libraries are in the PATH or LD_LIBRARY_PATH variables."
         ),
     )
+    parser.add_argument(
+        "--trt_plugins_precision",
+        type=str,
+        default=[],
+        nargs="+",
+        help=(
+            "A space-separated list indicating the precision for each custom op. "
+            "Each item should have the format <op_type>:<precision> (all inputs and outputs have the same precision) "
+            "or <op_type>:[<inp1_precision>,<inp2_precision>,...]:[<out1_precision>,<out2_precision>,...] "
+            "(inputs and outputs can have different precisions), where precision can be fp32 (default) or fp16."
+            "For example: op_type_1:fp16 op_type_2:[fp16,fp32]:[fp16]."
+        ),
+    )
 
     return parser
 
@@ -192,6 +205,7 @@ def main(argv=None):
         init_conversion_max_bytes=args.init_conversion_max_bytes,
         providers=args.providers,
         trt_plugins=args.trt_plugins,
+        trt_plugins_precision=args.trt_plugins_precision,
         max_depth_of_reduction=args.max_depth_of_reduction,
     )
 

modelopt/onnx/autocast/convert.py

@@ -58,6 +58,7 @@ def convert_to_mixed_precision(
     init_conversion_max_bytes: int | None = None,
     providers: list[str] = ["cpu"],
     trt_plugins: list[str] = [],
+    trt_plugins_precision: list[str] = [],
     max_depth_of_reduction: int | None = None,
 ) -> onnx.ModelProto:
     """Convert model to mixed precision.
@@ -78,6 +79,7 @@ def convert_to_mixed_precision(
                                    runtime.
         providers: List of ORT execution providers.
         trt_plugins: List of TensorRT plugin library paths in .so format (compiled shared library).
+        trt_plugins_precision: List indicating the precision for each custom op.
         max_depth_of_reduction: Maximum depth of reduction for node classification.
 
     Returns:
@@ -92,7 +94,11 @@ def convert_to_mixed_precision(
     # Otherwise, prefer to keep the original opset version unless it's very old
     min_opset = 22 if low_precision_type == "bf16" else 13
     graph_sanitizer = GraphSanitizer(
-        model, min_opset, trt_plugins=trt_plugins, max_ir_version=LATEST_IR_VERSION_SUPPORTED_BY_ORT
+        model,
+        min_opset,
+        trt_plugins=trt_plugins,
+        trt_plugins_precision=trt_plugins_precision,
+        max_ir_version=LATEST_IR_VERSION_SUPPORTED_BY_ORT,
     )
     graph_sanitizer.sanitize()
     model = graph_sanitizer.model
@@ -118,6 +124,7 @@ def convert_to_mixed_precision(
         init_max=init_max,
         custom_rule=custom_rule,
         max_depth_of_reduction=max_depth_of_reduction,
+        custom_ops_low_precision_nodes=graph_sanitizer.custom_ops_low_precision_nodes or [],
     )
 
     precision_converter = PrecisionConverter(

modelopt/onnx/autocast/graphsanitizer.py

@@ -23,6 +23,8 @@
 import modelopt.onnx.autocast.utils as utils
 import modelopt.onnx.utils as onnx_utils
 from modelopt.onnx.autocast.logging_config import logger
+from modelopt.onnx.quantization.graph_utils import cast_custom_ops
+from modelopt.onnx.trt_utils import interpret_trt_plugins_precision_flag
 
 
 class GraphSanitizer:
@@ -34,6 +36,7 @@ def __init__(
         min_opset: int = 13,
         max_ir_version: int | None = None,
         trt_plugins: list[str] | None = [],
+        trt_plugins_precision: list[str] | None = [],
     ) -> None:
         """Initialize GraphSanitizer.
 
@@ -48,7 +51,9 @@ def __init__(
         self.max_ir_version = max_ir_version
         self.standard_ops = {schema.name for schema in onnx.defs.get_all_schemas()}
         self.custom_ops = None
+        self.custom_ops_low_precision_nodes = []
         self.trt_plugins = trt_plugins
+        self.trt_plugins_precision = trt_plugins_precision or []
 
     def sanitize(self) -> None:
         """Sanitize the model graph.
@@ -67,6 +72,7 @@ def sanitize(self) -> None:
         self.cleanup_model()
         self.set_ir_version(self.max_ir_version)
         self.convert_fp64_to_fp32()
+        self.ensure_custom_ops_precision()
 
     def convert_fp64_to_fp32(self) -> None:
         """Convert FP64 initializers, I/O types, and specific nodes to FP32."""
@@ -88,6 +94,19 @@ def convert_fp64_to_fp32(self) -> None:
             logger.info("Converted FP64 initializers, I/O types, and nodes to FP32")
             self.model = onnx_utils.infer_shapes(self.model, strict_mode=True)
 
+    def ensure_custom_ops_precision(self) -> None:
+        """Ensure that custom ops run in the requested precision."""
+        custom_ops_to_cast, _ = interpret_trt_plugins_precision_flag(
+            self.model,
+            self.trt_plugins_precision,
+        )
+        if custom_ops_to_cast.get("fp16", {}):
+            self.model = cast_custom_ops(self.model, custom_ops_to_cast["fp16"])
+            self.custom_ops_low_precision_nodes = [
+                n.name for n in self.model.graph.node if n.op_type in custom_ops_to_cast["fp16"]
+            ]
+            logger.info("Ensured custom ops precision")
+
     def find_custom_nodes(self) -> None:
         """Find custom nodes in the model.
 

modelopt/onnx/autocast/nodeclassifier.py

@@ -360,6 +360,7 @@ def __init__(
         data_max: float | None = 1000.0,
         init_max: float | None = np.finfo(np.float16).max,
         max_depth_of_reduction: int | None = None,
+        custom_ops_low_precision_nodes: list[str] | None = None,
     ):
         """Initialize the node classifier.
 
@@ -375,6 +376,7 @@ def __init__(
             data_max: Maximum absolute value allowed for node I/O.
             init_max: Maximum absolute value allowed for initializers.
             max_depth_of_reduction: Maximum depth of reduction allowed in low precision.
+            custom_ops_low_precision_nodes: List of custom op node names to convert to low precision.
         """
         self.model = model
         self.node_to_init_map = node_to_init_map
@@ -387,6 +389,7 @@ def __init__(
         self.data_max = data_max
         self.init_max = init_max
         self.max_depth_of_reduction = max_depth_of_reduction
+        self.custom_ops_low_precision_nodes = custom_ops_low_precision_nodes
 
     def _gen_exclude_node_rules(self, reference_data):
         """Generate list of rules for blocking nodes from precision conversion.
@@ -446,12 +449,14 @@ def run(self, ref_outputs_dict=None):
         """
         exclude_node_rules = self._gen_exclude_node_rules(ref_outputs_dict)
         include_node_rules = self._gen_include_node_rules()
-        low_precision_nodes = []
+        low_precision_nodes = self.custom_ops_low_precision_nodes or []
         high_precision_nodes = []
         for node in self.model.graph.node:
             # If any condition is met - node will be executed in high precision
-            if any(rule.check(node) for rule in exclude_node_rules) and not any(
-                rule.check(node) for rule in include_node_rules
+            if (
+                node.name not in low_precision_nodes
+                and any(rule.check(node) for rule in exclude_node_rules)
+                and not any(rule.check(node) for rule in include_node_rules)
             ):
                 high_precision_nodes.append(node.name)
             else:

modelopt/onnx/trt_utils.py

@@ -349,7 +349,7 @@ def load_onnx_model(
 def interpret_trt_plugins_precision_flag(
     onnx_model: onnx.ModelProto,
     trt_plugins_precision: list[str],
-    quantize_mode: str,
+    quantize_mode: str = "int8",
 ) -> tuple[dict, dict]:
     """Convert custom ops precision flag to dictionaries with custom op and I/O indices to be cast/quantized.