[17/17] Opt-in FP16 strong typing for DLA

CHANGELOG.rst

@@ -35,7 +35,8 @@ Changelog
 
 - The ``evaluation`` agent skill can stop looping ("doom loop") reasoning early: ask it to avoid doom loops and it adds vLLM's ``repetition_detection`` request field (vLLM 0.17.0 or later) to every task, ending a request once its output repeats. Use the same setting for baseline and candidate runs.
 - Add the ordered DLA compatibility pipeline through the ``modelopt.onnx.graph_surgery.make_dla_compatible``
-  API and ``make-dla-compatible`` CLI command. Use ``--remove-all-qdq`` to remove Q/DQ pairs of every dtype.
+  API and ``make-dla-compatible`` CLI command. Use ``--remove-all-qdq`` to remove Q/DQ pairs of every dtype,
+  or ``--force-fp16-activations`` for FP16 internal computation with FP32 model boundaries.
 
 **Backward Breaking Changes**
 

modelopt/onnx/graph_surgery/__main__.py

@@ -284,6 +284,14 @@ def main():
         action="store_true",
         help="Remove QDQ pairs for every dtype, matching the C++ remove-all mode",
     )
+    dla_parser.add_argument(
+        "--force-fp16-activations",
+        action="store_true",
+        help=(
+            "Use FP16 internal activations and weights with FP32 model boundaries; "
+            "L2 normalization intermediates can overflow or underflow"
+        ),
+    )
 
     # Analyze subcommand
     analyze_parser = subparsers.add_parser(
@@ -367,6 +375,7 @@ def main():
             verbose=not args.quiet,
             qdq_quantized_dtypes=args.qdq_quantized_dtypes,
             remove_all_qdq=args.remove_all_qdq,
+            force_fp16_activations=args.force_fp16_activations,
         )
 
     elif args.command == "analyze":