Qualcomm AI Engine Direct - Support EPS for cli and fix mv2 to work in QNN2.45 (#18002)

Author: winskuo-quic

examples/qualcomm/oss_scripts/mobilevit_v2.py

@@ -22,6 +22,7 @@
     build_executorch_binary,
     get_backend_type,
     make_output_dir,
+    make_quantizer,
     parse_skip_delegation_node,
     setup_common_args_and_variables,
     SimpleADB,
@@ -90,9 +91,12 @@ def main(args):
 
     pte_filename = "mobilevit_v2_qnn"
     backend = get_backend_type(args.backend)
-    quant_dtype = {
+    quantizer = {
         QnnExecuTorchBackendType.kGpuBackend: None,
-        QnnExecuTorchBackendType.kHtpBackend: QuantDtype.use_16a8w,
+        QnnExecuTorchBackendType.kHtpBackend: make_quantizer(
+            quant_dtype=QuantDtype.use_16a8w,
+            eps=2**-10,
+        ),
     }[backend]
     build_executorch_binary(
         module,
@@ -102,7 +106,7 @@ def main(args):
         inputs,
         skip_node_id_set=skip_node_id_set,
         skip_node_op_set=skip_node_op_set,
-        quant_dtype=quant_dtype,
+        custom_quantizer=quantizer,
         backend=backend,
         shared_buffer=args.shared_buffer,
         online_prepare=args.online_prepare,

examples/qualcomm/util_scripts/cli.py

@@ -135,7 +135,6 @@ def __iter__(self):
 
 def quantize(args):
     logger = get_logger()
-
     # get corresponding QnnQuantizer
     try:
         quant_dtype = getattr(QuantDtype, args.config)
@@ -147,6 +146,7 @@ def quantize(args):
             act_observer=act_observer,
             backend=get_backend_type(args.backend),
             soc_model=args.model,
+            eps=args.eps,
         )
     except Exception:
         logger.error(
@@ -450,6 +450,12 @@ def main():
         default="htp",
         help="Backend to be deployed ('htp'/'gpu' are currently supported).",
     )
+    sub_quantize.add_argument(
+        "--eps",
+        help="EPS value for quantizer. Accepts floating‑point literal. E.g., 0.0009765625.",
+        type=float,
+        default=None,
+    )
     sub_quantize.set_defaults(callback=quantize)
 
     sub_compile = subparsers.add_parser(

examples/qualcomm/utils.py

@@ -371,9 +371,9 @@ def make_quantizer(
     act_symmetric=False,
     is_qat=False,
     submodule_qconfig_list: Optional[List[Tuple[Callable, ModuleQConfig]]] = None,
-    eps=None,
     backend=QnnExecuTorchBackendType.kHtpBackend,
     soc_model="SM8750",
+    eps=None,
 ):
     quantizer = QnnQuantizer(backend=backend, soc_model=getattr(QcomChipset, soc_model))
     quantizer.add_custom_quant_annotations(custom_annotations)