Add quant mode for qwen3.5

src/python/py/models/builder.py

@@ -51,6 +51,9 @@
 )
 
 
+VALID_QUANT_MODES = frozenset({"default", "hybrid", "int4"})
+
+
 def check_extra_options(kv_pairs, execution_provider):
     """
     Check key-value pairs and set values correctly
@@ -71,6 +74,23 @@ def check_extra_options(kv_pairs, execution_provider):
         "disable_qkv_fusion",
         "prune_lm_head",
     ]
+
+    # Validate quant_mode if provided. Empty or whitespace-only values are
+    # treated as unset so downstream builders can apply their default behavior.
+    if "quant_mode" in kv_pairs:
+        quant_mode = kv_pairs["quant_mode"]
+        if isinstance(quant_mode, str):
+            quant_mode = quant_mode.strip().lower()
+
+        if not quant_mode:
+            del kv_pairs["quant_mode"]
+        elif quant_mode not in VALID_QUANT_MODES:
+            valid_modes_display = ", ".join(sorted(VALID_QUANT_MODES))
+            raise ValueError(
+                f"quant_mode must be one of {valid_modes_display}, got '{kv_pairs['quant_mode']}'"
+            )
+        else:
+            kv_pairs["quant_mode"] = quant_mode
     for key in bools:
         if key in kv_pairs:
             if kv_pairs[key] in {"false", "False", "0"}:

src/python/py/models/builders/qwen.py

@@ -1019,15 +1019,27 @@ def __init__(self, config, io_dtype, onnx_dtype, ep, cache_dir, extra_options):
         #
         # Linear attention recurrence accumulates errors across the full sequence,
         # unlike softmax attention which normalizes per-step.
+        #
+        # Control via extra_options quant_mode:
+        #   "default"  - INT8 for all linear attn + MLP layers (most accurate)
+        #   "hybrid"   - INT8 for linear attn projections only, INT4 for MLPs (balanced)
+        #   "int4"     - INT4 for everything (fastest, may degrade quality)
+        quant_mode = extra_options.get("quant_mode", "").strip().lower() or "default"
+        if quant_mode not in ("default", "hybrid", "int4"):
+            raise ValueError(f"quant_mode must be one of default, hybrid, int4, got '{quant_mode}'")
+
+        linear_attn_projs = ("in_proj_a", "in_proj_b", "in_proj_qkv", "in_proj_z", "out_proj")
+        mlp_projs = ("gate_proj", "up_proj", "down_proj")
+
         int8_nodes = {}
-        for i, lt in enumerate(self.layer_types):
-            if lt == "linear_attention":
-                # All linear attention projections: INT8
-                for proj in ("in_proj_a", "in_proj_b", "in_proj_qkv", "in_proj_z", "out_proj"):
-                    int8_nodes[f"/model/layers.{i}/linear_attn/{proj}/MatMul"] = {"bits": 8}
-                # MLP projections in linear attention layers: INT8
-                for proj in ("gate_proj", "up_proj", "down_proj"):
-                    int8_nodes[f"/model/layers.{i}/mlp/{proj}/MatMul"] = {"bits": 8}
+        if quant_mode in ("default", "hybrid"):
+            for i, lt in enumerate(self.layer_types):
+                if lt == "linear_attention":
+                    for proj in linear_attn_projs:
+                        int8_nodes[f"/model/layers.{i}/linear_attn/{proj}/MatMul"] = {"bits": 8}
+                    if quant_mode == "default":
+                        for proj in mlp_projs:
+                            int8_nodes[f"/model/layers.{i}/mlp/{proj}/MatMul"] = {"bits": 8}
 
         if int8_nodes:
             algo_config = self.quant_attrs["int4"].get("algo_config")