Aanuf/sdpa v fp8

nncf/common/graph/operator_metatypes.py

@@ -64,6 +64,17 @@ def subtype_check(cls, metatype: type["OperatorMetatype"]) -> bool:
 
         return any(subtype.subtype_check(metatype) for subtype in subtypes)
 
+    @classmethod
+    def get_target_input_ports(cls, is_fp8: bool = False) -> list[int] | None:
+        """
+        Returns the target input ports for FP8.
+
+        :returns: A list of target input ports for FP8.
+        """
+        if is_fp8 and hasattr(cls, "target_input_ports_fp8"):
+            return cls.target_input_ports_fp8
+        return cls.target_input_ports
+
 
 class OperatorMetatypeRegistry(Registry):
     """

nncf/common/quantization/quantizer_propagation/solver.py

@@ -333,6 +333,7 @@ def __init__(
         post_processing_marker_metatypes: Optional[list[type[OperatorMetatype]]] = None,
         metatypes_to_ignore: Optional[list[type[OperatorMetatype]]] = None,
         scales_unification_map: Optional[dict[type[OperatorMetatype], list[type[OperatorMetatype]]]] = None,
+        is_fp8: bool = False,
     ):
         """
         Initializes the solver with parameters affecting the resulting quantizer setup.
@@ -386,6 +387,7 @@ def __init__(
             which should be automatically ignored.
         :param scales_unification_map: The framework-specific map with NNCF metatypes, which generating a quantizer
             that can be unified if it so requires based on metatype.
+        :param is_fp8: Whether the quantization is done in FP8 mode.
         """
         if default_trait_to_metatype_map is None:
             self._default_trait_to_metatype_map = {}
@@ -409,6 +411,7 @@ def __init__(
         self._weight_quantizable_node_names_vs_qconfigs = self._filter_by_weight_ignored_target_scopes(
             quantizable_layer_nodes, weight_ignored_scopes, weight_target_scopes
         )
+        self._is_fp8 = is_fp8
 
         if scope_overrides is None:
             self._scope_overrides: dict[str, Any] = {}
@@ -1147,7 +1150,8 @@ def _setup_initial_quantizers_for_operator_node(
             if input_port_id in metatype.ignored_input_ports:
                 continue
 
-            if metatype.target_input_ports is not None and input_port_id not in metatype.target_input_ports:
+            target_input_ports = metatype.get_target_input_ports(self._is_fp8)
+            if target_input_ports is not None and input_port_id not in target_input_ports:
                 continue
 
             edge = quant_prop_graph.edges[pred_ip_key, operator_node_key]

nncf/openvino/graph/metatypes/openvino_metatypes.py

@@ -731,6 +731,7 @@ class OVScaledDotProductAttentionMetatype(OVOpMetatype):
     op_names = ["ScaledDotProductAttention"]
     hw_config_names = [HWConfigOpName.SCALED_DOT_PRODUCT_ATTENTION]
     target_input_ports = [0, 1]
+    target_input_ports_fp8 = [0, 1, 2]
 
 
 @OV_OPERATOR_METATYPES.register()

nncf/quantization/algorithms/min_max/algorithm.py

@@ -289,11 +289,14 @@ def _set_mode_based_params(self) -> None:
             if getattr(self, self_name) is None:
                 setattr(self, self_name, default_value)
 
+    def _is_fp8(self):
+        return self._mode in (QuantizationMode.FP8_E4M3, QuantizationMode.FP8_E5M2)
+
     def _review_mode_based_params(self):
         """
         Reviews parameter values because mode option doesn't support them.
         """
-        if self._mode in (QuantizationMode.FP8_E4M3, QuantizationMode.FP8_E5M2):
+        if self._is_fp8():
             nncf_logger.warning(f"You're using experimental option mode with {self._mode} value.")
 
             if self._preset != QuantizationPreset.PERFORMANCE:
@@ -696,6 +699,7 @@ def _get_quantizer_setup(
             metatypes_to_ignore=metatypes_to_ignore,
             scales_unification_map=self._backend_entity.scales_unification_map,
             scope_overrides=scope_overrides,
+            is_fp8=self._is_fp8(),
         )
 
         quantization_proposal = solver.run_on_ip_graph(ip_graph, self._backend_entity.elementwise_metatypes)

nncf/torch/graph/operator_metatypes.py

@@ -1129,6 +1129,7 @@ class PTScaledDotProductAttentionMetatype(PTOperatorMetatype):
     }
     hw_config_names = [HWConfigOpName.SCALED_DOT_PRODUCT_ATTENTION]
     target_input_ports = [0, 1]
+    target_input_ports_fp8 = [0, 1, 2]
 
 
 @PT_OPERATOR_METATYPES.register()