Hunyuan-3b model support kvcache and gqa fusion

src/plugins/intel_gpu/src/plugin/transformations/kv_cache_fusion.cpp

@@ -24,6 +24,7 @@
 #include "openvino/pass/pattern/op/or.hpp"
 #include "openvino/pass/visualize_tree.hpp"
 #include "transformations/utils/utils.hpp"
+#include "openvino/opsets/opset8.hpp"
 
 namespace ov {
 namespace intel_gpu {
@@ -42,7 +43,8 @@ KVCacheFusionMatcher::KVCacheFusionMatcher() {
     auto gather_input = std::make_shared<ov::pass::pattern::op::Or>(OutputVector{past, convert_past});
     auto beam_idx = wrap_type<ov::op::v0::Parameter>();
     auto gather_past = wrap_type<ov::op::v8::Gather>({gather_input, beam_idx, wrap_type<ov::op::v0::Constant>()});
-    auto concat_past_input = std::make_shared<ov::pass::pattern::op::Or>(OutputVector{past, convert_past, gather_past});
+    auto gather_convert = wrap_type<ov::op::v0::Convert>({gather_past});
+    auto concat_past_input = std::make_shared<ov::pass::pattern::op::Or>(OutputVector{past, convert_past, gather_past, gather_convert});
     auto concat = wrap_type<ov::op::v0::Concat>({concat_past_input, any_input()});
     auto convert_present = wrap_type<ov::op::v0::Convert>({concat});
     auto present_input = std::make_shared<ov::pass::pattern::op::Or>(OutputVector{concat, convert_present});
@@ -62,9 +64,11 @@ KVCacheFusionMatcher::KVCacheFusionMatcher() {
         if (past_node->get_variable_id() != present_node->get_variable_id())
             return false;
 
-        // TODO: Support conversion internally
-        if (!concat_node || concat_node->get_output_element_type(0) != past_node->get_output_element_type(0))
-            return false;
+        // // TODO: Support conversion internally
+	if (ov::is_type<ov::opset8::Gather>(concat_past_input)) {
+            if (!concat_node || concat_node->get_output_element_type(0) != past_node->get_output_element_type(0))
+                return false;
+	}
 
         auto variable = past_node->get_variable();
         auto concat_axis = concat_node->get_axis();

src/plugins/intel_gpu/src/plugin/transformations/unsqueeze_broadcast_reshape_sdpa_fusion.cpp

@@ -52,8 +52,13 @@ UnsqueezeBroadcastReshapeSDPAFusion::UnsqueezeBroadcastReshapeSDPAFusion() {
     auto reshape_b_m = wrap_type<ov::op::v1::Reshape>({broadcast_b_m, any_input()}, reshape_predicate);
     auto reshape_c_m = wrap_type<ov::op::v1::Reshape>({broadcast_c_m, any_input()}, reshape_predicate);
 
+    auto convert_reshape_b_m = wrap_type<ov::op::v0::Convert>({reshape_b_m});
+    auto reshape_b_m_input = std::make_shared<ov::pass::pattern::op::Or>(OutputVector{reshape_b_m, convert_reshape_b_m});
+    auto convert_reshape_c_m = wrap_type<ov::op::v0::Convert>({reshape_c_m});
+    auto reshape_c_m_input = std::make_shared<ov::pass::pattern::op::Or>(OutputVector{reshape_c_m, convert_reshape_c_m});
+
     auto sdpa_without_attn_mask_m = wrap_type<op::SDPA>({ input_a_m, reshape_b_m, reshape_c_m });
-    auto sdpa_with_attn_mask_m = wrap_type<op::SDPA>({ input_a_m, reshape_b_m, reshape_c_m, input_attn_mask });
+    auto sdpa_with_attn_mask_m = wrap_type<op::SDPA>({ input_a_m, reshape_b_m_input, reshape_c_m_input, input_attn_mask });
     auto sdpa_with_attn_mask_and_scale_m = wrap_type<op::SDPA>({ input_a_m, reshape_b_m, reshape_c_m, input_attn_mask, input_scale });
 
     auto sdpa_m = std::make_shared<Or>(OutputVector{sdpa_without_attn_mask_m, sdpa_with_attn_mask_m, sdpa_with_attn_mask_and_scale_m});