[GPU] Enable KV-cache compression by default for non-systolic platforms

sshlyapn · Nov 6, 2024 · f599577 · f599577
1 parent 84e6837
commit f599577
Show file tree

Hide file tree

Showing 2 changed files with 9 additions and 1 deletion.
diff --git a/src/plugins/intel_gpu/src/runtime/execution_config.cpp b/src/plugins/intel_gpu/src/runtime/execution_config.cpp
@@ -246,6 +246,11 @@ void ExecutionConfig::apply_user_properties(const cldnn::device_info& info) {
         set_property(ov::intel_gpu::queue_type(QueueTypes::in_order));
     }
 
+    // Enable KV-cache compression by default for non-systolic platforms
+    if (!is_set_by_user(ov::hint::kv_cache_precision) && !info.supports_immad) {
+        set_property(ov::hint::kv_cache_precision(ov::element::i8));
+    }
+
     user_properties.clear();
 }
 

diff --git a/src/plugins/intel_gpu/tests/functional/subgraph_tests/dynamic/kv_cache_sdpa.cpp b/src/plugins/intel_gpu/tests/functional/subgraph_tests/dynamic/kv_cache_sdpa.cpp
@@ -47,8 +47,11 @@ class SDPAWithKVCacheTest : public ::testing::Test, public ::testing::WithParamI
         ov::AnyMap properties = {ov::hint::inference_precision(ov::element::f16),
                                  ov::intel_gpu::hint::enable_sdpa_optimization(true)};
 
-        if (p.compressed)
+        if (p.compressed) {
             properties.emplace(ov::hint::kv_cache_precision(ov::element::i8));
+        } else {
+            properties.emplace(ov::hint::kv_cache_precision(ov::element::undefined));
+        }
 
         const size_t n_heads = 16;
         const size_t n_features = 64;