openvinotoolkit
diff --git a/‎src/common/low_precision_transformations/include/low_precision/layer_transformation.hpp‎
Lines changed: 2 additions & 1 deletion b/‎src/common/low_precision_transformations/include/low_precision/layer_transformation.hpp‎
Lines changed: 2 additions & 1 deletion
diff --git a/‎src/common/low_precision_transformations/src/layer_transformation.cpp‎
Lines changed: 7 additions & 0 deletions b/‎src/common/low_precision_transformations/src/layer_transformation.cpp‎
Lines changed: 7 additions & 0 deletions
diff --git a/‎src/common/transformations/src/transformations/fp16_compression/mark_decompression_convert_constant_folding.cpp‎
Lines changed: 1 addition & 0 deletions b/‎src/common/transformations/src/transformations/fp16_compression/mark_decompression_convert_constant_folding.cpp‎
Lines changed: 1 addition & 0 deletions
diff --git a/‎src/plugins/intel_gpu/src/graph/debug_helper.cpp‎
Lines changed: 2 additions & 0 deletions b/‎src/plugins/intel_gpu/src/graph/debug_helper.cpp‎
Lines changed: 2 additions & 0 deletions
diff --git a/‎src/plugins/intel_gpu/src/graph/impls/ocl/dynamic_quantize.cpp‎
Lines changed: 3 additions & 0 deletions b/‎src/plugins/intel_gpu/src/graph/impls/ocl/dynamic_quantize.cpp‎
Lines changed: 3 additions & 0 deletions
diff --git a/‎src/plugins/intel_gpu/src/graph/impls/ocl/kernel_selector_helper.cpp‎
Lines changed: 8 additions & 0 deletions b/‎src/plugins/intel_gpu/src/graph/impls/ocl/kernel_selector_helper.cpp‎
Lines changed: 8 additions & 0 deletions
diff --git a/‎src/plugins/intel_gpu/src/graph/impls/ocl/primitive_base.hpp‎
Lines changed: 11 additions & 0 deletions b/‎src/plugins/intel_gpu/src/graph/impls/ocl/primitive_base.hpp‎
Lines changed: 11 additions & 0 deletions
diff --git a/‎src/plugins/intel_gpu/src/graph/impls/ocl/reorder.cpp‎
Lines changed: 3 additions & 0 deletions b/‎src/plugins/intel_gpu/src/graph/impls/ocl/reorder.cpp‎
Lines changed: 3 additions & 0 deletions
diff --git a/‎src/plugins/intel_gpu/src/graph/impls/ocl_v2/utils/jitter.cpp‎
Lines changed: 12 additions & 0 deletions b/‎src/plugins/intel_gpu/src/graph/impls/ocl_v2/utils/jitter.cpp‎
Lines changed: 12 additions & 0 deletions
diff --git a/‎src/plugins/intel_gpu/src/graph/impls/onednn/fully_connected_onednn.cpp‎
Lines changed: 3 additions & 3 deletions b/‎src/plugins/intel_gpu/src/graph/impls/onednn/fully_connected_onednn.cpp‎
Lines changed: 3 additions & 3 deletions
@@ -28,6 +28,7 @@ namespace low_precision {
 namespace precision_set {
     LP_TRANSFORMATIONS_API const std::vector<element::Type>& get_int8_support();
     LP_TRANSFORMATIONS_API const std::vector<element::Type>& get_fp8_support();
+    LP_TRANSFORMATIONS_API const std::vector<element::Type>& get_fp8_fp4_support();
     LP_TRANSFORMATIONS_API const std::vector<element::Type>& get_int8_int16_int32_support();
 } // namespace precision_set
 
@@ -59,7 +60,7 @@ class LP_TRANSFORMATIONS_API DataPrecision {
                 element::i8, element::u8,
                 element::i16, element::u16,
                 element::i32, element::u32,
-                element::f8e4m3, element::f8e5m2,
+                element::f8e4m3, element::f8e5m2, element::f4e2m1,
         };
         return lowPrecision.find(precision) != lowPrecision.end();
     }
 
@@ -36,6 +36,13 @@ const std::vector<element::Type>& precision_set::get_fp8_support() {
     return fp8_support;
 }
 
+const std::vector<element::Type>& precision_set::get_fp8_fp4_support() {
+    static const std::vector<element::Type> low_bit_float_support = {
+        ov::element::f8e4m3,  ov::element::f8e5m2, ov::element::f4e2m1,
+    };
+    return low_bit_float_support;
+}
+
 const std::vector<element::Type>& precision_set::get_int8_int16_int32_support() {
     static const std::vector<element::Type> int8_int16_int32_support = {
         ov::element::u8,  ov::element::i8,
 
@@ -143,6 +143,7 @@ MarkCompressedFloatConstants::MarkCompressedFloatConstants() {
             return false;
         if (const_node->get_output_element_type(0) != element::f16 &&
             const_node->get_output_element_type(0) != element::bf16 &&
+            const_node->get_output_element_type(0) != element::f4e2m1 &&
             const_node->get_output_element_type(0) != element::f8e4m3 &&
             const_node->get_output_element_type(0) != element::f8e5m2 &&
             const_node->get_output_element_type(0) != element::f8e8m0)
 
@@ -284,6 +284,8 @@ void log_memory_to_file(memory::ptr mem, layout data_layout, stream& stream, std
         dump<ov::float8_e5m2>(actual_mem, stream, file_stream, dump_raw);
     else if (mem_dt == cldnn::data_types::f8e4m3)
         dump<ov::float8_e4m3>(actual_mem, stream, file_stream, dump_raw);
+    else if (mem_dt == cldnn::data_types::f4e2m1)
+        dump<ov::float4_e2m1>(actual_mem, stream, file_stream, dump_raw);
     else if (mem_dt == cldnn::data_types::f8e8m0)
         dump<ov::float8_e8m0>(actual_mem, stream, file_stream, dump_raw);
     else if (mem_dt == cldnn::data_types::boolean)
 
@@ -55,6 +55,8 @@ struct dynamic_quantize_impl : typed_primitive_impl_ocl<dynamic_quantize> {
         params.combine_scales_and_zp = desc->attrs.output_storage_type != ov::op::internal::DynamicQuantize::OutputStorageType::Planar;
         params.generate_precomputed_reduction = desc->attrs.precomputed_reduction;
 
+        validate_f4e2m1_packed_output(impl_param.get_output_layout(0), "dynamic_quantize");
+
         return params;
     }
 
@@ -75,6 +77,7 @@ attach_dynamic_quantize_impl::attach_dynamic_quantize_impl() {
         data_types::f16,
         data_types::i8,
         data_types::u8,
+        data_types::f4e2m1,
         data_types::f8e4m3,
         data_types::f8e5m2,
         data_types::f8e8m0,
 
@@ -268,6 +268,8 @@ kernel_selector::data_type to_data_type(data_types dt) {
             return kernel_selector::data_type::F32;
         case cldnn::data_types::bf16:
             return kernel_selector::data_type::BF16;
+        case cldnn::data_types::f4e2m1:
+            return kernel_selector::data_type::F4E2M1;
         case cldnn::data_types::f8e4m3:
             return kernel_selector::data_type::F8E4M3;
         case cldnn::data_types::f8e5m2:
@@ -305,6 +307,8 @@ data_types from_data_type(kernel_selector::data_type dt) {
             return cldnn::data_types::f16;
         case kernel_selector::data_type::F32:
             return cldnn::data_types::f32;
+        case kernel_selector::data_type::F4E2M1:
+            return cldnn::data_types::f4e2m1;
         case kernel_selector::data_type::F8E4M3:
             return cldnn::data_types::f8e4m3;
         case kernel_selector::data_type::F8E5M2:
@@ -336,6 +340,8 @@ kernel_selector::weights_type to_weights_type(data_types dt) {
             return kernel_selector::weights_type::INT32;
         case cldnn::data_types::bf16:
             return kernel_selector::weights_type::BF16;
+        case cldnn::data_types::f4e2m1:
+            return kernel_selector::weights_type::F4E2M1;
         case cldnn::data_types::f8e4m3:
             return kernel_selector::weights_type::F8E4M3;
         case cldnn::data_types::f8e5m2:
@@ -365,6 +371,8 @@ data_types from_weights_type(kernel_selector::weights_type dt) {
             return data_types::f32;
         case kernel_selector::weights_type::INT32:
             return data_types::i32;
+        case kernel_selector::weights_type::F4E2M1:
+            return data_types::f4e2m1;
         case kernel_selector::weights_type::F8E4M3:
             return data_types::f8e4m3;
         case kernel_selector::weights_type::F8E5M2:
 
@@ -33,6 +33,17 @@
 namespace cldnn {
 namespace ocl {
 
+inline void validate_f4e2m1_packed_output(const layout& output_layout, const char* primitive_name) {
+    if (output_layout.data_type != ov::element::f4e2m1 || output_layout.is_dynamic()) {
+        return;
+    }
+
+    OPENVINO_ASSERT(output_layout.get_linear_size() % 8 == 0,
+                    "[GPU] ", primitive_name, ": f4e2m1 output size must be a multiple of 8 elements "
+                    "(32-bit atomic write granularity), but got: ",
+                    output_layout.get_linear_size());
+}
+
 /*
 Base class for all GPU implementation of specified primitive type.
 For example, all gpu convolution implementations should derive from typed_primitive_impl_ocl<convolution>.
 
@@ -112,6 +112,8 @@ struct reorder_impl : typed_primitive_impl_ocl<reorder> {
         params.winograd = impl_param.input_layouts[0].format.is_winograd() || output_layout.format.is_winograd();
         params.truncate = impl_param.typed_desc<reorder>()->truncate;
 
+        validate_f4e2m1_packed_output(output_layout, "reorder");
+
         return params;
     }
 
@@ -122,6 +124,7 @@ struct reorder_impl : typed_primitive_impl_ocl<reorder> {
         }
 
         update_shapes(*_kernel_data.params, impl_param);
+        validate_f4e2m1_packed_output(impl_param.get_output_layout(), "reorder");
         (_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);
     }
 
 
@@ -320,6 +320,18 @@ JitConstants make_type_jit_constants(const std::string& name, const ov::element:
         type_size = "4";
         is_fp = true;
         break;
+    case ov::element::f4e2m1:
+        type = "fp4e2m1_t";
+        max_val = "(fp4e2m1_t){as_uchar((uchar)0x7)}";  // 6.0
+        min_val = "(fp4e2m1_t){as_uchar((uchar)0xF)}";  // -6.0
+        val_one = "(fp4e2m1_t){as_uchar((uchar)0x2)}";
+        val_zero = "(fp4e2m1_t){as_uchar((uchar)0x0)}";
+        to_type = "_convert_fp4e2m1_t(v)";
+        to_type_sat = "_convert_fp4e2m1_t_sat(v)";
+        as_type = "as_fp4e2m1_t(v)";
+        type_size = "0.5f";
+        is_fp = true;
+        break;
     case ov::element::f8e4m3:
         type = "fp8e4m3_t";
         max_val = "(fp8e4m3_t){as_char((char)0x7E)}";  // 448.0
 
@@ -71,7 +71,7 @@ struct fully_connected_onednn : typed_primitive_onednn_impl<fully_connected> {
             }
 
             const auto input_dt = instance.get_input_layout(0).data_type;
-            const bool is_dyn_quan_input = cldnn::one_of(input_dt, {data_types::i8, data_types::u8, data_types::f8e4m3, data_types::f8e5m2});
+            const bool is_dyn_quan_input = cldnn::one_of(input_dt, {data_types::i8, data_types::u8, data_types::f4e2m1, data_types::f8e4m3, data_types::f8e5m2});
 
             if (is_dyn_quan_input && prim->activation_scale.is_valid()) {
                 const auto activation_scale_idx = idx++;
@@ -312,7 +312,7 @@ struct fully_connected_onednn : typed_primitive_onednn_impl<fully_connected> {
         }
 
         const auto input_dt = impl_params->get_input_layout(0).data_type;
-        const bool is_dyn_quan_input = cldnn::one_of(input_dt, {data_types::i8, data_types::u8, data_types::f8e4m3, data_types::f8e5m2});
+        const bool is_dyn_quan_input = cldnn::one_of(input_dt, {data_types::i8, data_types::u8, data_types::f4e2m1, data_types::f8e4m3, data_types::f8e5m2});
         if (is_dyn_quan_input && dynamic_quantized_activation) {
             auto src_scale_idx = ++idx;
             auto partial_shape = impl_params->get_input_layout(0).get_partial_shape();
@@ -360,7 +360,7 @@ struct fully_connected_onednn : typed_primitive_onednn_impl<fully_connected> {
 
         if (prim->compressed_weights) {
             const auto input_dt = impl_params.get_input_layout(0).data_type;
-            const bool is_dyn_quan_input = cldnn::one_of(input_dt, {data_types::i8, data_types::u8, data_types::f8e4m3, data_types::f8e5m2});
+            const bool is_dyn_quan_input = cldnn::one_of(input_dt, {data_types::i8, data_types::u8, data_types::f4e2m1, data_types::f8e4m3, data_types::f8e5m2});
             if (is_dyn_quan_input) {
                 OPENVINO_ASSERT(prim->input_size <= 3, "[GPU] Dynamic quantization for 4D matmul is not implemented");
             } else {
Original file line number	Diff line number	Diff line change
`@@ -112,6 +112,8 @@ struct reorder_impl : typed_primitive_impl_ocl<reorder> {`
`112`	`112`	`params.winograd = impl_param.input_layouts[0].format.is_winograd() \|\| output_layout.format.is_winograd();`
`113`	`113`	`params.truncate = impl_param.typed_desc<reorder>()->truncate;`
`114`	`114`
	`115`	`+ validate_f4e2m1_packed_output(output_layout, "reorder");`
	`116`	`+`
`115`	`117`	`return params;`
`116`	`118`	`}`
`117`	`119`
`@@ -122,6 +124,7 @@ struct reorder_impl : typed_primitive_impl_ocl<reorder> {`
`122`	`124`	`}`
`123`	`125`
`124`	`126`	`update_shapes(*_kernel_data.params, impl_param);`
	`127`	`+ validate_f4e2m1_packed_output(impl_param.get_output_layout(), "reorder");`
`125`	`128`	`(_kernel_data.update_dispatch_data_func)(*_kernel_data.params, _kernel_data);`
`126`	`129`	`}`
`127`	`130`