[NPU] Add datatype NF4 support (#27903)

### Details: - Add datatype NF4 support in NPU plugin - *...* --------- Signed-off-by: Stefania Hergane <[email protected]> Co-authored-by: Stepan <[email protected]>
openvinotoolkit · Feb 14, 2025 · d7ecf52 · d7ecf52
1 parent 9d92d9c
commit d7ecf52
Show file tree

Hide file tree

Showing 4 changed files with 9 additions and 4 deletions.
diff --git a/src/plugins/intel_npu/src/backend/src/zero_infer_request.cpp b/src/plugins/intel_npu/src/backend/src/zero_infer_request.cpp
@@ -726,6 +726,8 @@ void ZeroInferRequest::check_network_precision(const ov::element::Type_t precisi
         break;
     case ov::element::Type_t::bf16:
         break;
+    case ov::element::Type_t::nf4:
+        break;
     case ov::element::Type_t::u4:
         break;
     case ov::element::Type_t::i4:
@@ -749,8 +751,9 @@ void ZeroInferRequest::check_network_precision(const ov::element::Type_t precisi
     case ov::element::Type_t::f64:
         break;
     default:
-        OPENVINO_THROW("Unsupported tensor precision: " + ov::element::Type(precision).get_type_name() +
-                       "! Supported precisions: FP32, FP16, BF16, U4, I4, U8, I8, U16, I16, U32, I32, U64, I64, FP64");
+        OPENVINO_THROW(
+            "Unsupported tensor precision: " + ov::element::Type(precision).get_type_name() +
+            "! Supported precisions: FP32, FP16, BF16, NF4, U4, I4, U8, I8, U16, I16, U32, I32, U64, I64, FP64");
     }
 }
 

diff --git a/src/plugins/intel_npu/src/compiler_adapter/src/driver_compiler_adapter.cpp b/src/plugins/intel_npu/src/compiler_adapter/src/driver_compiler_adapter.cpp
@@ -78,6 +78,8 @@ std::string ovPrecisionToLegacyPrecisionString(const ov::element::Type& precisio
         return "FP64";
     case ov::element::Type_t::bf16:
         return "BF16";
+    case ov::element::Type_t::nf4:
+        return "NF4";
     case ov::element::Type_t::i4:
         return "I4";
     case ov::element::Type_t::i8:

diff --git a/src/plugins/intel_npu/src/utils/src/zero/zero_wrappers.cpp b/src/plugins/intel_npu/src/utils/src/zero/zero_wrappers.cpp
@@ -69,7 +69,7 @@ CommandList::CommandList(const std::shared_ptr<ZeroInitStructsHolder>& initStruc
     if (mtci_is_supported) {
         ze_mutable_command_id_exp_desc_t mutableCmdIdDesc = {ZE_STRUCTURE_TYPE_MUTABLE_COMMAND_ID_EXP_DESC,
                                                              nullptr,
-                                                             ZE_MUTABLE_COMMAND_EXP_FLAG_GRAPH_ARGUMENT};
+                                                             ZE_MUTABLE_COMMAND_EXP_FLAG_GRAPH_ARGUMENT_DEPRECATED};
         THROW_ON_FAIL_FOR_LEVELZERO("zeCommandListGetNextCommandIdExp",
                                     zeCommandListGetNextCommandIdExp(_handle, &mutableCmdIdDesc, &_command_id));
     }

diff --git a/src/plugins/intel_npu/thirdparty/level-zero-ext b/src/plugins/intel_npu/thirdparty/level-zero-ext