Introduce support for NF4 data type for OV weight compression (#988)

* Add NF4 weight format

* remove test

* Update optimum/intel/openvino/configuration.py

Co-authored-by: Nikita Savelyev <[email protected]>

* Update optimum/intel/openvino/configuration.py

Co-authored-by: Nikita Savelyev <[email protected]>

* Add extra checks

* apply black

---------

Co-authored-by: Nikita Savelyev <[email protected]>

Author: l-bat

optimum/commands/export/openvino.py

@@ -71,7 +71,7 @@ def parse_args_openvino(parser: "ArgumentParser"):
     optional_group.add_argument(
         "--weight-format",
         type=str,
-        choices=["fp32", "fp16", "int8", "int4", "mxfp4"],
+        choices=["fp32", "fp16", "int8", "int4", "mxfp4", "nf4"],
         default=None,
         help="The weight format of the exported model.",
     )

optimum/intel/openvino/configuration.py

@@ -347,7 +347,7 @@ class OVWeightQuantizationConfig(OVQuantizationConfigBase):
             Indicates whether to apply a scale estimation algorithm that minimizes the L2 error between the original and
             compressed layers. Providing a dataset is required to run scale estimation.
         weight_format (`str`, defaults to 'int'):
-            Data format weights are compressed to. Possible values: ['int4', 'int8', 'mxfp4'].
+            Data format weights are compressed to. Possible values: ['int4', 'int8', 'mxfp4', 'nf4'].
         qptq (`bool`, *optional*):
             Whether to apply GPTQ algorithm. GPTQ optimizes compressed weights in a layer-wise fashion to minimize the
             difference between activations of a compressed and original layer. Dataset is required to run GPTQ.
@@ -455,20 +455,22 @@ def post_init(self):
 
         if self.weight_format is None:
             self.weight_format = "int4" if self.bits == 4 else "int8"
-        if self.weight_format not in ["int4", "int8", "mxfp4"]:
+        if self.weight_format not in ["int4", "int8", "mxfp4", "nf4"]:
             raise ValueError(
-                f"Weight format must be one of the following: ['int4', 'int8', 'mxfp4'], but found: {self.weight_format}."
+                f"Weight format must be one of the following: ['int4', 'int8', 'mxfp4', 'nf4'], but found: {self.weight_format}."
             )
-        if self.weight_format == "mxfp4":
+        if self.weight_format in ["mxfp4", "nf4"]:
             if self.bits != 4:
                 raise ValueError(
-                    f"When applying weight compression with 'mxfp4' weight format the `bits` parameters must be set to 4, but found {self.bits}"
+                    f"When applying weight compression with '{self.weight_format}' weight format, the `bits` parameter must be set to 4, but found {self.bits}"
                 )
             if self.quant_method == OVQuantizationMethod.AWQ:
-                raise ValueError("The AWQ algorithm is not supported for 'mxfp4' weight format")
+                raise ValueError(f"The AWQ algorithm is not supported for '{self.weight_format}' weight format")
             if self.scale_estimation:
-                raise ValueError("The Scale Estimation algorithm is not supported for 'mxfp4' weight format")
-            if self.gptq:
+                raise ValueError(
+                    f"The Scale Estimation algorithm is not supported for '{self.weight_format}' weight format"
+                )
+            if self.weight_format == "mxfp4" and self.gptq:
                 raise ValueError("The GPTQ algorithm is not supported for 'mxfp4' weight format")
 
 

optimum/intel/openvino/quantization.py

@@ -930,6 +930,8 @@ def _weight_only_quantization(
 
     if config.weight_format == "mxfp4":
         mode = CompressWeightsMode.E2M1
+    elif config.weight_format == "nf4":
+        mode = CompressWeightsMode.NF4
     else:
         if config.bits == 8:
             mode = CompressWeightsMode.INT8_SYM if config.sym else CompressWeightsMode.INT8_ASYM

tests/openvino/test_exporters_cli.py

@@ -108,6 +108,7 @@ class OVCLIExportTestCase(unittest.TestCase):
         ("text-generation-with-past", "opt125m", "int4 --sym --group-size 128", {"int8": 4, "int4": 72}),
         ("text-generation-with-past", "opt125m", "int4 --group-size 64", {"int8": 4, "int4": 144}),
         ("text-generation-with-past", "opt125m", "mxfp4", {"int8": 4, "f4e2m1": 72, "f8e8m0": 72}),
+        ("text-generation-with-past", "opt125m", "nf4", {"int8": 4, "nf4": 72}),
         ("text-generation-with-past", "llama_awq", "int4 --ratio 1.0 --sym --group-size 8 --all-layers", {"int4": 16}),
         (
             "text-generation-with-past",
@@ -267,7 +268,7 @@ def test_exporters_cli_hybrid_quantization(self, model_type: str, exp_num_fq: in
             self.assertEqual(exp_num_fq, num_fq)
 
     @parameterized.expand(TEST_4BIT_CONFIGURATIONS)
-    def test_exporters_cli_int4(self, task: str, model_type: str, option: str, expected_num_weight_nodes: dict):
+    def test_exporters_cli_4bit(self, task: str, model_type: str, option: str, expected_num_weight_nodes: dict):
         with TemporaryDirectory() as tmpdir:
             result = subprocess.run(
                 f"optimum-cli export openvino --model {MODEL_NAMES[model_type]} --task {task} --weight-format {option} {tmpdir}",

tests/openvino/test_quantization.py

@@ -206,6 +206,13 @@ class OVWeightCompressionTest(unittest.TestCase):
             dict(bits=4, weight_format="mxfp4", group_size=32),
             {"f4e2m1": 20, "f8e8m0": 20, "int8": 4},
         ),
+        (
+            OVModelForCausalLM,
+            "gpt2",
+            False,
+            dict(bits=4, weight_format="nf4", group_size=32),
+            {"nf4": 20, "int8": 4},
+        ),
         (
             OVModelForCausalLM,
             "gpt2",

tests/openvino/utils_tests.py

@@ -195,6 +195,7 @@ def get_num_quantized_nodes(model):
         "int4": 0,
         "f4e2m1": 0,
         "f8e8m0": 0,
+        "nf4": 0,
     }
     ov_model = model if isinstance(model, ov.Model) else model.model
     for elem in ov_model.get_ops():
@@ -210,4 +211,6 @@ def get_num_quantized_nodes(model):
                 num_weight_nodes["f4e2m1"] += 1
             if type_name == "f8e8m0":
                 num_weight_nodes["f8e8m0"] += 1
+            if type_name == "nf4":
+                num_weight_nodes["nf4"] += 1
     return num_fake_quantize, num_weight_nodes