Add UT and remove unused code for torch MX quant (#1854)

mengniwang95 · web-flow · commit 455f1e1f0f02 · 2024-06-11T15:28:40.000+08:00
* Add UT and remove unused code for torch MX quant
---------

Signed-off-by: Mengni Wang &lt;mengni.wang@intel.com&gt;
diff --git a/neural_compressor/torch/algorithms/mx_quant/mx.py b/neural_compressor/torch/algorithms/mx_quant/mx.py
@@ -62,9 +62,6 @@ def apply_mx_specs(self):
                 axes=[-1],
             )
 
-    def append_name(self, postfix):
-        self.name += postfix
-
     def forward(self, input):
         if self.mx_none:
             return super().forward(input)
diff --git a/test/3x/torch/algorithms/mx_quant/test_mx_utility.py b/test/3x/torch/algorithms/mx_quant/test_mx_utility.py
@@ -0,0 +1,22 @@
+import pytest
+import torch
+
+from neural_compressor.torch.algorithms.mx_quant import utils
+
+
+def test_mx_quant_utility():
+    tensor = torch.rand((1, 30))
+    assert torch.equal(tensor, utils.quantize_mx_op(tensor, None, "nearest", 32))
+    assert torch.equal(tensor, utils._quantize_fp(tensor))
+    assert torch.equal(tensor, utils._quantize_bfloat(tensor, 0))
+    assert torch.equal(tensor, utils._quantize_mx(tensor, 8, None))
+
+    assert not torch.equal(utils._shared_exponents(tensor, "none"), utils._shared_exponents(tensor))
+    with pytest.raises(Exception):
+        utils._shared_exponents(tensor, None)
+    with pytest.raises(Exception):
+        utils._reshape_to_blocks(tensor, None, 32)
+    with pytest.raises(Exception):
+        utils.quantize_elemwise_op(tensor, "test")
+    with pytest.raises(Exception):
+        utils._round_mantissa(tensor, 3, "test")
diff --git a/test/3x/torch/quantization/test_mx_quant.py b/test/3x/torch/quantization/test_mx_quant.py
@@ -3,7 +3,7 @@
 import pytest
 import torch
 
-from neural_compressor.torch.quantization import MXQuantConfig, get_default_mx_config, quantize
+from neural_compressor.torch.quantization import MXQuantConfig, convert, get_default_mx_config, prepare
 
 
 def build_simple_torch_model():
@@ -40,20 +40,35 @@ def teardown_class(self):
     def test_mx_quant_default(self):
         fp32_model = copy.deepcopy(self.fp32_model)
         quant_config = get_default_mx_config()
-        q_model = quantize(fp32_model, quant_config=quant_config)
+        fp32_model = prepare(model=fp32_model, quant_config=quant_config)
+        q_model = convert(model=fp32_model)
         assert q_model is not None, "Quantization failed!"
 
     @pytest.mark.parametrize(
-        "w_dtype, weight_only",
+        "w_dtype, weight_only, round_method, out_dtype",
         [
-            ("fp4", True),
-            ("fp8_e5m2", False),
+            ("fp4", True, "dither", "float32"),
+            ("fp8_e5m2", False, "floor", "bfloat16"),
+            ("int8", False, "even", "float16"),
+            ("int4", False, "nearest", "float32"),
+            ("int2", False, "dither", "bfloat16"),
+            ("fp8_e4m3", False, "floor", "float16"),
+            ("fp6_e3m2", False, "even", "float32"),
+            ("fp6_e2m3", False, "nearest", "bfloat16"),
+            ("float16", False, "dither", "float16"),
+            ("bfloat16", False, "floor", "float32"),
         ],
     )
-    def test_mx_quant_params(self, w_dtype, weight_only):
+    def test_mx_quant_params(self, w_dtype, weight_only, round_method, out_dtype):
         fp32_model = copy.deepcopy(self.fp32_model)
-        quant_config = MXQuantConfig(w_dtype=w_dtype, weight_only=weight_only)
-        q_model = quantize(fp32_model, quant_config=quant_config)
+        quant_config = MXQuantConfig(
+            w_dtype=w_dtype,
+            weight_only=weight_only,
+            round_method=round_method,
+            out_dtype=out_dtype,
+        )
+        fp32_model = prepare(model=fp32_model, quant_config=quant_config)
+        q_model = convert(model=fp32_model)
         assert q_model is not None, "Quantization failed!"
 
     def test_mx_quant_accuracy(self):
@@ -72,8 +87,10 @@ def forward(self, x):
         fp32_model = copy.deepcopy(model)
         fp32_model.linear.weight = torch.nn.Parameter(torch.tensor([[0.0, 1.0], [1.0, 0.0]]))
         example_inputs = torch.zeros(3, 2)
+
         quant_config = MXQuantConfig()
-        q_model = quantize(fp32_model, quant_config=quant_config)
+        fp32_model = prepare(model=fp32_model, quant_config=quant_config)
+        q_model = convert(model=fp32_model)
         output1 = fp32_model(example_inputs)
         output2 = q_model(example_inputs)
         # set a big atol to avoid random issue

Original file line number	Diff line number	Diff line change
`@@ -62,9 +62,6 @@ def apply_mx_specs(self):`
`62`	`62`	`axes=[-1],`
`63`	`63`	`)`
`64`	`64`
`65`		`- def append_name(self, postfix):`
`66`		`- self.name += postfix`
`67`		`-`
`68`	`65`	`def forward(self, input):`
`69`	`66`	`if self.mx_none:`
`70`	`67`	`return super().forward(input)`