Revert "Arm backend: Add bfloat16 to relevant vgf tests." (#20663)

Reverts #20660 due to #20299 also getting reverted.



cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils
@Sebastian-Larsson @robell

Author: rascani

backends/arm/test/models/Qwen3_VL/test_qwen3_vl_layers.py

@@ -474,8 +474,6 @@ class Qwen3VLTestCase:
     "qk_norm": TOSA_FP_TEST_CASES["qk_norm"],
 }
 
-VGF_NO_QUANT_BF16_TEST_CASES: dict[str, Qwen3VLTestCase] = TOSA_BF16_TEST_CASES
-
 
 TOSA_MXFP8_TEST_CASES: dict[str, Qwen3VLTestCase] = {
     "text_attention": Qwen3VLTestCase(
@@ -604,25 +602,3 @@ def test_qwen3_vl_vgf_no_quant(test_case: Qwen3VLTestCase):
             transform_passes=list(test_case.transform_passes),
         )
         pipeline.run()
-
-
-@common.SkipIfNoModelConverter
-@common.parametrize(
-    "test_case",
-    VGF_NO_QUANT_BF16_TEST_CASES,
-)
-def test_qwen3_vl_vgf_no_quant_bf16(test_case: Qwen3VLTestCase):
-    model, inputs = test_case.model_cls.prepare_model_and_inputs()
-    model, inputs = _to_bfloat16(model, inputs)
-    with torch.no_grad():
-        pipeline = VgfPipeline[input_t](
-            model,
-            inputs,
-            aten_op=[],
-            exir_op=[],
-            quantize=False,
-            transform_passes=list(test_case.transform_passes),
-            atol=5e-2,
-            rtol=5e-2,
-        )
-        pipeline.run()

backends/arm/test/models/phi3/test_phi3_layers.py

@@ -180,16 +180,14 @@ def _dtype_cases() -> list:
     ]
 
 
-def _vgf_no_quant_dtype_cases() -> list:
-    return [
-        pytest.param(torch.float32, id="fp32"),
-        pytest.param(torch.bfloat16, id="bf16"),
-    ]
-
-
-def _vgf_quant_dtype_cases() -> list:
+def _vgf_dtype_cases() -> list:
     return [
         pytest.param(torch.float32, id="fp32"),
+        pytest.param(
+            torch.bfloat16,
+            marks=pytest.mark.xfail(reason="BF16 runtime support not ready for VGF."),
+            id="bf16",
+        ),
     ]
 
 
@@ -235,14 +233,12 @@ def test_phi3_tosa_INT_layers(name, module_factory, input_factory):
 
 
 @common.SkipIfNoModelConverter
-@pytest.mark.parametrize("dtype", _vgf_no_quant_dtype_cases())
+@pytest.mark.parametrize("dtype", _vgf_dtype_cases())
 @pytest.mark.parametrize("name,module_factory,input_factory", _module_cases())
 def test_phi3_vgf_no_quant_layers(name, module_factory, input_factory, dtype):
     config = _phi3_config()
     module = module_factory(config).to(dtype)
     inputs = input_factory(config, dtype)
-    atol = 1e-02 if dtype == torch.bfloat16 else 1e-03
-    rtol = 1e-02 if dtype == torch.bfloat16 else 1e-03
 
     pipeline = VgfPipeline[input_t1 if len(inputs) == 1 else input_t2](
         module,
@@ -254,14 +250,12 @@ def test_phi3_vgf_no_quant_layers(name, module_factory, input_factory, dtype):
             InsertInt32CastsAfterInt64PlaceholdersPass(),
         ],
         quantize=False,
-        atol=atol,
-        rtol=rtol,
     )
     pipeline.run()
 
 
 @common.SkipIfNoModelConverter
-@pytest.mark.parametrize("dtype", _vgf_quant_dtype_cases())
+@pytest.mark.parametrize("dtype", _vgf_dtype_cases())
 @pytest.mark.parametrize("name,module_factory,input_factory", _module_cases())
 def test_phi3_vgf_quant_layers(name, module_factory, input_factory, dtype):
     config = _phi3_config()

backends/arm/test/models/test_mobilenet_v2_arm.py

@@ -191,26 +191,6 @@ def test_mv2_vgf_no_quant():
     pipeline.run()
 
 
-@common.SkipIfNoModelConverter
-def test_mv2_vgf_no_quant_bf16():
-    bf16_model = models.mobilenetv2.mobilenet_v2(
-        weights=MobileNet_V2_Weights.DEFAULT
-    ).eval()
-    bf16_model = bf16_model.to(torch.bfloat16)
-    bf16_input = normalize(torch.rand((1, 3, 224, 224))).to(torch.bfloat16)
-    pipeline = VgfPipeline[input_t](
-        bf16_model,
-        (bf16_input,),
-        aten_op=[],
-        exir_op=[],
-        use_to_edge_transform_and_lower=True,
-        quantize=False,
-        atol=2.5e-01,
-        rtol=2.5e-01,
-    )
-    pipeline.run()
-
-
 def test_mv2_tosa_INT_FP_partial_quant():
     pipeline = TosaPipelineINT[input_t](
         mv2,

backends/arm/test/models/test_resnet18.py

@@ -14,7 +14,6 @@
     EthosU85PipelineINT,
     TosaPipelineFP,
     TosaPipelineINT,
-    VgfPipeline,
 )
 
 from torchvision import transforms  # type: ignore[import-untyped]
@@ -66,22 +65,6 @@ def test_resnet_18_tosa_FP_bf16():
     pipeline.run()
 
 
-@common.SkipIfNoModelConverter
-def test_resnet_18_vgf_no_quant_bf16():
-    bf16_model = resnet18(weights=ResNet18_Weights).eval()
-    bf16_model = bf16_model.to(torch.bfloat16)
-    bf16_input = normalize(torch.rand((1, 3, 224, 224)) * 2 - 1).to(torch.bfloat16)
-    pipeline = VgfPipeline[input_t](
-        bf16_model,
-        (bf16_input,),
-        aten_op=[],
-        quantize=False,
-        atol=6e-01,
-        rtol=6e-01,
-    )
-    pipeline.run()
-
-
 @common.parametrize("per_channel_quantization", quant_test_data)
 def test_resnet_18_tosa_INT(per_channel_quantization):
     pipeline = TosaPipelineINT[input_t](

backends/arm/test/ops/test_abs.py

@@ -88,10 +88,7 @@ def test_abs_u85_INT(test_data: torch.Tensor):
     pipeline.run()
 
 
-@common.parametrize(
-    "test_data",
-    Abs.test_parameters | Abs.test_parameters_bf16 | Abs.test_parameters_fp16,
-)
+@common.parametrize("test_data", Abs.test_parameters | Abs.test_parameters_fp16)
 @common.SkipIfNoModelConverter
 def test_abs_vgf_no_quant(test_data: input_t1):
     pipeline = VgfPipeline[input_t1](

backends/arm/test/ops/test_add.py

@@ -263,9 +263,7 @@ def test_add_tensor_u85_INT_2(test_data: input_t2):
     pipeline.run()
 
 
-@common.parametrize(
-    "test_data", Add.test_data | Add.test_data_fp16 | Add.test_data_bf16
-)
+@common.parametrize("test_data", Add.test_data | Add.test_data_fp16)
 @common.SkipIfNoModelConverter
 def test_add_tensor_vgf_no_quant(test_data: input_t1):
     pipeline = VgfPipeline[input_t1](

backends/arm/test/ops/test_alias_copy.py

@@ -98,10 +98,7 @@ def test_alias_u85_INT(test_data: input_t1):
     ).run()
 
 
-@common.parametrize(
-    "test_data",
-    AliasCopy.test_data | AliasCopy.test_data_bf16 | AliasCopy.test_data_fp16,
-)
+@common.parametrize("test_data", AliasCopy.test_data | AliasCopy.test_data_fp16)
 @common.SkipIfNoModelConverter
 def test_alias_vgf_no_quant(test_data: input_t1):
     pipeline = VgfPipeline[input_t1](

backends/arm/test/ops/test_amax.py

@@ -212,9 +212,7 @@ def test_max_dim_tosa_FP_not_delegated():
     pipeline.run()
 
 
-@common.parametrize(
-    "test_data", Amax.test_data | Amax.test_data_fp16 | Amax.test_data_bf16
-)
+@common.parametrize("test_data", Amax.test_data | Amax.test_data_fp16)
 @common.SkipIfNoModelConverter
 def test_amax_vgf_no_quant(test_data: Amax.input_t):
     data, dim, keep_dims = test_data()
@@ -242,9 +240,7 @@ def test_amax_vgf_quant(test_data: Amax.input_t):
     pipeline.run()
 
 
-@common.parametrize(
-    "test_data", Max.test_data | Max.test_data_fp16 | Max.test_data_bf16
-)
+@common.parametrize("test_data", Max.test_data | Max.test_data_fp16)
 @common.SkipIfNoModelConverter
 def test_max_dim_vgf_no_quant_to_amax(test_data: Max.input_t):
     data, dim = test_data()

backends/arm/test/ops/test_amin.py

@@ -224,9 +224,7 @@ def test_min_dim_tosa_FP_not_delegated():
     pipeline.run()
 
 
-@common.parametrize(
-    "test_data", Amin.test_data | Amin.test_data_fp16 | Amin.test_data_bf16
-)
+@common.parametrize("test_data", Amin.test_data | Amin.test_data_fp16)
 @common.SkipIfNoModelConverter
 def test_amin_vgf_no_quant(test_data: Amin.input_t):
     data, dim, keep_dims = test_data()
@@ -252,9 +250,7 @@ def test_amin_vgf_quant(test_data: Amin.input_t):
     pipeline.run()
 
 
-@common.parametrize(
-    "test_data", Min.test_data | Min.test_data_fp16 | Min.test_data_bf16
-)
+@common.parametrize("test_data", Min.test_data | Min.test_data_fp16)
 @common.SkipIfNoModelConverter
 def test_min_dim_vgf_no_quant_to_amin(test_data: Min.input_t):
     data, dim = test_data()

backends/arm/test/ops/test_avg_pool2d.py

@@ -286,31 +286,16 @@ def test_avg_pool2d_16a8w_u85_INT(test_module):
     pipeline.run()
 
 
-@common.parametrize(
-    "test_module",
-    test_modules | test_modules_bf16 | test_modules_fp16,
-    xfails={
-        "kernel_3x3_stride_1_pad_1_bf16": "'Unsupported BF16 PAD constant encoding' in emulation layer. MLCE-1887."
-    },
-)
+@common.parametrize("test_module", test_modules | test_modules_fp16)
 @common.SkipIfNoModelConverter
 def test_avg_pool2d_vgf_no_quant(test_module):
     model, input_tensor = test_module()
-    match input_tensor[0].dtype:
-        case torch.bfloat16:
-            atol = 5e-3
-            rtol = 5e-3
-        case _:
-            atol = 1e-3
-            rtol = 1e-3
     pipeline = VgfPipeline[input_t](
         model,
         input_tensor,
         aten_op,
         exir_op,
         quantize=False,
-        atol=atol,
-        rtol=rtol,
     )
     pipeline.run()