From 1bcd7170293e834af0d4e6b3baae867f78ba7546 Mon Sep 17 00:00:00 2001
From: Stephen Jia <ssjia@meta.com>
Date: Mon, 6 Jul 2026 14:59:21 -0700
Subject: [PATCH] Update

[ghstack-poisoned]
---
 backends/arm/runtime/VGFSetup.cpp             |  6 ++++
 backends/arm/scripts/vulkan_utils.sh          | 12 +++----
 .../models/Qwen3_VL/test_qwen3_vl_layers.py   | 24 ++++++++++++++
 .../arm/test/models/phi3/test_phi3_layers.py  | 22 ++++++++-----
 .../arm/test/models/test_mobilenet_v2_arm.py  | 20 ++++++++++++
 backends/arm/test/models/test_resnet18.py     | 17 ++++++++++
 backends/arm/test/ops/test_abs.py             |  5 ++-
 backends/arm/test/ops/test_add.py             |  4 ++-
 backends/arm/test/ops/test_alias_copy.py      |  5 ++-
 backends/arm/test/ops/test_amax.py            |  8 +++--
 backends/arm/test/ops/test_amin.py            |  8 +++--
 backends/arm/test/ops/test_avg_pool2d.py      | 17 +++++++++-
 backends/arm/test/ops/test_cat.py             |  5 ++-
 backends/arm/test/ops/test_ceil.py            |  2 +-
 backends/arm/test/ops/test_clamp.py           |  9 ++++--
 backends/arm/test/ops/test_constant_pad_nd.py |  4 ++-
 backends/arm/test/ops/test_conv2d.py          | 13 ++++++--
 backends/arm/test/ops/test_conv3d.py          | 11 ++++++-
 backends/arm/test/ops/test_cos.py             | 16 ++++++++--
 backends/arm/test/ops/test_depthwise_conv.py  | 15 ++++++++-
 backends/arm/test/ops/test_eq.py              |  8 +++--
 backends/arm/test/ops/test_erf.py             | 16 ++++++++--
 backends/arm/test/ops/test_exp.py             |  6 ++--
 backends/arm/test/ops/test_floor.py           |  2 +-
 backends/arm/test/ops/test_gather.py          |  2 +-
 backends/arm/test/ops/test_ge.py              |  8 +++--
 backends/arm/test/ops/test_gelu.py            |  4 ++-
 backends/arm/test/ops/test_gt.py              |  8 +++--
 backends/arm/test/ops/test_index_put.py       |  6 +++-
 backends/arm/test/ops/test_index_select.py    |  2 +-
 backends/arm/test/ops/test_log.py             | 16 ++++++++--
 backends/arm/test/ops/test_matmul.py          |  2 +-
 backends/arm/test/ops/test_max_pool.py        |  4 ++-
 backends/arm/test/ops/test_maximum.py         |  7 +++-
 backends/arm/test/ops/test_minimum.py         |  7 +++-
 backends/arm/test/ops/test_mul.py             |  1 +
 backends/arm/test/ops/test_permute.py         |  4 ++-
 backends/arm/test/ops/test_pow.py             | 11 +++++--
 backends/arm/test/ops/test_reciprocal.py      | 16 ++++++++--
 backends/arm/test/ops/test_repeat.py          |  4 ++-
 backends/arm/test/ops/test_round.py           |  2 +-
 backends/arm/test/ops/test_rsqrt.py           | 17 ++++++++--
 backends/arm/test/ops/test_select_scatter.py  |  2 +-
 backends/arm/test/ops/test_sigmoid.py         | 16 ++++++++--
 backends/arm/test/ops/test_sin.py             | 16 ++++++++--
 backends/arm/test/ops/test_slice.py           |  4 ++-
 backends/arm/test/ops/test_slice_scatter.py   |  6 +++-
 backends/arm/test/ops/test_sub.py             |  8 +++--
 backends/arm/test/ops/test_sum.py             | 17 ++++++++--
 backends/arm/test/ops/test_tanh.py            | 16 ++++++++--
 backends/arm/test/ops/test_to_copy.py         | 13 +++++++-
 .../arm/test/ops/test_transpose_conv2d.py     |  5 ++-
 backends/arm/test/ops/test_unfold_copy.py     |  2 +-
 .../arm/test/ops/test_upsample_bilinear2d.py  | 32 ++++++++++++++++---
 .../arm/test/ops/test_upsample_nearest2d.py   | 14 +++++---
 backends/arm/test/ops/test_where.py           |  2 +-
 backends/arm/test/runner_utils.py             |  3 +-
 backends/arm/test/tester/test_pipeline.py     | 14 ++++----
 58 files changed, 448 insertions(+), 98 deletions(-)

diff --git a/backends/arm/runtime/VGFSetup.cpp b/backends/arm/runtime/VGFSetup.cpp
index 88ed1758163..dd92af1d15b 100644
--- a/backends/arm/runtime/VGFSetup.cpp
+++ b/backends/arm/runtime/VGFSetup.cpp
@@ -66,6 +66,7 @@ enum class FormatScalarKind {
   Uint,
   Sint,
   Float,
+  BFloat,
 };
 
 struct FormatInfo {
@@ -157,6 +158,7 @@ static uint32_t get_format_component_count(VkFormat format) {
     case VK_FORMAT_R16_UINT:
     case VK_FORMAT_R16_SINT:
     case VK_FORMAT_R16_SFLOAT:
+    case VK_FORMAT_R16_SFLOAT_FPENCODING_BFLOAT16_ARM:
     case VK_FORMAT_R32_UINT:
     case VK_FORMAT_R32_SINT:
     case VK_FORMAT_R32_SFLOAT:
@@ -209,6 +211,9 @@ static bool get_format_info(VkFormat format, FormatInfo* info) {
     case VK_FORMAT_R16_SFLOAT:
       *info = FormatInfo{1, 2, FormatScalarKind::Float};
       return true;
+    case VK_FORMAT_R16_SFLOAT_FPENCODING_BFLOAT16_ARM:
+      *info = FormatInfo{1, 2, FormatScalarKind::BFloat};
+      return true;
     case VK_FORMAT_R32_UINT:
       *info = FormatInfo{1, 4, FormatScalarKind::Uint};
       return true;
@@ -3696,6 +3701,7 @@ static uint32_t get_format_size(VkFormat format) {
     case VK_FORMAT_R16_UINT:
     case VK_FORMAT_R16_SINT:
     case VK_FORMAT_R16_SFLOAT:
+    case VK_FORMAT_R16_SFLOAT_FPENCODING_BFLOAT16_ARM:
     case VK_FORMAT_R8G8_UINT:
     case VK_FORMAT_R8G8_SINT:
       return 2;
diff --git a/backends/arm/scripts/vulkan_utils.sh b/backends/arm/scripts/vulkan_utils.sh
index f81a0cd0468..e99693f2b17 100644
--- a/backends/arm/scripts/vulkan_utils.sh
+++ b/backends/arm/scripts/vulkan_utils.sh
@@ -26,21 +26,21 @@ vulkan_sdk_arch="${ARCH}"
 # macOS and Linux x86_64 use the official LunarG SDK tarballs. Linux ARM64
 # uses a separately repackaged mirror of the same SDK version.
 if [[ "${os_name}" == "Darwin" ]]; then
-    vulkan_sdk_version="1.4.341.1"
+    vulkan_sdk_version="1.4.350.0"
     vulkan_sdk_arch="macOS"
     vulkan_sdk_url="https://sdk.lunarg.com/sdk/download/${vulkan_sdk_version}/mac/vulkansdk-macos-${vulkan_sdk_version}.zip"
-    vulkan_sdk_sha256="632cbe96c8ed6ed00c6ce25e3a7738c466134f76586e1c51f1419410d7f9042e"
+    vulkan_sdk_sha256="7acc181b8fd9b4781bf51ed086222ec95d22004b85b3d0a6683a7e48ca5a1679"
 elif [[ "${os_name}" == "Linux" ]] && [[ "${ARCH}" == "x86_64" ]]; then
-    vulkan_sdk_version="1.4.341.1"
+    vulkan_sdk_version="1.4.350.0"
     vulkan_sdk_url="https://sdk.lunarg.com/sdk/download/${vulkan_sdk_version}/linux/vulkansdk-linux-x86_64-${vulkan_sdk_version}.tar.xz"
-    vulkan_sdk_sha256="3bf0f762afb6c79bc6a9d9fb5998745ccff928800a29619b501ed9de7fd9789b"
+    vulkan_sdk_sha256="b65f068ab36263559da49d7cacd7e7b9df23824ca8b68ccc522a2b06f5725df2"
 elif [[ "${os_name}" == "Linux" ]] && ([[ "${ARCH}" == "aarch64" ]] || [[ "${ARCH}" == "arm64" ]]); then
-    vulkan_sdk_version="1.4.341.1"
+    vulkan_sdk_version="1.4.350.0"
     if [[ "${vulkan_sdk_arch}" == "arm64" ]]; then
         vulkan_sdk_arch="aarch64"
     fi
     vulkan_sdk_url="https://github.com/jakoch/vulkan-sdk-arm/releases/download/${vulkan_sdk_version}/vulkansdk-ubuntu-22.04-arm-${vulkan_sdk_version}.tar.xz"
-    vulkan_sdk_sha256="345312aee2c835e128b30653278593f899a659a7ba287c571cafb22acb708b8f"
+    vulkan_sdk_sha256="9e403d444219bb7c17e9231b580d704453e2afa30a1c2fdd568d1776dc68790b"
 else
     log_step "vulkan" "Error: only macOS and Linux are supported (detected ${os_name}); architecture must be x86-64 or aarch64/arm64"
     exit 1
diff --git a/backends/arm/test/models/Qwen3_VL/test_qwen3_vl_layers.py b/backends/arm/test/models/Qwen3_VL/test_qwen3_vl_layers.py
index 6573371bb78..0be23f4dbc2 100644
--- a/backends/arm/test/models/Qwen3_VL/test_qwen3_vl_layers.py
+++ b/backends/arm/test/models/Qwen3_VL/test_qwen3_vl_layers.py
@@ -474,6 +474,8 @@ 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(
@@ -602,3 +604,25 @@ 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()
diff --git a/backends/arm/test/models/phi3/test_phi3_layers.py b/backends/arm/test/models/phi3/test_phi3_layers.py
index c598ff2f664..34e3a25a304 100644
--- a/backends/arm/test/models/phi3/test_phi3_layers.py
+++ b/backends/arm/test/models/phi3/test_phi3_layers.py
@@ -180,14 +180,16 @@ def _dtype_cases() -> list:
     ]
 
 
-def _vgf_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:
     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",
-        ),
     ]
 
 
@@ -233,12 +235,14 @@ def test_phi3_tosa_INT_layers(name, module_factory, input_factory):
 
 
 @common.SkipIfNoModelConverter
-@pytest.mark.parametrize("dtype", _vgf_dtype_cases())
+@pytest.mark.parametrize("dtype", _vgf_no_quant_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,
@@ -250,12 +254,14 @@ 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_dtype_cases())
+@pytest.mark.parametrize("dtype", _vgf_quant_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()
diff --git a/backends/arm/test/models/test_mobilenet_v2_arm.py b/backends/arm/test/models/test_mobilenet_v2_arm.py
index fb624f3eb3f..0616d34c90b 100644
--- a/backends/arm/test/models/test_mobilenet_v2_arm.py
+++ b/backends/arm/test/models/test_mobilenet_v2_arm.py
@@ -191,6 +191,26 @@ 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,
diff --git a/backends/arm/test/models/test_resnet18.py b/backends/arm/test/models/test_resnet18.py
index db1e021461a..0430873246f 100644
--- a/backends/arm/test/models/test_resnet18.py
+++ b/backends/arm/test/models/test_resnet18.py
@@ -14,6 +14,7 @@
     EthosU85PipelineINT,
     TosaPipelineFP,
     TosaPipelineINT,
+    VgfPipeline,
 )
 
 from torchvision import transforms  # type: ignore[import-untyped]
@@ -65,6 +66,22 @@ 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](
diff --git a/backends/arm/test/ops/test_abs.py b/backends/arm/test/ops/test_abs.py
index 956b0b12697..c9297b27b55 100644
--- a/backends/arm/test/ops/test_abs.py
+++ b/backends/arm/test/ops/test_abs.py
@@ -88,7 +88,10 @@ def test_abs_u85_INT(test_data: torch.Tensor):
     pipeline.run()
 
 
-@common.parametrize("test_data", Abs.test_parameters | Abs.test_parameters_fp16)
+@common.parametrize(
+    "test_data",
+    Abs.test_parameters | Abs.test_parameters_bf16 | Abs.test_parameters_fp16,
+)
 @common.SkipIfNoModelConverter
 def test_abs_vgf_no_quant(test_data: input_t1):
     pipeline = VgfPipeline[input_t1](
diff --git a/backends/arm/test/ops/test_add.py b/backends/arm/test/ops/test_add.py
index 632de5e999a..3e4a5510005 100644
--- a/backends/arm/test/ops/test_add.py
+++ b/backends/arm/test/ops/test_add.py
@@ -263,7 +263,9 @@ def test_add_tensor_u85_INT_2(test_data: input_t2):
     pipeline.run()
 
 
-@common.parametrize("test_data", Add.test_data | Add.test_data_fp16)
+@common.parametrize(
+    "test_data", Add.test_data | Add.test_data_fp16 | Add.test_data_bf16
+)
 @common.SkipIfNoModelConverter
 def test_add_tensor_vgf_no_quant(test_data: input_t1):
     pipeline = VgfPipeline[input_t1](
diff --git a/backends/arm/test/ops/test_alias_copy.py b/backends/arm/test/ops/test_alias_copy.py
index d57fffd15d6..a345b4d3b9e 100644
--- a/backends/arm/test/ops/test_alias_copy.py
+++ b/backends/arm/test/ops/test_alias_copy.py
@@ -98,7 +98,10 @@ def test_alias_u85_INT(test_data: input_t1):
     ).run()
 
 
-@common.parametrize("test_data", AliasCopy.test_data | AliasCopy.test_data_fp16)
+@common.parametrize(
+    "test_data",
+    AliasCopy.test_data | AliasCopy.test_data_bf16 | AliasCopy.test_data_fp16,
+)
 @common.SkipIfNoModelConverter
 def test_alias_vgf_no_quant(test_data: input_t1):
     pipeline = VgfPipeline[input_t1](
diff --git a/backends/arm/test/ops/test_amax.py b/backends/arm/test/ops/test_amax.py
index c1d8e047ac7..28f1627eed5 100644
--- a/backends/arm/test/ops/test_amax.py
+++ b/backends/arm/test/ops/test_amax.py
@@ -212,7 +212,9 @@ def test_max_dim_tosa_FP_not_delegated():
     pipeline.run()
 
 
-@common.parametrize("test_data", Amax.test_data | Amax.test_data_fp16)
+@common.parametrize(
+    "test_data", Amax.test_data | Amax.test_data_fp16 | Amax.test_data_bf16
+)
 @common.SkipIfNoModelConverter
 def test_amax_vgf_no_quant(test_data: Amax.input_t):
     data, dim, keep_dims = test_data()
@@ -240,7 +242,9 @@ def test_amax_vgf_quant(test_data: Amax.input_t):
     pipeline.run()
 
 
-@common.parametrize("test_data", Max.test_data | Max.test_data_fp16)
+@common.parametrize(
+    "test_data", Max.test_data | Max.test_data_fp16 | Max.test_data_bf16
+)
 @common.SkipIfNoModelConverter
 def test_max_dim_vgf_no_quant_to_amax(test_data: Max.input_t):
     data, dim = test_data()
diff --git a/backends/arm/test/ops/test_amin.py b/backends/arm/test/ops/test_amin.py
index 9068feceaf6..14ea58e0d2c 100644
--- a/backends/arm/test/ops/test_amin.py
+++ b/backends/arm/test/ops/test_amin.py
@@ -224,7 +224,9 @@ def test_min_dim_tosa_FP_not_delegated():
     pipeline.run()
 
 
-@common.parametrize("test_data", Amin.test_data | Amin.test_data_fp16)
+@common.parametrize(
+    "test_data", Amin.test_data | Amin.test_data_fp16 | Amin.test_data_bf16
+)
 @common.SkipIfNoModelConverter
 def test_amin_vgf_no_quant(test_data: Amin.input_t):
     data, dim, keep_dims = test_data()
@@ -250,7 +252,9 @@ def test_amin_vgf_quant(test_data: Amin.input_t):
     pipeline.run()
 
 
-@common.parametrize("test_data", Min.test_data | Min.test_data_fp16)
+@common.parametrize(
+    "test_data", Min.test_data | Min.test_data_fp16 | Min.test_data_bf16
+)
 @common.SkipIfNoModelConverter
 def test_min_dim_vgf_no_quant_to_amin(test_data: Min.input_t):
     data, dim = test_data()
diff --git a/backends/arm/test/ops/test_avg_pool2d.py b/backends/arm/test/ops/test_avg_pool2d.py
index 36b3d3983fc..0eb59f8d2b9 100644
--- a/backends/arm/test/ops/test_avg_pool2d.py
+++ b/backends/arm/test/ops/test_avg_pool2d.py
@@ -286,16 +286,31 @@ def test_avg_pool2d_16a8w_u85_INT(test_module):
     pipeline.run()
 
 
-@common.parametrize("test_module", test_modules | test_modules_fp16)
+@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.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()
 
diff --git a/backends/arm/test/ops/test_cat.py b/backends/arm/test/ops/test_cat.py
index 29738ddbe32..1286435ec89 100644
--- a/backends/arm/test/ops/test_cat.py
+++ b/backends/arm/test/ops/test_cat.py
@@ -201,7 +201,10 @@ def test_cat_u85_INT(test_data: Tuple):
     pipeline.run()
 
 
-@common.parametrize("test_data", Cat.test_parameters | Cat.test_parameters_fp16)
+@common.parametrize(
+    "test_data",
+    Cat.test_parameters | Cat.test_parameters_bf16 | Cat.test_parameters_fp16,
+)
 @common.SkipIfNoModelConverter
 def test_cat_vgf_no_quant(test_data: Tuple):
     pipeline = VgfPipeline[input_t1](
diff --git a/backends/arm/test/ops/test_ceil.py b/backends/arm/test/ops/test_ceil.py
index 7332d0f40ce..d5c8e3a3f09 100644
--- a/backends/arm/test/ops/test_ceil.py
+++ b/backends/arm/test/ops/test_ceil.py
@@ -122,7 +122,7 @@ def test_ceil_u85_INT(test_data: input_t1):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data | test_data_fp16)
+@common.parametrize("test_data", test_data | test_data_bf16 | test_data_fp16)
 @common.SkipIfNoModelConverter
 def test_ceil_vgf_no_quant(test_data: input_t1):
     module, data = test_data()
diff --git a/backends/arm/test/ops/test_clamp.py b/backends/arm/test/ops/test_clamp.py
index 0d7033df52e..61ee5361678 100644
--- a/backends/arm/test/ops/test_clamp.py
+++ b/backends/arm/test/ops/test_clamp.py
@@ -230,7 +230,9 @@ def test_clamp_u85_INT_16a8w(test_data):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_suite | test_data_suite_fp16)
+@common.parametrize(
+    "test_data", test_data_suite | test_data_suite_bf16 | test_data_suite_fp16
+)
 @common.SkipIfNoModelConverter
 def test_clamp_vgf_no_quant(test_data):
     input_tensor, min_val, max_val = test_data()
@@ -546,7 +548,10 @@ def test_clamp_u85_INT_16a8w_tensor(test_data):
 
 
 @common.parametrize(
-    "test_data", test_data_suite_tensor_FP | test_data_suite_tensor_fp16
+    "test_data",
+    test_data_suite_tensor_FP
+    | test_data_suite_tensor_bf16
+    | test_data_suite_tensor_fp16,
 )
 @common.SkipIfNoModelConverter
 def test_clamp_vgf_no_quant_tensor(test_data):
diff --git a/backends/arm/test/ops/test_constant_pad_nd.py b/backends/arm/test/ops/test_constant_pad_nd.py
index 96d829851ed..d465c21e6b8 100644
--- a/backends/arm/test/ops/test_constant_pad_nd.py
+++ b/backends/arm/test/ops/test_constant_pad_nd.py
@@ -344,7 +344,9 @@ def test_constant_pad_nd_tosa_INT_a16w8(test_data: Tuple):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_suite | test_data_suite_fp16)
+@common.parametrize(
+    "test_data", test_data_suite | test_data_suite_bf16 | test_data_suite_fp16
+)
 @common.SkipIfNoModelConverter
 def test_constant_pad_nd_vgf_no_quant(test_data: Tuple):
     inp, padding, value, mode = test_data()
diff --git a/backends/arm/test/ops/test_conv2d.py b/backends/arm/test/ops/test_conv2d.py
index db8a7c6b323..91898b8c32c 100644
--- a/backends/arm/test/ops/test_conv2d.py
+++ b/backends/arm/test/ops/test_conv2d.py
@@ -718,18 +718,25 @@ def test_convolution_2d_u85_INT_a8w4(test_data):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_FP | test_data_FP_fp16)
+@common.parametrize("test_data", test_data_FP | test_data_FP_bf16 | test_data_FP_fp16)
 @common.SkipIfNoModelConverter
 def test_convolution_2d_vgf_no_quant(test_data):
     model = test_data()
+    match model.dtype:
+        case torch.bfloat16:
+            atol = 2e-2
+            rtol = 2e-2
+        case _:
+            atol = 3e-3
+            rtol = 3e-3
     pipeline = VgfPipeline[input_t](
         model,
         model.get_inputs(),
         aten_op,
         exir_op,
         quantize=False,
-        atol=3e-3,
-        rtol=3e-3,
+        atol=atol,
+        rtol=rtol,
     )
     pipeline.run()
 
diff --git a/backends/arm/test/ops/test_conv3d.py b/backends/arm/test/ops/test_conv3d.py
index a1418f77b53..09cd44525c5 100644
--- a/backends/arm/test/ops/test_conv3d.py
+++ b/backends/arm/test/ops/test_conv3d.py
@@ -772,16 +772,25 @@ def test_convolution_3d_u85_INT_a8w4(test_data):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_FP | test_data_FP_fp16)
+@common.parametrize("test_data", test_data_FP | test_data_FP_bf16 | test_data_FP_fp16)
 @common.SkipIfNoModelConverter
 def test_convolution_3d_vgf_no_quant(test_data):
     model = test_data()
+    match model.dtype:
+        case torch.bfloat16:
+            atol = 1e-2
+            rtol = 1e-2
+        case _:
+            atol = 1e-3
+            rtol = 1e-3
     pipeline = VgfPipeline[input_t](
         model,
         model.get_inputs(),
         aten_op,
         exir_op,
         quantize=False,
+        atol=atol,
+        rtol=rtol,
     )
     pipeline.run()
 
diff --git a/backends/arm/test/ops/test_cos.py b/backends/arm/test/ops/test_cos.py
index e020c3de971..e3702d4caf3 100644
--- a/backends/arm/test/ops/test_cos.py
+++ b/backends/arm/test/ops/test_cos.py
@@ -98,15 +98,27 @@ def test_cos_u85_INT(test_data: Tuple):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_suite | test_data_suite_fp16)
+@common.parametrize(
+    "test_data", test_data_suite | test_data_suite_bf16 | test_data_suite_fp16
+)
 @common.SkipIfNoModelConverter
 def test_cos_vgf_no_quant(test_data: Tuple):
+    data = test_data()
+    match data.dtype:
+        case torch.bfloat16:
+            atol = 5e-3
+            rtol = 5e-3
+        case _:
+            atol = 1e-3
+            rtol = 1e-3
     pipeline = VgfPipeline[input_t1](
         Cos(),
-        (test_data(),),
+        (data,),
         aten_op,
         exir_op=[],
         quantize=False,
+        atol=atol,
+        rtol=rtol,
     )
     pipeline.run()
 
diff --git a/backends/arm/test/ops/test_depthwise_conv.py b/backends/arm/test/ops/test_depthwise_conv.py
index 95fe3b01826..a81a656017f 100644
--- a/backends/arm/test/ops/test_depthwise_conv.py
+++ b/backends/arm/test/ops/test_depthwise_conv.py
@@ -340,17 +340,30 @@ def test_convolution_2d_tosa_INT_a8w4_depthwise(test_data):
 
 
 @common.parametrize(
-    "test_data", test_data_conv1d_FP | test_data_conv2d_FP | test_data_conv2d_FP_fp16
+    "test_data",
+    test_data_conv1d_FP
+    | test_data_conv2d_FP
+    | test_data_conv2d_FP_bf16
+    | test_data_conv2d_FP_fp16,
 )
 @common.SkipIfNoModelConverter
 def test_convolution_2d_vgf_no_quant_depthwise(test_data: torch.nn.Module):
     model = test_data()
+    match model.dtype:
+        case torch.bfloat16:
+            atol = 2e-2
+            rtol = 2e-2
+        case _:
+            atol = 1e-3
+            rtol = 1e-3
     pipeline = VgfPipeline[input_t](
         model,
         model.get_inputs(),
         aten_op=[],
         exir_op=exir_op,
         quantize=False,
+        atol=atol,
+        rtol=rtol,
     )
     pipeline.run()
 
diff --git a/backends/arm/test/ops/test_eq.py b/backends/arm/test/ops/test_eq.py
index 3a28a8a061c..094f2ce7f45 100644
--- a/backends/arm/test/ops/test_eq.py
+++ b/backends/arm/test/ops/test_eq.py
@@ -285,7 +285,9 @@ def test_eq_scalar_u85_INT_16a8w(test_module):
     pipeline.run()
 
 
-@common.parametrize("test_module", test_data_tensor | test_data_tensor_fp16)
+@common.parametrize(
+    "test_module", test_data_tensor | test_data_tensor_bf16 | test_data_tensor_fp16
+)
 @common.SkipIfNoModelConverter
 def test_eq_scalar_vgf_no_quant_tensor(test_module):
     pipeline = VgfPipeline[input_t](
@@ -298,7 +300,9 @@ def test_eq_scalar_vgf_no_quant_tensor(test_module):
     pipeline.run()
 
 
-@common.parametrize("test_module", test_data_scalar | test_data_scalar_fp16)
+@common.parametrize(
+    "test_module", test_data_scalar | test_data_scalar_bf16 | test_data_scalar_fp16
+)
 @common.SkipIfNoModelConverter
 def test_eq_scalar_vgf_no_quant(test_module):
     pipeline = VgfPipeline[input_t](
diff --git a/backends/arm/test/ops/test_erf.py b/backends/arm/test/ops/test_erf.py
index a5501fbe87d..1d3c5d5e79b 100644
--- a/backends/arm/test/ops/test_erf.py
+++ b/backends/arm/test/ops/test_erf.py
@@ -83,15 +83,27 @@ def test_erf_u85_INT(test_data: input_t1):
     pipeline.run()
 
 
-@common.parametrize("test_data", Erf.test_data | Erf.test_data_fp16)
+@common.parametrize(
+    "test_data", Erf.test_data | Erf.test_data_bf16 | Erf.test_data_fp16
+)
 @common.SkipIfNoModelConverter
 def test_erf_vgf_no_quant(test_data: input_t1):
+    data = test_data()
+    match data[0].dtype:
+        case torch.bfloat16:
+            atol = 5e-3
+            rtol = 5e-3
+        case _:
+            atol = 1e-3
+            rtol = 1e-3
     pipeline = VgfPipeline[input_t1](
         Erf(),
-        test_data(),
+        data,
         aten_op,
         exir_op,
         quantize=False,
+        atol=atol,
+        rtol=rtol,
     )
     pipeline.run()
 
diff --git a/backends/arm/test/ops/test_exp.py b/backends/arm/test/ops/test_exp.py
index 1661e5f5e79..300d5fc8a56 100644
--- a/backends/arm/test/ops/test_exp.py
+++ b/backends/arm/test/ops/test_exp.py
@@ -96,11 +96,13 @@ def test_exp_u85_INT(test_data: Tuple):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_suite | test_data_suite_fp16)
+@common.parametrize(
+    "test_data", test_data_suite | test_data_suite_bf16 | test_data_suite_fp16
+)
 @common.SkipIfNoModelConverter
 def test_exp_vgf_no_quant(test_data: Tuple):
     data = test_data()
-    if data.dtype == torch.float16:
+    if data.dtype in (torch.float16, torch.bfloat16):
         atol = 2e-2
         rtol = 2e-2
     else:
diff --git a/backends/arm/test/ops/test_floor.py b/backends/arm/test/ops/test_floor.py
index 736f2e5d3ee..606a61a2ff5 100644
--- a/backends/arm/test/ops/test_floor.py
+++ b/backends/arm/test/ops/test_floor.py
@@ -118,7 +118,7 @@ def test_floor_u85_INT(test_data: input_t1):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data | test_data_fp16)
+@common.parametrize("test_data", test_data | test_data_bf16 | test_data_fp16)
 @common.SkipIfNoModelConverter
 def test_floor_vgf_no_quant(test_data: input_t1):
     module, data = test_data()
diff --git a/backends/arm/test/ops/test_gather.py b/backends/arm/test/ops/test_gather.py
index 22eada92c07..08112e39667 100644
--- a/backends/arm/test/ops/test_gather.py
+++ b/backends/arm/test/ops/test_gather.py
@@ -230,7 +230,7 @@ def test_gather_u85_INT(test_data: input_params):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_fp | test_data_int)
+@common.parametrize("test_data", test_data_fp | test_data_fp_bf16 | test_data_int)
 @common.SkipIfNoModelConverter
 def test_gather_vgf_no_quant(test_data: input_params):
     pipeline = VgfPipeline[input_params](
diff --git a/backends/arm/test/ops/test_ge.py b/backends/arm/test/ops/test_ge.py
index 454e8e4269c..130727a0a56 100644
--- a/backends/arm/test/ops/test_ge.py
+++ b/backends/arm/test/ops/test_ge.py
@@ -293,7 +293,9 @@ def test_ge_scalar_16a8w_u85_INT(test_module):
     pipeline.run()
 
 
-@common.parametrize("test_module", test_data_tensor | test_data_tensor_fp16)
+@common.parametrize(
+    "test_module", test_data_tensor | test_data_tensor_bf16 | test_data_tensor_fp16
+)
 @common.SkipIfNoModelConverter
 def test_ge_tensor_vgf_no_quant(test_module):
     pipeline = VgfPipeline[input_t](
@@ -319,7 +321,9 @@ def test_ge_tensor_vgf_quant(test_module):
     pipeline.run()
 
 
-@common.parametrize("test_module", test_data_scalar | test_data_scalar_fp16)
+@common.parametrize(
+    "test_module", test_data_scalar | test_data_scalar_bf16 | test_data_scalar_fp16
+)
 @common.SkipIfNoModelConverter
 def test_ge_scalar_vgf_no_quant(test_module):
     pipeline = VgfPipeline[input_t](
diff --git a/backends/arm/test/ops/test_gelu.py b/backends/arm/test/ops/test_gelu.py
index 84626c70b4d..12a6cf49974 100644
--- a/backends/arm/test/ops/test_gelu.py
+++ b/backends/arm/test/ops/test_gelu.py
@@ -151,7 +151,7 @@ def test_gelu_u85_INT(test_data: input_t1):
     ).run()
 
 
-@common.parametrize("test_data", Gelu.test_data)
+@common.parametrize("test_data", Gelu.test_data | Gelu.test_data_bf16)
 @common.SkipIfNoModelConverter
 def test_gelu_vgf_no_quant(test_data: input_t1):
     approximate, data = test_data()
@@ -162,6 +162,8 @@ def test_gelu_vgf_no_quant(test_data: input_t1):
         Gelu.exir_op,
         quantize=False,
     )
+    if data.dtype == torch.bfloat16:
+        pipeline.change_args("run_method_and_compare_outputs", atol=1e-2, rtol=1e-2)
     pipeline.run()
 
 
diff --git a/backends/arm/test/ops/test_gt.py b/backends/arm/test/ops/test_gt.py
index bd07407d0af..963f9d427b8 100644
--- a/backends/arm/test/ops/test_gt.py
+++ b/backends/arm/test/ops/test_gt.py
@@ -298,7 +298,9 @@ def test_gt_scalar_16a8w_u85_INT(test_module):
     pipeline.run()
 
 
-@common.parametrize("test_module", test_data_tensor | test_data_tensor_fp16)
+@common.parametrize(
+    "test_module", test_data_tensor | test_data_tensor_bf16 | test_data_tensor_fp16
+)
 @common.SkipIfNoModelConverter
 def test_gt_tensor_vgf_no_quant(test_module):
     pipeline = VgfPipeline[input_t](
@@ -311,7 +313,9 @@ def test_gt_tensor_vgf_no_quant(test_module):
     pipeline.run()
 
 
-@common.parametrize("test_module", test_data_scalar | test_data_scalar_fp16)
+@common.parametrize(
+    "test_module", test_data_scalar | test_data_scalar_bf16 | test_data_scalar_fp16
+)
 @common.SkipIfNoModelConverter
 def test_gt_scalar_vgf_no_quant(test_module):
     pipeline = VgfPipeline[input_t](
diff --git a/backends/arm/test/ops/test_index_put.py b/backends/arm/test/ops/test_index_put.py
index 5d6bc10e2af..93b2262d6fc 100644
--- a/backends/arm/test/ops/test_index_put.py
+++ b/backends/arm/test/ops/test_index_put.py
@@ -463,7 +463,11 @@ def test_index_put_u85_INT(test_module: input_t):
 
 
 @common.SkipIfNoModelConverter
-@common.parametrize("test_module", test_data_suite_fp | test_data_int, xfails=xfails)
+@common.parametrize(
+    "test_module",
+    test_data_suite_fp | test_data_suite_bf16 | test_data_int,
+    xfails=xfails,
+)
 def test_index_put_vgf_no_quant(test_module: input_t):
     pipeline = VgfPipeline[input_t](
         IndexPut(),
diff --git a/backends/arm/test/ops/test_index_select.py b/backends/arm/test/ops/test_index_select.py
index 110b3429d75..5410bc09a4e 100644
--- a/backends/arm/test/ops/test_index_select.py
+++ b/backends/arm/test/ops/test_index_select.py
@@ -205,7 +205,7 @@ def test_index_select_u85_INT(test_data: input_params):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_fp | test_data_int)
+@common.parametrize("test_data", test_data_fp | test_data_fp_bf16 | test_data_int)
 @common.SkipIfNoModelConverter
 def test_index_select_vgf_no_quant(test_data: input_params):
     pipeline = VgfPipeline[input_params](
diff --git a/backends/arm/test/ops/test_log.py b/backends/arm/test/ops/test_log.py
index e5fb4bacdc5..1a37610b082 100644
--- a/backends/arm/test/ops/test_log.py
+++ b/backends/arm/test/ops/test_log.py
@@ -88,15 +88,27 @@ def test_log_u85_INT(test_data: input_t1):
     ).run()
 
 
-@common.parametrize("test_data", test_data_suite | test_data_suite_fp16)
+@common.parametrize(
+    "test_data", test_data_suite | test_data_suite_bf16 | test_data_suite_fp16
+)
 @common.SkipIfNoModelConverter
 def test_log_vgf_no_quant(test_data: input_t1):
+    data = test_data()
+    match data.dtype:
+        case torch.bfloat16:
+            atol = 2e-2
+            rtol = 2e-2
+        case _:
+            atol = 1e-3
+            rtol = 1e-3
     pipeline = VgfPipeline[input_t1](
         Log(),
-        (test_data(),),
+        (data,),
         aten_op,
         exir_op,
         quantize=False,
+        atol=atol,
+        rtol=rtol,
     )
     pipeline.run()
 
diff --git a/backends/arm/test/ops/test_matmul.py b/backends/arm/test/ops/test_matmul.py
index 9fc93bfd9b2..4b9ab93ad56 100644
--- a/backends/arm/test/ops/test_matmul.py
+++ b/backends/arm/test/ops/test_matmul.py
@@ -455,7 +455,7 @@ def test_matmul_u85_INT(test_case: test_case_t):
     pipeline.run()
 
 
-@common.parametrize("test_case", test_suite | test_suite_fp16)
+@common.parametrize("test_case", test_suite | test_suite_fp16 | test_suite_bf16)
 @common.SkipIfNoModelConverter
 def test_matmul_vgf_no_quant(test_case: test_case_t):
     test_data = test_case()
diff --git a/backends/arm/test/ops/test_max_pool.py b/backends/arm/test/ops/test_max_pool.py
index eed12eac06e..22dfe09b070 100644
--- a/backends/arm/test/ops/test_max_pool.py
+++ b/backends/arm/test/ops/test_max_pool.py
@@ -373,7 +373,9 @@ def test_max_pool2d_tosa_INT_dilation(test_data):
 
 
 # VGF tests
-@common.parametrize("test_data", test_data_suite | test_data_suite_fp16)
+@common.parametrize(
+    "test_data", test_data_suite | test_data_suite_bf16 | test_data_suite_fp16
+)
 @common.SkipIfNoModelConverter
 def test_max_pool2d_vgf_no_quant(test_data: torch.Tensor):
     test_data, model_params = test_data()
diff --git a/backends/arm/test/ops/test_maximum.py b/backends/arm/test/ops/test_maximum.py
index 0ee3ead794f..c9ef999f86b 100644
--- a/backends/arm/test/ops/test_maximum.py
+++ b/backends/arm/test/ops/test_maximum.py
@@ -100,7 +100,12 @@ def test_maximum_u85_INT(test_data: Tuple):
     ).run()
 
 
-@common.parametrize("test_data", Maximum.test_parameters | Maximum.test_parameters_fp16)
+@common.parametrize(
+    "test_data",
+    Maximum.test_parameters
+    | Maximum.test_parameters_bf16
+    | Maximum.test_parameters_fp16,
+)
 @common.SkipIfNoModelConverter
 def test_maximum_vgf_no_quant(test_data: Tuple):
     pipeline = VgfPipeline[test_t](
diff --git a/backends/arm/test/ops/test_minimum.py b/backends/arm/test/ops/test_minimum.py
index 4cf8744975b..b3df52e227f 100644
--- a/backends/arm/test/ops/test_minimum.py
+++ b/backends/arm/test/ops/test_minimum.py
@@ -100,7 +100,12 @@ def test_minimum_u85_INT(test_data: Tuple):
     ).run()
 
 
-@common.parametrize("test_data", Minimum.test_parameters | Minimum.test_parameters_fp16)
+@common.parametrize(
+    "test_data",
+    Minimum.test_parameters
+    | Minimum.test_parameters_bf16
+    | Minimum.test_parameters_fp16,
+)
 @common.SkipIfNoModelConverter
 def test_minimum_vgf_no_quant(test_data: test_t):
     pipeline = VgfPipeline[test_t](
diff --git a/backends/arm/test/ops/test_mul.py b/backends/arm/test/ops/test_mul.py
index a9fce9f2e91..adfcb1f6d58 100644
--- a/backends/arm/test/ops/test_mul.py
+++ b/backends/arm/test/ops/test_mul.py
@@ -263,6 +263,7 @@ def test_mul_tensor_u85_INT_int32(test_data: torch.Tensor):
 @common.parametrize(
     "test_data",
     test_data_suite
+    | test_data_suite_bf16
     | test_data_suite_fp16
     | test_data_suite_2
     | test_data_int32_without_broadcasting,
diff --git a/backends/arm/test/ops/test_permute.py b/backends/arm/test/ops/test_permute.py
index 6819929104e..29e134e464b 100644
--- a/backends/arm/test/ops/test_permute.py
+++ b/backends/arm/test/ops/test_permute.py
@@ -161,7 +161,9 @@ def test_permute_u85_INT(test_data: torch.Tensor):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_suite | test_data_suite_fp16)
+@common.parametrize(
+    "test_data", test_data_suite | test_data_suite_bf16 | test_data_suite_fp16
+)
 @common.SkipIfNoModelConverter
 def test_permute_vgf_no_quant(test_data):
     test_data, dims = test_data()
diff --git a/backends/arm/test/ops/test_pow.py b/backends/arm/test/ops/test_pow.py
index 28f95f2c622..2cde848c7ef 100644
--- a/backends/arm/test/ops/test_pow.py
+++ b/backends/arm/test/ops/test_pow.py
@@ -128,7 +128,12 @@ def test_pow_tensor_tensor_tosa_FP(test_data: Pow_TensorTensor.input_t):
     pipeline.run()
 
 
-@common.parametrize("test_data", Pow_TensorTensor.test_data, x_fail, strict=False)
+@common.parametrize(
+    "test_data",
+    Pow_TensorTensor.test_data | Pow_TensorTensor.test_data_bf16,
+    x_fail,
+    strict=False,
+)
 @common.SkipIfNoModelConverter
 def test_pow_tensor_tensor_vgf_no_quant(test_data: Pow_TensorTensor.input_t):
     pipeline = VgfPipeline[Pow_TensorTensor.input_t](
@@ -199,7 +204,9 @@ def test_pow_tensor_scalar_u85_INT(test_data: Pow_TensorScalar.input_t):
 
 @common.parametrize(
     "test_data",
-    Pow_TensorScalar.test_data | Pow_TensorScalar.test_data_fp16,
+    Pow_TensorScalar.test_data
+    | Pow_TensorScalar.test_data_bf16
+    | Pow_TensorScalar.test_data_fp16,
 )
 @common.SkipIfNoModelConverter
 def test_pow_tensor_scalar_vgf_no_quant(test_data: Pow_TensorScalar.input_t):
diff --git a/backends/arm/test/ops/test_reciprocal.py b/backends/arm/test/ops/test_reciprocal.py
index b49d5913034..1344f86ac1e 100644
--- a/backends/arm/test/ops/test_reciprocal.py
+++ b/backends/arm/test/ops/test_reciprocal.py
@@ -101,14 +101,26 @@ def test_reciprocal_u85_INT(test_data: torch.Tensor):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_suite | test_data_suite_fp16)
+@common.parametrize(
+    "test_data", test_data_suite | test_data_suite_bf16 | test_data_suite_fp16
+)
 @common.SkipIfNoModelConverter
 def test_reciprocal_vgf_no_quant(test_data: torch.Tensor):
+    data = test_data()
+    match data.dtype:
+        case torch.bfloat16:
+            atol = 5e-2
+            rtol = 5e-2
+        case _:
+            atol = 1e-3
+            rtol = 1e-3
     pipeline = VgfPipeline[input_t1](
         Reciprocal(),
-        (test_data(),),
+        (data,),
         aten_op,
         quantize=False,
+        atol=atol,
+        rtol=rtol,
     )
     pipeline.run()
 
diff --git a/backends/arm/test/ops/test_repeat.py b/backends/arm/test/ops/test_repeat.py
index 3368864564d..cedf9f739d4 100644
--- a/backends/arm/test/ops/test_repeat.py
+++ b/backends/arm/test/ops/test_repeat.py
@@ -179,7 +179,9 @@ def test_repeat_u85_INT(test_data: Tuple):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_suite | test_data_suite_fp16)
+@common.parametrize(
+    "test_data", test_data_suite | test_data_suite_bf16 | test_data_suite_fp16
+)
 @common.SkipIfNoModelConverter
 def test_repeat_vgf_no_quant(test_data: Tuple):
     module, args = test_data()
diff --git a/backends/arm/test/ops/test_round.py b/backends/arm/test/ops/test_round.py
index bcc71b70725..1f4470cebb3 100644
--- a/backends/arm/test/ops/test_round.py
+++ b/backends/arm/test/ops/test_round.py
@@ -88,7 +88,7 @@ def test_round_u85_INT(test_data: torch.Tensor):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_suite)
+@common.parametrize("test_data", test_data_suite | test_data_suite_bf16)
 @common.SkipIfNoModelConverter
 def test_round_vgf_no_quant(test_data: torch.Tensor):
     pipeline = VgfPipeline[input_t1](
diff --git a/backends/arm/test/ops/test_rsqrt.py b/backends/arm/test/ops/test_rsqrt.py
index 3d7a51a99a0..4fd8442ba4a 100644
--- a/backends/arm/test/ops/test_rsqrt.py
+++ b/backends/arm/test/ops/test_rsqrt.py
@@ -107,14 +107,27 @@ def test_rsqrt_u85_INT(test_tensor: torch.Tensor):
     pipeline.run()
 
 
-@common.parametrize("test_tensor", Rsqrt.test_parameters | Rsqrt.test_parameters_fp16)
+@common.parametrize(
+    "test_tensor",
+    Rsqrt.test_parameters | Rsqrt.test_parameters_bf16 | Rsqrt.test_parameters_fp16,
+)
 @common.SkipIfNoModelConverter
 def test_rsqrt_vgf_no_quant(test_tensor: torch.Tensor):
+    data = test_tensor()
+    match data[0].dtype:
+        case torch.bfloat16:
+            atol = 5e-2
+            rtol = 5e-2
+        case _:
+            atol = 1e-3
+            rtol = 1e-3
     pipeline = VgfPipeline[input_t1](
         Rsqrt(),
-        test_tensor(),
+        data,
         aten_op,
         quantize=False,
+        atol=atol,
+        rtol=rtol,
     )
     pipeline.run()
 
diff --git a/backends/arm/test/ops/test_select_scatter.py b/backends/arm/test/ops/test_select_scatter.py
index 04c9c5d8422..6f3c08bce80 100644
--- a/backends/arm/test/ops/test_select_scatter.py
+++ b/backends/arm/test/ops/test_select_scatter.py
@@ -163,7 +163,7 @@ def test_select_scatter_u85_INT(test_module: input_t):
 
 
 @common.SkipIfNoModelConverter
-@common.parametrize("test_module", test_data_suite)
+@common.parametrize("test_module", test_data_suite | test_data_suite_bf16)
 def test_select_scatter_vgf_no_quant(test_module: input_t):
     pipeline = VgfPipeline[input_t](
         SelectScatter(),
diff --git a/backends/arm/test/ops/test_sigmoid.py b/backends/arm/test/ops/test_sigmoid.py
index 33d8659472a..5f88728ce7d 100644
--- a/backends/arm/test/ops/test_sigmoid.py
+++ b/backends/arm/test/ops/test_sigmoid.py
@@ -176,15 +176,27 @@ def test_sigmoid_u85_INT(test_data: Tuple):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_suite | test_data_suite_fp16)
+@common.parametrize(
+    "test_data", test_data_suite | test_data_suite_bf16 | test_data_suite_fp16
+)
 @common.SkipIfNoModelConverter
 def test_sigmoid_vgf_no_quant(test_data: Tuple):
+    data = test_data()
+    match data.dtype:
+        case torch.bfloat16:
+            atol = 5e-3
+            rtol = 5e-3
+        case _:
+            atol = 1e-3
+            rtol = 1e-3
     pipeline = VgfPipeline[input_t1](
         Sigmoid(),
-        (test_data(),),
+        (data,),
         aten_op,
         exir_op,
         quantize=False,
+        atol=atol,
+        rtol=rtol,
     )
     pipeline.run()
 
diff --git a/backends/arm/test/ops/test_sin.py b/backends/arm/test/ops/test_sin.py
index f9b3c2585e6..38afee6a31b 100644
--- a/backends/arm/test/ops/test_sin.py
+++ b/backends/arm/test/ops/test_sin.py
@@ -94,14 +94,26 @@ def test_sin_u85_INT(test_data: Tuple):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_suite | test_data_suite_fp16)
+@common.parametrize(
+    "test_data", test_data_suite | test_data_suite_bf16 | test_data_suite_fp16
+)
 @common.SkipIfNoModelConverter
 def test_sin_vgf_no_quant(test_data: Tuple):
+    data = test_data()
+    match data.dtype:
+        case torch.bfloat16:
+            atol = 5e-3
+            rtol = 5e-3
+        case _:
+            atol = 1e-3
+            rtol = 1e-3
     pipeline = VgfPipeline[input_t1](
         Sin(),
-        (test_data(),),
+        (data,),
         aten_op,
         quantize=False,
+        atol=atol,
+        rtol=rtol,
     )
     pipeline.run()
 
diff --git a/backends/arm/test/ops/test_slice.py b/backends/arm/test/ops/test_slice.py
index 28c9731a6aa..94a69130893 100644
--- a/backends/arm/test/ops/test_slice.py
+++ b/backends/arm/test/ops/test_slice.py
@@ -144,7 +144,9 @@ def test_slice_tensor_u85_INT(test_data: torch.Tensor):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_suite | test_data_suite_fp16)
+@common.parametrize(
+    "test_data", test_data_suite | test_data_suite_bf16 | test_data_suite_fp16
+)
 @common.SkipIfNoModelConverter
 def test_slice_tensor_vgf_no_quant(test_data: torch.Tensor):
     pipeline = VgfPipeline[input_t1](
diff --git a/backends/arm/test/ops/test_slice_scatter.py b/backends/arm/test/ops/test_slice_scatter.py
index ae3ceadadac..d367adc8622 100644
--- a/backends/arm/test/ops/test_slice_scatter.py
+++ b/backends/arm/test/ops/test_slice_scatter.py
@@ -299,7 +299,11 @@ def test_slice_scatter_u85_INT_stepN(test_module: input_t):
 @common.SkipIfNoModelConverter
 @common.parametrize(
     "test_module",
-    test_data_int_step1 | test_data_int_stepN | test_data_fp_step1 | test_data_fp_stepN,
+    test_data_int_step1
+    | test_data_int_stepN
+    | test_data_fp_step1
+    | test_data_fp_stepN
+    | test_data_bf16,
 )
 def test_slice_scatter_vgf_no_quant(test_module: input_t):
     pipeline = VgfPipeline[input_t](
diff --git a/backends/arm/test/ops/test_sub.py b/backends/arm/test/ops/test_sub.py
index 88b2816b503..c7127de645c 100644
--- a/backends/arm/test/ops/test_sub.py
+++ b/backends/arm/test/ops/test_sub.py
@@ -257,7 +257,9 @@ def test_sub_tensor_u85_INT(test_data: Tuple[torch.Tensor, torch.Tensor]):
     pipeline.run()
 
 
-@common.parametrize("test_data", sub_test_data | sub_test_data_fp16)
+@common.parametrize(
+    "test_data", sub_test_data | sub_test_data_bf16 | sub_test_data_fp16
+)
 @common.SkipIfNoModelConverter
 def test_sub_tensor_vgf_no_quant(test_data: Tuple[torch.Tensor]):
     """Test Subtraction (VGF FP)"""
@@ -271,7 +273,9 @@ def test_sub_tensor_vgf_no_quant(test_data: Tuple[torch.Tensor]):
     pipeline.run()
 
 
-@common.parametrize("test_data", sub2_test_data | sub2_test_data_fp16)
+@common.parametrize(
+    "test_data", sub2_test_data | sub2_test_data_bf16 | sub2_test_data_fp16
+)
 @common.SkipIfNoModelConverter
 def test_sub_tensor_vgf_no_quant_2(test_data: Tuple[torch.Tensor, torch.Tensor]):
     """Test Two-Operand Subtraction (VGF FP)"""
diff --git a/backends/arm/test/ops/test_sum.py b/backends/arm/test/ops/test_sum.py
index d727eb0408a..71af0d4014b 100644
--- a/backends/arm/test/ops/test_sum.py
+++ b/backends/arm/test/ops/test_sum.py
@@ -131,15 +131,28 @@ def test_sum_u85_INT_1_0(test_data: Tuple):
     pipeline.run()
 
 
-@common.parametrize("test_data", Sum.test_parameters | Sum.test_parameters_fp16)
+@common.parametrize(
+    "test_data",
+    Sum.test_parameters | Sum.test_parameters_bf16 | Sum.test_parameters_fp16,
+)
 @common.SkipIfNoModelConverter
 def test_sum_dim_intlist_vgf_no_quant(test_data: input_t1):
+    data = test_data()
+    match data[0].dtype:
+        case torch.bfloat16:
+            atol = 1e-1
+            rtol = 1e-1
+        case _:
+            atol = 1e-3
+            rtol = 1e-3
     pipeline = VgfPipeline[input_t1](
         Sum(),
-        test_data(),
+        data,
         aten_op,
         run_on_vulkan_runtime=True,
         quantize=False,
+        atol=atol,
+        rtol=rtol,
     )
     pipeline.run()
 
diff --git a/backends/arm/test/ops/test_tanh.py b/backends/arm/test/ops/test_tanh.py
index b3931f73389..7543cd20526 100644
--- a/backends/arm/test/ops/test_tanh.py
+++ b/backends/arm/test/ops/test_tanh.py
@@ -97,14 +97,26 @@ def test_tanh_u85_INT(test_data: Tuple):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_suite | test_data_suite_fp16)
+@common.parametrize(
+    "test_data", test_data_suite | test_data_suite_bf16 | test_data_suite_fp16
+)
 @common.SkipIfNoModelConverter
 def test_tanh_vgf_no_quant(test_data: Tuple):
+    data = test_data()
+    match data.dtype:
+        case torch.bfloat16:
+            atol = 5e-3
+            rtol = 5e-3
+        case _:
+            atol = 1e-3
+            rtol = 1e-3
     pipeline = VgfPipeline[input_t1](
         Tanh(),
-        (test_data(),),
+        (data,),
         aten_op,
         quantize=False,
+        atol=atol,
+        rtol=rtol,
     )
     pipeline.run()
 
diff --git a/backends/arm/test/ops/test_to_copy.py b/backends/arm/test/ops/test_to_copy.py
index 77e08a7995c..699e1f6460b 100644
--- a/backends/arm/test/ops/test_to_copy.py
+++ b/backends/arm/test/ops/test_to_copy.py
@@ -104,6 +104,17 @@ def forward(self, x: torch.Tensor, y: torch.Tensor):
     ),
 }
 
+_TO_COPY_TEST_DATA_VGF_BF16 = {
+    "rand_fp32_bf16": lambda: (
+        torch.rand((1, 2, 3, 4), dtype=torch.float32),
+        torch.bfloat16,
+    ),
+    "rand_bf16_fp32": lambda: (
+        torch.rand((1, 2, 3, 4), dtype=torch.bfloat16),
+        torch.float32,
+    ),
+}
+
 
 @common.parametrize("test_data", _TO_COPY_TEST_DATA_FP)
 def test_to_tosa_FP(test_data: Tuple):
@@ -275,7 +286,7 @@ def test_to_tosa_FP_bf16_fp8_with_extensions(test_data: Tuple):
     pipeline.run()
 
 
-@common.parametrize("test_data", _TO_COPY_TEST_DATA_FP)
+@common.parametrize("test_data", _TO_COPY_TEST_DATA_FP | _TO_COPY_TEST_DATA_VGF_BF16)
 @common.SkipIfNoModelConverter
 def test_to_vgf_no_quant(test_data: Tuple):
     test_tensor, new_dtype = test_data()
diff --git a/backends/arm/test/ops/test_transpose_conv2d.py b/backends/arm/test/ops/test_transpose_conv2d.py
index d34679f4bcb..9c53e59464b 100644
--- a/backends/arm/test/ops/test_transpose_conv2d.py
+++ b/backends/arm/test/ops/test_transpose_conv2d.py
@@ -537,12 +537,15 @@ def test_conv_transpose2d_tosa_INT_a16w8(test_data):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_FP | test_data_FP_fp16)
+@common.parametrize("test_data", test_data_FP | test_data_BF16 | test_data_FP_fp16)
 @common.SkipIfNoModelConverter
 def test_conv_transpose2d_vgf_no_quant(test_data):
     model = test_data()
     inputs = model.get_inputs()
     match inputs[0].dtype:
+        case torch.bfloat16:
+            atol = 1e-2
+            rtol = 1e-2
         case torch.float16:
             atol = 5e-3
             rtol = 5e-3
diff --git a/backends/arm/test/ops/test_unfold_copy.py b/backends/arm/test/ops/test_unfold_copy.py
index 446575d92ff..4b235ffa720 100644
--- a/backends/arm/test/ops/test_unfold_copy.py
+++ b/backends/arm/test/ops/test_unfold_copy.py
@@ -236,7 +236,7 @@ def test_unfold_copy_u85_INT(test_data: input_params):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_fp | test_data_int)
+@common.parametrize("test_data", test_data_fp | test_data_bf16 | test_data_int)
 @common.SkipIfNoModelConverter
 def test_unfold_copy_vgf_no_quant(test_data: input_params):
     pipeline = VgfPipeline[input_params](
diff --git a/backends/arm/test/ops/test_upsample_bilinear2d.py b/backends/arm/test/ops/test_upsample_bilinear2d.py
index 705f9ba4d20..1ebb38c2368 100644
--- a/backends/arm/test/ops/test_upsample_bilinear2d.py
+++ b/backends/arm/test/ops/test_upsample_bilinear2d.py
@@ -502,31 +502,47 @@ def test_upsample_bilinear2d_vec_u85_INT_a16w8(
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_suite_tosa | test_data_suite_tosa_fp16)
+@common.parametrize(
+    "test_data",
+    test_data_suite_tosa | test_data_suite_tosa_bf16 | test_data_suite_tosa_fp16,
+)
 @common.SkipIfNoModelConverter
 def test_upsample_bilinear2d_vec_vgf_no_quant_UpsamplingBilinear2d(
     test_data: torch.Tensor,
 ):
     data, size, scale_factor, compare = test_data()
+    match data.dtype:
+        case torch.bfloat16:
+            atol = 1e-2
+            rtol = 1e-2
+        case _:
+            atol = 2e-3
+            rtol = 2e-3
     pipeline = VgfPipeline[input_t1](
         UpsamplingBilinear2d(size, scale_factor),
         (data,),
         aten_op,
         exir_op,
         quantize=False,
-        atol=2e-3,
-        rtol=2e-3,
+        atol=atol,
+        rtol=rtol,
     )
     if not compare:
         pipeline.pop_stage(-1)
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_suite_tosa | test_data_suite_tosa_fp16)
+@common.parametrize(
+    "test_data",
+    test_data_suite_tosa | test_data_suite_tosa_bf16 | test_data_suite_tosa_fp16,
+)
 @common.SkipIfNoModelConverter
 def test_upsample_bilinear2d_vec_vgf_no_quant_Upsample(test_data: torch.Tensor):
     data, size, scale_factor, compare = test_data()
     match data.dtype:
+        case torch.bfloat16:
+            atol = 1e-2
+            rtol = 1e-2
         case torch.float16:
             atol = 1e-2
             rtol = 1e-2
@@ -547,11 +563,17 @@ def test_upsample_bilinear2d_vec_vgf_no_quant_Upsample(test_data: torch.Tensor):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_suite_tosa | test_data_suite_tosa_fp16)
+@common.parametrize(
+    "test_data",
+    test_data_suite_tosa | test_data_suite_tosa_bf16 | test_data_suite_tosa_fp16,
+)
 @common.SkipIfNoModelConverter
 def test_upsample_bilinear2d_vec_vgf_no_quant_Interpolate(test_data: torch.Tensor):
     data, size, scale_factor, compare = test_data()
     match data.dtype:
+        case torch.bfloat16:
+            atol = 1e-2
+            rtol = 1e-2
         case torch.float16:
             atol = 1e-2
             rtol = 1e-2
diff --git a/backends/arm/test/ops/test_upsample_nearest2d.py b/backends/arm/test/ops/test_upsample_nearest2d.py
index d8bf4d7dbd5..ea17175221e 100644
--- a/backends/arm/test/ops/test_upsample_nearest2d.py
+++ b/backends/arm/test/ops/test_upsample_nearest2d.py
@@ -282,7 +282,9 @@ def test_upsample_nearest2d_vec_tosa_INT_a16w8(test_data: torch.Tensor):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_suite | test_data_suite_fp16)
+@common.parametrize(
+    "test_data", test_data_suite | test_data_suite_bf16 | test_data_suite_fp16
+)
 @common.SkipIfNoModelConverter
 def test_upsample_nearest2d_vec_vgf_no_quant(test_data: torch.Tensor):
     data, size, scale_factor, compare = test_data()
@@ -298,7 +300,9 @@ def test_upsample_nearest2d_vec_vgf_no_quant(test_data: torch.Tensor):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_suite | test_data_suite_fp16)
+@common.parametrize(
+    "test_data", test_data_suite | test_data_suite_bf16 | test_data_suite_fp16
+)
 @common.SkipIfNoModelConverter
 def test_upsample_nearest2d_vec_vgf_no_quant_nearest(test_data: torch.Tensor):
     data, size, scale_factor, compare = test_data()
@@ -314,7 +318,9 @@ def test_upsample_nearest2d_vec_vgf_no_quant_nearest(test_data: torch.Tensor):
     pipeline.run()
 
 
-@common.parametrize("test_data", test_data_suite | test_data_suite_fp16)
+@common.parametrize(
+    "test_data", test_data_suite | test_data_suite_bf16 | test_data_suite_fp16
+)
 @common.SkipIfNoModelConverter
 def test_upsample_nearest2d_vec_vgf_no_quant_interpolate(test_data: torch.Tensor):
     data, size, scale_factor, compare = test_data()
@@ -325,7 +331,7 @@ def test_upsample_nearest2d_vec_vgf_no_quant_interpolate(test_data: torch.Tensor
         exir_op,
         quantize=False,
         # Override tosa version to test FP-only path
-        tosa_version="TOSA-1.0+FP",
+        tosa_version="TOSA-1.0+FP+bf16",
     )
     if not compare:
         pipeline.pop_stage(-1)
diff --git a/backends/arm/test/ops/test_where.py b/backends/arm/test/ops/test_where.py
index 234203f1cc4..b5c201931cb 100644
--- a/backends/arm/test/ops/test_where.py
+++ b/backends/arm/test/ops/test_where.py
@@ -292,7 +292,7 @@ def test_where_self_u85_INT(test_module):
     pipeline.run()
 
 
-@common.parametrize("test_module", test_modules_FP)
+@common.parametrize("test_module", test_modules_FP | test_modules_FP_bf16)
 @common.SkipIfNoModelConverter
 def test_where_self_vgf_no_quant(test_module):
     module = test_module()
diff --git a/backends/arm/test/runner_utils.py b/backends/arm/test/runner_utils.py
index 9a63452e325..b59fab917fc 100644
--- a/backends/arm/test/runner_utils.py
+++ b/backends/arm/test/runner_utils.py
@@ -243,8 +243,7 @@ def is_concrete_shape(shape_like) -> bool:
         return all(isinstance(dim, numbers.Integral) for dim in shape_like)
 
     def to_torch_tensor() -> torch.Tensor:
-        if array.dtype.type is np.void:
-            # If dtype is void, "cheat" and use the output_tensor dtype.
+        if output_tensor.dtype == torch.bfloat16 or array.dtype.type is np.void:
             return torch.frombuffer(array, dtype=output_tensor.dtype)
         return torch.from_numpy(array)
 
diff --git a/backends/arm/test/tester/test_pipeline.py b/backends/arm/test/tester/test_pipeline.py
index 589c0a851a3..6538f80cc28 100644
--- a/backends/arm/test/tester/test_pipeline.py
+++ b/backends/arm/test/tester/test_pipeline.py
@@ -1258,13 +1258,15 @@ def __init__(
     ):
         if tosa_spec is None:
             if tosa_version is None:
-                tosa_spec = VgfCompileSpec().tosa_spec
-            else:
-                if tosa_extensions is None:
+                tosa_version = str(VgfCompileSpec().tosa_spec)
+            if tosa_extensions is None:
+                if "FP" in tosa_version:
+                    tosa_extensions = ["bf16"]
+                else:
                     tosa_extensions = []
-                tosa_spec = TosaSpecification.create_from_string(
-                    tosa_version + "".join([f"+{ext}" for ext in tosa_extensions])
-                )
+            tosa_spec = TosaSpecification.create_from_string(
+                tosa_version + "".join([f"+{ext}" for ext in tosa_extensions])
+            )
         elif isinstance(tosa_spec, str):
             tosa_spec = TosaSpecification.create_from_string(tosa_spec)
         compile_spec = common.get_vgf_compile_spec(