Arm backend: Add Qwen3 VL FP32 MXFP8 layer tests (pytorch#20485)

Add TOSA MXFP8 coverage for Qwen3 VL text attention, MLP, and decoder
layer wrappers. Keep the tests in FP32 while converting linear layers to
MXFP through the MXFP pipeline.

Change-Id: Ib94bfdc114031a7558d42ea9e347142d721f7403

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

Signed-off-by: Yufeng Shi <yufeng.shi@arm.com>

Author: YufengShi-dudu

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

@@ -9,10 +9,12 @@
 from typing import Tuple
 
 import torch
+from executorch.backends.arm.ao_ext import MXFPOpConfig
 from executorch.backends.arm.test import common
 from executorch.backends.arm.test.models.Qwen3_VL.qwen3_vl_test_config import (
     get_qwen3_vl_2b_instruct_checkpoint_config,
 )
+from executorch.backends.arm.test.ops.mxfp.common import MXFPTosaPipelineFP
 from executorch.backends.arm.test.tester.test_pipeline import (
     TosaPipelineFP,
     VgfPipeline,
@@ -34,6 +36,7 @@
 )
 
 input_t = Tuple[torch.Tensor | int, ...]
+aten_op_mxfp_linear = "torch.ops.tosa_mxfp.linear.default"
 
 
 def _make_qwen3_vl_2b_instruct_layer_config():
@@ -112,6 +115,10 @@ def _to_bfloat16(
     )
 
 
+def _is_linear(module: torch.nn.Module, _fqn: str) -> bool:
+    return isinstance(module, torch.nn.Linear)
+
+
 class Qwen3VLVisionMLPModel(Qwen3VLTestModule):
     def __init__(self, config) -> None:
         super().__init__()
@@ -468,6 +475,19 @@ class Qwen3VLTestCase:
 }
 
 
+TOSA_MXFP8_TEST_CASES: dict[str, Qwen3VLTestCase] = {
+    "text_attention": Qwen3VLTestCase(
+        model_cls=TextAttentionModel,
+    ),
+    "text_mlp": Qwen3VLTestCase(
+        model_cls=TextMLPModel,
+    ),
+    "text_decoder_layer": Qwen3VLTestCase(
+        model_cls=TextDecoderLayerModel,
+    ),
+}
+
+
 @common.parametrize(
     "test_case",
     TOSA_FP_TEST_CASES,
@@ -516,6 +536,30 @@ def test_qwen3_vl_tosa_FP_bf16(test_case: Qwen3VLTestCase):
         pipeline.run()
 
 
+@common.parametrize(
+    "test_case",
+    TOSA_MXFP8_TEST_CASES,
+)
+def test_qwen3_vl_tosa_mxfp8_fp32(test_case: Qwen3VLTestCase):
+    model, inputs = test_case.model_cls.prepare_model_and_inputs()
+    mxfp_config = MXFPOpConfig(weight_dtype=torch.float8_e4m3fn)
+
+    with torch.no_grad():
+        pipeline = MXFPTosaPipelineFP[input_t](
+            model,
+            inputs,
+            aten_op=aten_op_mxfp_linear,
+            exir_op=[],
+            filter_fn=_is_linear,
+            frobenius_threshold=0.05,
+            cosine_threshold=0.995,
+            mxfp_config=mxfp_config,
+            tosa_version="1.1",
+            tosa_extensions=["mxfp"],
+        )
+        pipeline.run()
+
+
 @common.SkipIfNoModelConverter
 @common.parametrize(
     "test_case",