From 58f14049038a757d34098a908b18b7a4dac3a87c Mon Sep 17 00:00:00 2001 From: Sebastian Larsson Date: Mon, 6 Jul 2026 12:03:04 +0200 Subject: [PATCH] Arm backend: Relax Qwen3-VL BF16 text tolerance Raise the aarch64 TOSA BF16 tolerance for the Qwen3-VL text full-model test only. Change-Id: Ifd15f99ceb5895f7161018380b919d7c3c1fce08 Signed-off-by: Sebastian Larsson --- .../arm/test/models/Qwen3_VL/test_qwen3_vl_model.py | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/backends/arm/test/models/Qwen3_VL/test_qwen3_vl_model.py b/backends/arm/test/models/Qwen3_VL/test_qwen3_vl_model.py index fa7a69ceb6e..f62b4da5030 100644 --- a/backends/arm/test/models/Qwen3_VL/test_qwen3_vl_model.py +++ b/backends/arm/test/models/Qwen3_VL/test_qwen3_vl_model.py @@ -256,12 +256,12 @@ def _test_qwen3_vl_full_models_tosa_FP_bf16( model, inputs = test_case.model_cls.prepare_model_and_inputs(config_factory) model, inputs = _to_bfloat16_model_and_floating_inputs(model, inputs) # Slightly higher atol for TOSA BF16 on aarch64 (MLETORCH-2048: numeric mismatch) - atol = ( - 0.4 - if common.is_aarch64_host() - and test_case.model_cls is LowerableVisionModelWrapper - else 0.1 - ) + if common.is_aarch64_host() and test_case.model_cls is LowerableVisionModelWrapper: + atol = 0.4 + elif common.is_aarch64_host() and test_case.model_cls is TextModelWrapper: + atol = 0.15 + else: + atol = 0.1 with torch.no_grad(): pipeline = TosaPipelineFP[input_t]( model,