Arm backend: Add Qwen3-VL_2B_IT FP32 layer tests (#19628)

Change-Id: I62d3848e0a6546e21d508b4ed565c2403b63f72d



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

---------

Signed-off-by: Tom Allsop <tom.allsop@arm.com>
Co-authored-by: Baris Demir <baris.demir@arm.com>

Author: tom-arm

backends/arm/MODELS.md

@@ -10,6 +10,7 @@
 - Inception v3 (IC3)
 - Llama
 - Gemma3n
+- Qwen3-VL
 - Long Short-Term Memory (LSTM)
 - MobileNet V1 0.25
 - MobileNet v2 (MV2)

backends/arm/test/models/Qwen3_VL/qwen3_vl_test_config.py

@@ -0,0 +1,64 @@
+# Copyright 2026 Arm Limited and/or its affiliates.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from transformers.models.qwen3_vl.configuration_qwen3_vl import (
+    Qwen3VLConfig,
+    Qwen3VLTextConfig,
+    Qwen3VLVisionConfig,
+)
+
+
+def get_qwen3_vl_2b_instruct_checkpoint_config() -> Qwen3VLConfig:
+    text_config = Qwen3VLTextConfig(
+        attention_bias=False,
+        attention_dropout=0.0,
+        bos_token_id=151643,  # type: ignore[call-arg]
+        dtype="bfloat16",
+        eos_token_id=151645,  # type: ignore[call-arg]
+        head_dim=128,
+        hidden_act="silu",
+        hidden_size=2048,
+        initializer_range=0.02,
+        intermediate_size=6144,
+        max_position_embeddings=262144,
+        num_attention_heads=16,
+        num_hidden_layers=28,
+        num_key_value_heads=8,
+        rms_norm_eps=1e-6,
+        rope_parameters={
+            "mrope_interleaved": True,  # type: ignore[dict-item]
+            "mrope_section": [24, 20, 20],  # type: ignore[dict-item]
+            "rope_type": "default",  # type: ignore[dict-item]
+            "rope_theta": 5_000_000,  # type: ignore[dict-item]
+        },
+        tie_word_embeddings=True,  # type: ignore[call-arg]
+        use_cache=True,
+        vocab_size=151936,
+    )
+    vision_config = Qwen3VLVisionConfig(
+        deepstack_visual_indexes=[5, 11, 17],
+        depth=24,
+        hidden_act="gelu_pytorch_tanh",
+        hidden_size=1024,
+        in_channels=3,
+        initializer_range=0.02,
+        intermediate_size=4096,
+        num_heads=16,
+        num_position_embeddings=2304,
+        out_hidden_size=2048,
+        patch_size=16,
+        spatial_merge_size=2,
+        temporal_patch_size=2,
+    )
+    return Qwen3VLConfig(
+        architectures=["Qwen3VLForConditionalGeneration"],
+        image_token_id=151655,
+        text_config=text_config.to_dict(),
+        tie_word_embeddings=True,
+        video_token_id=151656,
+        vision_config=vision_config.to_dict(),
+        vision_end_token_id=151653,
+        vision_start_token_id=151652,
+    )