Add quantization tests for llava_next_video

Author: nikita-savelyevv

tests/openvino/test_exporters_cli.py

@@ -341,6 +341,19 @@ class OVCLIExportTestCase(unittest.TestCase):
             ]
         )
 
+    if is_transformers_version(">=", "4.42.0"):
+        TEST_4BIT_CONFIGURATIONS.extend(
+            [
+                (
+                    "image-text-to-text",
+                    "llava_next_video",
+                    'int4 --group-size 16 --ratio 0.8 --sensitivity-metric "hessian_input_activation" '
+                    "--dataset contextual --num-samples 1",
+                    [{"int8": 6, "int4": 24}, {"int8": 1}, {"int8": 7}, {}, {"int8": 2}],
+                ),
+            ]
+        )
+
     if is_transformers_version(">=", "4.45.0"):
         TEST_4BIT_CONFIGURATIONS.extend(
             [

tests/openvino/test_quantization.py

@@ -581,6 +581,27 @@ class OVWeightCompressionTest(unittest.TestCase):
             ]
         )
 
+    if is_transformers_version(">=", "4.42.0"):
+        LOAD_IN_4_BITS_SCOPE.extend(
+            [
+                (
+                    OVModelForVisualCausalLM,
+                    "llava_next_video",
+                    False,
+                    dict(
+                        bits=4,
+                        group_size=16,
+                        dataset="contextual",
+                        ratio=0.8,
+                        sensitivity_metric="hessian_input_activation",
+                        num_samples=1,
+                        processor=MODEL_NAMES["llava_next"],
+                    ),
+                    [{"int8": 6, "int4": 24}, {"int8": 1}, {"int8": 7}, {}, {"int8": 2}],
+                ),
+            ]
+        )
+
     if is_transformers_version(">=", "4.45.0"):
         LOAD_IN_4_BITS_SCOPE.extend(
             [
@@ -668,6 +689,9 @@ class OVWeightCompressionTest(unittest.TestCase):
     if is_transformers_version(">=", "4.40.0"):
         SUPPORTED_ARCHITECTURES_WITH_AUTO_COMPRESSION.append((OVModelForVisualCausalLM, "nanollava", True))
 
+    if is_transformers_version(">=", "4.42.0"):
+        SUPPORTED_ARCHITECTURES_WITH_AUTO_COMPRESSION.append((OVModelForVisualCausalLM, "llava_next_video", False))
+
     if is_transformers_version(">=", "4.45.0"):
         SUPPORTED_ARCHITECTURES_WITH_AUTO_COMPRESSION.append((OVModelForVisualCausalLM, "minicpmv", True))
         SUPPORTED_ARCHITECTURES_WITH_AUTO_COMPRESSION.append((OVModelForVisualCausalLM, "qwen2_vl", False))

tests/openvino/utils_tests.py

@@ -202,6 +202,7 @@
     "flux-fill": (56, 28, 24, 64, 64),
     "llava": (30, 1, 9),
     "llava_next": (30, 1, 9),
+    "llava_next_video": (30, 1, 7, 0, 2),
     "minicpmv": (30, 1, 26, 6),
     "nanollava": (30, 1, 15),
     "qwen2_vl": (30, 1, 1, 10),