Various test fixes for AMD (#39978)

* Add amd expectation in internvl

* Add amd expectation to llama

* Added bnb decorator for a llava test that requires bnb

* Added amd expectation for mistral3

* Style

Author: remi-or

tests/models/internvl/test_modeling_internvl.py

@@ -652,10 +652,11 @@ def test_llama_small_model_integration_forward(self):
 
         expected_logits_all = Expectations(
             {
-                ("xpu", 3): torch.tensor([-9.8750, -0.5703, 1.4297, -10.3125, -10.3125], dtype=torch.float16),
-                ("cuda", 7): torch.tensor([-9.8750,  -0.4861,   1.4648, -10.3359, -10.3359], dtype=torch.float16),
-                ("cuda", 8): torch.tensor([-9.8906,  -0.4995,   1.4473, -10.3359, -10.3438], dtype=torch.float16),
-                ("rocm", (9, 5)): torch.tensor([ -9.8906,  -0.4976,   1.4502, -10.3359, -10.3438], dtype=torch.float16),
+                ("xpu", 3): [-9.8750, -0.5703, 1.4297, -10.3125, -10.3125],
+                ("cuda", 7): [-9.8750,  -0.4861,   1.4648, -10.3359, -10.3359],
+                ("cuda", 8): [-9.8906,  -0.4995,   1.4473, -10.3359, -10.3438],
+                ("rocm", (9, 4)): [ -9.8750,  -0.4885,   1.4668, -10.3359, -10.3359],
+                ("rocm", (9, 5)): [ -9.8906,  -0.4976,   1.4502, -10.3359, -10.3438],
             }
         )  # fmt: skip
         expected_logits = torch.tensor(expected_logits_all.get_expectation(), dtype=torch.float16)

tests/models/llama/test_modeling_llama.py

@@ -151,14 +151,16 @@ def test_model_7b_logits_bf16(self):
             {
             ("xpu", 3): torch.tensor([[-6.5208, -4.1218, -4.9377, -3.2536,  0.8127, -2.9811,  1.2918, -3.3848]]),
             ("cuda", 7): torch.tensor([[-6.5061, -4.1147, -4.9669, -3.2038, 0.8069, -2.9694, 1.2864, -3.3786]]),
-            ("cuda", 8): torch.tensor([[-6.5208, -4.1218, -4.9377, -3.2536,  0.8127, -2.9811,  1.2918, -3.3848]])
-         })
+            ("cuda", 8): torch.tensor([[-6.5208, -4.1218, -4.9377, -3.2536,  0.8127, -2.9811,  1.2918, -3.3848]]),
+            ("rocm", (9, 4)): torch.tensor([[-6.5094, -4.1329, -4.9754, -3.5042,  0.8082, -2.9443,  1.2830, -3.3539]]),
+        })
 
-        expected_mean = expected_means.get_expectation()
+        expected_mean = expected_means.get_expectation().to(torch_device)
+        actual_mean = out.logits.float().mean(-1)
         self.assertTrue(
             torch.allclose(
-                expected_mean.to(torch_device),
-                out.logits.float().mean(-1),
+                expected_mean,
+                actual_mean,
                 atol=1e-2,
                 rtol=1e-2
             )
@@ -169,18 +171,13 @@ def test_model_7b_logits_bf16(self):
             {
             ("xpu", 3): torch.tensor([[-12.5625,  -7.1250,  -0.6289,  -7.8750,  -6.9688,  -7.8125,  -6.5000, -7.4375,  -7.6562,  -6.9688,  -6.0312,  -7.0312,  -1.8203,   1.8750, -8.5000]]),
             ("cuda", 7): torch.tensor([[-12.5000, -7.0625, -0.6289, -7.8750, -6.9688, -7.8125, -6.4688, -7.4375, -7.6875, -6.9375, -6.0312, -7.0000, -1.8594, 1.8438, -8.5000]]),
-            ("cuda", 8): torch.tensor([[-12.5625,  -7.1250,  -0.6289,  -7.8750,  -6.9688,  -7.8125,  -6.5000, -7.4375,  -7.6562,  -6.9688,  -6.0312,  -7.0312,  -1.8203,   1.8750, -8.5000]])
+            ("cuda", 8): torch.tensor([[-12.5625,  -7.1250,  -0.6289,  -7.8750,  -6.9688,  -7.8125,  -6.5000, -7.4375,  -7.6562,  -6.9688,  -6.0312,  -7.0312,  -1.8203,   1.8750, -8.5000]]),
+            ("rocm", (9, 4)): torch.tensor([[-12.5000,  -7.0625,  -0.6289,  -7.8750,  -6.9688,  -7.8125,  -6.5000, -7.4375,  -7.6562,  -6.9375,  -6.0312,  -7.0312,  -1.8594,   1.8438, -8.5000]])
         })
         # fmt: on
-        expected_slice = expected_slices.get_expectation()
-        self.assertTrue(
-            torch.allclose(
-                expected_slice.to(torch_device),
-                out.logits[0, 0, :15].float(),
-                atol=1e-2,
-                rtol=1e-2,
-            )
-        )
+        expected_slice = expected_slices.get_expectation().to(torch_device)
+        actual_slice = out.logits[0, 0, :15].float()
+        self.assertTrue(torch.allclose(expected_slice, actual_slice, atol=1e-2, rtol=1e-2))
 
     @slow
     def test_model_7b_logits(self):

tests/models/llava/test_modeling_llava.py

@@ -476,6 +476,7 @@ def test_small_model_integration_test_llama_batched_regression(self):
     @slow
     @require_torch
     @require_vision
+    @require_bitsandbytes
     def test_batched_generation(self):
         model = LlavaForConditionalGeneration.from_pretrained("llava-hf/llava-1.5-7b-hf", load_in_4bit=True)
 

tests/models/mistral3/test_modeling_mistral3.py

@@ -317,6 +317,7 @@ def test_mistral3_integration_generate(self):
             {
                 ("xpu", 3): "The image features two cats resting on a pink blanket. The cat on the left is a kitten",
                 ("cuda", 8): 'The image features two cats lying on a pink surface, which appears to be a couch or a bed',
+                ("rocm", (9, 4)): "The image features two cats lying on a pink surface, which appears to be a couch or a bed",
                 ("rocm", (9, 5)): "The image features two tabby cats lying on a pink surface, which appears to be a cushion or"
             }
         )  # fmt: skip