# Copyright 2025 The HuggingFace Inc. team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Testing suite for the PyTorch Gemma3 model.""" import logging import tempfile import unittest import pytest from parameterized import parameterized from pytest import mark from transformers import ( AutoModelForCausalLM, AutoTokenizer, Gemma3Config, Gemma3TextConfig, is_torch_available, ) from transformers.testing_utils import ( Expectations, cleanup, is_flash_attn_2_available, require_deterministic_for_xpu, require_flash_attn, require_flash_attn_3, require_read_token, require_torch, require_torch_accelerator, require_torch_gpu, require_torch_large_accelerator, slow, torch_device, ) from ...causal_lm_tester import CausalLMModelTest, CausalLMModelTester from ...generation.test_utils import GenerationTesterMixin from ...test_configuration_common import ConfigTester from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor if is_torch_available(): import torch from transformers import ( Gemma3ForCausalLM, Gemma3ForConditionalGeneration, Gemma3ForSequenceClassification, Gemma3Model, Gemma3Processor, Gemma3TextForSequenceClassification, Gemma3TextModel, ) from transformers.pytorch_utils import is_torch_greater_or_equal class Gemma3TextModelTester(CausalLMModelTester): if is_torch_available(): base_model_class = Gemma3TextModel causal_lm_class = Gemma3ForCausalLM sequence_classification_class = Gemma3TextForSequenceClassification @require_torch class Gemma3TextModelTest(CausalLMModelTest, unittest.TestCase): model_tester_class = Gemma3TextModelTester _is_stateful = True model_split_percents = [0.5, 0.6] @unittest.skip("Gemma3 applies key/query norm which doesn't work with packing") def test_flash_attention_2_padding_matches_padding_free_with_position_ids(self): pass @unittest.skip("Gemma3 applies key/query norm which doesn't work with packing") def test_flash_attention_2_padding_matches_padding_free_with_position_ids_and_fa_kwargs(self): pass @unittest.skip("Gemma3 applies key/query norm which doesn't work with packing") def test_eager_padding_matches_padding_free_with_position_ids(self): pass @unittest.skip("Gemma3 applies key/query norm which doesn't work with packing") def test_sdpa_padding_matches_padding_free_with_position_ids(self): pass @unittest.skip( "Gemma3 has no base model prefix which causes issues when loading base model from saved task model checkpoint" ) def test_load_with_mismatched_shapes(self): pass def test_generation_beyond_sliding_window_tiny_model(self): """Test generation with a tiny randomly initialised model whose input length is larger than the `sliding_window`. The model is configured with both `full_attention` and `sliding_attention` layers to make sure the hybrid cache and mask slicing logic is covered. """ config = Gemma3TextConfig.from_pretrained("hf-internal-testing/tiny-random-Gemma3ForCausalLM") config.attn_implementation = "eager" config.layer_types = ["full_attention", "sliding_attention"] config.sliding_window = 8 config.max_position_embeddings = 128 config.rope_parameters = { "full_attention": {"rope_type": "default", "rope_theta": 1000000}, "sliding_attention": {"rope_type": "default", "rope_theta": 10000}, } model = AutoModelForCausalLM.from_pretrained( "hf-internal-testing/tiny-random-Gemma3ForCausalLM", config=config ).to(torch_device) input_len = 10 input_ids = torch.tensor( [ [42300, 241087, 255445, 81315, 193760, 184471, 67719, 98191, 210651, 124725], [102294, 205314, 226646, 62020, 60245, 68025, 251839, 114053, 4695, 175511], ], device=torch_device, ) attention_mask = torch.ones_like(input_ids).to(torch_device) with torch.no_grad(): _ = model.generate( input_ids, attention_mask=attention_mask, max_new_tokens=1, do_sample=False, use_cache=True, disable_compile=True, ) # 2 generations are needed to trigger https://github.com/huggingface/transformers/issues/39711 # Since it requires model._cache to have been previously initialized output = model.generate( input_ids, attention_mask=attention_mask, max_new_tokens=5, do_sample=False, use_cache=True, disable_compile=True, ) generated_sequences = output[:, input_len:].cpu() EXPECTED_OUTPUT = torch.tensor([[90109, 90109, 90109, 83191, 83191], [246901, 69832, 69832, 69832, 62288]]) torch.testing.assert_close(generated_sequences, EXPECTED_OUTPUT) @parameterized.expand([("linear",), ("dynamic",), ("yarn",)]) @unittest.skip("TODO (joao): check why this is failing") def test_model_rope_scaling_from_config(self): pass def test_model_rope_scaling_frequencies(self): """Tests the frequency properties of the different RoPE scaling types on the model RoPE layer.""" config, _ = self.model_tester.prepare_config_and_inputs_for_common() # Retrieves the RoPE layer class from the base model class. Uses `.named_modules()` to avoid hardcoding the # named location of the RoPE layer class. base_model = self.model_tester.base_model_class(config) possible_rope_attributes = [ "pos_emb", "rotary_emb", # most common case "global_rotary_emb", "local_rotary_emb", ] for name, module in base_model.named_modules(): if any(potential_name in name for potential_name in possible_rope_attributes): rope_class = type(module) break scaling_factor = 10 short_input_length = 10 long_input_length = int(config.max_position_embeddings * 1.5) # Inputs x = torch.randn( 1, dtype=torch.float32, device=torch_device ) # used exclusively to get the dtype and the device position_ids_short = torch.arange(short_input_length, dtype=torch.long, device=torch_device) position_ids_short = position_ids_short.unsqueeze(0) position_ids_long = torch.arange(long_input_length, dtype=torch.long, device=torch_device) position_ids_long = position_ids_long.unsqueeze(0) # Sanity check original RoPE rope_params = {"rope_type": "default", "rope_theta": 10_000.0} config.rope_parameters = {"full_attention": rope_params, "sliding_attention": rope_params} original_rope = rope_class(config=config).to(torch_device) original_cos_short, original_sin_short = original_rope(x, position_ids_short, layer_type="sliding_attention") original_cos_long, original_sin_long = original_rope(x, position_ids_long, layer_type="sliding_attention") torch.testing.assert_close(original_cos_short, original_cos_long[:, :short_input_length, :]) torch.testing.assert_close(original_sin_short, original_sin_long[:, :short_input_length, :]) # Sanity check linear RoPE scaling # New position "x" should match original position with index "x/scaling_factor" rope_params = {"rope_type": "linear", "factor": scaling_factor, "rope_theta": 10_000.0} config.rope_parameters = {"full_attention": rope_params, "sliding_attention": rope_params} linear_scaling_rope = rope_class(config=config).to(torch_device) linear_cos_short, linear_sin_short = linear_scaling_rope(x, position_ids_short, layer_type="sliding_attention") linear_cos_long, linear_sin_long = linear_scaling_rope(x, position_ids_long, layer_type="sliding_attention") torch.testing.assert_close(linear_cos_short, linear_cos_long[:, :short_input_length, :]) torch.testing.assert_close(linear_sin_short, linear_sin_long[:, :short_input_length, :]) for new_position in range(0, long_input_length, scaling_factor): original_position = int(new_position // scaling_factor) torch.testing.assert_close(linear_cos_long[:, new_position, :], original_cos_long[:, original_position, :]) torch.testing.assert_close(linear_sin_long[:, new_position, :], original_sin_long[:, original_position, :]) # Sanity check Dynamic NTK RoPE scaling # Scaling should only be observed after a long input is fed. We can observe that the frequencies increase # with scaling_factor (or that `inv_freq` decreases) rope_params = {"rope_type": "dynamic", "factor": scaling_factor, "rope_theta": 10_000.0} config.rope_parameters = {"full_attention": rope_params, "sliding_attention": rope_params} ntk_scaling_rope = rope_class(config=config).to(torch_device) ntk_cos_short, ntk_sin_short = ntk_scaling_rope(x, position_ids_short, layer_type="sliding_attention") ntk_cos_long, ntk_sin_long = ntk_scaling_rope(x, position_ids_long, layer_type="sliding_attention") torch.testing.assert_close(ntk_cos_short, original_cos_short) torch.testing.assert_close(ntk_sin_short, original_sin_short) with self.assertRaises(AssertionError): torch.testing.assert_close(ntk_cos_long, original_cos_long) with self.assertRaises(AssertionError): torch.testing.assert_close(ntk_sin_long, original_sin_long) self.assertTrue( (ntk_scaling_rope.sliding_attention_inv_freq <= original_rope.sliding_attention_inv_freq).all() ) # Sanity check Yarn RoPE scaling # Scaling should be over the entire input rope_params = {"rope_type": "yarn", "factor": scaling_factor, "rope_theta": 10_000.0} config.rope_parameters = {"full_attention": rope_params, "sliding_attention": rope_params} yarn_scaling_rope = rope_class(config=config).to(torch_device) yarn_cos_short, yarn_sin_short = yarn_scaling_rope(x, position_ids_short, layer_type="sliding_attention") yarn_cos_long, yarn_sin_long = yarn_scaling_rope(x, position_ids_long, layer_type="sliding_attention") torch.testing.assert_close(yarn_cos_short, yarn_cos_long[:, :short_input_length, :]) torch.testing.assert_close(yarn_sin_short, yarn_sin_long[:, :short_input_length, :]) with self.assertRaises(AssertionError): torch.testing.assert_close(yarn_cos_short, original_cos_short) with self.assertRaises(AssertionError): torch.testing.assert_close(yarn_sin_short, original_sin_short) with self.assertRaises(AssertionError): torch.testing.assert_close(yarn_cos_long, original_cos_long) with self.assertRaises(AssertionError): torch.testing.assert_close(yarn_sin_long, original_sin_long) class Gemma3Vision2TextModelTester: def __init__( self, parent, mm_tokens_per_image=2, image_token_index=4, boi_token_index=5, eoi_token_index=6, seq_length=25, is_training=True, vision_config={ "use_labels": True, "image_size": 20, "patch_size": 5, "num_channels": 3, "is_training": True, "hidden_size": 32, "num_key_value_heads": 1, "num_hidden_layers": 2, "num_attention_heads": 4, "intermediate_size": 37, "dropout": 0.1, "attention_dropout": 0.1, "initializer_range": 0.02, }, use_cache=False, ): self.parent = parent # `image_token_index` is set to 0 to pass "resize_embeddings" test, do not modify self.mm_tokens_per_image = mm_tokens_per_image self.image_token_index = image_token_index self.boi_token_index = boi_token_index self.eoi_token_index = eoi_token_index self.llm_tester = Gemma3TextModelTester(self.parent) self.text_config = self.llm_tester.get_config() self.vision_config = vision_config self.seq_length = seq_length self.pad_token_id = self.text_config.pad_token_id self.num_hidden_layers = self.text_config.num_hidden_layers self.vocab_size = self.text_config.vocab_size self.hidden_size = self.text_config.hidden_size self.num_attention_heads = self.text_config.num_attention_heads self.is_training = is_training self.batch_size = 3 self.num_channels = vision_config["num_channels"] self.image_size = vision_config["image_size"] self.encoder_seq_length = seq_length self.use_cache = use_cache def get_config(self): return Gemma3Config( text_config=self.text_config, vision_config=self.vision_config, image_token_index=self.image_token_index, boi_token_index=self.boi_token_index, eoi_token_index=self.eoi_token_index, mm_tokens_per_image=self.mm_tokens_per_image, ) def prepare_config_and_inputs(self): pixel_values = floats_tensor( [ self.batch_size, self.vision_config["num_channels"], self.vision_config["image_size"], self.vision_config["image_size"], ] ) config = self.get_config() return config, pixel_values def prepare_config_and_inputs_for_common(self): config_and_inputs = self.prepare_config_and_inputs() config, pixel_values = config_and_inputs input_ids = ids_tensor([self.batch_size, self.seq_length], config.text_config.vocab_size - 1) + 1 attention_mask = input_ids.ne(self.pad_token_id).to(torch_device) # set the 3 first tokens to be image, and ensure that no other tokens are image tokens # do not change this unless you modified image size or patch size input_ids[input_ids == config.image_token_index] = self.pad_token_id input_ids[:, :1] = config.image_token_index token_type_ids = torch.zeros_like(input_ids) token_type_ids[input_ids == config.image_token_index] = 1 inputs_dict = { "pixel_values": pixel_values, "input_ids": input_ids, "attention_mask": attention_mask, "token_type_ids": token_type_ids, } return config, inputs_dict @require_torch class Gemma3Vision2TextModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase): all_model_classes = ( ( Gemma3Model, Gemma3ForConditionalGeneration, Gemma3ForSequenceClassification, ) if is_torch_available() else () ) all_generative_model_classes = (Gemma3ForConditionalGeneration,) if is_torch_available() else () test_missing_keys = False _is_stateful = True model_split_percents = [0.5, 0.6] additional_model_inputs = ["token_type_ids"] # MP works but offload doesn't work when the SigLIP MultiheadAttention is offloaded # TODO: One potential solution would be to add to set preload_module_classes = ["SiglipMultiheadAttentionPoolingHead"] # in the dispatch_model function test_cpu_offload = False test_disk_offload_safetensors = False test_disk_offload_bin = False def setUp(self): self.model_tester = Gemma3Vision2TextModelTester(self) self.config_tester = ConfigTester(self, config_class=Gemma3Config, hidden_size=37) def test_bidirectional_image_attention(self): """ Tests that each image can attend to itself bidirectionally. However an image cannot attend to future images, even within the same batch. """ config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() config._attn_implementation = "eager" model = Gemma3Model(config).to(torch_device) # First let's pass inputs without change which is one image per text and manipulate # `token_type_ids` to make sure bidirectional mask is applied where it has to be inputs_dict["token_type_ids"] = torch.zeros_like(inputs_dict["token_type_ids"]) inputs_dict["token_type_ids"][:, :4] = 1 # unmask first 4 tokens with torch.no_grad(): out = model(**inputs_dict, output_attentions=True) # We expect a non-causal mask on first 4 tokens, thus no zeros for attention in out.attentions: self.assertTrue((attention[..., :4, :4] != 0).all().item()) # Now when removing `token_type_ids`, we will get simple causal mask inputs_dict["token_type_ids"][:, :4] = 0 # mask back first 4 tokens with torch.no_grad(): out = model(**inputs_dict, output_attentions=True) # We expect a causal mask on first 4 tokens, thus no zeros for attention in out.attentions: self.assertFalse((attention[..., :4, :4] != 0).all().item()) # Let's add two "images" per text, first one spanning 4 tokens and last one 3 tokens inputs_dict["token_type_ids"][:, :4] = 1 inputs_dict["token_type_ids"][:, 7:10] = 1 with torch.no_grad(): out = model(**inputs_dict, output_attentions=True) for attention in out.attentions: self.assertTrue((attention[..., :4, :4] != 0).all().item()) self.assertTrue((attention[..., 7:10, 7:10] != 0).all().item()) # We expect a non-causal mask only within same image and no looking ahead to the future self.assertTrue((attention[..., :4, 7:10] == 0).all().item()) @unittest.skip(reason="SiglipVisionModel (vision backbone) does not support standalone training") def test_training_gradient_checkpointing(self): pass @unittest.skip(reason="SiglipVisionModel (vision backbone) does not support standalone training") def test_training_gradient_checkpointing_use_reentrant(self): pass @unittest.skip(reason="SiglipVisionModel (vision backbone) does not support standalone training") def test_training_gradient_checkpointing_use_reentrant_false(self): pass @unittest.skip("Loading nested configs with overwritten `kwargs` isn't supported yet, FIXME @raushan.") def test_load_with_mismatched_shapes(self): pass def test_automodelforcausallm(self): """ Regression test for #36741/#36917 -- make sure `AutoModelForCausalLM` works with a Gemma3 config, i.e. that `AutoModelForCausalLM.from_pretrained` pulls the text config before loading the model """ config = self.model_tester.get_config() model = Gemma3ForConditionalGeneration(config) with tempfile.TemporaryDirectory() as tmp_dir: model.save_pretrained(tmp_dir) for_causal_lm = AutoModelForCausalLM.from_pretrained(tmp_dir) self.assertIsInstance(for_causal_lm, Gemma3ForConditionalGeneration) @require_flash_attn @require_torch_accelerator @mark.flash_attn_test @slow def test_flash_attn_2_from_config(self): self.flash_attn_from_config(attn_implementation="flash_attention_2", test_fwd_in_train=False) @require_flash_attn_3 @require_torch_gpu @mark.flash_attn_3_test @slow def test_flash_attn_3_from_config(self): self.flash_attn_from_config(attn_implementation="flash_attention_3", test_fwd_in_train=False) @slow @require_torch_accelerator @require_read_token class Gemma3IntegrationTest(unittest.TestCase): def setUp(self): self.processor = Gemma3Processor.from_pretrained("google/gemma-3-4b-it", padding_side="left") url = "https://huggingface.co/datasets/hf-internal-testing/fixtures-captioning/resolve/main/cow_beach_1.png" self.messages = [ {"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]}, { "role": "user", "content": [ {"type": "image", "url": url}, {"type": "text", "text": "What is shown in this image?"}, ], }, ] def tearDown(self): cleanup(torch_device, gc_collect=True) @require_deterministic_for_xpu def test_model_4b_bf16(self): model_id = "google/gemma-3-4b-it" model = Gemma3ForConditionalGeneration.from_pretrained(model_id, dtype=torch.bfloat16).to(torch_device) inputs = self.processor.apply_chat_template( self.messages, tokenize=True, return_dict=True, return_tensors="pt", add_generation_prompt=True, ).to(torch_device) output = model.generate(**inputs, max_new_tokens=30, do_sample=False) output_text = self.processor.batch_decode(output, skip_special_tokens=True) EXPECTED_TEXTS = Expectations( { ("xpu", 3): ['user\nYou are a helpful assistant.\n\n\n\n\n\nWhat is shown in this image?\nmodel\nCertainly! \n\nThe image shows a brown cow standing on a sandy beach with turquoise water and a blue sky in the background. It looks like a'], ("cuda", (8, 0)): ['user\nYou are a helpful assistant.\n\n\n\n\n\nWhat is shown in this image?\nmodel\nCertainly! \n\nThe image shows a brown cow standing on a sandy beach with clear turquoise water and a blue sky in the background. It looks like'], ("cuda", (8, 6)): ['user\nYou are a helpful assistant.\n\n\n\n\n\nWhat is shown in this image?\nmodel\nCertainly! \n\nThe image shows a brown cow standing on a sandy beach with clear blue water and a blue sky in the background. It looks like'], ("rocm", (9, 4)): ['user\nYou are a helpful assistant.\n\n\n\n\n\nWhat is shown in this image?\nmodel\nCertainly! \n\nThe image shows a brown cow standing on a sandy beach with clear blue water and a blue sky in the background. It looks like'], ("rocm", (9, 5)): ['user\nYou are a helpful assistant.\n\n\n\n\n\nWhat is shown in this image?\nmodel\nCertainly! \n\nThe image shows a brown and white cow standing on a sandy beach with turquoise water and a distant coastline in the background. It looks'], } ) # fmt: skip EXPECTED_TEXT = EXPECTED_TEXTS.get_expectation() self.assertEqual(output_text, EXPECTED_TEXT) @require_torch_large_accelerator @require_deterministic_for_xpu def test_model_4b_batch(self): model_id = "google/gemma-3-4b-it" model = Gemma3ForConditionalGeneration.from_pretrained(model_id, dtype=torch.bfloat16).to(torch_device) messages_2 = [ {"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]}, { "role": "user", "content": [ { "type": "image", "url": "https://huggingface.co/datasets/hf-internal-testing/fixtures-captioning/resolve/main/cow_beach_1.png", }, { "type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/australia.jpg", }, {"type": "text", "text": "Are these images identical?"}, ], }, ] inputs = self.processor.apply_chat_template( [self.messages, messages_2], tokenize=True, return_dict=True, return_tensors="pt", padding=True, add_generation_prompt=True, ).to(torch_device) output = model.generate(**inputs, max_new_tokens=30, do_sample=False) output_text = self.processor.batch_decode(output, skip_special_tokens=True) EXPECTED_TEXTS = Expectations( { ("xpu", 3): [ 'user\nYou are a helpful assistant.\n\n\n\n\n\nWhat is shown in this image?\nmodel\nCertainly! \n\nThe image shows a brown cow standing on a sandy beach with turquoise water and a blue sky in the background. It looks like a', "user\nYou are a helpful assistant.\n\n\n\n\n\n\n\n\n\nAre these images identical?\nmodel\nNo, these images are not identical. \n\nHere's a breakdown of the differences:\n\n* **Image 1:** Shows a brown", ], ("cuda", (8,0)): [ 'user\nYou are a helpful assistant.\n\n\n\n\n\nWhat is shown in this image?\nmodel\nCertainly! \n\nThe image shows a brown cow standing on a sandy beach with clear turquoise water and a blue sky in the background. It looks like', "user\nYou are a helpful assistant.\n\n\n\n\n\n\n\n\n\nAre these images identical?\nmodel\nNo, these images are not identical. \n\nHere's a breakdown of the differences:\n\n* **Image 1:** Shows a brown" ], ("cuda", (8,6)): [ 'user\nYou are a helpful assistant.\n\n\n\n\n\nWhat is shown in this image?\nmodel\nCertainly! \n\nThe image shows a brown cow standing on a sandy beach with clear turquoise water and a blue sky in the background. It looks like', "user\nYou are a helpful assistant.\n\n\n\n\n\n\n\n\n\nAre these images identical?\nmodel\nNo, these images are not identical. \n\nHere's a breakdown of the differences:\n\n* **Image 1:** Shows a brown" ], ("rocm", (9, 4)): [ 'user\nYou are a helpful assistant.\n\n\n\n\n\nWhat is shown in this image?\nmodel\nCertainly! \n\nThe image shows a brown cow standing on a sandy beach with turquoise water and a blue sky in the background. It looks like a', "user\nYou are a helpful assistant.\n\n\n\n\n\n\n\n\n\nAre these images identical?\nmodel\nNo, these images are not identical. \n\nHere's a breakdown of the differences:\n\n* **Image 1:** Shows a cow" ], ("rocm", (9, 5)): [ 'user\nYou are a helpful assistant.\n\n\n\n\n\nWhat is shown in this image?\nmodel\nCertainly! \n\nThe image shows a brown and white cow standing on a sandy beach next to a turquoise ocean. There are some clouds in the blue', 'user\nYou are a helpful assistant.\n\n\n\n\n\n\n\n\n\nAre these images identical?\nmodel\nNo, these images are not identical. They depict very different scenes. \n\n* **Image 1** shows a cow standing on a beach', ], } ) # fmt: skip EXPECTED_TEXT = EXPECTED_TEXTS.get_expectation() self.assertEqual(output_text, EXPECTED_TEXT) @require_torch_large_accelerator def test_model_4b_crops(self): model_id = "google/gemma-3-4b-it" model = Gemma3ForConditionalGeneration.from_pretrained(model_id, dtype=torch.bfloat16).to(torch_device) crop_config = { "images_kwargs": { "do_pan_and_scan": True, "pan_and_scan_max_num_crops": 448, "pan_and_scan_min_crop_size": 32, "pan_and_scan_min_ratio_to_activate": 0.3, } } inputs = self.processor.apply_chat_template( self.messages, tokenize=True, return_dict=True, return_tensors="pt", add_generation_prompt=True, **crop_config, ).to(torch_device) output = model.generate(**inputs, max_new_tokens=30, do_sample=False, cache_implementation="static") output_text = self.processor.batch_decode(output, skip_special_tokens=True) EXPECTED_NUM_IMAGES = 3 # one for the origin image and two crops of images EXPECTED_TEXTS = Expectations( { ("xpu", 3): ["user\nYou are a helpful assistant.\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nWhat is shown in this image?\nmodel\nThe image shows a brown cow standing on a sandy beach next to a turquoise ocean. There's a bright blue sky with some white clouds in the"], ("cuda", (8, 0)): ["user\nYou are a helpful assistant.\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nWhat is shown in this image?\nmodel\nThe image shows a brown cow standing on a sandy beach next to a turquoise ocean. There's a blue sky with some white clouds in the background"], ("cuda", (8, 6)): ["user\nYou are a helpful assistant.\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nWhat is shown in this image?\nmodel\nThe image shows a brown cow standing on a sandy beach next to a turquoise ocean. There's a bright blue sky with some white clouds in the"], ("cuda", (9, 0)): ["user\nYou are a helpful assistant.\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nWhat is shown in this image?\nmodel\nThe image shows a brown cow standing on a sandy beach next to a turquoise ocean. There's a bright blue sky with some white clouds in the"], ("rocm", (9, 4)): ["user\nYou are a helpful assistant.\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nWhat is shown in this image?\nmodel\nThe image shows a brown cow standing on a sandy beach next to a turquoise ocean. There's a bright blue sky with some white clouds in the"], ("rocm", (9, 5)): ["user\nYou are a helpful assistant.\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nWhat is shown in this image?\nmodel\nThe image shows a brown cow standing on a sandy beach next to a turquoise ocean. There's a blue sky with some white clouds in the background"] } ) # fmt: skip EXPECTED_TEXT = EXPECTED_TEXTS.get_expectation() self.assertEqual(len(inputs["pixel_values"]), EXPECTED_NUM_IMAGES) self.assertEqual(output_text, EXPECTED_TEXT) @require_torch_large_accelerator @require_deterministic_for_xpu def test_model_4b_batch_crops(self): model_id = "google/gemma-3-4b-it" model = Gemma3ForConditionalGeneration.from_pretrained(model_id, dtype=torch.bfloat16).to(torch_device) crop_config = { "images_kwargs": { "do_pan_and_scan": True, "pan_and_scan_max_num_crops": 448, "pan_and_scan_min_crop_size": 32, "pan_and_scan_min_ratio_to_activate": 0.3, } } messages_2 = [ {"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]}, { "role": "user", "content": [ { "type": "image", "url": "https://huggingface.co/datasets/hf-internal-testing/fixtures-captioning/resolve/main/cow_beach_1.png", }, { "type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/australia.jpg", }, {"type": "text", "text": "Are these images identical?"}, ], }, ] inputs = self.processor.apply_chat_template( [self.messages, messages_2], tokenize=True, return_dict=True, return_tensors="pt", padding=True, add_generation_prompt=True, **crop_config, ).to(torch_device) output = model.generate(**inputs, max_new_tokens=30, do_sample=False, cache_implementation="static") output_text = self.processor.batch_decode(output, skip_special_tokens=True) EXPECTED_NUM_IMAGES = 9 # 3 * (one for the origin image and two crops of images) = 9 EXPECTED_TEXTS = Expectations( { ("xpu", 3): [ "user\nYou are a helpful assistant.\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nWhat is shown in this image?\nmodel\nThe image shows a brown cow standing on a sandy beach next to a turquoise ocean. There's a bright blue sky with some white clouds in the", 'user\nYou are a helpful assistant.\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nAre these images identical?\nmodel\nNo, the images are not identical. \n\nThe first image shows a cow on a beach, while the second image shows a street scene with a'], ("cuda", 7): [], ("cuda", (8,0)): [ "user\nYou are a helpful assistant.\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nWhat is shown in this image?\nmodel\nThe image shows a brown cow standing on a sandy beach next to a turquoise ocean. There's a blue sky with some white clouds in the background", 'user\nYou are a helpful assistant.\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nAre these images identical?\nmodel\nNo, the images are not identical. \n\nThe first image shows a cow on a beach, while the second image shows a street scene with a' ], ("cuda", (8, 6)): [ "user\nYou are a helpful assistant.\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nWhat is shown in this image?\nmodel\nThe image shows a brown cow standing on a sandy beach next to a turquoise ocean. There's a bright blue sky with some white clouds in the", 'user\nYou are a helpful assistant.\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nAre these images identical?\nmodel\nNo, the images are not identical. \n\nThe first image shows a cow on a beach, while the second image shows a street scene with a' ], ("rocm", (9, 4)) : [ "user\nYou are a helpful assistant.\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nWhat is shown in this image?\nmodel\nThe image shows a brown cow standing on a sandy beach next to a turquoise ocean. There's a bright blue sky with some white clouds in the", 'user\nYou are a helpful assistant.\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nAre these images identical?\nmodel\nNo, the images are not identical. \n\nThe first image shows a cow on a beach, while the second image shows a street scene with a' ], ("rocm", (9, 5)) : [ 'user\nYou are a helpful assistant.\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nWhat is shown in this image?\nmodel\nThe image shows a brown cow standing on a sandy beach next to a turquoise ocean. There are clouds in the blue sky above.', 'user\nYou are a helpful assistant.\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nHere is the original image \n\n\n\n and here are some crops to help you see better \n\n\n\n \n\n\n\nAre these images identical?\nmodel\nNo, the images are not identical. \n\nThe first image shows a cow on a beach, while the second image shows a street scene with a', ], } ) # fmt: skip EXPECTED_TEXT = EXPECTED_TEXTS.get_expectation() self.assertEqual(len(inputs["pixel_values"]), EXPECTED_NUM_IMAGES) self.assertEqual(output_text, EXPECTED_TEXT) @require_torch_large_accelerator def test_model_4b_multiimage(self): model_id = "google/gemma-3-4b-it" model = Gemma3ForConditionalGeneration.from_pretrained(model_id, dtype=torch.bfloat16).to(torch_device) messages = [ {"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]}, { "role": "user", "content": [ { "type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/australia.jpg", }, {"type": "text", "text": "What do you see here?"}, ], }, ] inputs = self.processor.apply_chat_template( messages, tokenize=True, return_dict=True, return_tensors="pt", padding=True, add_generation_prompt=True, ).to(torch_device) output = model.generate(**inputs, max_new_tokens=30, do_sample=False, cache_implementation="static") output_text = self.processor.batch_decode(output, skip_special_tokens=True) EXPECTED_TEXTS = Expectations( { ("xpu", 3): ["user\nYou are a helpful assistant.\n\n\n\n\n\nWhat do you see here?\nmodel\nOkay, let's break down what I see in this image:\n\n**Overall Scene:**\n\nIt looks like a street scene in a city with"], ("cuda", (8, 0)): ["user\nYou are a helpful assistant.\n\n\n\n\n\nWhat do you see here?\nmodel\nOkay, let's break down what I see in this image:\n\n**Overall Scene:**\n\nIt looks like a street scene in a vibrant,"], ("cuda", (8, 6)): ["user\nYou are a helpful assistant.\n\n\n\n\n\nWhat do you see here?\nmodel\nOkay, let's break down what I see in this image:\n\n**Overall Scene:**\n\nIt appears to be a street scene in a city"], ("cuda", (9, 0)): ["user\nYou are a helpful assistant.\n\n\n\n\n\nWhat do you see here?\nmodel\nOkay, let's break down what I see in this image!\n\nHere's a description of the scene:\n\n* **Location:**"], ("rocm", (9, 4)): ["user\nYou are a helpful assistant.\n\n\n\n\n\nWhat do you see here?\nmodel\nOkay, let's break down what I see in this image:\n\n**Overall Scene:**\n\nIt appears to be a street scene in a vibrant"], ("rocm", (9, 5)): ["user\nYou are a helpful assistant.\n\n\n\n\n\nWhat do you see here?\nmodel\nOkay, let's break down what I see in this image:\n\n**Main Features:**\n\n* **Chinese Archway:** The most prominent"], } ) # fmt: skip EXPECTED_TEXT = EXPECTED_TEXTS.get_expectation() self.assertEqual(output_text, EXPECTED_TEXT) @require_deterministic_for_xpu def test_model_1b_text_only(self): model_id = "google/gemma-3-1b-it" model = Gemma3ForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16).to(torch_device) tokenizer = AutoTokenizer.from_pretrained(model_id, padding_side="left") inputs = tokenizer("Write a poem about Machine Learning.", return_tensors="pt").to(torch_device) output = model.generate(**inputs, max_new_tokens=30, do_sample=False, cache_implementation="static") output_text = tokenizer.batch_decode(output, skip_special_tokens=True) EXPECTED_TEXTS = Expectations( { ("xpu", 3): ['Write a poem about Machine Learning.\n\n---\n\nThe data flows, a river deep,\nWith patterns hidden, secrets sleep.\nA neural net, a watchful eye,\nLearning'], ("cuda", 7): ['Write a poem about Machine Learning.\n\n---\n\nThe data flows, a silent stream,\nInto the neural net, a waking dream.\nAlgorithms hum, a coded grace,\n'], ("cuda", 8): ['Write a poem about Machine Learning.\n\n---\n\nThe data flows, a silent stream,\nInto the neural net, a waking dream.\nAlgorithms hum, a coded grace,\n'], ("rocm", (9, 4)): ['Write a poem about Machine Learning.\n\n---\n\nThe data flows, a silent stream,\nInto the neural net, a waking dream.\nAlgorithms hum, a coded grace,\n'], ("rocm", (9, 5)): ['Write a poem about Machine Learning.\n\n---\n\nThe data flows, a river deep,\nWith patterns hidden, secrets sleep.\nA neural net, a watchful eye,\nLearning'], } ) # fmt: skip EXPECTED_TEXT = EXPECTED_TEXTS.get_expectation() self.assertEqual(output_text, EXPECTED_TEXT) # TODO: raushan FA2 generates gibberish for no reason, check later @require_flash_attn @require_torch_large_accelerator @pytest.mark.flash_attn_test def test_model_4b_flash_attn(self): model_id = "google/gemma-3-4b-it" model = Gemma3ForConditionalGeneration.from_pretrained( model_id, dtype=torch.bfloat16, attn_implementation="flash_attention_2" ).to(torch_device) inputs = self.processor.apply_chat_template( self.messages, tokenize=True, return_dict=True, return_tensors="pt", add_generation_prompt=True, ).to(torch_device) output = model.generate(**inputs, max_new_tokens=30, do_sample=False) output_text = self.processor.batch_decode(output, skip_special_tokens=True) EXPECTED_TEXTS = Expectations( { ("xpu", 3): ['user\nYou are a helpful assistant.\n\n\n\n\n\nWhat is shown in this image?\nmodel\nThe image shows a brown and white cow standing on a sandy beach with turquoise water and a distant island in the background. It looks like a sunny day'], ("cuda", 7): [], ("cuda", 8): ['user\nYou are a helpful assistant.\n\n\n\n\n\nWhat is shown in this image?\nmodel\nThe image shows a brown and white cow standing on a sandy beach with turquoise water and a distant island in the background. It looks like a sunny day'], ("rocm", (9, 4)): ['user\nYou are a helpful assistant.\n\n\n\n\n\nWhat is shown in this image?\nmodel\nThe image shows a brown and white cow standing on a sandy beach with turquoise water and a distant island in the background. It looks like a sunny day'], ("rocm", (9, 5)): ['user\nYou are a helpful assistant.\n\n\n\n\n\nWhat is shown in this image?\nmodel\nThe image shows a brown and white cow standing on a sandy beach with a turquoise ocean and a distant island in the background. It looks like a sunny'], } ) # fmt: skip EXPECTED_TEXT = EXPECTED_TEXTS.get_expectation() self.assertEqual(output_text, EXPECTED_TEXT) @parameterized.expand([("flash_attention_2",), ("sdpa",), ("eager",)]) def test_generation_beyond_sliding_window(self, attn_implementation: str): """Test that we can correctly generate beyond the sliding window. This is non trivial as we need to correctly slice the attention mask in all cases (because we use a hybrid cache). Outputs for every attention functions should be coherent and identical. """ model_id = "google/gemma-3-1b-it" if attn_implementation == "flash_attention_2" and not is_flash_attn_2_available(): self.skipTest("FlashAttention2 is required for this test.") input_text = [ "This is a nice place. " * 800 + "I really enjoy the scenery,", # This is larger than 4096 tokens "A list of colors: red, blue", # This will almost all be padding tokens ] tokenizer = AutoTokenizer.from_pretrained(model_id, padding="left") inputs = tokenizer(input_text, padding=True, return_tensors="pt").to(torch_device) model = AutoModelForCausalLM.from_pretrained( model_id, attn_implementation=attn_implementation, dtype=torch.float16 ).to(torch_device) # Make sure prefill is larger than sliding window input_size = inputs.input_ids.shape[-1] self.assertTrue(input_size > model.config.sliding_window) out = model.generate(**inputs, max_new_tokens=20, do_sample=False)[:, input_size:] output_text = tokenizer.batch_decode(out) EXPECTED_COMPLETIONS = [ " and I'm going to take a walk.\n\nI really enjoy the scenery, and I'", ", green, yellow, orange, purple, brown, black, white, gray.\n\nI'", ] self.assertEqual(output_text, EXPECTED_COMPLETIONS) @pytest.mark.torch_export_test def test_export_text_only(self): if not is_torch_greater_or_equal("2.6.0"): self.skipTest(reason="This test requires torch >= 2.6 to run.") from transformers.integrations.executorch import TorchExportableModuleForDecoderOnlyLM model_id = "google/gemma-3-1b-it" model = AutoModelForCausalLM.from_pretrained(model_id) self.assertEqual(model.config.cache_implementation, "hybrid") # Export model.eval() exportable_module = TorchExportableModuleForDecoderOnlyLM(model, batch_size=1, max_cache_len=1024) exported_program = exportable_module.export( input_ids=torch.tensor([[1]], dtype=torch.long, device=model.device), cache_position=torch.tensor([0], dtype=torch.long, device=model.device), ) logging.info(f"\nExported program: {exported_program}") # Test generation with the exported model prompt = "What is the capital of France?" max_new_tokens_to_generate = 20 # Generate text with the exported model tokenizer = AutoTokenizer.from_pretrained(model_id) export_generated_text = TorchExportableModuleForDecoderOnlyLM.generate( exported_program, tokenizer, prompt, max_new_tokens=max_new_tokens_to_generate ) logging.info(f"\nExport generated texts: '{export_generated_text}'") input_text = tokenizer(prompt, return_tensors="pt") with torch.no_grad(): eager_outputs = model.generate( **input_text, max_new_tokens=max_new_tokens_to_generate, do_sample=False, # Use greedy decoding to match the exported model ) eager_generated_text = tokenizer.decode(eager_outputs[0], skip_special_tokens=True) logging.info(f"\nEager generated texts: '{eager_generated_text}'") self.assertEqual(export_generated_text, eager_generated_text) def test_dynamic_sliding_window_is_default(self): """ Test that the dynamic sliding window cache (added in #40039) is the default cache implementation for Gemma3 models, despite the fact that Hub checkpoints may have `cache_implementation="hybrid"` (static sliding window). """ model_id = "google/gemma-3-1b-it" model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto") # the default cache is static sliding window self.assertEqual(model.config.cache_implementation, "hybrid") self.assertEqual(model.generation_config.cache_implementation, "hybrid") tokenizer = AutoTokenizer.from_pretrained(model_id) prompt = "What is the capital of France?" model_inputs = tokenizer(prompt, return_tensors="pt").to(model.device) forward_outputs = model(**model_inputs) self.assertIn("DynamicSlidingWindowLayer", str(forward_outputs.past_key_values)) generate_outputs = model.generate( **model_inputs, max_new_tokens=2, do_sample=False, return_dict_in_generate=True ) self.assertIn("DynamicSlidingWindowLayer", str(generate_outputs.past_key_values)) # If we manually specify the cache implementation = "hybrid", it will use the static sliding window cache generate_outputs = model.generate( **model_inputs, max_new_tokens=2, do_sample=False, return_dict_in_generate=True, cache_implementation="hybrid", ) self.assertNotIn("DynamicSlidingWindowLayer", str(generate_outputs.past_key_values))