declare-lab/tango2

1import gc2import unittest3 4from parameterized import parameterized5 6from diffusers import FlaxUNet2DConditionModel7from diffusers.utils import is_flax_available8from diffusers.utils.testing_utils import load_hf_numpy, require_flax, slow9 10 11if is_flax_available():12    import jax13    import jax.numpy as jnp14 15 16@slow17@require_flax18class FlaxUNet2DConditionModelIntegrationTests(unittest.TestCase):19    def get_file_format(self, seed, shape):20        return f"gaussian_noise_s={seed}_shape={'_'.join([str(s) for s in shape])}.npy"21 22    def tearDown(self):23        # clean up the VRAM after each test24        super().tearDown()25        gc.collect()26 27    def get_latents(self, seed=0, shape=(4, 4, 64, 64), fp16=False):28        dtype = jnp.bfloat16 if fp16 else jnp.float3229        image = jnp.array(load_hf_numpy(self.get_file_format(seed, shape)), dtype=dtype)30        return image31 32    def get_unet_model(self, fp16=False, model_id="CompVis/stable-diffusion-v1-4"):33        dtype = jnp.bfloat16 if fp16 else jnp.float3234        revision = "bf16" if fp16 else None35 36        model, params = FlaxUNet2DConditionModel.from_pretrained(37            model_id, subfolder="unet", dtype=dtype, revision=revision38        )39        return model, params40 41    def get_encoder_hidden_states(self, seed=0, shape=(4, 77, 768), fp16=False):42        dtype = jnp.bfloat16 if fp16 else jnp.float3243        hidden_states = jnp.array(load_hf_numpy(self.get_file_format(seed, shape)), dtype=dtype)44        return hidden_states45 46    @parameterized.expand(47        [48            # fmt: off49            [83, 4, [-0.2323, -0.1304, 0.0813, -0.3093, -0.0919, -0.1571, -0.1125, -0.5806]],50            [17, 0.55, [-0.0831, -0.2443, 0.0901, -0.0919, 0.3396, 0.0103, -0.3743, 0.0701]],51            [8, 0.89, [-0.4863, 0.0859, 0.0875, -0.1658, 0.9199, -0.0114, 0.4839, 0.4639]],52            [3, 1000, [-0.5649, 0.2402, -0.5518, 0.1248, 1.1328, -0.2443, -0.0325, -1.0078]],53            # fmt: on54        ]55    )56    def test_compvis_sd_v1_4_flax_vs_torch_fp16(self, seed, timestep, expected_slice):57        model, params = self.get_unet_model(model_id="CompVis/stable-diffusion-v1-4", fp16=True)58        latents = self.get_latents(seed, fp16=True)59        encoder_hidden_states = self.get_encoder_hidden_states(seed, fp16=True)60 61        sample = model.apply(62            {"params": params},63            latents,64            jnp.array(timestep, dtype=jnp.int32),65            encoder_hidden_states=encoder_hidden_states,66        ).sample67 68        assert sample.shape == latents.shape69 70        output_slice = jnp.asarray(jax.device_get((sample[-1, -2:, -2:, :2].flatten())), dtype=jnp.float32)71        expected_output_slice = jnp.array(expected_slice, dtype=jnp.float32)72 73        # Found torch (float16) and flax (bfloat16) outputs to be within this tolerance, in the same hardware74        assert jnp.allclose(output_slice, expected_output_slice, atol=1e-2)75 76    @parameterized.expand(77        [78            # fmt: off79            [83, 4, [0.1514, 0.0807, 0.1624, 0.1016, -0.1896, 0.0263, 0.0677, 0.2310]],80            [17, 0.55, [0.1164, -0.0216, 0.0170, 0.1589, -0.3120, 0.1005, -0.0581, -0.1458]],81            [8, 0.89, [-0.1758, -0.0169, 0.1004, -0.1411, 0.1312, 0.1103, -0.1996, 0.2139]],82            [3, 1000, [0.1214, 0.0352, -0.0731, -0.1562, -0.0994, -0.0906, -0.2340, -0.0539]],83            # fmt: on84        ]85    )86    def test_stabilityai_sd_v2_flax_vs_torch_fp16(self, seed, timestep, expected_slice):87        model, params = self.get_unet_model(model_id="stabilityai/stable-diffusion-2", fp16=True)88        latents = self.get_latents(seed, shape=(4, 4, 96, 96), fp16=True)89        encoder_hidden_states = self.get_encoder_hidden_states(seed, shape=(4, 77, 1024), fp16=True)90 91        sample = model.apply(92            {"params": params},93            latents,94            jnp.array(timestep, dtype=jnp.int32),95            encoder_hidden_states=encoder_hidden_states,96        ).sample97 98        assert sample.shape == latents.shape99 100        output_slice = jnp.asarray(jax.device_get((sample[-1, -2:, -2:, :2].flatten())), dtype=jnp.float32)101        expected_output_slice = jnp.array(expected_slice, dtype=jnp.float32)102 103        # Found torch (float16) and flax (bfloat16) outputs to be within this tolerance, on the same hardware104        assert jnp.allclose(output_slice, expected_output_slice, atol=1e-2)105