import gc
import os

import numpy as np
import pytest
import torch
from PIL import Image
from transformers import Qwen2TokenizerFast, Qwen3Config, Qwen3ForCausalLM

from diffusers import (
    AutoencoderKLFlux2,
    FlowMatchEulerDiscreteScheduler,
    Flux2KleinPipeline,
    Flux2Transformer2DModel,
)

from ...testing_utils import (
    assert_tensors_close,
    backend_empty_cache,
    backend_synchronize,
    require_torch_neuron,
    torch_device,
)
from ..testing_utils import (
    BasePipelineTesterConfig,
    MemoryTesterMixin,
    PipelineTesterMixin,
    check_qkv_fused_layers_exist,
)


class Flux2KleinPipelineTesterConfig(BasePipelineTesterConfig):
    pipeline_class = Flux2KleinPipeline
    required_input_params_in_call_signature = frozenset(
        ["prompt", "height", "width", "guidance_scale", "prompt_embeds"]
    )
    batch_input_params = frozenset(["prompt"])
    output_shape = (3, 8, 8)

    def get_dummy_components(self, num_layers: int = 1, num_single_layers: int = 1):
        torch.manual_seed(0)
        transformer = Flux2Transformer2DModel(
            patch_size=1,
            in_channels=4,
            num_layers=num_layers,
            num_single_layers=num_single_layers,
            attention_head_dim=16,
            num_attention_heads=2,
            joint_attention_dim=16,
            timestep_guidance_channels=256,
            axes_dims_rope=[4, 4, 4, 4],
            guidance_embeds=False,
        )

        # Create minimal Qwen3 config
        config = Qwen3Config(
            intermediate_size=16,
            hidden_size=16,
            num_hidden_layers=2,
            num_attention_heads=2,
            num_key_value_heads=2,
            vocab_size=151936,
            max_position_embeddings=512,
        )
        torch.manual_seed(0)
        text_encoder = Qwen3ForCausalLM(config)

        # Use a simple tokenizer for testing
        tokenizer = Qwen2TokenizerFast.from_pretrained(
            "hf-internal-testing/tiny-random-Qwen2VLForConditionalGeneration"
        )

        torch.manual_seed(0)
        vae = AutoencoderKLFlux2(
            sample_size=32,
            in_channels=3,
            out_channels=3,
            down_block_types=("DownEncoderBlock2D",),
            up_block_types=("UpDecoderBlock2D",),
            block_out_channels=(4,),
            layers_per_block=1,
            latent_channels=1,
            norm_num_groups=1,
            use_quant_conv=False,
            use_post_quant_conv=False,
        )

        scheduler = FlowMatchEulerDiscreteScheduler()

        return {
            "scheduler": scheduler,
            "text_encoder": text_encoder,
            "tokenizer": tokenizer,
            "transformer": transformer,
            "vae": vae,
        }

    def get_dummy_inputs(self):
        inputs = {
            "prompt": "a dog is dancing",
            "generator": self.get_generator(0),
            "num_inference_steps": 2,
            "guidance_scale": 4.0,
            "height": 8,
            "width": 8,
            "max_sequence_length": 64,
            # Request torch outputs so tests compare torch tensors directly (see `BasePipelineTesterConfig`).
            # Note `"pt"` images are `(batch, channels, height, width)`, unlike `"np"` (`(batch, h, w, c)`).
            "output_type": "pt",
            "text_encoder_out_layers": (1,),
        }
        return inputs


class TestFlux2KleinPipeline(Flux2KleinPipelineTesterConfig, PipelineTesterMixin):
    def test_fused_qkv_projections(self):
        pipe = self.get_pipeline()

        inputs = self.get_dummy_inputs()
        image = pipe(**inputs).images
        original_image_slice = image[0, -1, -3:, -3:]

        pipe.transformer.fuse_qkv_projections()
        assert check_qkv_fused_layers_exist(pipe.transformer, ["to_qkv"]), (
            "Something wrong with the fused attention layers. Expected all the attention projections to be fused."
        )

        inputs = self.get_dummy_inputs()
        image = pipe(**inputs).images
        image_slice_fused = image[0, -1, -3:, -3:]

        pipe.transformer.unfuse_qkv_projections()
        inputs = self.get_dummy_inputs()
        image = pipe(**inputs).images
        image_slice_disabled = image[0, -1, -3:, -3:]

        assert_tensors_close(
            original_image_slice,
            image_slice_fused,
            atol=1e-3,
            rtol=1e-3,
            msg="Fusion of QKV projections shouldn't affect the outputs.",
        )
        assert_tensors_close(
            image_slice_fused,
            image_slice_disabled,
            atol=1e-3,
            rtol=1e-3,
            msg="Outputs, with QKV projection fusion enabled, shouldn't change when fused QKV projections are disabled.",
        )
        assert_tensors_close(
            original_image_slice,
            image_slice_disabled,
            atol=1e-2,
            rtol=1e-2,
            msg="Original outputs should match when fused QKV projections are disabled.",
        )

    def test_image_output_shape(self):
        pipe = self.get_pipeline().to(torch_device)
        inputs = self.get_dummy_inputs()

        height_width_pairs = [(32, 32), (72, 57)]
        for height, width in height_width_pairs:
            expected_height = height - height % (pipe.vae_scale_factor * 2)
            expected_width = width - width % (pipe.vae_scale_factor * 2)

            inputs.update({"height": height, "width": width})
            image = pipe(**inputs).images[0]
            _, output_height, output_width = image.shape
            assert (output_height, output_width) == (expected_height, expected_width), (
                f"Output shape {image.shape} does not match expected shape {(expected_height, expected_width)}"
            )

    def test_image_input(self):
        pipe = self.get_pipeline()
        inputs = self.get_dummy_inputs()

        inputs["image"] = Image.new("RGB", (64, 64))
        # Permute the `"pt"` output to the `"np"` layout before flattening so the slice matches the recorded values.
        image = pipe(**inputs).images.permute(0, 2, 3, 1).flatten()
        generated_slice = torch.cat([image[:8], image[-8:]])
        # fmt: off
        expected_slice = torch.tensor(
            [
                0.8255048 , 0.66054785, 0.6643694 , 0.67462724, 0.5494932 , 0.3480271 , 0.52535003, 0.44510138, 0.23549396, 0.21372932, 0.21166152, 0.63198495, 0.49942136, 0.39147034, 0.49156153, 0.3713916
            ]
        )
        # fmt: on
        assert_tensors_close(generated_slice, expected_slice, atol=1e-4, rtol=1e-4)

    @pytest.mark.skip("Needs to be revisited")
    def test_encode_prompt_works_in_isolation(self):
        pass


class TestFlux2KleinPipelineMemory(Flux2KleinPipelineTesterConfig, MemoryTesterMixin):
    """Memory optimization tests (CPU offload, group offload, layerwise casting) for the Flux2 Klein pipeline."""


@require_torch_neuron
class TestFlux2KleinPipelineIntegration:
    ckpt_id = "black-forest-labs/FLUX.2-klein-4B"
    prompt = "A small cactus with a happy face in the Sahara desert."

    @pytest.fixture(autouse=True)
    def neuron_env(self):
        saved_env = {}
        neff_cache_dir = "/tmp/neff_cache"
        os.makedirs(neff_cache_dir, exist_ok=True)
        for key in ("TORCH_NEURONX_NEFF_CACHE_DIR", "TORCH_NEURONX_ENABLE_NKI_SDPA"):
            saved_env[key] = os.environ.get(key)
        os.environ["TORCH_NEURONX_NEFF_CACHE_DIR"] = neff_cache_dir
        os.environ.setdefault("TORCH_NEURONX_ENABLE_NKI_SDPA", "0")
        gc.collect()
        backend_empty_cache(torch_device)
        yield
        for key, original in saved_env.items():
            if original is None:
                os.environ.pop(key, None)
            else:
                os.environ[key] = original
        gc.collect()
        backend_empty_cache(torch_device)

    def test_flux2_klein_inference_512(self):
        generator = torch.Generator("cpu").manual_seed(0)

        pipe = Flux2KleinPipeline.from_pretrained(self.ckpt_id, torch_dtype=torch.bfloat16)
        pipe.to(torch_device)
        backend_synchronize(torch_device)
        pipe.set_progress_bar_config(disable=None)

        image = pipe(
            prompt=self.prompt,
            height=512,
            width=512,
            num_inference_steps=4,
            guidance_scale=1.0,
            generator=generator,
            output_type="np",
        ).images

        image_slice = image[0, -3:, -3:, -1]
        assert image.shape == (1, 512, 512, 3)
        assert np.all((image >= 0.0) & (image <= 1.0)), "Pixel values must be in [0, 1]"
        expected_slice = np.array([0.3652, 0.3574, 0.3633, 0.4102, 0.4062, 0.4043, 0.4453, 0.4355, 0.4570])
        assert np.abs(image_slice.flatten() - expected_slice).max() < 5e-2

    def test_flux2_klein_neuron_compile_128(self):
        from torch_neuronx.neuron_dynamo_backend import set_model_name

        device = torch.neuron.current_device()
        generator = torch.Generator("cpu").manual_seed(0)

        pipe = Flux2KleinPipeline.from_pretrained(self.ckpt_id, torch_dtype=torch.bfloat16)
        pipe = pipe.to(device)
        backend_synchronize(torch_device)

        pipe.transformer.eval()
        pipe.vae.eval()
        pipe.text_encoder.eval()

        # Keep the text encoder eager: it reads intermediate hidden_states, which
        # transformers only materializes outside of torch.compile(fullgraph=True).
        # It runs once per generation, so leaving it uncompiled is negligible.
        set_model_name("flux2_klein_transformer")
        pipe.transformer = torch.compile(pipe.transformer, backend="neuron", fullgraph=True)

        set_model_name("flux2_klein_vae")
        pipe.vae = torch.compile(pipe.vae, backend="neuron", fullgraph=True)

        image = pipe(
            prompt=self.prompt,
            height=128,
            width=128,
            num_inference_steps=4,
            guidance_scale=1.0,
            generator=generator,
            output_type="np",
        ).images

        assert image.shape == (1, 128, 128, 3)
        assert not np.isnan(image).any(), "Output contains NaN values"
        assert (image >= 0.0).all() and (image <= 1.0).all(), "Output pixel values outside [0, 1]"