"""
Spectrum acceleration for MOSS-Image-Base in diffusers.

Spectrum: "Adaptive Spectral Feature Forecasting for Diffusion Sampling Acceleration"
(Han et al., arXiv:2603.01623)

Based on hanjq17/Spectrum



Usage
-----
    from spectrum_moss import apply_spectrum, SpectrumConfig

    pipe = DiffusionPipeline.from_pretrained(
        "OpenMOSS-Team/MOSS-Image-Base",
        custom_pipeline="pipeline_moss_image",
        trust_remote_code=True,
        torch_dtype={"default": torch.bfloat16, "vae": torch.float32},
    )
    apply_spectrum(pipe)                       # defaults below
    pipe.enable_model_cpu_offload()

    image = pipe("1girl", guidance_scale=3.0).images[0]
    print(pipe.transformer._spectrum_state.summary())

    apply_spectrum(pipe, SpectrumConfig(max_consecutive_forecasts=1))  # safer
    remove_spectrum(pipe)                      # restore exact behaviour

No extra forward is ever spent: error statistics are measured at refresh steps by
comparing the incoming real feature against what the forecaster had predicted for it.
"""


import math
import types
from dataclasses import dataclass, field
from typing import Any, Optional

import torch
from diffusers.models.modeling_outputs import Transformer2DModelOutput

@dataclass
class SpectrumConfig:
    warmup_steps: int = 8            # always-real steps at the start
    tail_actual_steps: int = 2       # always-real steps at the end
    window_size: float = 2.0         # initial refresh interval
    flex_window: float = 0.75        # interval growth per real step
    max_consecutive_forecasts: int = 2
    history_points: int = 8          # K anchors kept
    chebyshev_degree: int = 4        # M
    ridge_lambda: float = 0.1        # lambda
    blend_weight: float = 0.5        # w: spectral vs linear
    cache_device: str = "same"       # "same" | "cpu"
    debug: bool = False

    def validate(self) -> None:
        if self.warmup_steps < 0 or self.tail_actual_steps < 0:
            raise ValueError("warmup_steps / tail_actual_steps must be >= 0")
        if self.window_size < 1.0:
            raise ValueError("window_size must be >= 1.0")
        if self.flex_window < 0.0:
            raise ValueError("flex_window must be >= 0.0")
        if self.max_consecutive_forecasts < 0:
            raise ValueError("max_consecutive_forecasts must be >= 0")
        if self.history_points < self.chebyshev_degree + 1:
            raise ValueError("history_points must be >= chebyshev_degree + 1")
        if self.chebyshev_degree < 1:
            raise ValueError("chebyshev_degree must be >= 1")
        if self.ridge_lambda < 0.0:
            raise ValueError("ridge_lambda must be >= 0")
        if not 0.0 <= self.blend_weight <= 1.0:
            raise ValueError("blend_weight must be in [0, 1]")
        if self.cache_device not in ("same", "cpu"):
            raise ValueError("cache_device must be 'same' or 'cpu'")

    @property
    def window_cap(self) -> float:
        return max(1.0, self.window_size, float(self.history_points))

def cheb_basis(u, degree: int) -> torch.Tensor:
    u = torch.as_tensor(u, dtype=torch.float64).reshape(-1)
    cols = [torch.ones_like(u)]
    if degree >= 1:
        cols.append(u)
    for _ in range(2, degree + 1):
        cols.append(2.0 * u * cols[-1] - cols[-2])
    return torch.stack(cols[: degree + 1], dim=-1)          # [n, P]

class Forecaster:
    """Chebyshev ridge forecaster (prediction = weighted sum of anchors)"""
    def __init__(self, cfg: SpectrumConfig):
        self.cfg = cfg
        self.us: list[float] = []
        self.feats: list[torch.Tensor] = []
        self._chol: Optional[torch.Tensor] = None
        self._last: Optional[tuple[float, torch.Tensor]] = None

    def reset(self) -> None:
        self.us.clear()
        self.feats.clear()
        self._chol = None
        self._last = None

    def ready(self) -> bool:
        return len(self.us) >= self.cfg.chebyshev_degree + 1

    def update(self, u: float, feat: torch.Tensor) -> None:
        if self.feats and tuple(feat.shape) != tuple(self.feats[0].shape):
            self.reset()
        self.us.append(float(u))
        self.feats.append(feat.detach())
        if len(self.us) > self.cfg.history_points:
            self.us.pop(0)
            self.feats.pop(0)
        self._chol = None

    def _design(self) -> torch.Tensor:
        return cheb_basis(self.us, self.cfg.chebyshev_degree)   # [K, P]

    def _factor(self) -> torch.Tensor:
        if self._chol is not None:
            return self._chol
        X = self._design()
        G = X.T @ X
        P = G.shape[0]
        eye = torch.eye(P, dtype=torch.float64)
        lhs = G + self.cfg.ridge_lambda * eye
        try:
            chol = torch.linalg.cholesky(lhs)
        except RuntimeError:
            scale = max(float(G.diagonal().abs().mean().item()), 1.0)
            chol = None
            for eps in (1e-8, 1e-7, 1e-6, 1e-5):
                try:
                    chol = torch.linalg.cholesky(lhs + scale * eps * eye)
                    break
                except RuntimeError:
                    pass
            if chol is None:
                raise
        self._chol = chol
        return chol

    def _linear_weights(self, u: float) -> torch.Tensor:
        k = len(self.us)
        w = torch.zeros(k, dtype=torch.float64)
        if k == 0:
            return w
        if k == 1:
            w[0] = 1.0
            return w
        spacing = self.us[-1] - self.us[-2]
        if abs(spacing) < 1e-12:
            w[-1] = 1.0
            return w
        ratio = (float(u) - self.us[-1]) / spacing
        w[-2] = -ratio
        w[-1] = 1.0 + ratio
        return w

    def weights(self, u: float) -> torch.Tensor:
        phi = cheb_basis([u], self.cfg.chebyshev_degree)        # [1, P]
        solved = torch.cholesky_solve(self._design().T, self._factor())  # [P, K]
        spectral = (phi @ solved).reshape(-1)
        bw = self.cfg.blend_weight
        if bw >= 1.0 - 1e-9:
            return spectral
        if bw <= 1e-9:
            return self._linear_weights(u)
        return bw * spectral + (1.0 - bw) * self._linear_weights(u)

    def predict(self, u: float) -> torch.Tensor:
        ws = self.weights(u).tolist()
        acc: Optional[torch.Tensor] = None
        for w, f in zip(ws, self.feats):
            if w == 0.0:
                continue
            term = f.to(torch.float32) * float(w)
            acc = term if acc is None else acc + term
        if acc is None:
            acc = torch.zeros_like(self.feats[-1], dtype=torch.float32)
        out = acc.to(self.feats[-1].dtype)
        self._last = (float(u), out)
        return out

    def relative_error(self, u: float, actual: torch.Tensor) -> Optional[float]:
        if self._last is None:
            return None
        _, pred = self._last
        if tuple(pred.shape) != tuple(actual.shape):
            return None
        num = (pred.float() - actual.float()).norm()
        den = actual.float().norm().clamp_min(1e-6)
        return float((num / den).item())

class Branch:
    """Per-branch scheduler state"""
    def __init__(self, cfg: SpectrumConfig):
        self.cfg = cfg
        self.window = float(cfg.window_size)
        self.consecutive = 0
        self.actual = 0
        self.forecast = 0
        self.forecasters: dict[int, Forecaster] = {}
        self.errors: list[float] = []

    def forecaster(self, sample: int) -> Forecaster:
        f = self.forecasters.get(sample)
        if f is None:
            f = self.forecasters[sample] = Forecaster(self.cfg)
        return f

    def decide(self, step: int, total: int) -> bool:
        """Return True for a real forward, False for a forecast."""
        cfg = self.cfg
        if step < cfg.warmup_steps:
            return True
        if cfg.tail_actual_steps > 0 and step >= max(0, total - cfg.tail_actual_steps):
            return True
        if self.consecutive >= cfg.max_consecutive_forecasts:
            return True
        interval = max(1, int(math.floor(self.window)))
        return (self.consecutive + 1) % interval == 0

    def note(self, actual: bool, step: int) -> None:
        if actual:
            self.actual += 1
            self.consecutive = 0
            if step >= self.cfg.warmup_steps:
                self.window = min(self.window + self.cfg.flex_window, self.cfg.window_cap)
        else:
            self.forecast += 1
            self.consecutive += 1


class State:
    def __init__(self, cfg: SpectrumConfig, scheduler: Any):
        self.cfg = cfg
        self.scheduler = scheduler
        self.original_forward = None
        self.step_index = -1
        self.branches: dict[Any, Branch] = {}
        self._sig_counts: dict[Any, int] = {}
        self._log: list[str] = []

    def branch(self, key: Any) -> Branch:
        b = self.branches.get(key)
        if b is None:
            b = self.branches[key] = Branch(self.cfg)
        return b

    def reset_run(self) -> None:
        self.branches.clear()
        self._sig_counts.clear()
        self.step_index = -1

    def summary(self) -> str:
        a = sum(b.actual for b in self.branches.values())
        f = sum(b.forecast for b in self.branches.values())
        errs = [e for b in self.branches.values() for e in b.errors]
        mean_err = sum(errs) / len(errs) if errs else 0.0
        speed = (a + f) / a if a else 0.0
        return (
            f"Spectrum: actual={a} forecast={f} transformer_calls_saved="
            f"{f}/{a + f} ({speed:.2f}x) mean_forecast_rel_err={mean_err:.4f}"
        )

def validate_skip_layers(layers, depth: int, name: str = "skip_layers") -> tuple:
    layers = () if layers is None else tuple(layers)
    if any(type(i) is not int or not 0 <= i < depth for i in layers):
        raise ValueError(f"{name} must contain integer block indices in [0, {depth}).")
    return layers

def resolve_step(state: State, timestep: torch.Tensor) -> tuple[int, int]:
    sched = state.scheduler
    timesteps = getattr(sched, "timesteps", None) if sched is not None else None
    if timesteps is None or len(timesteps) == 0:
        return -1, 0
    ts = timesteps if torch.is_tensor(timesteps) else torch.as_tensor(timesteps)
    t = float(torch.as_tensor(timestep).reshape(-1)[0])
    idx = int((ts.float() - t).abs().argmin().item())
    return idx, len(ts)

def spectrum_forward(
    self,
    hidden_states: torch.Tensor,
    timestep,
    encoder_hidden_states: torch.Tensor,
    encoder_attention_mask: Optional[torch.Tensor] = None,
    return_dict: bool = True,
    skip_layers=None,
):
    """
    License: Apache-2.0
    https://huggingface.co/nebulette/moss-spectrum-diffusers
    """
    state: Optional[State] = getattr(self, "_spectrum_state", None)
    if state is None or state.scheduler is None:
        return state.original_forward(
            hidden_states, timestep, encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            return_dict=return_dict, skip_layers=skip_layers,
        )

    cfg = state.cfg
    skip = validate_skip_layers(skip_layers, len(self.layers))

    if hidden_states.ndim != 4 or hidden_states.shape[1] != self.config.latent_channels:
        raise ValueError("hidden_states must be [B, latent_channels, H, W].")
    batch, _, height, width = hidden_states.shape
    if encoder_hidden_states.ndim != 3 or encoder_hidden_states.shape[0] != batch:
        raise ValueError("encoder_hidden_states must be [B, S, text_dim].")

    mask = encoder_attention_mask
    if mask is None:
        mask = torch.ones(encoder_hidden_states.shape[:2], dtype=torch.bool, device=hidden_states.device)
    if mask.shape != encoder_hidden_states.shape[:2] or not torch.all((mask == 0) | (mask == 1)):
        raise ValueError("encoder_attention_mask must be a binary [B, S] mask.")

    # --- step bookkeeping ------------------------------------------------- #
    step, total = resolve_step(state, timestep)
    if step < 0:  # no schedule available -> stay exact
        return state.original_forward(
            hidden_states, timestep, encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            return_dict=return_dict, skip_layers=skip_layers,
        )
    if step == 0 and state.step_index > 0:
        state.reset_run()
    if step != state.step_index:
        state.step_index = step
        state._sig_counts = {}

    sig = skip
    occurrence = state._sig_counts.get(sig, 0)
    state._sig_counts[sig] = occurrence + 1
    branch = state.branch((sig, occurrence))

    actual = branch.decide(step, total)
    u = 0.0 if total <= 1 else max(-1.0, min(1.0, 2.0 * step / (total - 1) - 1.0))

    # shared timestep conditioning (identical to the original)
    times = torch.as_tensor(timestep, device=hidden_states.device, dtype=torch.float32).reshape(-1).expand(batch)
    half = self.config.freq_dim // 2
    freq = 10000 ** (-torch.arange(half, device=times.device, dtype=torch.float64) / half)
    phase = times.double()[:, None] * freq
    embedding = torch.cat((phase.cos(), phase.sin()), -1).float()
    time = self.time_embedding(embedding.to(self.time_embedding[0].weight.dtype)).float()
    head = self.video_head_modulation(time.to(self.video_head_modulation.weight.dtype)).view(batch, 2, -1)
    mod = (
        self.shared_block_modulation(time.to(self.shared_block_modulation.weight.dtype)).view(batch, 6, -1)
        if actual else None
    )

    image_tokens = height * width
    outputs = []

    for index in range(batch):
        context = encoder_hidden_states[index:index + 1, mask[index].bool()]
        if context.shape[1] == 0:
            raise ValueError("Each sample needs at least one valid text token.")

        forecaster = branch.forecaster(index)

        if not actual:
            # skip embeddings, text blocks and all layers
            try:
                feat = forecaster.predict(u)
            except Exception:
                # numerical trouble -> fall back to a real step
                actual = True
                branch.note(True, step)
            else:
                y = self.final_norm_video(feat.to(self.final_norm_video.weight.dtype))
                y = y * (1 + head[index:index + 1, 1, None]) + head[index:index + 1, 0, None]
                y = self.final_linear_video(y.to(self.final_linear_video.weight.dtype))
                outputs.append(y.transpose(1, 2).reshape(1, -1, height, width))
                continue

        # real path
        video = hidden_states[index:index + 1].flatten(2).transpose(1, 2)
        video = self.video_embedder(video.to(self.video_embedder.weight.dtype).contiguous())
        rope = self.rope(height, width, context.shape[1], hidden_states.device).to(video.dtype)
        text = self.text_embedder(context, rope[:, image_tokens:])
        x = torch.cat((video, text), 1)

        for layer_index, layer in enumerate(self.layers):
            if layer_index not in skip:
                x = layer(x, rope, mod[index:index + 1])
                # make the block compatible with tuple outputs
                if isinstance(x, tuple):
                    x = x[0]

        feature = x[:, :image_tokens]

        # measure the error of the forecast that was made for this step
        err = forecaster.relative_error(u, feature)
        if err is not None:
            branch.errors.append(err)
            if cfg.debug:
                state._log.append(f"step {step + 1}/{total} branch={sig} rel_err={err:.4f}")

        cache = feature
        if cfg.cache_device == "cpu":
            cache = feature.detach().to("cpu", non_blocking=False)
        forecaster.update(u, cache)

        y = self.final_norm_video(feature.to(self.final_norm_video.weight.dtype))
        y = y * (1 + head[index:index + 1, 1, None]) + head[index:index + 1, 0, None]
        y = self.final_linear_video(y.to(self.final_linear_video.weight.dtype))
        outputs.append(y.transpose(1, 2).reshape(1, -1, height, width))

    branch.note(actual, step)

    sample = torch.cat(outputs)
    if cfg.debug and step >= total - 1:
        state._log.append(state.summary())

    if not return_dict:
        return (sample,)
    return Transformer2DModelOutput(sample=sample)


def apply_spectrum(pipe, config: Optional[SpectrumConfig] = None):
    """
    Patch ``pipe`` in place so the MOSS DiT runs under Spectrum scheduling.

    License: Apache-2.0
    https://huggingface.co/nebulette/moss-spectrum-diffusers
    """
    cfg = config or SpectrumConfig()
    cfg.validate()

    transformer = getattr(pipe, "transformer", None)
    if transformer is None:
        raise ValueError("pipe has no `transformer`.")

    if getattr(transformer, "_spectrum_state", None) is not None:
        remove_spectrum(pipe)

    state = State(cfg, getattr(pipe, "scheduler", None))
    state.original_forward = transformer.forward

    transformer._spectrum_state = state
    transformer.forward = types.MethodType(spectrum_forward, transformer)
    return pipe


def remove_spectrum(pipe) -> None:
    """Undo :func:`apply_spectrum`, restoring the untouched forward."""
    transformer = getattr(pipe, "transformer", None)
    if transformer is None:
        return
    state = getattr(transformer, "_spectrum_state", None)
    if state is None:
        return
    if state.original_forward is not None:
        transformer.forward = state.original_forward
    try:
        del transformer._spectrum_state
    except AttributeError:
        pass