FlowCache/FlowCache4MAGI-1-dev5-history/inference/pipeline/cache/historycache.py

# Copyright (c) 2025 SandAI. 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.

"""
HistoryAwareCache (dev5): MotionDetailCache + AR historical chunk signals.

Idea 1: cross-chunk spatial accumulator with decay
Idea 2: clean latent anchor distance weights
Idea 3: active streak propagation across denoise steps
"""

from typing import Dict, List, Optional, Tuple

import torch

from .motiondetailcache import MotionDetailCache


class HistoryAwareCache(MotionDetailCache):
    """Motion + detail + AR history aware token cache."""

    def __init__(
        self,
        use_history_cache: bool = True,
        history_decay: float = 0.7,
        history_anchor_horizon: int = 3,
        history_streak_len: int = 5,
        history_anchor_lambda: float = 0.3,
        history_streak_gamma: float = 0.2,
        history_anchor_alpha: float = 0.5,
        **kwargs,
    ):
        super().__init__(**kwargs)
        self.use_history_cache = use_history_cache
        self.history_decay = history_decay
        self.history_anchor_horizon = history_anchor_horizon
        self.history_streak_len = max(1, history_streak_len)
        self.history_anchor_lambda = history_anchor_lambda
        self.history_streak_gamma = history_streak_gamma
        self.history_anchor_alpha = history_anchor_alpha

        self.cross_chunk_accumulator: Dict[int, torch.Tensor] = {}
        self.clean_latent_history: Dict[int, torch.Tensor] = {}
        self._clean_chunk_order: List[int] = []
        self.active_streak: Dict[int, torch.Tensor] = {}
        self.token_anchor_weights: Dict[int, torch.Tensor] = {}
        self.token_history_weights: Dict[int, torch.Tensor] = {}

    def reset(self):
        super().reset()
        self.cross_chunk_accumulator.clear()
        self.clean_latent_history.clear()
        self._clean_chunk_order.clear()
        self.active_streak.clear()
        self.token_anchor_weights.clear()
        self.token_history_weights.clear()

    def _shape_mask(self, x_chunk: torch.Tensor) -> Tuple[int, ...]:
        return (x_chunk.size(0), x_chunk.size(2), x_chunk.size(3), x_chunk.size(4))

    def _init_cross_chunk_state(
        self,
        chunk_id: int,
        x_chunk: torch.Tensor,
        chunk_offset: int,
    ) -> torch.Tensor:
        shape = self._shape_mask(x_chunk)
        device, dtype = x_chunk.device, x_chunk.dtype
        if chunk_id in self.cross_chunk_accumulator:
            return self.cross_chunk_accumulator[chunk_id]

        carried = torch.zeros(shape, device=device, dtype=dtype)
        if chunk_id > chunk_offset and (chunk_id - 1) in self.cross_chunk_accumulator:
            prev = self.cross_chunk_accumulator[chunk_id - 1]
            if prev.shape == shape:
                carried = self.history_decay * prev
            else:
                prev_last = prev[:, -1:, :, :]
                carried = self.history_decay * prev_last.expand(shape[0], shape[1], shape[2], shape[3])
        self.cross_chunk_accumulator[chunk_id] = carried
        return carried

    def _init_active_streak(self, chunk_id: int, x_chunk: torch.Tensor, chunk_offset: int) -> torch.Tensor:
        shape = self._shape_mask(x_chunk)
        device, dtype = x_chunk.device, x_chunk.dtype
        if chunk_id in self.active_streak:
            return self.active_streak[chunk_id]

        streak = torch.zeros(shape, device=device, dtype=dtype)
        if chunk_id > chunk_offset and (chunk_id - 1) in self.active_streak:
            prev = self.active_streak[chunk_id - 1]
            if prev.shape == shape:
                streak = torch.clamp(prev * self.history_decay, max=float(self.history_streak_len))
            else:
                prev_last = prev[:, -1:, :, :]
                streak = torch.clamp(
                    prev_last.expand(shape[0], shape[1], shape[2], shape[3]) * self.history_decay,
                    max=float(self.history_streak_len),
                )
        self.active_streak[chunk_id] = streak
        return streak

    def register_clean_chunk(self, chunk_id: int, x_clean: torch.Tensor):
        """Store clean latent for anchor reference (Idea 2)."""
        if not self.use_history_cache:
            return
        # Keep conditional branch only when CFG duplicates batch.
        if x_clean.size(0) > 1:
            x_clean = x_clean[:1].clone()
        else:
            x_clean = x_clean.detach().clone()
        self.clean_latent_history[chunk_id] = x_clean
        if chunk_id in self._clean_chunk_order:
            self._clean_chunk_order.remove(chunk_id)
        self._clean_chunk_order.append(chunk_id)
        while len(self._clean_chunk_order) > self.history_anchor_horizon:
            old_id = self._clean_chunk_order.pop(0)
            self.clean_latent_history.pop(old_id, None)

    def _get_recent_clean_refs(self, chunk_id: int) -> List[torch.Tensor]:
        refs = []
        for cid in reversed(self._clean_chunk_order):
            if cid < chunk_id and cid in self.clean_latent_history:
                refs.append(self.clean_latent_history[cid])
            if len(refs) >= self.history_anchor_horizon:
                break
        return refs

    def compute_anchor_weights(
        self,
        x_chunk: torch.Tensor,
        chunk_id: int,
        chunk_offset: int,
    ) -> torch.Tensor:
        """Idea 2: rel-L1 distance to recent clean latent history."""
        n, t, h, w = self._shape_mask(x_chunk)
        device, dtype = x_chunk.device, x_chunk.dtype
        refs = self._get_recent_clean_refs(chunk_id)

        importance = torch.zeros(n, t, h, w, device=device, dtype=dtype)
        cur_mag = x_chunk.float().abs().mean(dim=1)[:n]

        if refs:
            for ref in refs:
                ref_mag = ref.float().abs().mean(dim=1)[:n]
                rt = ref_mag.size(1)
                for frame_idx in range(t):
                    ref_frame = ref_mag[:, min(frame_idx, rt - 1)]
                    diff = (cur_mag[:, frame_idx] - ref_frame).abs()
                    denom = ref_frame.abs().mean(dim=(1, 2), keepdim=True) + self.eps
                    rel = diff / denom
                    importance[:, frame_idx] = torch.maximum(importance[:, frame_idx], rel)
        elif chunk_id > chunk_offset and (chunk_id - 1) in self.prev_latent_chunks:
            prev = self.prev_latent_chunks[chunk_id - 1].float().abs().mean(dim=1)[:n]
            prev_last = prev[:, -1]
            diff = (cur_mag[:, 0] - prev_last).abs()
            denom = prev_last.abs().mean(dim=(1, 2), keepdim=True) + self.eps
            importance[:, 0] = diff / denom
            if t > 1:
                importance[:, 1] = importance[:, 0]

        weights = torch.zeros_like(importance)
        for frame_idx in range(t):
            frame_importance = importance[:, frame_idx]
            min_val = frame_importance.amin(dim=(1, 2), keepdim=True)
            max_val = frame_importance.amax(dim=(1, 2), keepdim=True)
            normalized = (frame_importance - min_val) / (max_val - min_val + self.eps)
            weights[:, frame_idx] = self.history_anchor_alpha + (1.0 - self.history_anchor_alpha) * normalized
        return weights.to(dtype=dtype)

    def compute_streak_boost(self, chunk_id: int) -> Optional[torch.Tensor]:
        """Idea 3: boost weight for tokens with sustained activity."""
        streak = self.active_streak.get(chunk_id)
        if streak is None:
            return None
        return self.history_streak_gamma * torch.clamp(
            streak / float(self.history_streak_len), max=1.0
        )

    def fuse_history_weights(
        self,
        base_weights: torch.Tensor,
        anchor_weights: torch.Tensor,
        streak_boost: Optional[torch.Tensor],
    ) -> torch.Tensor:
        lam = self.history_anchor_lambda
        fused = (1.0 - lam) * base_weights + lam * torch.maximum(base_weights, anchor_weights)
        if streak_boost is not None:
            fused = fused * (1.0 + streak_boost)
        return fused

    def update_active_streak(self, chunk_id: int, token_mask: torch.Tensor):
        if chunk_id not in self.active_streak:
            return
        streak = self.active_streak[chunk_id]
        active = token_mask.to(dtype=streak.dtype)
        streak.copy_(torch.where(active > 0, streak + 1, torch.clamp(streak - 1, min=0)))

    def update_token_policy(
        self,
        chunk_id: int,
        x_chunk: torch.Tensor,
        current_features: torch.Tensor,
        chunk_offset: int,
        chunk_denoise_count: Optional[Dict[int, int]] = None,
    ) -> torch.Tensor:
        if not self.use_history_cache:
            return super().update_token_policy(
                chunk_id, x_chunk, current_features, chunk_offset, chunk_denoise_count
            )

        if (
            chunk_denoise_count is not None
            and chunk_denoise_count.get(chunk_id, 0) == self.phase1_steps
        ):
            mask = torch.ones(self._shape_mask(x_chunk), device=x_chunk.device, dtype=torch.bool)
            self.token_active_mask[chunk_id] = mask
            self.token_accumulator[chunk_id] = torch.zeros(
                self._shape_mask(x_chunk), device=x_chunk.device, dtype=x_chunk.dtype
            )
            self._init_cross_chunk_state(chunk_id, x_chunk, chunk_offset)
            self._init_active_streak(chunk_id, x_chunk, chunk_offset)
            return mask

        prev_features = self.prev_metric_chunks.get(chunk_id)
        if prev_features is None:
            mask = torch.ones(self._shape_mask(x_chunk), device=x_chunk.device, dtype=torch.bool)
            self.token_active_mask[chunk_id] = mask
            return mask

        delta_chunk = self.compute_chunk_delta_l1(current_features, prev_features)
        motion_weights = self.compute_motion_weights(x_chunk, chunk_id, chunk_offset)
        detail_weights = self.compute_detail_weights(x_chunk)
        base_weights = self.combine_motion_detail_weights(motion_weights, detail_weights)

        anchor_weights = self.compute_anchor_weights(x_chunk, chunk_id, chunk_offset)
        self._init_active_streak(chunk_id, x_chunk, chunk_offset)
        streak_boost = self.compute_streak_boost(chunk_id)
        final_weights = self.fuse_history_weights(base_weights, anchor_weights, streak_boost)

        self.token_motion_weights[chunk_id] = motion_weights
        self.token_detail_weights[chunk_id] = detail_weights
        self.token_combined_weights[chunk_id] = base_weights
        self.token_anchor_weights[chunk_id] = anchor_weights
        self.token_history_weights[chunk_id] = final_weights

        if chunk_id not in self.token_accumulator:
            self.token_accumulator[chunk_id] = torch.zeros_like(final_weights)
        cross_acc = self._init_cross_chunk_state(chunk_id, x_chunk, chunk_offset)

        step_mass = final_weights * delta_chunk
        self.token_accumulator[chunk_id] = self.token_accumulator[chunk_id] + step_mass
        self.cross_chunk_accumulator[chunk_id] = cross_acc + step_mass

        local_active = self.token_accumulator[chunk_id] > self.rel_l1_thresh
        cross_active = self.cross_chunk_accumulator[chunk_id] > self.rel_l1_thresh
        mask = local_active | cross_active
        self.token_active_mask[chunk_id] = mask
        return mask

    def reset_token_accumulator(self, chunk_id: int, mask: torch.Tensor):
        super().reset_token_accumulator(chunk_id, mask)
        if not self.use_history_cache:
            return
        if chunk_id in self.cross_chunk_accumulator:
            self.cross_chunk_accumulator[chunk_id] = torch.where(
                mask,
                torch.zeros_like(self.cross_chunk_accumulator[chunk_id]),
                self.cross_chunk_accumulator[chunk_id],
            )

    def record_motion_decision(
        self,
        chunk_id: int,
        reused: bool,
        active_ratio: Optional[float] = None,
        **kwargs,
    ):
        if not self.metric_stats_path:
            return
        anchor_ratio = None
        anchor_w = self.token_anchor_weights.get(chunk_id)
        if anchor_w is not None:
            anchor_ratio = float((anchor_w > self.history_anchor_alpha + 1e-6).float().mean().item())

        record = {
            "infer_idx": kwargs.get("infer_idx"),
            "cur_denoise_step": kwargs.get("cur_denoise_step"),
            "denoise_stage": kwargs.get("denoise_stage"),
            "denoise_idx": kwargs.get("denoise_idx"),
            "chunk_idx": chunk_id,
            "generated_chunk_idx": chunk_id - kwargs.get("chunk_offset", 0),
            "chunk_denoise_count": kwargs.get("chunk_denoise_count_value"),
            "phase": (
                "phase1_chunk"
                if self.in_phase1(chunk_id, kwargs.get("chunk_denoise_count", {}))
                else "phase2_token"
            ),
            "reused": bool(reused),
            "execution": "reuse" if reused else "compute",
            "active_token_ratio": active_ratio,
            "high_anchor_token_ratio": anchor_ratio,
            "use_history_cache": self.use_history_cache,
            "history_decay": self.history_decay,
            "history_anchor_lambda": self.history_anchor_lambda,
            "history_streak_gamma": self.history_streak_gamma,
            "rel_l1_thresh": self.rel_l1_thresh,
        }
        self.execution_records.append(record)

    def save_metric_stats(self):
        if not self.metric_stats_path:
            return
        import json
        import os

        save_dir = os.path.dirname(self.metric_stats_path)
        if save_dir:
            os.makedirs(save_dir, exist_ok=True)

        payload = {
            "description": "HistoryAwareCache: motion + detail + AR history (cross-chunk acc, anchor, streak).",
            "hyperparameters": {
                "alpha": self.alpha,
                "detail_alpha": self.detail_alpha,
                "detail_window_size": self.detail_window_size,
                "detail_lambda": self.detail_lambda,
                "weight_combine_mode": self.weight_combine_mode,
                "use_history_cache": self.use_history_cache,
                "history_decay": self.history_decay,
                "history_anchor_horizon": self.history_anchor_horizon,
                "history_streak_len": self.history_streak_len,
                "history_anchor_lambda": self.history_anchor_lambda,
                "history_streak_gamma": self.history_streak_gamma,
                "phase1_steps": self.phase1_steps,
                "warmup_steps": self.warmup_steps,
                "rel_l1_thresh": self.rel_l1_thresh,
            },
            "chunk_execution_summary": self.get_execution_summary(),
            "execution_records": self.execution_records,
            "records": self.metric_records,
        }
        if self.metric_stats_path.endswith((".pt", ".pth")):
            torch.save(payload, self.metric_stats_path)
        else:
            with open(self.metric_stats_path, "w") as f:
                json.dump(payload, f, indent=2)
        print(f"Saved HistoryAwareCache metric stats to {self.metric_stats_path}")