"""Deterministic sequential ring loader (PRIVATE to Hardware_Module) - Tier 3 (1-2 GB RAM).

The model skeleton is built with EMPTY parameters (zero-size tensors, real CPU buffers such as the
rotary inv_freq). Forward pre/post hooks then stream weights from the safetensors files in the exact
order  embed -> layer 0 -> layer 1 -> ... -> layer N-1 -> norm -> lm_head  (and wrap around to layer 0 for
the next token). At most `window` (default 2) decoder layers are resident at any time.

Parameters are always REPLACED by fresh nn.Parameter objects (never mutated in place), so autograd
graphs / checkpoint recomputation that still reference an evicted weight stay valid.

Extra capabilities used by the Elevator trainer:
  pin(idxs)/unpin(idxs, save=True)   keep trainable layers resident, persist them as overlay shards
  export(dest)                       write the full (base + overlay) model as safetensors shards
"""
from __future__ import annotations

import gc
import json
import shutil
import time
from collections import deque
from pathlib import Path
from typing import Dict, Iterator, List, Optional, Tuple

import torch
from torch import nn
from safetensors import safe_open
from safetensors.torch import save_file


def _from_config(config, dtype):
    from transformers import AutoModelForCausalLM
    try:
        return AutoModelForCausalLM.from_config(config, dtype=dtype)
    except TypeError:
        return AutoModelForCausalLM.from_config(config, torch_dtype=dtype)


class SequentialRingLoader:
    EMBED_KEY = "model.embed_tokens.weight"
    HEAD_KEY = "lm_head.weight"

    def __init__(self, model_dir, dtype=torch.bfloat16, window: int = 2, evict_ends: bool = True,
                 overlay_dir: Optional[str] = None, row_recipe: Optional[Dict] = None, verbose: bool = True):
        self.model_dir = Path(model_dir)
        self.dtype = dtype
        self.window = max(1, int(window))
        self.evict_ends = bool(evict_ends)
        self.overlay_dir = Path(overlay_dir) if overlay_dir else self.model_dir / ".rss_overlay"
        self.row_recipe = {int(k): [int(x) for x in v] for k, v in (row_recipe or {}).items()}
        self.verbose = verbose
        self.trace: deque = deque(maxlen=256)
        self.pinned: set = set()
        self.loaded: set = set()
        self.overlay_map: Dict[str, Path] = {}
        self.weight_map: Dict[str, str] = {}
        self._handles: Dict[str, object] = {}
        self.model = None
        self.segments: List[Dict] = []
        self.by_name: Dict[str, Dict] = {}
        self.num_layers = 0
        self.original_tied = False

    # ------------------------------------------------------------ indexing
    def _index_weights(self) -> None:
        index = self.model_dir / "model.safetensors.index.json"
        if index.exists():
            self.weight_map = dict(json.loads(index.read_text(encoding="utf-8"))["weight_map"])
        else:
            single = self.model_dir / "model.safetensors"
            if not single.exists():
                raise FileNotFoundError(f"no model.safetensors / index in {self.model_dir}")
            with safe_open(str(single), framework="pt", device="cpu") as handle:
                self.weight_map = {k: "model.safetensors" for k in handle.keys()}

    def _handle(self, filename: str):
        if filename not in self._handles:
            self._handles[filename] = safe_open(str(self.model_dir / filename), framework="pt", device="cpu")
        return self._handles[filename]

    def _raw_get(self, key: str) -> torch.Tensor:
        if key in self.overlay_map:
            with safe_open(str(self.overlay_map[key]), framework="pt", device="cpu") as handle:
                return handle.get_tensor(key)
        return self._handle(self.weight_map[key]).get_tensor(key)

    def _get(self, key: str) -> torch.Tensor:
        tensor = self._raw_get(key)
        if self.row_recipe and key in (self.EMBED_KEY, self.HEAD_KEY):
            tensor = tensor.clone()
            for new_id, sources in self.row_recipe.items():
                if new_id < tensor.shape[0] and sources:
                    idx = torch.tensor([s for s in sources if s < tensor.shape[0]], dtype=torch.long)
                    if len(idx):
                        tensor[new_id] = tensor[idx].float().mean(0).to(tensor.dtype)
        return tensor

    # ---------------------------------------------------------------- build
    def build(self):
        from accelerate import init_empty_weights
        from transformers import AutoConfig

        self._index_weights()
        config = AutoConfig.from_pretrained(str(self.model_dir))
        self.original_tied = bool(getattr(config, "tie_word_embeddings", False))
        config.tie_word_embeddings = False          # embed and lm_head are streamed independently
        with init_empty_weights():                   # parameters on meta, buffers stay real
            model = _from_config(config, self.dtype)
        for module in model.modules():
            for name, param in list(module._parameters.items()):
                if param is not None:
                    module._parameters[name] = nn.Parameter(torch.empty(0, dtype=self.dtype), requires_grad=False)
        model.eval()
        model.config.use_cache = True

        inner = model.model
        self.num_layers = len(inner.layers)
        self.segments = [{"name": "embed", "module": inner.embed_tokens, "prefix": "model.embed_tokens", "idx": None}]
        for i, layer in enumerate(inner.layers):
            self.segments.append({"name": f"layer{i}", "module": layer, "prefix": f"model.layers.{i}", "idx": i})
        self.segments.append({"name": "norm", "module": inner.norm, "prefix": "model.norm", "idx": None})
        self.segments.append({"name": "head", "module": model.lm_head, "prefix": "lm_head", "idx": None})
        self.by_name = {s["name"]: s for s in self.segments}
        self.model = model

        for seg in self.segments:
            self._install_hooks(seg)
        if not self.evict_ends:
            for name in ("embed", "norm", "head"):
                self._ensure(name)
        if self.verbose:
            print(f"   [ring] built skeleton: {self.num_layers} layers on disk, window={self.window}, "
                  f"evict_ends={self.evict_ends}, dtype={str(self.dtype).replace('torch.', '')}")
        return model

    # ------------------------------------------------------ key resolution
    def _key_for(self, seg: Dict, rel: str) -> Optional[str]:
        rel = rel.replace(".base_layer", "")
        if "lora_" in rel or "modules_to_save" in rel or "original_module" in rel:
            return None
        key = f"{seg['prefix']}.{rel}"
        if key in self.weight_map:
            return key
        if seg["name"] == "head" and rel == "weight" and self.EMBED_KEY in self.weight_map:
            return self.EMBED_KEY                    # tied checkpoint: head shares the embedding matrix
        return None

    def _managed(self, seg: Dict) -> Iterator[Tuple[str, nn.Module, str]]:
        for mod_name, mod in seg["module"].named_modules():
            for pname, param in list(mod._parameters.items()):
                if param is None:
                    continue
                rel = f"{mod_name}.{pname}" if mod_name else pname
                key = self._key_for(seg, rel)
                if key:
                    yield key, mod, pname

    # ----------------------------------------------------------- load/evict
    def _ensure(self, name: str) -> None:
        if name in self.loaded:
            return
        started = time.time()
        for key, mod, pname in self._managed(self.by_name[name]):
            old = mod._parameters[pname]
            tensor = self._get(key).to(old.dtype)
            mod._parameters[pname] = nn.Parameter(tensor, requires_grad=old.requires_grad)
        self.loaded.add(name)
        self.trace.append(("load", name, round(time.time() - started, 3)))

    def _evict(self, name: str) -> None:
        if name not in self.loaded or name in self.pinned:
            return
        for _key, mod, pname in self._managed(self.by_name[name]):
            old = mod._parameters[pname]
            mod._parameters[pname] = nn.Parameter(torch.empty(0, dtype=self.dtype), requires_grad=old.requires_grad)
        self.loaded.discard(name)
        self.trace.append(("evict", name, 0.0))

    # ---------------------------------------------------------------- hooks
    def _install_hooks(self, seg: Dict) -> None:
        name, idx = seg["name"], seg["idx"]
        module = seg["module"]
        if idx is not None:
            def pre(_m, _args, i=idx):
                self._enter_layer(i)

            def post(_m, _args, _out, i=idx):
                if self.window == 1:
                    self._evict(f"layer{i}")
            module.register_forward_pre_hook(pre)
            module.register_forward_hook(post)
        else:
            def pre_end(_m, _args, n=name):
                self._ensure(n)

            def post_end(_m, _args, _out, n=name):
                if self.evict_ends:
                    self._evict(n)
            module.register_forward_pre_hook(pre_end)
            module.register_forward_hook(post_end)

    def _enter_layer(self, i: int) -> None:
        keep = {f"layer{i}"}
        self._ensure(f"layer{i}")
        if self.window >= 2 and self.num_layers > 1:
            nxt = f"layer{(i + 1) % self.num_layers}"          # ring: last layer prefetches layer 0
            self._ensure(nxt)
            keep.add(nxt)
        for name in list(self.loaded):
            if name.startswith("layer") and name not in keep and name not in self.pinned:
                self._evict(name)

    # --------------------------------------------------- pin / overlay / io
    def pin(self, idxs) -> None:
        for i in idxs:
            self._ensure(f"layer{i}")
            self.pinned.add(f"layer{i}")

    def _save_overlay(self, i: int) -> None:
        seg = self.by_name[f"layer{i}"]
        tensors = {}
        for key, mod, pname in self._managed(seg):
            tensors[key] = mod._parameters[pname].detach().to(self.dtype).contiguous().clone()
        self.overlay_dir.mkdir(parents=True, exist_ok=True)
        path = self.overlay_dir / f"layer_{i}_{int(time.time() * 1000)}.safetensors"
        save_file(tensors, str(path), metadata={"format": "pt"})
        for key in tensors:
            old = self.overlay_map.get(key)
            self.overlay_map[key] = path
            if old and old != path and old.exists() and old not in set(self.overlay_map.values()):
                old.unlink()

    def unpin(self, idxs, save: bool = True) -> None:
        for i in idxs:
            name = f"layer{i}"
            if save:
                self._save_overlay(i)
            self.pinned.discard(name)
            for _key, mod, pname in self._managed(self.by_name[name]):
                mod._parameters[pname].requires_grad_(False)
            self._evict(name)
        gc.collect()

    def export(self, dest) -> Path:
        """Write base + overlay weights as per-segment safetensors shards (1 segment in RAM at a time)."""
        dest = Path(dest)
        dest.mkdir(parents=True, exist_ok=True)
        covered = set()
        order = [s["name"] for s in self.segments]
        keys_by_segment: Dict[str, List[str]] = {}
        for name in order:
            seg = self.by_name[name]
            keys = sorted({key for key, _m, _p in self._managed(seg)})
            if name == "head" and self.original_tied:
                keys = []                                     # keep the checkpoint tied: no duplicate matrix
            keys_by_segment[name] = keys
            covered.update(keys)
        leftovers = sorted(k for k in self.weight_map if k not in covered
                           and not (self.original_tied and k == self.HEAD_KEY)
                           and not (k == self.EMBED_KEY))
        total_files = len([n for n in order if keys_by_segment[n]]) or 1
        weight_map_out: Dict[str, str] = {}
        total_size = 0
        shard_no = 0
        for name in order:
            keys = keys_by_segment[name]
            if name == order[-1]:
                keys = keys + leftovers
            if not keys:
                continue
            shard_no += 1
            fname = f"model-{shard_no:05d}-of-{total_files:05d}.safetensors"
            tensors = {k: self._get(k).to(self.dtype).contiguous() for k in keys}
            save_file(tensors, str(dest / fname), metadata={"format": "pt"})
            for k, t in tensors.items():
                weight_map_out[k] = fname
                total_size += t.numel() * t.element_size()
            del tensors
            gc.collect()
        (dest / "model.safetensors.index.json").write_text(
            json.dumps({"metadata": {"total_size": total_size}, "weight_map": weight_map_out}, indent=1), encoding="utf-8")
        for fname in ("config.json", "generation_config.json"):
            if (self.model_dir / fname).exists():
                shutil.copy(self.model_dir / fname, dest / fname)
        if self.verbose:
            print(f"   [ring] exported {shard_no} shards ({total_size / 1e9:.2f} GB) -> {dest}")
        return dest

    def describe(self) -> str:
        layers = sorted(int(n[5:]) for n in self.loaded if n.startswith("layer"))
        try:
            import psutil
            rss = psutil.Process().memory_info().rss / 1024 ** 3
            mem = f"{rss:.2f} GiB RSS"
        except Exception:
            mem = "n/a"
        return f"ring: resident layers={layers} pinned={sorted(self.pinned)} | {mem}"