monet_code/src/online_teacher.py

"""In-process teacher wrappers that produce the same tensors as offline precompute.

`OnlineTeacherStage2` mirrors `src/precompute_teacher_reps.py` (no-text branch):
  forward base model → pool aux-image hidden states → [L+1, N*latent_size, D] per sample.

`OnlineTeacherStage3` mirrors `src/precompute_teacher_latents.py`:
  forward Stage-2 ckpt with latent_mode=True → [L+1, N*latent_size, D] per sample.

Both write each computed tensor to disk under the same filename layout used by the
offline scripts, so subsequent runs (or pure-offline Stage-2/3 runs) can reuse them
via `load_offline_tensor`. The training loss path is unchanged: the trainer hands the
returned list[CPU Tensor] straight to `inputs['teacher_hidden_states_for_alignment']`.
"""

import hashlib
import logging
import os

import torch
import torch.nn.functional as F
from transformers import Qwen3VLConfig

from monet_qwen3_model.modeling_qwen3_vl_monet import Qwen3VLMonetForConditionalGeneration


def _compute_teacher_fingerprint(model_path: str) -> str:
    """Cheap stable fingerprint of a teacher checkpoint.

    Used to namespace cache files so that re-training Stage 2 (and pointing the
    Stage 3 online teacher at the new ckpt) cannot silently consume stale latents
    from a previous Stage 2 run. Captures:
      - absolute path (catches different ckpt directories)
      - mtime + size of the key weight/config files (catches in-place re-saves)
    Does not require reading any weights — runs in O(stat).
    """
    abs_path = os.path.abspath(model_path)
    parts = [abs_path]
    for name in (
        "config.json",
        "model.safetensors.index.json",
        "model.safetensors",
        "pytorch_model.bin.index.json",
        "pytorch_model.bin",
    ):
        p = os.path.join(abs_path, name)
        if os.path.isfile(p):
            try:
                st = os.stat(p)
                parts.append(f"{name}:{int(st.st_mtime)}:{st.st_size}")
            except Exception:
                pass
    raw = "|".join(parts).encode("utf-8")
    return hashlib.sha1(raw).hexdigest()[:12]


def _per_sample_hidden_states(hidden_states, b, batch_size):
    """Robust per-sample extraction matching `_get_sample_hidden_states` in precompute."""
    if not hidden_states:
        raise RuntimeError("hidden_states is empty; model must be called with output_hidden_states=True")
    first = hidden_states[0]
    if torch.is_tensor(first) and first.dim() == 3 and len(hidden_states) == batch_size:
        return hidden_states[b]
    if torch.is_tensor(first) and first.dim() == 3 and first.size(0) == batch_size:
        return torch.stack([layer[b] for layer in hidden_states], dim=0)
    raise RuntimeError(f"Unsupported hidden_states layout: type={type(first)}")


class _OnlineTeacherBase:
    rep_prefix = "rep"

    def __init__(
        self,
        model_path,
        tokenizer_len,
        special_token_ids,
        device,
        dtype=torch.bfloat16,
        cache_dir=None,
        answer_start_pattern=None,
        alignment_layer="all_layers",
    ):
        if alignment_layer not in ("all_layers", "last_layer"):
            raise ValueError(
                f"alignment_layer must be 'all_layers' or 'last_layer', got {alignment_layer!r}"
            )
        self.alignment_layer = alignment_layer
        config = Qwen3VLConfig.from_pretrained(model_path)
        try:
            setattr(config, "use_cache", False)
        except Exception:
            pass
        model = Qwen3VLMonetForConditionalGeneration.from_pretrained(
            model_path, config=config, dtype=dtype, attn_implementation="sdpa",
        )

        try:
            model.resize_token_embeddings(tokenizer_len)
            model.config.vocab_size = tokenizer_len
        except Exception as e:
            logging.warning(f"[online_teacher] resize_token_embeddings failed: {e}")

        model.config.latent_token_id = int(special_token_ids["abs_pad"])
        model.config.latent_start_id = int(special_token_ids["abs_start"])
        model.config.latent_end_id = int(special_token_ids["abs_end"])
        if answer_start_pattern is not None:
            try:
                model.config.answer_start_pattern = (
                    answer_start_pattern.tolist()
                    if hasattr(answer_start_pattern, "tolist")
                    else list(answer_start_pattern)
                )
            except Exception:
                pass

        for p in model.parameters():
            p.requires_grad = False
        try:
            model.gradient_checkpointing_disable()
        except Exception:
            pass
        model.eval()
        model.to(device)

        self.model = model
        self.device = device
        self.special_token_ids = special_token_ids

        # Per-teacher fingerprint: namespaces the cache so different ckpts can't
        # silently share files even when the user passes the same --teacher_*_dir.
        self.teacher_fingerprint = _compute_teacher_fingerprint(model_path)
        self.cache_root = cache_dir
        if cache_dir:
            self.cache_dir = os.path.join(cache_dir, self.teacher_fingerprint)
            os.makedirs(self.cache_dir, exist_ok=True)
            # Drop a sidecar so the subdir is self-describing (which teacher made these files).
            try:
                manifest = os.path.join(self.cache_dir, "TEACHER_INFO.txt")
                if not os.path.isfile(manifest):
                    with open(manifest, "w") as f:
                        f.write(
                            f"fingerprint: {self.teacher_fingerprint}\n"
                            f"model_path:  {os.path.abspath(model_path)}\n"
                            f"class:       {self.__class__.__name__}\n"
                        )
            except Exception as e:
                logging.warning(f"[online_teacher] failed to write TEACHER_INFO.txt: {e}")
        else:
            self.cache_dir = None
        logging.info(
            f"[online_teacher] loaded {self.__class__.__name__} from {model_path} "
            f"on {device}; fingerprint={self.teacher_fingerprint}; "
            f"cache_dir={self.cache_dir or 'none'}"
        )

    def _cache_path(self, metadata):
        # Filename matches load_offline_tensor's expectation:
        #   f"{rep_type}_{alignment_layer}_{dataset_name}_{sample_id}.pt"
        info = f"{self.alignment_layer}_{metadata['dataset_name']}_{metadata['sample_id']}"
        if not self.cache_dir:
            return None, info
        return os.path.join(self.cache_dir, f"{self.rep_prefix}_{info}.pt"), info

    def _is_valid_teacher_tensor(self, tensor, metadata, source):
        if not torch.is_tensor(tensor):
            logging.warning(
                f"[online_teacher] invalid {source} teacher tensor for "
                f"{metadata.get('dataset_name')}/{metadata.get('sample_id')}: type={type(tensor)}"
            )
            return False
        if not torch.isfinite(tensor).all().item():
            logging.warning(
                f"[online_teacher] non-finite {source} teacher tensor for "
                f"{metadata.get('dataset_name')}/{metadata.get('sample_id')}; recomputing or failing fast"
            )
            return False
        return True

    def _try_load_cache(self, metadata):
        path, _ = self._cache_path(metadata)
        if not path or not os.path.isfile(path):
            return None
        try:
            data = torch.load(path, map_location="cpu")
        except Exception as e:
            logging.warning(f"[online_teacher] cache load failed for {path}: {e}")
            return None
        # Defense in depth: even with the fingerprinted subdir, refuse to use a file
        # whose embedded fingerprint disagrees (e.g., manually copied across teachers).
        cached_fp = data.get("teacher_fingerprint") if isinstance(data, dict) else None
        if cached_fp is not None and cached_fp != self.teacher_fingerprint:
            logging.warning(
                f"[online_teacher] cache fingerprint mismatch at {path} "
                f"(cached={cached_fp}, current={self.teacher_fingerprint}); recomputing"
            )
            return None
        if not isinstance(data, dict) or "latent" not in data:
            logging.warning(f"[online_teacher] malformed cache at {path}; recomputing")
            return None
        latent = data["latent"]
        if not self._is_valid_teacher_tensor(latent, metadata, f"cached file {path}"):
            return None
        return latent

    def _write_cache(self, metadata, tensor):
        path, info = self._cache_path(metadata)
        if not path:
            return
        if not self._is_valid_teacher_tensor(tensor, metadata, "generated"):
            logging.warning(f"[online_teacher] skip writing invalid cache for {info}")
            return
        tmp = f"{path}.tmp.{os.getpid()}"
        try:
            torch.save(
                {
                    "metadata_info": info,
                    "latent": tensor.detach().cpu(),
                    "teacher_fingerprint": self.teacher_fingerprint,
                },
                tmp,
            )
            os.replace(tmp, path)
        except Exception as e:
            logging.warning(f"[online_teacher] cache write failed for {path}: {e}")
            try:
                if os.path.isfile(tmp):
                    os.remove(tmp)
            except Exception:
                pass


class OnlineTeacherStage2(_OnlineTeacherBase):
    """Replaces offline Step 1: pool aux-image hidden states from a frozen base model."""

    rep_prefix = "rep"

    def __init__(self, *args, latent_size=8, **kwargs):
        super().__init__(*args, **kwargs)
        self.latent_size = int(latent_size)

    @torch.inference_mode()
    def __call__(self, inputs):
        """Returns list[CPU Tensor [L+1, N*latent_size, D]], matching load_offline_tensor."""
        B = inputs["teacher_input_ids"].size(0)
        results = [None] * B
        misses = []
        for b in range(B):
            cached = self._try_load_cache(inputs["metadata"][b])
            if cached is not None:
                results[b] = cached
            else:
                misses.append(b)

        if not misses:
            return results

        aux_blocks = inputs["teacher_aux_image_blocks"]

        fwd_inputs = {
            "input_ids": inputs["teacher_input_ids"].to(self.device, non_blocking=True),
            "attention_mask": inputs["teacher_attention_mask"].to(self.device, non_blocking=True),
            "pixel_values": inputs["teacher_pixel_values"].to(self.device, non_blocking=True),
            "image_grid_thw": inputs["teacher_image_grid_thw"].to(self.device, non_blocking=True),
            "output_hidden_states": True,
            "return_dict": True,
            "latent_mode": False,
            "labels": None,
            "loss_type": [],
            "alignment_poss": [[] for _ in range(B)],
        }
        outputs = self.model(**fwd_inputs)
        hidden_states = outputs.hidden_states

        for b in misses:
            blocks = aux_blocks[b]
            hs_sample = _per_sample_hidden_states(hidden_states, b, B)  # [L+1, T, D]
            # last_layer: pool only the last hidden state → output is 2D [N*latent_size, D].
            # all_layers: pool every layer → 3D [L+1, N*latent_size, D].
            if self.alignment_layer == "last_layer":
                hs_for_pool = hs_sample[-1:, :, :]  # [1, T, D]
            else:
                hs_for_pool = hs_sample  # [L+1, T, D]
            pooled_per_aux = []
            for idx in blocks or []:
                if idx.numel() == 0:
                    continue
                idx_dev = idx.to(hs_for_pool.device)
                h_aux = hs_for_pool[:, idx_dev, :].permute(0, 2, 1)  # [layers, D, T_aux]
                h_aux = F.adaptive_avg_pool1d(h_aux, self.latent_size)
                pooled_per_aux.append(h_aux.permute(0, 2, 1))        # [layers, latent_size, D]
            if pooled_per_aux:
                pooled = torch.cat(pooled_per_aux, dim=1)            # [layers, N*latent_size, D]
            else:
                pooled = torch.zeros(hs_for_pool.size(0), 0, hs_for_pool.size(-1),
                                     device=hs_for_pool.device)
            if self.alignment_layer == "last_layer":
                pooled = pooled.squeeze(0)                            # [N*latent_size, D]
            pooled_cpu = pooled.detach().cpu()
            if not self._is_valid_teacher_tensor(pooled_cpu, inputs["metadata"][b], "generated"):
                raise RuntimeError(f"[online_teacher] generated non-finite teacher reps for sample {b}")
            results[b] = pooled_cpu
            self._write_cache(inputs["metadata"][b], pooled_cpu)

        return results


class OnlineTeacherStage3(_OnlineTeacherBase):
    """Replaces offline Step 3: dump latent_mode-generated reps from a frozen Stage-2 ckpt."""

    rep_prefix = "latent"

    @torch.inference_mode()
    def __call__(self, inputs):
        """Returns list[CPU Tensor [L+1, N*latent_size, D]], matching load_offline_tensor."""
        B = inputs["teacher_input_ids"].size(0)
        results = [None] * B
        misses = []
        for b in range(B):
            cached = self._try_load_cache(inputs["metadata"][b])
            if cached is not None:
                results[b] = cached
            else:
                misses.append(b)

        if not misses:
            return results

        fwd_inputs = {
            "input_ids": inputs["teacher_input_ids"].to(self.device, non_blocking=True),
            "attention_mask": inputs["teacher_attention_mask"].to(self.device, non_blocking=True),
            "pixel_values": inputs["teacher_pixel_values"].to(self.device, non_blocking=True),
            "image_grid_thw": inputs["teacher_image_grid_thw"].to(self.device, non_blocking=True),
            "return_dict": True,
            "latent_mode": True,
            "labels": None,
            "loss_type": [],
        }
        if "teacher_attention_mask_4d" in inputs and inputs["teacher_attention_mask_4d"] is not None:
            fwd_inputs["attention_mask_4d"] = inputs["teacher_attention_mask_4d"]

        if self.alignment_layer == "last_layer":
            # Mirror precompute_teacher_latents.py with --output_latent_embeds:
            # the model returns outputs.latent_embeds[b] of shape [N*latent_size, D] (2D).
            # It needs alignment_poss = latent-pad positions in each teacher sample.
            abs_pad_id = int(self.special_token_ids["abs_pad"])
            teacher_input_ids_cpu = inputs["teacher_input_ids"]
            teacher_alignment_poss = [
                (teacher_input_ids_cpu[b] == abs_pad_id)
                .nonzero(as_tuple=False)
                .flatten()
                .tolist()
                for b in range(B)
            ]
            fwd_inputs["alignment_poss"] = teacher_alignment_poss
            fwd_inputs["output_latent_embeds"] = True
        else:
            fwd_inputs["output_hidden_states"] = True

        outputs = self.model(**fwd_inputs)
        if self.alignment_layer == "last_layer":
            teacher_reps = outputs.latent_embeds      # list[Tensor [N*latent_size, D]]
        else:
            teacher_reps = outputs.hidden_states      # list[Tensor [L+1, N*latent_size, D]]

        for b in misses:
            latent_b = teacher_reps[b].detach().cpu()
            if not self._is_valid_teacher_tensor(latent_b, inputs["metadata"][b], "generated"):
                raise RuntimeError(f"[online_teacher] generated non-finite teacher latents for sample {b}")
            results[b] = latent_b
            self._write_cache(inputs["metadata"][b], latent_b)

        return results