lhallee's picture
Update FastPLMs runtime and model cards
443b6bc verified
Raw
History Blame Contribute Delete
37.8 kB
from __future__ import annotations
import contextlib
import math
import numbers
import torch
import torch.nn as nn
import torch.nn.functional as F
from collections.abc import Iterator, Mapping
from dataclasses import asdict, dataclass, fields
from typing import Any
_STANDARD_AMINO_ACIDS = "ACDEFGHIKLMNPQRSTVWY"
_TTT_SERIALIZATION_VERSION = 1
@dataclass
class TTTConfig:
lr: float = 4e-4
steps: int = 30
ags: int = 16
batch_size: int = 2
mask_ratio: float = 0.15
crop_size: int = 1024
bert_leave_prob: float = 0.1
bert_replace_prob: float = 0.1
optimizer: str = "sgd"
momentum: float = 0.0
weight_decay: float = 0.0
seed: int | None = 0
lora_rank: int = 8
lora_alpha: float = 32.0
lora_target_replace_module: str | None = None
lora_target_modules: tuple[str, ...] | None = None
initial_state_reset: bool = True
automatic_best_state_reset: bool = False
eval_each_step: bool = False
gradient_clip: bool = False
gradient_clip_max_norm: float = 1.0
def __post_init__(self) -> None:
self.verify()
@classmethod
def from_kwargs(cls, **kwargs: Any) -> TTTConfig:
valid_names = {field.name for field in fields(cls)}
unknown_names = set(kwargs) - valid_names
if unknown_names:
raise ValueError(f"Unknown TTTConfig fields: {sorted(unknown_names)}")
# JSON has no tuple type. Normalize the serialized representation while
# keeping the public constructor and runtime overrides type-strict.
if isinstance(kwargs.get("lora_target_modules"), list):
kwargs["lora_target_modules"] = tuple(kwargs["lora_target_modules"])
return cls(**kwargs)
def merged(self, overrides: Mapping[str, Any] | TTTConfig | None) -> TTTConfig:
if overrides is None:
return self
if isinstance(overrides, TTTConfig):
return overrides
values = {field.name: self.__dict__[field.name] for field in fields(self)}
for name, value in overrides.items():
if name not in values:
raise ValueError(f"Unknown TTTConfig field: {name}")
values[name] = value
return TTTConfig(**values)
def to_dict(self) -> dict[str, Any]:
return asdict(self)
def verify(self) -> None:
numeric_fields = {
"lr": self.lr,
"mask_ratio": self.mask_ratio,
"lora_alpha": self.lora_alpha,
"bert_leave_prob": self.bert_leave_prob,
"bert_replace_prob": self.bert_replace_prob,
"gradient_clip_max_norm": self.gradient_clip_max_norm,
"momentum": self.momentum,
"weight_decay": self.weight_decay,
}
for name, value in numeric_fields.items():
if isinstance(value, bool) or not isinstance(value, numbers.Real):
raise TypeError(f"TTT {name} must be a real number.")
if not math.isfinite(float(value)):
raise ValueError(f"TTT {name} must be finite.")
integer_fields = {
"steps": self.steps,
"ags": self.ags,
"batch_size": self.batch_size,
"crop_size": self.crop_size,
"lora_rank": self.lora_rank,
}
for name, value in integer_fields.items():
if isinstance(value, bool) or not isinstance(value, int):
raise TypeError(f"TTT {name} must be an integer.")
if self.seed is not None and (
isinstance(self.seed, bool) or not isinstance(self.seed, int)
):
raise TypeError("TTT seed must be None or an integer.")
boolean_fields = {
"initial_state_reset": self.initial_state_reset,
"automatic_best_state_reset": self.automatic_best_state_reset,
"eval_each_step": self.eval_each_step,
"gradient_clip": self.gradient_clip,
}
for name, value in boolean_fields.items():
if type(value) is not bool:
raise TypeError(f"TTT {name} must be a boolean.")
if self.lr <= 0.0:
raise ValueError("TTT learning rate must be positive.")
if self.steps < 1:
raise ValueError("TTT steps must be >= 1.")
if self.ags < 1:
raise ValueError("TTT gradient accumulation steps must be >= 1.")
if self.batch_size < 1:
raise ValueError("TTT batch_size must be >= 1.")
if not 0.0 < self.mask_ratio <= 1.0:
raise ValueError("TTT mask_ratio must be in (0, 1].")
if self.crop_size < 1:
raise ValueError("TTT crop_size must be >= 1.")
if self.lora_rank < 1:
raise ValueError("TTT v1 is LoRA-only, so lora_rank must be >= 1.")
if self.lora_alpha <= 0.0:
raise ValueError("TTT lora_alpha must be positive.")
if not isinstance(self.optimizer, str):
raise TypeError("TTT optimizer must be a string.")
if self.optimizer not in {"adamw", "sgd"}:
raise ValueError("TTT optimizer must be 'adamw' or 'sgd'.")
if self.momentum < 0.0:
raise ValueError("TTT momentum must be non-negative.")
if self.weight_decay < 0.0:
raise ValueError("TTT weight_decay must be non-negative.")
if not 0.0 <= self.bert_leave_prob <= 1.0:
raise ValueError("bert_leave_prob must be in [0, 1].")
if not 0.0 <= self.bert_replace_prob <= 1.0:
raise ValueError("bert_replace_prob must be in [0, 1].")
if self.bert_leave_prob + self.bert_replace_prob > 1.0:
raise ValueError("bert_leave_prob + bert_replace_prob must be <= 1.")
if self.gradient_clip and self.gradient_clip_max_norm <= 0.0:
raise ValueError("gradient_clip_max_norm must be positive.")
if self.lora_target_replace_module is not None:
if not isinstance(self.lora_target_replace_module, str):
raise TypeError("lora_target_replace_module must be None or a string.")
if not self.lora_target_replace_module.strip():
raise ValueError("lora_target_replace_module must not be empty.")
if self.lora_target_modules is not None:
if not isinstance(self.lora_target_modules, tuple):
raise TypeError("lora_target_modules must be None or a tuple of strings.")
if not self.lora_target_modules:
raise ValueError("lora_target_modules must not be empty.")
if any(not isinstance(name, str) for name in self.lora_target_modules):
raise TypeError("lora_target_modules must contain only strings.")
if any(not name.strip() for name in self.lora_target_modules):
raise ValueError(
"lora_target_modules must contain only non-empty strings."
)
if len(set(self.lora_target_modules)) != len(self.lora_target_modules):
raise ValueError("lora_target_modules must not contain duplicates.")
class LoraInjectedLinear(nn.Module):
"""ProteinTTT-compatible low-rank adapter.
``alpha`` is the direct adapter-output multiplier used by the pinned
ProteinTTT ``inject_trainable_lora(..., scale=lora_alpha)`` contract. It
is intentionally not divided by ``rank`` as it would be in the common
PEFT LoRA convention.
"""
def __init__(
self,
linear: nn.Module,
rank: int,
alpha: float,
generator: torch.Generator | None = None,
) -> None:
super().__init__()
weight = linear._parameters.get("weight")
if not isinstance(weight, torch.Tensor):
raise TypeError("LoRA targets must expose a tensor weight parameter.")
if weight.ndim != 2:
raise ValueError("LoRA can only wrap 2D linear weights.")
self.linear = linear
self.linear.requires_grad_(False)
self.rank = rank
# ProteinTTT names this setting ``lora_alpha`` but passes it directly
# to cloneofsimo/lora's ``scale`` argument. Preserve that numerical
# contract for parity and for saved FastPLMs TTT configurations.
self.scale = alpha
in_features = weight.shape[1]
out_features = weight.shape[0]
# ``nn.Linear`` initializes from the process-global CPU generator. Preserve
# that state when TTT supplies its own generator so lazy adapter injection
# is reproducible without perturbing the caller's RNG stream.
with torch.random.fork_rng(devices=[], enabled=generator is not None):
self.lora_down = nn.Linear(in_features, rank, bias=False, dtype=torch.float32)
self.lora_up = nn.Linear(rank, out_features, bias=False, dtype=torch.float32)
nn.init.normal_(self.lora_down.weight, std=1.0 / rank, generator=generator)
nn.init.zeros_(self.lora_up.weight)
self.lora_down.to(device=weight.device)
self.lora_up.to(device=weight.device)
self.register_buffer(
"_ttt_initial_lora_down",
self.lora_down.weight.detach().clone(),
persistent=True,
)
self.register_buffer(
"_ttt_initial_lora_up",
self.lora_up.weight.detach().clone(),
persistent=True,
)
@property
def weight(self) -> torch.Tensor:
return self.linear._parameters["weight"]
@property
def bias(self) -> torch.Tensor | None:
return self.linear._parameters["bias"]
def forward(self, x: torch.Tensor) -> torch.Tensor:
# x: (..., d_in)
base = self.linear(x) # (..., d_out)
delta = ( # (..., d_out)
self.lora_up(self.lora_down(x.to(dtype=torch.float32))) * self.scale
)
return base + delta.to(dtype=base.dtype) # (..., d_out)
def reset_lora_parameters(self) -> None:
with torch.no_grad():
self.lora_down.weight.copy_(self._ttt_initial_lora_down)
self.lora_up.weight.copy_(self._ttt_initial_lora_up)
class FastPLMTestTimeTrainingMixin:
def init_ttt(self, ttt_config: TTTConfig | Mapping[str, Any] | None = None) -> None:
base_config = self.__dict__.get("_ttt_cfg")
if base_config is None:
base_config = TTTConfig()
if not isinstance(base_config, TTTConfig):
raise TypeError("Existing TTT configuration must be a TTTConfig instance.")
configured = base_config.merged(ttt_config)
serialized = getattr(getattr(self, "config", None), "fastplms_ttt", None)
serialized_initialized = False
if serialized is not None:
if not isinstance(serialized, Mapping):
raise ValueError("config.fastplms_ttt must be a mapping.")
version = serialized.get("version")
if version != _TTT_SERIALIZATION_VERSION:
raise ValueError(
"Unsupported FastPLMs TTT serialization version "
f"{version!r}; expected {_TTT_SERIALIZATION_VERSION}."
)
serialized_config = serialized.get("config")
if not isinstance(serialized_config, Mapping):
raise ValueError("Serialized FastPLMs TTT state is missing its config mapping.")
configured = TTTConfig.from_kwargs(**dict(serialized_config))
initialized_value = serialized.get("initialized", False)
if type(initialized_value) is not bool:
raise ValueError("Serialized FastPLMs TTT initialized flag must be a boolean.")
serialized_initialized = initialized_value
self._ttt_cfg = configured
self._ttt_cfg.verify()
self._ttt_initialized = False
if serialized_initialized:
self._ttt_inject_lora()
self._ttt_initialized = True
@property
def ttt_config(self) -> TTTConfig:
if "_ttt_cfg" not in self.__dict__:
self.init_ttt()
return self._ttt_cfg
def _ttt_get_trainable_modules(self) -> list[nn.Module]:
return [self]
def _ttt_get_frozen_modules(self) -> list[nn.Module]:
return []
def _ttt_tokenize(
self,
seq: str | list[str] | None = None,
input_ids: torch.Tensor | None = None,
**kwargs: Any,
) -> torch.Tensor | dict[str, torch.Tensor]:
del kwargs
if input_ids is not None:
return input_ids # (b, l)
if seq is None:
raise ValueError("Pass either seq or input_ids for TTT.")
tokenized = self.tokenizer(seq, return_tensors="pt", padding=True)
return tokenized["input_ids"] # (b, l)
def _ttt_mask_token(self) -> int:
return int(self.tokenizer.mask_token_id)
def _ttt_padding_token(self) -> int:
return int(self.tokenizer.pad_token_id)
def _ttt_replacement_tokens(self, input_ids: torch.Tensor) -> torch.Tensor:
# input_ids: (b, l)
tokenizer = self.tokenizer
special_ids = set(tokenizer.all_special_ids)
vocab_size = int(self.config.vocab_size)
unknown_id = getattr(tokenizer, "unk_token_id", None)
if unknown_id is not None:
special_ids.add(int(unknown_id))
vocab: Mapping[str, Any] = {}
get_vocab = getattr(tokenizer, "get_vocab", None)
if callable(get_vocab):
vocab = get_vocab()
elif isinstance(getattr(tokenizer, "vocab", None), Mapping):
vocab = tokenizer.vocab
elif isinstance(getattr(tokenizer, "_token_to_id", None), Mapping):
vocab = tokenizer._token_to_id
ids: list[int] = []
convert = getattr(tokenizer, "convert_tokens_to_ids", None)
for amino_acid in _STANDARD_AMINO_ACIDS:
token_id = convert(amino_acid) if callable(convert) else vocab.get(amino_acid)
if (
isinstance(token_id, int)
and 0 <= token_id < vocab_size
and token_id not in special_ids
and token_id not in ids
):
ids.append(token_id)
if not ids:
raise ValueError(
"TTT could not resolve any canonical amino-acid token IDs from the tokenizer; "
"refusing to sample arbitrary or reserved vocabulary entries."
)
return torch.tensor(ids, device=input_ids.device, dtype=input_ids.dtype) # (c_aa,)
def _ttt_predict_logits(
self,
batch: torch.Tensor | dict[str, torch.Tensor],
**kwargs: Any,
) -> torch.Tensor:
del kwargs
if isinstance(batch, dict):
output = self(**batch)
return output.logits # (b, l, c)
attention_mask = batch.ne(self._ttt_padding_token()) # (b, l)
output = self(input_ids=batch, attention_mask=attention_mask)
return output.logits # (b, l, c)
def _ttt_eval_step(
self,
step: int,
loss: float,
seq: str | list[str] | None = None,
input_ids: torch.Tensor | None = None,
**kwargs: Any,
) -> tuple[dict[str, Any], float | None]:
del step, loss, seq, input_ids, kwargs
return {}, None
def _ttt_is_lora_target(
self,
name: str,
full_name: str,
module: nn.Module,
active: bool,
target_modules: tuple[str, ...] | None,
) -> bool:
if not active:
return False
if isinstance(module, LoraInjectedLinear):
return False
if (
target_modules is not None
and name not in target_modules
and full_name not in target_modules
):
return False
if isinstance(module, nn.Linear):
return True
if "weight" not in module._parameters:
return False
weight = module._parameters["weight"]
if weight is None or weight.ndim != 2:
return False
return "Linear" in module.__class__.__name__
def _ttt_inject_lora(self) -> int:
cfg = self.ttt_config
cfg.verify()
target_class = cfg.lora_target_replace_module
target_modules = cfg.lora_target_modules
wrapped = 0
generator = None
if cfg.seed is not None:
generator = torch.Generator(device="cpu")
generator.manual_seed(cfg.seed)
def inject(module: nn.Module, prefix: str, active: bool) -> None:
nonlocal wrapped
for name, child in list(module.named_children()):
full_name = f"{prefix}.{name}" if prefix else name
child_active = active
if target_class is not None:
child_active = active or child.__class__.__name__ == target_class
if self._ttt_is_lora_target(name, full_name, child, child_active, target_modules):
setattr(
module,
name,
LoraInjectedLinear(
child,
rank=cfg.lora_rank,
alpha=cfg.lora_alpha,
generator=generator,
),
)
wrapped += 1
continue
inject(child, full_name, child_active)
for trainable_module in self._ttt_get_trainable_modules():
inject(trainable_module, "", target_class is None)
if wrapped == 0:
raise ValueError("TTT LoRA injection did not find any target modules.")
return wrapped
def _ttt_lora_modules(self) -> list[LoraInjectedLinear]:
return [module for module in self.modules() if isinstance(module, LoraInjectedLinear)]
def _ttt_lora_parameters(self) -> list[nn.Parameter]:
params: list[nn.Parameter] = []
for module in self._ttt_lora_modules():
params.extend(module.lora_down.parameters())
params.extend(module.lora_up.parameters())
if not params:
raise RuntimeError("TTT has no LoRA parameters.")
return params
def _ttt_snapshot_lora_state(self) -> list[dict[str, torch.Tensor]]:
snapshot = []
for module in self._ttt_lora_modules():
snapshot.append(
{
"lora_down.weight": module.lora_down.weight.detach().clone(), # (r, d_in)
"lora_up.weight": module.lora_up.weight.detach().clone(), # (d_out, r)
}
)
if not snapshot:
raise RuntimeError("TTT has no LoRA state to snapshot.")
return snapshot
def _ttt_restore_lora_state(self, state: list[dict[str, torch.Tensor]]) -> None:
modules = self._ttt_lora_modules()
if len(modules) != len(state):
raise RuntimeError("TTT LoRA state/module count mismatch.")
with torch.no_grad():
for module, module_state in zip(modules, state, strict=True):
module.lora_down.weight.copy_(module_state["lora_down.weight"])
module.lora_up.weight.copy_(module_state["lora_up.weight"])
def _ttt_ensure_initialized(self) -> None:
if "_ttt_cfg" not in self.__dict__:
self.init_ttt()
if self._ttt_initialized:
return
self._ttt_inject_lora()
self._ttt_initialized = True
def ttt_reset(self) -> None:
self._ttt_ensure_initialized()
for module in self._ttt_lora_modules():
module.reset_lora_parameters()
def _ttt_serialized_contract(self) -> dict[str, Any]:
return {
"version": _TTT_SERIALIZATION_VERSION,
"initialized": bool(self._ttt_initialized),
"config": self.ttt_config.to_dict(),
}
def save_pretrained(self, save_directory: Any, *args: Any, **kwargs: Any) -> Any:
"""Save initialized adapters, their reset baseline, and the TTT config.
Adapter injection changes the module tree, so the serialized config must
reconstruct that tree before Transformers loads the state dict. Models
whose own state-dict hooks omit their trainable TTT modules fail closed
instead of producing an artifact that cannot restore the adaptation.
"""
if self._ttt_initialized:
state_keys = set(self.state_dict())
missing_adapter_keys = [
name
for name, _ in self.named_parameters()
if ".lora_" in name and name not in state_keys
]
if missing_adapter_keys:
raise RuntimeError(
"This model attaches TTT adapters to transient modules that its "
"checkpoint excludes, so save_pretrained cannot persist the adapted "
"state safely. Reset the model or use a model-specific adapter export."
)
self.config.fastplms_ttt = self._ttt_serialized_contract()
return super().save_pretrained(save_directory, *args, **kwargs)
def _ttt_make_optimizer(self) -> torch.optim.Optimizer:
cfg = self.ttt_config
params = self._ttt_lora_parameters()
if cfg.optimizer == "sgd":
return torch.optim.SGD(
params,
lr=cfg.lr,
momentum=cfg.momentum,
weight_decay=cfg.weight_decay,
)
return torch.optim.AdamW(params, lr=cfg.lr, weight_decay=cfg.weight_decay)
def _ttt_to_device(
self,
batch: torch.Tensor | dict[str, torch.Tensor],
device: torch.device,
) -> torch.Tensor | dict[str, torch.Tensor]:
if isinstance(batch, dict):
return {name: tensor.to(device) for name, tensor in batch.items()} # unchanged shapes
return batch.to(device) # unchanged shape
def _ttt_input_ids_from_batch(
self,
batch: torch.Tensor | dict[str, torch.Tensor],
) -> torch.Tensor:
if isinstance(batch, dict):
return batch["input_ids"] # (b, l)
return batch # (b, l)
def _ttt_set_input_ids(
self,
batch: torch.Tensor | dict[str, torch.Tensor],
input_ids: torch.Tensor,
) -> torch.Tensor | dict[str, torch.Tensor]:
if isinstance(batch, dict):
updated = dict(batch)
updated["input_ids"] = input_ids # (b, l)
return updated
return input_ids # (b, l)
def _ttt_non_special_mask(self, input_ids: torch.Tensor) -> torch.Tensor:
# input_ids: (b, l)
residue_ids = self._ttt_replacement_tokens(input_ids) # (c_aa,)
return torch.isin(input_ids, residue_ids) # (b, l)
def _ttt_validate_tokenized_batch(
self,
batch: torch.Tensor | dict[str, torch.Tensor],
) -> None:
input_ids = self._ttt_input_ids_from_batch(batch)
if input_ids.ndim != 2 or input_ids.shape[0] == 0 or input_ids.shape[1] == 0:
raise ValueError(
"TTT input_ids must have non-empty shape (batch, sequence); got "
f"{tuple(input_ids.shape)}."
)
if str(getattr(self.config, "model_type", "")) == "dplm2":
tokenizer = self.tokenizer
token_to_id = getattr(tokenizer, "_token_to_id", {})
struct_cls_token = getattr(tokenizer, "struct_cls_token", None)
struct_boundary = token_to_id.get(struct_cls_token)
if struct_boundary is None:
raise ValueError(
"DPLM2 TTT could not resolve the structure-token boundary safely."
)
pad_token = self._ttt_padding_token()
generic_aa_special_ids = torch.tensor( # (4,)
[int(self.config.vocab_size) + offset for offset in range(4)],
device=input_ids.device,
dtype=input_ids.dtype,
)
is_structure = input_ids.ge(int(struct_boundary)) & input_ids.ne( # (b, l)
pad_token
)
is_structure &= ~torch.isin(input_ids, generic_aa_special_ids)
if bool(is_structure.any()):
raise ValueError(
"DPLM2 TTT currently supports amino-acid-only inputs. Packed or "
"structure-token inputs require a modality-specific corruption objective."
)
if isinstance(batch, dict) and "type_ids" in batch:
type_ids = batch["type_ids"] # (b, l)
attention_mask = batch.get( # (b, l)
"attention_mask",
input_ids.ne(pad_token),
).bool()
if bool(((type_ids == int(self.config.struct_type)) & attention_mask).any()):
raise ValueError(
"DPLM2 TTT currently supports amino-acid-only inputs; structure "
"type_ids are not accepted."
)
if not bool(self._ttt_non_special_mask(input_ids).any()):
raise ValueError(
"TTT input contains no trainable biological residue tokens after excluding "
"padding, boundary, mask, and reserved tokens."
)
def _ttt_sample_crop(
self,
batch: torch.Tensor | dict[str, torch.Tensor],
generator: torch.Generator,
) -> torch.Tensor | dict[str, torch.Tensor]:
input_ids = self._ttt_input_ids_from_batch(batch)
cfg = self.ttt_config
if input_ids.shape[1] <= cfg.crop_size:
return batch
position_has_residue = ( # (l,)
self._ttt_non_special_mask(input_ids).any(dim=0).to(torch.int64)
)
prefix = F.pad(position_has_residue.cumsum(dim=0), (1, 0)) # (l + 1,)
window_counts = prefix[cfg.crop_size :] - prefix[: -cfg.crop_size] # (l-crop+1,)
valid_starts = torch.where(window_counts > 0)[0] # (n_valid,)
if valid_starts.numel() == 0:
raise ValueError("TTT could not find a crop containing a biological residue token.")
selected = torch.randint( # (1,)
valid_starts.numel(),
(1,),
generator=generator,
device=input_ids.device,
)
start = int(valid_starts[selected].item())
end = start + cfg.crop_size
if isinstance(batch, dict):
cropped = {}
for name, tensor in batch.items():
if tensor.ndim >= 2 and tensor.shape[1] == input_ids.shape[1]:
cropped[name] = tensor[:, start:end] # (b, crop_size, ...)
else:
cropped[name] = tensor
return cropped
return input_ids[:, start:end] # (b, crop_size)
def _ttt_sample_batch(
self,
tokenized: torch.Tensor | dict[str, torch.Tensor],
generator: torch.Generator,
) -> tuple[torch.Tensor | dict[str, torch.Tensor], torch.Tensor]:
cfg = self.ttt_config
batch = self._ttt_sample_crop(tokenized, generator)
input_ids = self._ttt_input_ids_from_batch(batch) # (b, l)
row_has_residue = self._ttt_non_special_mask(input_ids).any(dim=1) # (b,)
eligible_rows = torch.where(row_has_residue)[0] # (n_eligible,)
if eligible_rows.numel() == 0:
raise ValueError(
"TTT sampled batch contains no trainable biological residue tokens."
)
sampled_row_indices = torch.randint( # (b_sample,)
eligible_rows.numel(),
(cfg.batch_size,),
generator=generator,
device=input_ids.device,
)
rows = eligible_rows[sampled_row_indices] # (b_sample,)
if isinstance(batch, dict):
sampled: torch.Tensor | dict[str, torch.Tensor] = {}
for name, tensor in batch.items():
if tensor.ndim >= 1 and tensor.shape[0] == input_ids.shape[0]:
sampled[name] = tensor.index_select(0, rows) # (b_sample, ...)
else:
sampled[name] = tensor
else:
sampled = input_ids.index_select(0, rows) # (b_sample, l)
sampled_ids = self._ttt_input_ids_from_batch(sampled) # (b_sample, l)
labels = sampled_ids.clone() # (b_sample, l)
non_special = self._ttt_non_special_mask(sampled_ids) # (b_sample, l)
label_mask = torch.zeros_like(non_special) # (b_sample, l)
for row_idx in range(sampled_ids.shape[0]):
candidate_positions = torch.where(non_special[row_idx])[0] # (n_candidates,)
if candidate_positions.numel() == 0:
continue
num_mask = max(1, round(candidate_positions.numel() * cfg.mask_ratio))
order = torch.randperm( # (n_candidates,)
candidate_positions.numel(),
generator=generator,
device=sampled_ids.device,
)
chosen = candidate_positions[order[:num_mask]] # (n_mask,)
label_mask[row_idx, chosen] = True
labels = labels.masked_fill(~label_mask, -100) # (b_sample, l)
masked_ids = sampled_ids.clone() # (b_sample, l)
chosen_positions = torch.where(label_mask) # two (n_chosen,) tensors
if chosen_positions[0].numel() > 0:
random_values = torch.rand( # (n_chosen,)
chosen_positions[0].shape,
generator=generator,
device=sampled_ids.device,
)
leave = random_values < cfg.bert_leave_prob # (n_chosen,)
replace = (random_values >= cfg.bert_leave_prob) & ( # (n_chosen,)
random_values < cfg.bert_leave_prob + cfg.bert_replace_prob
)
mask = ~(leave | replace) # (n_chosen,)
if mask.any():
masked_ids[
chosen_positions[0][mask],
chosen_positions[1][mask],
] = self._ttt_mask_token()
if replace.any():
replacement_tokens = self._ttt_replacement_tokens(sampled_ids) # (c_aa,)
replacement_idx = torch.randint( # (n_replace,)
replacement_tokens.shape[0],
(int(replace.sum().item()),),
generator=generator,
device=sampled_ids.device,
)
masked_ids[
chosen_positions[0][replace],
chosen_positions[1][replace],
] = replacement_tokens[replacement_idx]
return self._ttt_set_input_ids(sampled, masked_ids), labels # batch, (b_sample, l)
@contextlib.contextmanager
def _ttt_seed_scope(self, seed: int | None) -> Iterator[None]:
if seed is None:
yield
return
cuda_devices = sorted(
{
parameter.device.index
for parameter in self.parameters()
if parameter.device.type == "cuda" and parameter.device.index is not None
}
)
with torch.random.fork_rng(devices=cuda_devices):
torch.random.default_generator.manual_seed(seed)
for device_index in cuda_devices:
with torch.cuda.device(device_index):
torch.cuda.manual_seed(seed)
yield
def ttt(
self,
seq: str | list[str] | None = None,
input_ids: torch.Tensor | None = None,
ttt_config: TTTConfig | Mapping[str, Any] | None = None,
**kwargs: Any,
) -> dict[str, Any]:
if ttt_config is not None:
if "_ttt_initialized" in self.__dict__ and self._ttt_initialized:
next_cfg = self.ttt_config.merged(ttt_config)
current_cfg = self.ttt_config
if next_cfg.lora_rank != current_cfg.lora_rank:
raise ValueError(
"Changing lora_rank after TTT initialization is not supported."
)
if next_cfg.lora_alpha != current_cfg.lora_alpha:
raise ValueError(
"Changing lora_alpha after TTT initialization is not supported."
)
if (
next_cfg.lora_target_replace_module
!= current_cfg.lora_target_replace_module
):
raise ValueError(
"Changing LoRA target class after TTT initialization is not supported."
)
if next_cfg.lora_target_modules != current_cfg.lora_target_modules:
raise ValueError(
"Changing LoRA target modules after TTT initialization is not supported."
)
self._ttt_cfg = next_cfg
else:
# Family constructors preconfigure the attention class that may
# receive LoRA adapters. A first-call mapping changes only the
# requested fields; rebuilding from TTTConfig defaults here
# would erase that family target immediately before injection.
self._ttt_cfg = self.ttt_config.merged(ttt_config)
self._ttt_cfg.verify()
cfg = self.ttt_config
device = next(self.parameters()).device
tokenized = self._ttt_tokenize(seq=seq, input_ids=input_ids, **kwargs)
tokenized = self._ttt_to_device(tokenized, device)
self._ttt_validate_tokenized_batch(tokenized)
self._ttt_ensure_initialized()
if cfg.initial_state_reset:
self.ttt_reset()
generator_device = device if device.type == "cuda" else torch.device("cpu")
generator = torch.Generator(device=generator_device)
if cfg.seed is not None:
generator.manual_seed(cfg.seed)
module_modes = {module: module.training for module in self.modules()}
requires_grad = {param: param.requires_grad for param in self.parameters()}
losses: list[float] = []
step_metrics: list[dict[str, Any]] = []
best_state: list[dict[str, torch.Tensor]] | None = None
best_metric: float | None = None
best_step = 0
with self._ttt_seed_scope(cfg.seed):
try:
self.train()
for param in self.parameters():
param.requires_grad_(False)
for param in self._ttt_lora_parameters():
param.requires_grad_(True)
optimizer = self._ttt_make_optimizer()
optimizer.zero_grad(set_to_none=True)
total_micro_steps = cfg.steps * cfg.ags
for micro_step in range(total_micro_steps):
batch, labels = self._ttt_sample_batch( # batch, (b_sample, l)
tokenized,
generator,
)
if not bool(labels.ne(-100).any()):
raise RuntimeError(
"TTT produced an all-ignored label batch; refusing a NaN update."
)
logits = self._ttt_predict_logits(batch, **kwargs) # (b_sample, l, c)
labels = labels.to(device=logits.device) # (b_sample, l)
loss = F.cross_entropy( # ()
logits.reshape(-1, logits.shape[-1]),
labels.reshape(-1),
ignore_index=-100,
)
if not bool(torch.isfinite(loss)):
raise FloatingPointError(
f"TTT loss is non-finite at micro-step {micro_step + 1}."
)
(loss / cfg.ags).backward()
if (micro_step + 1) % cfg.ags != 0:
continue
if cfg.gradient_clip:
torch.nn.utils.clip_grad_norm_(
self._ttt_lora_parameters(),
cfg.gradient_clip_max_norm,
)
optimizer.step()
optimizer.zero_grad(set_to_none=True)
step = (micro_step + 1) // cfg.ags
loss_value = float(loss.detach().item())
losses.append(loss_value)
if cfg.eval_each_step:
metrics, metric = self._ttt_eval_step(
step=step,
loss=loss_value,
seq=seq,
input_ids=input_ids,
**kwargs,
)
if len(metrics) > 0:
step_metrics.append(metrics)
if metric is not None and (best_metric is None or metric > best_metric):
best_metric = metric
best_step = step
best_state = self._ttt_snapshot_lora_state()
if cfg.automatic_best_state_reset and best_state is not None:
self._ttt_restore_lora_state(best_state)
finally:
for param, value in requires_grad.items():
param.requires_grad_(value)
for module, training in module_modes.items():
module.train(training)
return {
"losses": losses,
"step_metrics": step_metrics,
"best_step": best_step,
"best_metric": best_metric,
}