jisue's picture
Experiment backup 2026-11-10T01:58:18.622752+00:00
0f5d220 verified
Raw History Blame Contribute Delete
35.3 kB
# Copyright 2026 Tensor Auto Inc. 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.
"""Test-Time Training (TTT) layer with an MLP fast model.
This is the sequence-modeling primitive behind RoboTTT (`arXiv:2607.15275
<https://arxiv.org/abs/2607.15275>`_): a recurrent layer whose recurrent state
is a set of *fast weights* — the parameters of a small two-layer MLP that is
updated by gradient descent on a self-supervised reconstruction loss at every
step, during training *and* inference.
The math follows the reference implementation in
`test-time-training/ttt-video-dit <https://github.com/test-time-training/ttt-video-dit>`_
(``ttt/models/ssm/ops/ttt_mlp.py``), which is itself the "dual form" of the
update in Sun et al., *Learning to (Learn at Test Time)*. Two things differ
here, both required by the robot-policy setting:
1. **Fast weights are carried in and out.** The reference always starts a
forward from the learned initialization :math:`W_0`. Truncated
backpropagation through time (TBPTT) needs the *final* fast weights of
segment *n* to become the *initial* fast weights of segment *n+1*, detached
at the boundary, and closed-loop inference needs them to persist across
policy calls. So :meth:`TTTMLPLayer.forward` accepts an optional incoming
:class:`TTTFastWeights` and returns the outgoing one.
2. **Positional embeddings are 1-D over robot timesteps**, not 3-D over video
latents, and they take an explicit offset so a TBPTT segment that resumes
mid-trajectory does not restart its RoPE phase at zero.
Terminology used throughout, matching the paper:
* *slow weights* — ordinary ``nn.Parameter``s (Q/K/V/O projections, the fast
model's initialization :math:`W_0`, the inner learning-rate gate). Updated by
the outer optimizer.
* *fast weights* — ``W1/b1/W2/b2``, the two-layer MLP that *is* the recurrent
state. Updated by the inner gradient step, once per mini-batch.
* *mini-batch* — the group of tokens processed by one inner gradient step. For
a robot policy this is exactly one timestep's worth of tokens, so one policy
call performs one fast-weight update (the paper's "one mini batch per
inference").
The dual form computes a whole mini-batch's outputs *and* the end-of-mini-batch
fast weights with matrix multiplies rather than a token-by-token loop, which is
what makes the layer trainable at sequence lengths in the thousands.
"""
from __future__ import annotations
from dataclasses import dataclass, field
from typing import Any
import torch
import torch.nn.functional as F # noqa: N812
from einops import einsum, rearrange, repeat
from torch import Tensor, nn
# The inner reconstruction loss normalizes over the head dimension. This
# epsilon matches the reference implementation's `ln_fwd` / `ln_fused_l2_bwd`
# (1e-8) and is deliberately *not* the policy-level normalization epsilon.
TTT_LN_EPS = 1e-8
@dataclass
class TTTFastWeights:
"""The recurrent state of a :class:`TTTMLPLayer`: one fast MLP per head.
Attributes:
w1: First-layer weights, shape ``(B, H, head_dim, hidden_dim)``.
b1: First-layer bias, shape ``(B, H, 1, hidden_dim)``.
w2: Second-layer weights, shape ``(B, H, hidden_dim, head_dim)``.
b2: Second-layer bias, shape ``(B, H, 1, head_dim)``.
"""
w1: Tensor
b1: Tensor
w2: Tensor
b2: Tensor
def detach(self) -> TTTFastWeights:
"""Returns a copy holding the same values with the autograd graph cut.
This is the TBPTT boundary operation: the *values* cross into the next
segment so TTT keeps running over the whole trajectory, while the graph
does not, so activation memory is bounded by the segment length rather
than the trajectory length.
Returns:
A new :class:`TTTFastWeights` whose tensors are detached leaves.
"""
return TTTFastWeights(
w1=self.w1.detach(),
b1=self.b1.detach(),
w2=self.w2.detach(),
b2=self.b2.detach(),
)
def to(self, *args: Any, **kwargs: Any) -> TTTFastWeights:
"""Applies ``Tensor.to`` to every field.
Args:
*args: Positional arguments forwarded to ``Tensor.to``.
**kwargs: Keyword arguments forwarded to ``Tensor.to``.
Returns:
A new :class:`TTTFastWeights` with converted tensors.
"""
return TTTFastWeights(
w1=self.w1.to(*args, **kwargs),
b1=self.b1.to(*args, **kwargs),
w2=self.w2.to(*args, **kwargs),
b2=self.b2.to(*args, **kwargs),
)
def _layer_norm_forward(x: Tensor, gamma: Tensor, beta: Tensor, eps: float = TTT_LN_EPS) -> Tensor:
"""Per-head layer norm over the last dimension.
Args:
x: Input, shape ``(..., head_dim)``.
gamma: Scale, broadcastable to ``x``.
beta: Shift, broadcastable to ``x``.
eps: Variance epsilon.
Returns:
The normalized, scaled and shifted tensor, same shape as ``x``.
"""
mu = x.mean(dim=-1, keepdim=True)
var = x.var(dim=-1, keepdim=True, unbiased=False)
x_hat = (x - mu) / torch.sqrt(var + eps)
return gamma * x_hat + beta
def _layer_norm_fused_l2_backward(
x: Tensor, l2_target: Tensor, gamma: Tensor, beta: Tensor, eps: float = TTT_LN_EPS
) -> Tensor:
"""Gradient of ``0.5 * ||LN(x) - l2_target||^2`` with respect to ``x``.
This is the inner loss's backward pass, written in closed form rather than
obtained from autograd. Doing it by hand is what keeps the inner gradient
step cheap enough to run once per token group: autograd would have to
retain a graph per mini-batch across the whole sequence.
Args:
x: Pre-norm activations, shape ``(..., head_dim)``.
l2_target: Reconstruction target, same shape as ``x``.
gamma: Layer-norm scale, broadcastable to ``x``.
beta: Layer-norm shift, broadcastable to ``x``.
eps: Variance epsilon.
Returns:
The gradient with respect to ``x``, same shape as ``x``.
"""
dim = x.shape[-1]
mu = x.mean(dim=-1, keepdim=True)
var = x.var(dim=-1, keepdim=True, unbiased=False)
std = torch.sqrt(var + eps)
x_hat = (x - mu) / std
grad_output = (gamma * x_hat + beta) - l2_target
grad_x_hat = grad_output * gamma
return (
(1.0 / dim)
* (
dim * grad_x_hat
- grad_x_hat.sum(dim=-1, keepdim=True)
- x_hat * (grad_x_hat * x_hat).sum(dim=-1, keepdim=True)
)
/ std
)
def _gelu_backward(x: Tensor) -> Tensor:
"""Derivative of the tanh-approximated GeLU used by the fast model.
Args:
x: Pre-activation input.
Returns:
``d/dx gelu_tanh(x)``, same shape as ``x``.
"""
tanh_out = torch.tanh(0.79788456 * x * (1 + 0.044715 * x * x))
return 0.5 * x * ((1 - tanh_out * tanh_out) * (0.79788456 + 0.1070322243 * x * x)) + 0.5 * (1 + tanh_out)
def _rotary_1d(x: Tensor, positions: Tensor, theta: float) -> Tensor:
"""Applies 1-D rotary position embeddings over the sequence axis.
Adjacent channel pairs ``(0, 1), (2, 3), ...`` are treated as one complex
number and rotated, matching the interleaved convention the reference
implementation gets from ``torch.view_as_complex`` on a ``(..., d/2, 2)``
view. Computed with real arithmetic instead, because ``view_as_complex``
requires a float32/float64 input and a contiguous final stride — both
awkward inside a bf16 model.
Args:
x: Input, shape ``(B, L, H, head_dim)``. ``head_dim`` must be even.
positions: Integer positions, shape ``(B, L)`` or ``(L,)``.
theta: RoPE base. RoboTTT uses 10000.
Returns:
The rotated tensor, same shape and dtype as ``x``.
Raises:
ValueError: If ``head_dim`` is odd.
"""
head_dim = x.shape[-1]
if head_dim % 2 != 0:
raise ValueError(f"RoPE needs an even head_dim, got {head_dim}")
if positions.ndim == 1:
positions = repeat(positions, "l -> b l", b=x.shape[0])
# At least float32 for the trig: the angle grows linearly with position, and
# at 8K timesteps a bf16 angle carries ~1e-2 absolute error. Promote rather
# than pin, matching the dtype policy in `TTTMLPLayer.forward` — pinning
# float32 here would cap a float64 module at float32 through this path,
# which is what the unit tests use float64 to avoid.
angle_dtype = torch.float32 if x.dtype in (torch.float16, torch.bfloat16) else x.dtype
inv_freq = 1.0 / (theta ** (torch.arange(0, head_dim, 2, device=x.device, dtype=angle_dtype) / head_dim))
angles = rearrange(positions.to(angle_dtype), "b l -> b l 1") * inv_freq
cos = rearrange(torch.cos(angles), "b l d -> b l 1 d")
sin = rearrange(torch.sin(angles), "b l d -> b l 1 d")
pairs = rearrange(x.to(angle_dtype), "b l h (d two) -> b l h d two", two=2)
even, odd = pairs[..., 0], pairs[..., 1]
rotated = torch.stack((even * cos - odd * sin, even * sin + odd * cos), dim=-1)
return rearrange(rotated, "b l h d two -> b l h (d two)", two=2).to(x.dtype)
def _ttt_mini_batch_step(
fast_weights: TTTFastWeights,
xq: Tensor,
xk: Tensor,
xv: Tensor,
eta: Tensor,
ln_weight: Tensor,
ln_bias: Tensor,
) -> tuple[TTTFastWeights, Tensor]:
"""Runs one inner gradient step and produces the mini-batch's outputs.
This is the dual form. It takes **one** gradient step over the whole
mini-batch and applies the resulting weights to every query in it — it is
*not* a causal per-token cascade, and there is no causal mask anywhere in
this function. :meth:`TTTMLPLayer._inner_learning_rates` builds a
row-constant ``eta`` precisely so that this holds: the tokens of one
mini-batch are one observation, arriving together, with no ordering among
them. Verified in float64 to agree with "one step over all C tokens, then
apply to every query" to 4.4e-16, and to differ from a strictly causal
per-token loop by ~0.03-0.09.
**Why that makes the mini-batch size load-bearing.** Because a mini-batch is
non-causal, grouping two timesteps into one inner step lets the later
timestep influence the earlier one's output — measured at 0.66 versus
exactly 0.0 when one mini-batch is one timestep. The caller must therefore
keep ``mini_batch_size`` equal to the token count of a single timestep;
:class:`TTTMLPLayer` enforces it rather than trusting the call site.
Shapes use ``B`` batch, ``H`` heads, ``C`` mini-batch size (tokens per
inner step), ``D`` head dim, ``F`` fast-MLP hidden dim.
Args:
fast_weights: Incoming fast weights (the recurrent state).
xq: Query projections, shape ``(B, H, C, D)``.
xk: Key projections, shape ``(B, H, C, D)``.
xv: Value projections (already turned into a reconstruction target),
shape ``(B, H, C, D)``.
eta: Per-token inner learning rates, shape ``(B, H, C, C)``.
ln_weight: Inner layer-norm scale, shape ``(H, 1, D)``.
ln_bias: Inner layer-norm shift, shape ``(H, 1, D)``.
Returns:
A tuple of the outgoing fast weights and the mini-batch output of shape
``(B, H, C, D)``.
"""
w1, b1, w2, b2 = fast_weights.w1, fast_weights.b1, fast_weights.w2, fast_weights.b2
# Forward pass of the fast model on the keys.
z1 = xk @ w1 + b1
x2 = F.gelu(z1, approximate="tanh")
z2 = x2 @ w2 + b2
# The self-supervised inner loss reconstructs (V - K) from K, so the fast
# model learns the *residual* association rather than the identity.
reconstruction_target = xv - xk
grad_z2 = _layer_norm_fused_l2_backward(z2, reconstruction_target, ln_weight, ln_bias)
grad_z1 = grad_z2 @ w2.transpose(-2, -1) * _gelu_backward(z1)
# Apply the updated weights to the queries. With a row-constant eta the
# `eta * Attn` terms reduce exactly to `xq @ W_after + b_after`, which is
# what the update-then-apply unit test pins. They are NOT a per-token causal
# cascade; see this function's docstring.
attn1 = xq @ xk.transpose(-2, -1)
b1_bar = b1 - eta @ grad_z1
z1_bar = xq @ w1 - (eta * attn1) @ grad_z1 + b1_bar
x2_bar = F.gelu(z1_bar, approximate="tanh")
attn2 = x2_bar @ x2.transpose(-2, -1)
b2_bar = b2 - eta @ grad_z2
z2_bar = x2_bar @ w2 - (eta * attn2) @ grad_z2 + b2_bar
# End-of-mini-batch fast weights: the cumulative effect of every token's
# gradient step, read off the last row of eta.
last_eta = eta[..., -1, :, None]
next_fast_weights = TTTFastWeights(
w1=w1 - (last_eta * xk).transpose(-1, -2) @ grad_z1,
b1=b1 - torch.sum(last_eta * grad_z1, dim=-2, keepdim=True),
w2=w2 - (last_eta * x2).transpose(-1, -2) @ grad_z2,
b2=b2 - torch.sum(last_eta * grad_z2, dim=-2, keepdim=True),
)
# Residual around the fast model, so a freshly initialized layer is close
# to a pass-through of the queries.
output = xq + _layer_norm_forward(z2_bar, ln_weight, ln_bias)
return next_fast_weights, output
def _scan_mini_batches(
fast_weights: TTTFastWeights,
xq: Tensor,
xk: Tensor,
xv: Tensor,
eta: Tensor,
ln_weight: Tensor,
ln_bias: Tensor,
checkpoint_group_size: int = 0,
) -> tuple[TTTFastWeights, Tensor]:
"""Scans :func:`_ttt_mini_batch_step` over the mini-batch axis.
Args:
fast_weights: Incoming fast weights.
xq: Queries, shape ``(NC, B, H, C, D)`` with ``NC`` mini-batches.
xk: Keys, same shape as ``xq``.
xv: Values, same shape as ``xq``.
eta: Inner learning rates, shape ``(NC, B, H, C, C)``.
ln_weight: Inner layer-norm scale, shape ``(H, 1, D)``.
ln_bias: Inner layer-norm shift, shape ``(H, 1, D)``.
checkpoint_group_size: When > 0, gradient-checkpoint the scan in groups
of this many mini-batches, trading recompute for activation memory.
0 disables checkpointing.
Returns:
A tuple of the final fast weights and the stacked outputs of shape
``(NC, B, H, C, D)``.
"""
num_mini_batch = xq.shape[0]
def run_group(carry: TTTFastWeights, start: int, end: int) -> tuple[TTTFastWeights, Tensor]:
outputs = []
for i in range(start, end):
carry, out = _ttt_mini_batch_step(carry, xq[i], xk[i], xv[i], eta[i], ln_weight, ln_bias)
outputs.append(out)
return carry, torch.stack(outputs)
if checkpoint_group_size <= 0:
return run_group(fast_weights, 0, num_mini_batch)
# torch.utils.checkpoint cannot carry a dataclass through its
# pack/unpack, so flatten to a tuple of tensors at the boundary.
def run_group_flat(w1: Tensor, b1: Tensor, w2: Tensor, b2: Tensor, start: int, end: int):
carry, out = run_group(TTTFastWeights(w1=w1, b1=b1, w2=w2, b2=b2), start, end)
return carry.w1, carry.b1, carry.w2, carry.b2, out
carry = fast_weights
chunks = []
for group_start in range(0, num_mini_batch, checkpoint_group_size):
group_end = min(group_start + checkpoint_group_size, num_mini_batch)
w1, b1, w2, b2, out = torch.utils.checkpoint.checkpoint(
run_group_flat,
carry.w1,
carry.b1,
carry.w2,
carry.b2,
group_start,
group_end,
use_reentrant=False,
)
carry = TTTFastWeights(w1=w1, b1=b1, w2=w2, b2=b2)
chunks.append(out)
return carry, torch.cat(chunks, dim=0)
class TTTMLPLayer(nn.Module):
"""A TTT layer whose fast model is a two-layer MLP.
One instance is attached to each action-expert decoder layer. It is the
only component in the policy that carries information *across* timesteps —
attention stays strictly within a timestep — which is the division of
labour RoboTTT's architecture is built on.
The layer is a no-op contributor at initialization only in combination with
:class:`TanhGate`; on its own its output is nonzero (its fast model has a
random :math:`W_0`).
Args:
width: Model width, i.e. the action expert's hidden size.
num_heads: Number of TTT heads. ``width`` must divide evenly.
mlp_hidden_multiplier: Fast-MLP hidden dim as a multiple of
``head_dim``. The reference video model uses 4; RoboTTT's own
budget (~10M per layer on a 1536-wide backbone) works out to
roughly 2, which is also what keeps this layer from doubling the
size of a 1024-wide action expert.
base_lr: Base inner learning rate, scaled per token by a learned gate.
RoboTTT uses 0.1.
rope_theta: RoPE base for the 1-D positional embedding. RoboTTT uses
10000.
scan_checkpoint_group_size: Gradient-checkpoint group size for the
mini-batch scan; 0 disables.
Raises:
ValueError: If ``width`` is not divisible by ``num_heads``, or the
resulting head dim is odd (RoPE needs pairs).
"""
def __init__(
self,
width: int,
num_heads: int,
mlp_hidden_multiplier: int = 2,
base_lr: float = 0.1,
rope_theta: float = 10000.0,
scan_checkpoint_group_size: int = 0,
expected_mini_batch_size: int | None = None,
):
super().__init__()
if width % num_heads != 0:
raise ValueError(f"width={width} must be divisible by num_heads={num_heads}")
head_dim = width // num_heads
if head_dim % 2 != 0:
raise ValueError(f"head_dim={head_dim} must be even for rotary embeddings")
self.width = width
self.num_heads = num_heads
self.head_dim = head_dim
self.hidden_dim = head_dim * mlp_hidden_multiplier
self.base_lr = base_lr
self.rope_theta = rope_theta
self.scan_checkpoint_group_size = scan_checkpoint_group_size
# When set, `forward` refuses any other mini-batch size. A mini-batch is
# processed non-causally, so one holding two timesteps would let the
# later one influence the earlier one's output — a silent leak that
# looks like a harmless efficiency win at the call site.
self.expected_mini_batch_size = expected_mini_batch_size
self.wq = nn.Linear(width, width, bias=True)
self.wk = nn.Linear(width, width, bias=True)
self.wv = nn.Linear(width, width, bias=True)
self.wo = nn.Linear(width, width, bias=True)
# Learned per-head, per-token inner learning-rate gate. Kept as raw
# parameters rather than an nn.Linear because it is applied per head
# with a single einsum over the pre-projection hidden states.
self.ttt_lr_weight = nn.Parameter(torch.normal(0.0, 0.02, size=(num_heads, 1, width)))
self.ttt_lr_bias = nn.Parameter(torch.zeros(num_heads, 1))
# Inner-loss layer norm, one per head.
self.ttt_norm_weight = nn.Parameter(torch.ones(num_heads, head_dim))
self.ttt_norm_bias = nn.Parameter(torch.zeros(num_heads, head_dim))
# The fast model's learned initialization W_0. These are slow weights:
# meta-learned through the inner update by the outer task loss, which
# is what tailors the fast-weight dynamics to robot trajectories. They
# only receive gradient through the *first* TBPTT segment, whose update
# originates directly from them.
self.w1_init = nn.Parameter(torch.normal(0.0, 0.02, size=(num_heads, head_dim, self.hidden_dim)))
self.b1_init = nn.Parameter(torch.zeros(num_heads, 1, self.hidden_dim))
self.w2_init = nn.Parameter(torch.normal(0.0, 0.02, size=(num_heads, self.hidden_dim, head_dim)))
self.b2_init = nn.Parameter(torch.zeros(num_heads, 1, head_dim))
self.post_norm = nn.LayerNorm(width, eps=1e-6)
def initial_fast_weights(self, batch_size: int) -> TTTFastWeights:
"""Materializes :math:`W_0` for a batch, starting a fresh rollout.
Args:
batch_size: Number of independent trajectories.
Returns:
Per-sample copies of the learned initialization. These are *not*
detached: the outer loss must reach ``w1_init`` and friends through
the first segment, otherwise the initialization never trains and
stays a random draw.
"""
return TTTFastWeights(
w1=repeat(self.w1_init, "h d f -> b h d f", b=batch_size),
b1=repeat(self.b1_init, "h one f -> b h one f", b=batch_size),
w2=repeat(self.w2_init, "h f d -> b h f d", b=batch_size),
b2=repeat(self.b2_init, "h one d -> b h one d", b=batch_size),
)
def _inner_learning_rates(self, hidden_states: Tensor, mini_batch_size: int) -> Tensor:
"""Computes the per-token inner learning rate matrix ``eta``.
Args:
hidden_states: Layer input, shape ``(B, L, width)``.
mini_batch_size: Tokens per inner gradient step.
Returns:
``eta`` of shape ``(B, H, NC, C, C)``, already divided by the
mini-batch size so a mini-batch's tokens share one gradient step
(they are one observation, so no causal ordering within it).
"""
grouped = rearrange(hidden_states, "b (nc c) w -> b nc c w", c=mini_batch_size)
gate = einsum(grouped, self.ttt_lr_weight.to(grouped.dtype), "b nc c w, h one w -> b h nc c one")
gate = gate + rearrange(self.ttt_lr_bias, "h one -> 1 h 1 1 one")
lr = self.base_lr * torch.sigmoid(gate) / self.head_dim
# (B, H, NC, C, 1) -> (B, H, NC, C, C): every token in the mini-batch
# sees the same learning-rate column, matching the non-causal
# mini-batch treatment in the reference implementation.
lr = rearrange(lr, "b h nc c one -> b h nc one c")
return repeat(lr, "b h nc one c -> b h nc (one r) c", r=mini_batch_size) / mini_batch_size
def forward(
self,
hidden_states: Tensor,
mini_batch_size: int,
fast_weights: TTTFastWeights | None = None,
position_offset: int = 0,
timestep_is_pad: Tensor | None = None,
) -> tuple[Tensor, TTTFastWeights]:
"""Runs the layer over a sequence, updating and returning fast weights.
Args:
hidden_states: Input, shape ``(B, L, width)``, where ``L`` is
``num_timesteps * mini_batch_size``. Position ``0`` of the
sequence axis is the oldest timestep.
mini_batch_size: Tokens per timestep, i.e. per inner gradient step.
``L`` must be an exact multiple.
fast_weights: Recurrent state to resume from. ``None`` starts from
the learned :math:`W_0` — correct for the first TBPTT segment
and for the first call of a rollout, wrong (a memory reset) in
the middle of either.
position_offset: Absolute index of the first token in the sequence,
used so a resumed segment continues its RoPE phase instead of
restarting at zero.
Returns:
A tuple of the layer output, shape ``(B, L, width)`` and dtype
matching the input, and the outgoing fast weights.
Raises:
ValueError: If ``L`` is not a multiple of ``mini_batch_size``, or if
``mini_batch_size`` disagrees with ``expected_mini_batch_size``.
"""
batch_size, seq_len, _ = hidden_states.shape
if seq_len % mini_batch_size != 0:
raise ValueError(
f"sequence length {seq_len} must be a multiple of mini_batch_size {mini_batch_size}"
)
if self.expected_mini_batch_size is not None and mini_batch_size != self.expected_mini_batch_size:
raise ValueError(
f"mini_batch_size={mini_batch_size} but this layer was built for "
f"{self.expected_mini_batch_size} (one timestep's tokens). A mini-batch is "
"processed non-causally, so grouping several timesteps into one inner step "
"lets a later timestep change an earlier timestep's output."
)
input_dtype = hidden_states.dtype
# Dtype policy, in two stages, because the linear projections and the
# inner loop want different things:
#
# * The projections run in the module's own parameter dtype. This model
# is served in bfloat16, so `wq`/`wk`/`wv` hold bf16 weights and
# handing them float32 activations is a hard dtype error, not a slow
# path.
# * The inner loop then runs in float32. It takes an explicit gradient
# step and divides by a standard deviation, thousands of times over a
# long trajectory; in bf16 that drifts enough to change the
# recurrence. The reference implementation gets away with low
# precision only by fusing the whole update into a CUDA kernel.
#
# "Promote", not "pin to float32": a float64 module — which the unit
# tests use, to get tolerances tight enough to catch a real algebra
# error rather than hide it under float32 noise — must not be silently
# downcast.
param_dtype = self.wq.weight.dtype
compute_dtype = torch.float32 if param_dtype in (torch.float16, torch.bfloat16) else param_dtype
projected = hidden_states.to(param_dtype)
xq = rearrange(self.wq(projected), "b l (h d) -> b l h d", h=self.num_heads).to(compute_dtype)
xk = rearrange(self.wk(projected), "b l (h d) -> b l h d", h=self.num_heads).to(compute_dtype)
xv = rearrange(self.wv(projected), "b l (h d) -> b l h d", h=self.num_heads).to(compute_dtype)
# L2-normalizing Q and K bounds the inner loss's curvature, so one
# gradient step with a shared base learning rate behaves consistently
# across layers and sequence positions.
xq = F.normalize(xq, p=2, dim=-1)
xk = F.normalize(xk, p=2, dim=-1)
positions = torch.arange(position_offset, position_offset + seq_len, device=hidden_states.device)
xq = _rotary_1d(xq, positions, self.rope_theta)
xk = _rotary_1d(xk, positions, self.rope_theta)
# Turn the values into the reconstruction target the inner loss uses.
#
# Deliberate deviation from the reference: ttt-video-dit normalizes this
# target with an *unbiased* std and divides by ``std + eps``
# (`ln_reconstruction_target`), while normalizing the fast model's own
# output with a biased variance and ``sqrt(var + eps)`` (`ln_fwd`). The
# two are inconsistent with each other upstream. We use the `ln_fwd`
# convention for both, so the target lives on the same scale as the
# prediction the inner loss compares it against. The difference is
# O(1/head_dim) and vanishes as head_dim grows, but a reviewer diffing
# against the reference will see it, so it is called out here rather
# than looking like a porting slip.
ln_weight = rearrange(self.ttt_norm_weight, "h d -> h 1 d").to(compute_dtype)
ln_bias = rearrange(self.ttt_norm_bias, "h d -> h 1 d").to(compute_dtype)
xv = xk + _layer_norm_forward(xv - xk, ln_weight.squeeze(1), ln_bias.squeeze(1))
eta = self._inner_learning_rates(hidden_states.to(compute_dtype), mini_batch_size)
# Padded timesteps must not be learned from. A timestep whose window
# position falls before the episode's first frame carries a repeated
# boundary frame; taking an inner gradient step on it writes that
# repetition into the fast weights, which is exactly the signal the
# demonstration is supposed to supply.
#
# Zeroing `eta` is sufficient and needs no change to the step itself:
# with `eta == 0`, `b1_bar == b1`, `z1_bar` reduces to a forward pass
# under the incoming weights, and `last_eta == 0` leaves
# `next_fast_weights` equal to `fast_weights`. The timestep is read
# with the current memory and contributes nothing to it.
#
# Masking here rather than skipping mini-batches keeps the scan shape
# static (padding differs per sample within a batch) and leaves the
# gradient-checkpoint grouping untouched.
if timestep_is_pad is not None:
num_mini_batch = hidden_states.shape[1] // mini_batch_size
if tuple(timestep_is_pad.shape) != (batch_size, num_mini_batch):
raise ValueError(
f"timestep_is_pad must have shape ({batch_size}, {num_mini_batch}), "
f"got {tuple(timestep_is_pad.shape)}"
)
keep = rearrange((~timestep_is_pad).to(eta.dtype), "b nc -> b 1 nc 1 1")
eta = eta * keep
to_mini_batches = "b (nc c) h d -> nc b h c d"
xq = rearrange(xq, to_mini_batches, c=mini_batch_size)
xk = rearrange(xk, to_mini_batches, c=mini_batch_size)
xv = rearrange(xv, to_mini_batches, c=mini_batch_size)
eta = rearrange(eta, "b h nc c1 c2 -> nc b h c1 c2")
if fast_weights is None:
fast_weights = self.initial_fast_weights(batch_size)
fast_weights = fast_weights.to(compute_dtype)
fast_weights, outputs = _scan_mini_batches(
fast_weights,
xq,
xk,
xv,
eta,
ln_weight,
ln_bias,
checkpoint_group_size=self.scan_checkpoint_group_size,
)
# Back to the parameter dtype for the output projections, then to the
# caller's dtype so this layer never changes the dtype of the residual
# stream it feeds.
out = rearrange(outputs, "nc b h c d -> b (nc c) (h d)").to(param_dtype)
out = self.post_norm(out)
out = self.wo(out)
return out.to(input_dtype), fast_weights
class TanhGate(nn.Module):
"""Blends a TTT output into an attention output through ``tanh(alpha)``.
Implements ``O = tanh(alpha) * O_ttt + O_attn`` with a learned per-channel
``alpha`` initialized near zero. This is what makes it safe to bolt TTT
onto an already-pretrained VLA: at step 0 the gate is ~0.001, so the model
reproduces the base policy's behavior, and training decides how far to open
it. Without it, a randomly initialized memory layer injects noise into
every layer of a pretrained action expert from the first step.
Args:
width: Number of channels; ``alpha`` is learned per channel.
init_value: Initial ``alpha``. RoboTTT uses 0.001.
"""
def __init__(self, width: int, init_value: float = 0.001):
super().__init__()
self.alpha = nn.Parameter(torch.full((width,), init_value))
# Diagnostic multiplier on the gate, applied only when the module is in
# eval mode — which includes in-training validation, so leave it at 1.0
# in training configs. 0.0 silences the memory contribution entirely so
# an existing checkpoint can be evaluated with the TTT pathway off.
# When a policy config is bound (`bind_diagnostics_config`), the scale
# is read live from it so flipping the knob on an already-built policy
# takes effect; the plain attribute remains the unbound fallback.
self.inference_alpha_scale: float = 1.0
self._diagnostics_config = None
def bind_diagnostics_config(self, config) -> None:
"""Binds a policy config whose ``ttt_inference_alpha_scale`` is read live.
Object identity makes later mutations of the config visible to the gate,
so the diagnostic scale can be flipped on an already-built policy — the
natural paired-eval flow on one loaded checkpoint. Stored as a plain
attribute (not a submodule/buffer), so ``load_state_dict`` never touches it.
Args:
config: The owning policy's config object (must expose
``ttt_inference_alpha_scale``).
"""
self._diagnostics_config = config
def forward(self, attention_output: Tensor, ttt_output: Tensor) -> Tensor:
"""Adds the gated TTT contribution to the attention output.
Args:
attention_output: The attention branch, shape ``(..., width)``.
ttt_output: The TTT branch, same shape.
Returns:
The blended output, same shape and dtype as ``attention_output``.
"""
gate = torch.tanh(self.alpha).to(ttt_output.dtype)
if not self.training:
scale = (
self._diagnostics_config.ttt_inference_alpha_scale
if self._diagnostics_config is not None
else self.inference_alpha_scale
)
if scale != 1.0:
gate = gate * scale
return attention_output + gate * ttt_output
@dataclass
class TTTSequenceState:
"""Per-forward runtime state threaded into the action expert's TTT layers.
:class:`~opentau.policies.pi05.paligemma_with_expert.PaliGemmaWithExpertModel`
folds the sequence axis into the batch axis so attention runs per timestep
(which is the whole point of the architecture: attention is spatial, TTT is
temporal). The TTT layers need the axis back, plus the carried recurrent
state, so this object travels alongside the hidden states.
``incoming`` and ``outgoing`` are keyed by decoder-layer index because each
layer owns an independent fast model. A missing ``incoming`` key means
"start this layer from its learned :math:`W_0`" — correct for the first
TBPTT segment and the first call of a rollout, and a silent memory reset
anywhere else.
Attributes:
num_timesteps: Number of policy calls folded into the batch axis. The
expert's hidden states arrive as ``(B * num_timesteps, S, W)`` and
are unfolded to ``(B, num_timesteps * S, W)`` for TTT.
position_offset: Absolute token index of the first token in this
segment, so a resumed segment continues its RoPE phase.
incoming: Carried fast weights per layer index.
outgoing: Filled by the forward with each layer's end-of-segment fast
weights, for the caller to detach and pass to the next segment.
"""
num_timesteps: int
position_offset: int = 0
incoming: dict[int, TTTFastWeights] = field(default_factory=dict)
outgoing: dict[int, TTTFastWeights] = field(default_factory=dict)
timestep_is_pad: Tensor | None = None