titans-pytorch 0.0.14__py3-none-any.whl

titans_pytorch/__init__.py

@@ -0,0 +1,3 @@
+from titans_pytorch.titans import (
+    NeuralMemory
+)

titans_pytorch/associative_scan.py

@@ -0,0 +1,90 @@
+from __future__ import annotations
+from typing import Callable
+
+import torch
+from torch import Tensor
+import torch.nn.functional as F
+
+# taken from S5-pytorch repository
+# https://github.com/i404788/s5-pytorch/blob/74e2fdae00b915a62c914bf3615c0b8a4279eb84/s5/jax_compat.py#L51-L134
+
+# helper functions
+
+def pad_at_dim(t, pad, dim = -1, value = 0.):
+    dims_from_right = (- dim - 1) if dim < 0 else (t.ndim - dim - 1)
+    zeros = ((0, 0) * dims_from_right)
+    return F.pad(t, (*zeros, *pad), value = value)
+
+# the operator that is needed
+
+@torch.jit.script
+def binary_operator(
+    a: tuple[Tensor, Tensor],
+    b: tuple[Tensor, Tensor]
+):
+    a_i, kv_i = a
+    a_j, kv_j = b
+    return a_j * a_i, torch.addcmul(kv_j, a_j, kv_i)
+
+# Pytorch impl. of jax.lax.associative_scan
+# made specifically for axis of 1 (sequence of tokens for autoregressive modeling)
+
+def associative_scan(
+    operator: Callable,
+    elems: tuple[Tensor, Tensor]
+):
+    num_elems = int(elems[0].shape[1])
+
+    if not all(int(elem.shape[1]) == num_elems for elem in elems[1:]):
+        raise ValueError('Array inputs to associative_scan must have the same '
+                         'first dimension. (saw: {})'
+                         .format([elem.shape for elem in elems]))
+
+    def _scan(elems):
+        """Perform scan on `elems`."""
+        num_elems = elems[0].shape[1]
+
+        if num_elems < 2:
+            return elems
+
+        # Combine adjacent pairs of elements.
+
+        reduced_elems = operator(
+          [elem[:, :-1:2] for elem in elems],
+          [elem[:, 1::2] for elem in elems])
+
+        # Recursively compute scan for partially reduced tensors.
+
+        odd_elems = _scan(reduced_elems)
+
+        if num_elems % 2 == 0:
+            even_elems = operator(
+                [e[:, :-1] for e in odd_elems],
+                [e[:, 2::2] for e in elems])
+        else:
+            even_elems = operator(
+                odd_elems,
+                [e[:, 2::2] for e in elems])
+
+        # The first element of a scan is the same as the first element
+        # of the original `elems`.
+
+        even_elems = [
+          torch.cat([elem[:, :1], result], dim=1)
+          for (elem, result) in zip(elems, even_elems)]
+
+        return list(map(_interleave, even_elems, odd_elems))
+
+    return _scan(elems)
+
+def _interleave(a, b):
+    a_axis_len, b_axis_len = a.shape[1], b.shape[1]
+    output_axis_len = a_axis_len + b_axis_len
+
+    if (a_axis_len == (b_axis_len + 1)):
+        b = pad_at_dim(b, (0, 1), dim = 1)
+
+    stacked = torch.stack([a, b], dim=2)
+    interleaved = torch.flatten(stacked, start_dim=1, end_dim=2)
+
+    return interleaved[:, :output_axis_len]

titans_pytorch/titans.py

@@ -0,0 +1,408 @@
+from __future__ import annotations
+import math
+from functools import partial
+
+import torch
+from torch import nn, Tensor
+import torch.nn.functional as F
+from torch.nn import Linear, Module
+from torch.func import functional_call, vmap, grad_and_value
+
+from tensordict import TensorDict
+
+from titans_pytorch.associative_scan import (
+    associative_scan,
+    binary_operator,
+    pad_at_dim
+)
+
+import einx
+from einops import rearrange, pack, unpack
+from einops.layers.torch import Rearrange, Reduce
+
+"""
+ein notation:
+b - batch
+n - sequence
+d - feature dimension
+c - intra-chunk
+"""
+
+# constants
+
+LinearNoBias = partial(Linear, bias = False)
+
+# functions
+
+def exists(v):
+    return v is not None
+
+def default(v, d):
+    return v if exists(v) else d
+
+def round_down_multiple(seq, mult):
+    return seq // mult * mult
+
+def round_up_multiple(seq, mult):
+    return math.ceil(seq / mult) * mult
+
+def pack_one_with_inverse(t, pattern):
+    packed, packed_shape = pack([t], pattern)
+
+    def inverse(out, inv_pattern = None):
+        inv_pattern = default(inv_pattern, pattern)
+        return unpack(out, packed_shape, inv_pattern)[0]
+
+    return packed, inverse
+
+# classes
+
+class MLP(Module):
+    def __init__(
+        self,
+        dim,
+        depth
+    ):
+        super().__init__()
+        self.weights = nn.ParameterList([nn.Parameter(torch.randn(dim, dim)) for _ in range(depth)])
+
+    def forward(
+        self,
+        x
+    ):
+        for ind, weight in enumerate(self.weights):
+            is_first = ind == 0
+
+            if not is_first:
+                x = F.silu(x)
+
+            x = x @ weight
+
+        return x
+
+# main neural memory
+
+def default_loss_fn(pred, target):
+    return (pred - target).pow(2).mean(dim = -1).sum()
+
+class NeuralMemory(Module):
+    def __init__(
+        self,
+        dim,
+        chunk_size = 1,
+        dim_head = None,
+        heads = 1,
+        model: Module | None = None,
+        store_memory_loss_fn: Callable = default_loss_fn,
+        pre_rmsnorm = True,
+        post_rmsnorm = True,
+        use_accelerated_scan = False,
+        default_mlp_kwargs: dict = dict(
+            depth = 4
+        )
+    ):
+        super().__init__()
+
+        # norms
+
+        self.retrieve_norm = nn.RMSNorm(dim) if pre_rmsnorm else nn.Identity()
+        self.store_norm = nn.RMSNorm(dim) if pre_rmsnorm else nn.Identity()
+
+        self.post_rmsnorm = nn.RMSNorm(dim) if post_rmsnorm else nn.Identity()
+
+        # maybe multi-headed
+
+        dim_head = default(dim_head, dim)
+        dim_inner = dim_head * heads
+
+        self.split_heads = Rearrange('b n (h d) -> (b h) n d', h = heads)
+        self.merge_heads = Rearrange('(b h) n d -> b n (h d)', h = heads)
+        self.combine_heads = LinearNoBias(dim_inner, dim) if heads > 1 else nn.Identity()
+
+        # memory mlp
+
+        if not exists(model):
+            model = MLP(dim_head, **default_mlp_kwargs)
+
+        assert not exists(next(model.buffers(), None)), 'model cannot have buffers for now'
+
+        # the memory is the weights of the model
+
+        self.memory_model = model
+
+        # the chunk size within the paper where adaptive step, momentum, weight decay are shared
+
+        self.chunk_size = chunk_size
+
+        # prepare function for per sample gradients from model above, using torch.func
+
+        def forward_and_loss(params, inputs, target):
+            pred = functional_call(self.memory_model, params, inputs)
+            loss = self.store_memory_loss_fn(pred, target) # simple mse loss in paper - eq (12) - |M(k) - v|²
+            return loss
+
+        self.per_sample_grad_and_value_fn = vmap(grad_and_value(forward_and_loss), in_dims = (None, 0, 0))
+
+        # queries for retrieving from the model
+
+        self.to_queries = LinearNoBias(dim, dim_inner)
+
+        # keys and values for storing to the model
+
+        self.to_keys_values = LinearNoBias(dim, dim_inner * 2)
+        self.store_memory_loss_fn = store_memory_loss_fn
+
+        # learned adaptive learning rate and momentum
+        # todo - explore mlp layerwise learned lr / momentum
+
+        self.to_momentum = nn.Sequential(
+            Reduce('b (n c) ... -> b n ...', 'mean', c = chunk_size),
+            LinearNoBias(dim, heads),
+            Rearrange('b n h -> (b h) n 1')
+        )
+
+        self.to_adaptive_step = nn.Sequential(
+            Reduce('b (n c) ... -> b n ...', 'mean', c = chunk_size),
+            LinearNoBias(dim, heads),
+            Rearrange('b n h -> (b h) n')
+        )
+
+        # weight decay factor
+
+        self.to_decay_factor = nn.Sequential(
+            Reduce('b (n c) ... -> b n ...', 'mean', c = chunk_size),
+            LinearNoBias(dim, heads),
+            Rearrange('b n h -> (b h) n 1')
+        )
+
+        # maybe use accelerated scan
+
+        self.use_accelerated_scan = use_accelerated_scan
+
+    def init_weights_and_momentum(self):
+        params = TensorDict(dict(self.memory_model.named_parameters()))
+
+        init_weights = params.clone().zero_()
+        init_momentum = params.clone().zero_()
+
+        return init_weights, init_momentum
+
+    def store_memories(
+        self,
+        seq,
+        past_state: tuple[dict[str, Tensor], dict[str, Tensor]]
+    ):
+
+        seq = self.store_norm(seq)
+
+        # curtail sequence by multiple of the chunk size
+        # only a complete chunk of the sequence provides the memory for the next chunk
+
+        seq_len, chunk_size = seq.shape[-2], self.chunk_size
+        round_down_seq_len = round_down_multiple(seq_len, self.chunk_size)
+
+        seq = seq[:, :round_down_seq_len]
+
+        # curr weights + past weights, in the case that the initial weights are learned
+
+        curr_weights = TensorDict(dict(self.memory_model.named_parameters()))
+
+        past_state = tuple(TensorDict(d) for d in past_state)
+        past_weights, past_momentum = past_state
+
+        curr_weights = curr_weights + past_weights
+
+        # pack batch and sequence dimension
+
+        adaptive_lr = (self.to_adaptive_step(seq).sigmoid() * -15).exp() # from 1. - 1e-7
+
+        adaptive_momentum = self.to_momentum(seq).sigmoid()
+        decay_factor = self.to_decay_factor(seq).sigmoid()
+
+        # keys and values
+
+        keys, values = self.to_keys_values(seq).chunk(2, dim = -1)
+
+        # maybe multi head
+
+        keys, values = map(self.split_heads, (keys, values))
+
+        batch = keys.shape[0]
+
+        # take care of chunking
+
+        keys, values = tuple(rearrange(t, 'b (n c) d -> (b n) c d', c = self.chunk_size) for t in (keys, values))
+
+        # get grads and extra auxiliary loss (for backwarding through qkv projection in base neural memory module)
+
+        grads, aux_store_loss = self.per_sample_grad_and_value_fn(dict(curr_weights), keys, values)
+
+        grads = TensorDict(grads)
+
+        # restore batch and sequence dimension
+
+        grads = grads.apply(lambda t: rearrange(t, '(b n) ... -> b n ...', b = batch))
+
+        # multiply gradients with learned adaptive step size
+
+        surprises = grads.apply(lambda t: einx.multiply('b n ..., b n -> b n ...', t, -adaptive_lr))
+
+        # determine scan function
+
+        def default_associative_scan(gates, inputs):
+            _, outputs = associative_scan(binary_operator, (gates, inputs))
+            return outputs
+
+        if self.use_accelerated_scan:
+            from accelerated_scan.triton import scan as triton_scan
+            from accelerated_scan.warp import scan as warp_scan
+
+            scan = triton_scan if seq.is_cuda else warp_scan
+
+            def accelerate_scan_fn(gates, inputs):
+                gates = gates.expand_as(inputs)
+                gates, inputs = tuple(rearrange(t, 'b n d -> b d n') for t in (gates, inputs))
+
+                seq_len = gates.shape[-1]
+                next_power_two_seq_len = 2 ** max(5, int(math.ceil(math.log2(seq_len))))
+
+                gates = F.pad(gates, (0, next_power_two_seq_len - seq_len))
+                inputs = F.pad(inputs, (0, next_power_two_seq_len - seq_len))
+
+                outputs = scan(gates, inputs)
+
+                outputs = outputs[..., :seq_len]
+                outputs = rearrange(outputs, 'b d n -> b n d')
+                return outputs
+
+            scan_fn = accelerate_scan_fn
+        else:
+            scan_fn = default_associative_scan
+
+        # momentum + weight decay - momentum is the new contribution, as most linear RNNs have learned forgetting gates
+
+        next_momentum = TensorDict()
+        updates = TensorDict()
+
+        for param_name, surprise in surprises.items():
+
+            surprise, inverse_pack = pack_one_with_inverse(surprise, 'b n *')
+
+            # derive momentum with associative scan - eq (10)
+
+            momentum = scan_fn(adaptive_momentum, surprise) # momentum is S / surprise in the paper
+
+            # use associative scan again for learned forgetting (weight decay) - eq (13)
+
+            update = scan_fn(1. - decay_factor, momentum) # momentum is S / surprise in the paper
+
+            updates[param_name] = inverse_pack(update)
+            next_momentum[param_name] = inverse_pack(momentum)
+
+        # compute the next weight per batch
+
+        last_update = updates.apply(lambda t: t[:, -1])
+
+        next_state = (curr_weights + last_update, next_momentum)
+
+        return updates, next_state, aux_store_loss.mean() / chunk_size
+
+    def retrieve_memories(
+        self,
+        seq,
+        past_weights: dict[str, Tensor] | None = None,
+    ):
+        chunk_size = self.chunk_size
+        seq_len = seq.shape[1]
+
+        seq = self.retrieve_norm(seq)
+
+        assert seq_len >= chunk_size
+
+        seq = seq[:, (chunk_size - 1):]
+        curtailed_seq_len = seq.shape[-2]
+
+        next_seq_len = round_up_multiple(curtailed_seq_len, chunk_size)
+
+        padding = next_seq_len - curtailed_seq_len
+
+        seq = pad_at_dim(seq, (0, padding), dim = 1)
+
+        # the parameters of the memory model stores the memories of the key / values
+        # when the MLP has only 1 weight matrix, it is equivalent to `kv` fast weight memories from linear attention literature (recall fetching of memories is q @ (kv)) / schmidhuber's paper
+
+        curr_weights = TensorDict(dict(self.memory_model.named_parameters()))
+
+        if exists(past_weights):
+            past_weights = TensorDict(past_weights)
+            assert past_weights.keys() == curr_weights.keys()
+
+            curr_weights = curr_weights + past_weights
+
+        # sequence Float['b n d'] to queries
+
+        queries = self.to_queries(seq)
+
+        # maybe multihead
+
+        queries = self.split_heads(queries)
+
+        batch = queries.shape[0]
+
+        # fetch values from memory model
+
+        curr_weights = curr_weights.apply(lambda t: rearrange(t, 'b n ... -> (b n) ...'))
+        queries = rearrange(queries, 'b (n c) d -> (b n) c d', c = chunk_size)
+
+        # forward functional call
+
+        values = functional_call(self.memory_model, dict(curr_weights), queries)
+
+        # reconstitute batch dimension
+
+        values = rearrange(values, '(b n) c d -> b (n c) d', b = batch)
+
+        # maybe merge heads and combine
+
+        values = self.merge_heads(values)
+
+        values = self.combine_heads(values)
+
+        # post norm, somehow could not stabilize this without it, not in paper
+
+        values = self.post_rmsnorm(values)
+
+        # restore
+
+        values = pad_at_dim(values, (chunk_size - 1, 0), dim = 1, value = 0.) # todo, used a learned null memory embedding instead of 0s for retrieving from empty neural memory
+        values = values[:, :-padding]
+
+        return values
+
+    def forward(
+        self,
+        seq,
+        past_state: tuple[dict[str, Tensor], dict[str, Tensor]] | None = None,
+        return_next_memories = False
+    ):
+        batch, seq_len = seq.shape[:2]
+
+        if seq_len < self.chunk_size:
+            return torch.zeros_like(seq)
+
+        if exists(past_state):
+            past_state = tuple(TensorDict(d) for d in past_state)
+
+        if not exists(past_state):
+            past_state = self.init_weights_and_momentum()
+
+        updates, next_memories, aux_kv_mse_loss = self.store_memories(seq, past_state)
+
+        past_weights, _ = past_state
+
+        retrieved = self.retrieve_memories(seq, past_weights + updates)
+
+        if not return_next_memories:
+            return retrieved
+
+        return retrieved, next_memories, aux_kv_mse_loss

titans_pytorch-0.0.14.dist-info/METADATA

@@ -0,0 +1,111 @@
+Metadata-Version: 2.4
+Name: titans-pytorch
+Version: 0.0.14
+Summary: Titans
+Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
+Project-URL: Repository, https://github.com/lucidrains/titans-pytorch
+Author-email: Phil Wang <lucidrains@gmail.com>
+License: MIT License
+        
+        Copyright (c) 2025 Phil Wang
+        
+        Permission is hereby granted, free of charge, to any person obtaining a copy
+        of this software and associated documentation files (the "Software"), to deal
+        in the Software without restriction, including without limitation the rights
+        to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+        copies of the Software, and to permit persons to whom the Software is
+        furnished to do so, subject to the following conditions:
+        
+        The above copyright notice and this permission notice shall be included in all
+        copies or substantial portions of the Software.
+        
+        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+        IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+        FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+        AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+        LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+        OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+        SOFTWARE.
+License-File: LICENSE
+Keywords: artificial intelligence,deep learning,linear attention,neural memory module,test time training
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3.9
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.9
+Requires-Dist: accelerated-scan>=0.2.0
+Requires-Dist: einops>=0.8.0
+Requires-Dist: einx>=0.3.0
+Requires-Dist: ninja
+Requires-Dist: tensordict
+Requires-Dist: torch>=2.2
+Provides-Extra: examples
+Requires-Dist: local-attention>=1.10.1; extra == 'examples'
+Requires-Dist: taylor-series-linear-attention; extra == 'examples'
+Requires-Dist: tqdm; extra == 'examples'
+Requires-Dist: wandb; extra == 'examples'
+Provides-Extra: test
+Requires-Dist: pytest; extra == 'test'
+Description-Content-Type: text/markdown
+
+<img src="./fig2.png" width="400px"></img>
+
+<img src="./fig1.png" width="400px"></img>
+
+## Titans - Pytorch (wip)
+
+Unofficial implementation of [Titans](https://arxiv.org/abs/2501.00663) in Pytorch. Will also contain some explorations into architectures beyond their simple 1-4 layer MLP for the neural memory module, if it works well to any degree.
+
+## Install
+
+```bash
+$ pip install titans-pytorch
+```
+
+## Usage
+
+```python
+import torch
+from titans_pytorch import NeuralMemory
+
+mem = NeuralMemory(
+    dim = 384,
+    chunk_size = 64,
+    pre_rmsnorm = True
+).cuda()
+
+seq = torch.randn(2, 1024, 384).cuda()
+retrieved = mem(seq)
+
+assert seq.shape == retrieved.shape
+```
+
+## Experiments
+
+```bash
+$ pip install .[examples]
+```
+
+For the SOTA linear attention, you will also need to run
+
+```bash
+$ pip install -r requirements.txt
+```
+
+Then modify `train.py` and run it to query nature
+
+```bash
+$ python train.py
+```
+
+## Citations
+
+```bibtex
+@inproceedings{Behrouz2024TitansLT,
+    title   = {Titans: Learning to Memorize at Test Time},
+    author  = {Ali Behrouz and Peilin Zhong and Vahab S. Mirrokni},
+    year    = {2024},
+    url     = {https://api.semanticscholar.org/CorpusID:275212078}
+}
+```

titans_pytorch-0.0.14.dist-info/RECORD

@@ -0,0 +1,7 @@
+titans_pytorch/__init__.py,sha256=QKuJPCOJCdgtaPeKoHEkYkiQe65_LV9_8-cIMbBPU30,55
+titans_pytorch/associative_scan.py,sha256=Y-iYqmFuG-NoCKu6kgql1mhowXTeJfyawi3eUIXamp0,2650
+titans_pytorch/titans.py,sha256=O4WO2I6GhxyZobqbgfFx01saFEKUxhA0BBwt70m2yeQ,12306
+titans_pytorch-0.0.14.dist-info/METADATA,sha256=HKDSJ3sWc54sN1_fOEYU7i5TjiQff49vcZG9G8EU6z4,3598
+titans_pytorch-0.0.14.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+titans_pytorch-0.0.14.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+titans_pytorch-0.0.14.dist-info/RECORD,,

titans_pytorch-0.0.14.dist-info/WHEEL

@@ -0,0 +1,4 @@
+Wheel-Version: 1.0
+Generator: hatchling 1.27.0
+Root-Is-Purelib: true
+Tag: py3-none-any

titans_pytorch-0.0.14.dist-info/licenses/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 Phil Wang
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.