跳转至

embeddings

embeddings 提供了一系列预定义的位置编码。

multimolecule.modules.embeddings

ALiBi

Bases: Module

Attention with Linear Biases (ALiBi).

ALiBi adds a per-head linear penalty on the query-key distance directly to the attention scores: score[h, i, j] -= slope[h] * |i - j|. Unlike :class:RotaryEmbedding (a per-token transform on the query/key) ALiBi is a bias on the score matrix, so the NestedTensor-native form is a FlexAttention score_mod rather than a materialized [seq, seq] tensor.

The module exposes the per-head slopes; consumers apply the bias from those slopes. The NestedTensor-native flash path passes them as alibi_slopes (the fused kernel applies -slope[h] * |i - j| directly); :meth:build_alibi_tensor materializes the equivalent dense [1, num_heads, seq, seq] additive bias for paths that take an explicit attn_mask.

Example

alibi = ALiBi(num_heads=8) bias = alibi.build_alibi_tensor(16) bias.shape torch.Size([1, 8, 16, 16])

Source code in multimolecule/modules/embeddings/alibi.py
Python
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
@POSITION_EMBEDDINGS.register("alibi")
@POSITION_EMBEDDINGS_HF.register("alibi")
class ALiBi(nn.Module):
    """
    Attention with Linear Biases (ALiBi).

    ALiBi adds a per-head linear penalty on the query-key distance directly to the attention
    scores: ``score[h, i, j] -= slope[h] * |i - j|``. Unlike :class:`RotaryEmbedding` (a per-token
    transform on the query/key) ALiBi is a bias on the *score matrix*, so the NestedTensor-native
    form is a [FlexAttention](https://pytorch.org/blog/flexattention/) ``score_mod`` rather than a
    materialized ``[seq, seq]`` tensor.

    The module exposes the per-head ``slopes``; consumers apply the bias from those slopes. The
    NestedTensor-native flash path passes them as ``alibi_slopes`` (the fused kernel applies
    ``-slope[h] * |i - j|`` directly); :meth:`build_alibi_tensor` materializes the equivalent dense
    ``[1, num_heads, seq, seq]`` additive bias for paths that take an explicit ``attn_mask``.

    Example:
        >>> alibi = ALiBi(num_heads=8)
        >>> bias = alibi.build_alibi_tensor(16)
        >>> bias.shape
        torch.Size([1, 8, 16, 16])
    """

    _is_hf_initialized = True

    def __init__(self, num_heads: int, device: torch.device | None = None, dtype: torch.dtype = torch.float32):
        super().__init__()
        self.num_heads = num_heads
        slopes = torch.tensor(self.get_slopes(num_heads), device=device, dtype=dtype)
        self.register_buffer("slopes", slopes, persistent=False)

    @staticmethod
    def get_slopes(num_heads: int) -> list[float]:
        """Geometric sequence of per-head slopes, matching the original ALiBi paper."""

        def slopes_power_of_2(n: int) -> list[float]:
            start = 2 ** (-(2 ** -(math.log2(n) - 3)))
            return [start * start**i for i in range(n)]

        if math.log2(num_heads).is_integer():
            return slopes_power_of_2(num_heads)
        closest = 2 ** math.floor(math.log2(num_heads))
        extra = ALiBi.get_slopes(2 * closest)[0::2][: num_heads - closest]
        return slopes_power_of_2(closest) + extra

    def build_alibi_tensor(
        self, size: int, device: torch.device | None = None, dtype: torch.dtype | None = None
    ) -> Tensor:
        """Dense ``[1, num_heads, size, size]`` additive ALiBi bias for SDPA / eager fallbacks."""
        slopes = self.slopes.to(device=device, dtype=dtype)
        position = torch.arange(size, device=slopes.device)
        relative = (position[None, :] - position[:, None]).abs().to(slopes.dtype)  # (size, size)
        return (-slopes[:, None, None] * relative).unsqueeze(0)  # (1, num_heads, size, size)

get_slopes staticmethod

Python
get_slopes(num_heads: int) -> list[float]

Geometric sequence of per-head slopes, matching the original ALiBi paper.

Source code in multimolecule/modules/embeddings/alibi.py
Python
65
66
67
68
69
70
71
72
73
74
75
76
77
@staticmethod
def get_slopes(num_heads: int) -> list[float]:
    """Geometric sequence of per-head slopes, matching the original ALiBi paper."""

    def slopes_power_of_2(n: int) -> list[float]:
        start = 2 ** (-(2 ** -(math.log2(n) - 3)))
        return [start * start**i for i in range(n)]

    if math.log2(num_heads).is_integer():
        return slopes_power_of_2(num_heads)
    closest = 2 ** math.floor(math.log2(num_heads))
    extra = ALiBi.get_slopes(2 * closest)[0::2][: num_heads - closest]
    return slopes_power_of_2(closest) + extra

build_alibi_tensor

Python
build_alibi_tensor(size: int, device: device | None = None, dtype: dtype | None = None) -> Tensor

Dense [1, num_heads, size, size] additive ALiBi bias for SDPA / eager fallbacks.

Source code in multimolecule/modules/embeddings/alibi.py
Python
79
80
81
82
83
84
85
86
def build_alibi_tensor(
    self, size: int, device: torch.device | None = None, dtype: torch.dtype | None = None
) -> Tensor:
    """Dense ``[1, num_heads, size, size]`` additive ALiBi bias for SDPA / eager fallbacks."""
    slopes = self.slopes.to(device=device, dtype=dtype)
    position = torch.arange(size, device=slopes.device)
    relative = (position[None, :] - position[:, None]).abs().to(slopes.dtype)  # (size, size)
    return (-slopes[:, None, None] * relative).unsqueeze(0)  # (1, num_heads, size, size)

RotaryEmbedding

Bases: Module

Rotary position embeddings based on those in RoFormer.

Query and keys are transformed by rotation matrices which depend on their relative positions.

Cache

The inverse frequency buffer is cached and updated only when the sequence length changes or the device changes.

Sequence Length

Rotary Embedding is irrespective of the sequence length and can be used for any sequence length. Use the scale parameter to extend context length beyond training (e.g., scale=2.0 doubles effective context).

Example

embedding = RotaryEmbedding(embedding_dim=64) query, key = torch.randn(2, 4, 28, 64), torch.randn(2, 4, 28, 64) query, key = embedding(query, key) query.shape torch.Size([2, 4, 28, 64])

For extended context length

embedding_extended = RotaryEmbedding(embedding_dim=64, scale=2.0) embedding.state_dict() # no weight in state_dict OrderedDict()

Source code in multimolecule/modules/embeddings/rotary.py
Python
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
@POSITION_EMBEDDINGS.register("rotary")
@POSITION_EMBEDDINGS_HF.register("rotary")
class RotaryEmbedding(nn.Module):
    """
    Rotary position embeddings based on those in
    [RoFormer](https://huggingface.co/docs/transformers/model_doc/roformer).

    Query and keys are transformed by rotation
    matrices which depend on their relative positions.

    Tip: **Cache**
        The inverse frequency buffer is cached and updated only when the sequence length changes or the device changes.

    Success: **Sequence Length**
        Rotary Embedding is irrespective of the sequence length and can be used for any sequence length.
        Use the `scale` parameter to extend context length beyond training (e.g., scale=2.0 doubles effective context).

    Example:
        >>> embedding = RotaryEmbedding(embedding_dim=64)
        >>> query, key = torch.randn(2, 4, 28, 64), torch.randn(2, 4, 28, 64)
        >>> query, key = embedding(query, key)
        >>> query.shape
        torch.Size([2, 4, 28, 64])
        >>> # For extended context length
        >>> embedding_extended = RotaryEmbedding(embedding_dim=64, scale=2.0)
        >>> embedding.state_dict()  # no weight in state_dict
        OrderedDict()
    """

    _is_hf_initialized = True
    _seq_len_cached: int | None = None
    _cos_cached: Tensor | None = None
    _sin_cached: Tensor | None = None

    def __init__(
        self,
        embedding_dim: int,
        base: float = 10000.0,
        scale: float = 1.0,
        interleaved: bool = False,
        device: torch.device | None = None,
        dtype: torch.dtype = torch.float32,
    ):
        """
        Initialize rotary position embeddings.

        Args:
            embedding_dim: Dimension of the embeddings (must be even)
            base: Base for computing inverse frequencies. Defaults to 10000.0.
            scale: Scaling factor for frequencies. Values > 1.0 extend context length
                   (e.g., scale=2.0 doubles the effective context). Defaults to 1.0.
            interleaved: Whether to rotate adjacent feature pairs. Defaults to False, which preserves the
                   existing split-half layout.
            dtype: Data type for computations. Defaults to torch.float32.
        """
        super().__init__()
        self.embedding_dim = embedding_dim
        self.base = base
        self.scale = scale
        self.interleaved = interleaved
        inv_freq_exponent = torch.arange(0, self.embedding_dim, 2, device=device, dtype=dtype) / self.embedding_dim
        self.register_buffer("inv_freq", 1.0 / (self.base**inv_freq_exponent), persistent=False)
        self._initialized = False

    def forward(self, q: Tensor, k: Tensor, offset: int = 0, seq_length: int | None = None) -> Tuple[Tensor, Tensor]:
        """
        Apply rotary position embeddings to query and key tensors.

        Args:
            q: Query tensor of shape `(batch_size, num_heads, seq_length, embedding_dim)`
            k: Key tensor of shape `(batch_size, num_heads, seq_length, embedding_dim)`
            offset: Position offset for the start of the sequence (used with past_key_values).
                    Defaults to 0.
            seq_length: Full sequence length including offset. If None, uses the sequence length
                    from the input tensors. Required when offset > 0.

        Returns:
            Tuple of (rotated_query, rotated_key) tensors with the same shapes as inputs.
        """
        if not self._initialized:
            inv_freq_exponent = (
                torch.arange(0, self.embedding_dim, 2, device=q.device, dtype=self.inv_freq.dtype) / self.embedding_dim
            )
            self.register_buffer("inv_freq", 1.0 / (self.base**inv_freq_exponent), persistent=False)
            self._initialized = True
        if offset > 0 and seq_length is None:
            raise ValueError("seq_length must be provided when offset > 0")

        if isinstance(q, NestedTensor):
            reference = q[..., 0, :, 0]
            positions = torch.cumsum(torch.ones_like(reference, dtype=torch.long), dim=-1) - 1 + offset
            cos, sin = self.cos_sin_tables(positions, dtype=q.dtype)
            return (
                self.apply_rotary(q, cos, sin, interleaved=self.interleaved),
                self.apply_rotary(k, cos, sin, interleaved=self.interleaved),
            )

        if seq_length is None:
            seq_length = k.shape[-2]

        self._update_cos_sin_tables(k, seq_len_dim=-2, seq_length=seq_length)
        return self.apply_rotary_pos_emb(q, offset=offset), self.apply_rotary_pos_emb(k, offset=offset)

    def _update_cos_sin_tables(
        self, x: Tensor, seq_len_dim: int = 2, seq_length: int | None = None
    ) -> Tuple[Tensor, Tensor]:
        """
        Update cached cos/sin tables for rotary embeddings.

        Args:
            x: Input tensor to determine device and dtype
            seq_len_dim: Dimension containing sequence length (default: -2)
            seq_length: Full sequence length to cache. If None, uses x.shape[seq_len_dim]
        """
        if seq_length is None:
            seq_length = x.shape[seq_len_dim]

        needs_update = (
            seq_length != self._seq_len_cached or self._cos_cached is None or self._cos_cached.device != x.device
        )
        if needs_update:
            self._seq_len_cached = seq_length
            t = torch.arange(seq_length, device=x.device, dtype=self.inv_freq.dtype)
            # Apply scaling: divide frequencies by scale to extend context length
            freqs = torch.outer(t, self.inv_freq) / self.scale
            if self.interleaved:
                emb = freqs.repeat_interleave(2, dim=-1).to(x.device)
            else:
                emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
            self._cos_cached = emb.cos()[None, None, :, :]
            self._sin_cached = emb.sin()[None, None, :, :]
        # At this point, _cos_cached and _sin_cached are guaranteed to be Tensor
        assert self._cos_cached is not None and self._sin_cached is not None
        return self._cos_cached, self._sin_cached

    def apply_rotary_pos_emb(self, x: Tensor, offset: int = 0) -> Tensor:
        """
        Apply rotary position embeddings to a tensor.

        Args:
            x: Input tensor of shape `(batch_size, num_heads, seq_length, embedding_dim)`
            offset: Position offset for the start of the sequence (used with past_key_values).
                    Defaults to 0.

        Returns:
            Rotated tensor with the same shape as input.
        """
        if self._cos_cached is None or self._sin_cached is None:
            raise RuntimeError("Cos/sin tables not initialized. Call forward() or _update_cos_sin_tables() first.")

        cos = self._cos_cached[:, :, offset : offset + x.shape[-2], :]
        sin = self._sin_cached[:, :, offset : offset + x.shape[-2], :]
        rotated = (x * cos) + (self.rotate_half(x, interleaved=self.interleaved) * sin)
        return rotated.type_as(x)

    @staticmethod
    def rotate_half(x: Tensor, interleaved: bool = False) -> Tensor:
        if interleaved:
            return torch.stack((-x[..., 1::2], x[..., ::2]), dim=-1).flatten(-2)
        half = x.shape[-1] // 2
        return torch.stack((-x[..., half:], x[..., :half]), dim=-2).flatten(-2)

    def cos_sin_tables(self, position_ids: Tensor, dtype: torch.dtype | None = None) -> Tuple[Tensor, Tensor]:
        """Compute ``(cos, sin)`` rotary tables for explicit ``position_ids``.

        Unlike :meth:`forward`, positions come from ``position_ids`` rather than an internal
        ``arange`` over the sequence. This is what autoregressive models (KV-cache offsets,
        generation) and packed ``NestedTensor`` inputs (per-sequence positions) require. Works for
        dense ``position_ids`` of shape ``(batch, seq)`` and ragged ``NestedTensor`` positions alike;
        the result broadcasts over the head dimension and is *not* cached (positions vary per call).

        Args:
            position_ids: Integer positions, shape ``(batch, seq)`` (dense or ``NestedTensor``).
            dtype: Optional output dtype for ``cos``/``sin``. Frequencies are always accumulated in the
                   ``inv_freq`` dtype (float32 by default) for numerical stability, then cast.

        Returns:
            Tuple of ``(cos, sin)`` each shaped ``(batch, 1, seq, embedding_dim)``.
        """
        if not self._initialized:
            inv_freq_exponent = (
                torch.arange(0, self.embedding_dim, 2, device=position_ids.device, dtype=self.inv_freq.dtype)
                / self.embedding_dim
            )
            self.register_buffer("inv_freq", 1.0 / (self.base**inv_freq_exponent), persistent=False)
            self._initialized = True
        freqs = position_ids.unsqueeze(-1).to(self.inv_freq.dtype) * self.inv_freq / self.scale
        emb = torch.repeat_interleave(freqs, 2, dim=-1) if self.interleaved else torch.cat((freqs, freqs), dim=-1)
        cos = emb.cos().unsqueeze(1)
        sin = emb.sin().unsqueeze(1)
        if dtype is not None:
            cos, sin = cos.to(dtype), sin.to(dtype)
        return cos, sin

    @staticmethod
    def apply_rotary(x: Tensor, cos: Tensor, sin: Tensor, interleaved: bool = False) -> Tensor:
        """Apply precomputed ``cos``/``sin`` tables to ``x`` (no internal cache).

        Companion to :meth:`cos_sin_tables` for models that precompute the tables once at the model level and
        apply them inside attention (e.g. autoregressive models passing ``position_embeddings`` down).
        """
        return (x * cos) + (RotaryEmbedding.rotate_half(x, interleaved=interleaved) * sin)

__init__

Python
__init__(embedding_dim: int, base: float = 10000.0, scale: float = 1.0, interleaved: bool = False, device: device | None = None, dtype: dtype = float32)

Initialize rotary position embeddings.

Parameters:

Name Type Description Default
embedding_dim
int

Dimension of the embeddings (must be even)

 required
base
float

Base for computing inverse frequencies. Defaults to 10000.0.

 10000.0
scale
float

Scaling factor for frequencies. Values > 1.0 extend context length (e.g., scale=2.0 doubles the effective context). Defaults to 1.0.

 1.0
interleaved
bool

Whether to rotate adjacent feature pairs. Defaults to False, which preserves the existing split-half layout.

 False
dtype
dtype

Data type for computations. Defaults to torch.float32.

 float32
Source code in multimolecule/modules/embeddings/rotary.py
Python
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
def __init__(
    self,
    embedding_dim: int,
    base: float = 10000.0,
    scale: float = 1.0,
    interleaved: bool = False,
    device: torch.device | None = None,
    dtype: torch.dtype = torch.float32,
):
    """
    Initialize rotary position embeddings.

    Args:
        embedding_dim: Dimension of the embeddings (must be even)
        base: Base for computing inverse frequencies. Defaults to 10000.0.
        scale: Scaling factor for frequencies. Values > 1.0 extend context length
               (e.g., scale=2.0 doubles the effective context). Defaults to 1.0.
        interleaved: Whether to rotate adjacent feature pairs. Defaults to False, which preserves the
               existing split-half layout.
        dtype: Data type for computations. Defaults to torch.float32.
    """
    super().__init__()
    self.embedding_dim = embedding_dim
    self.base = base
    self.scale = scale
    self.interleaved = interleaved
    inv_freq_exponent = torch.arange(0, self.embedding_dim, 2, device=device, dtype=dtype) / self.embedding_dim
    self.register_buffer("inv_freq", 1.0 / (self.base**inv_freq_exponent), persistent=False)
    self._initialized = False

forward

Python
forward(q: Tensor, k: Tensor, offset: int = 0, seq_length: int | None = None) -> Tuple[Tensor, Tensor]

Apply rotary position embeddings to query and key tensors.

Parameters:

Name Type Description Default
q
Tensor

Query tensor of shape (batch_size, num_heads, seq_length, embedding_dim)

 required
k
Tensor

Key tensor of shape (batch_size, num_heads, seq_length, embedding_dim)

 required
offset
int

Position offset for the start of the sequence (used with past_key_values). Defaults to 0.

 0
seq_length
int | None

Full sequence length including offset. If None, uses the sequence length from the input tensors. Required when offset > 0.

 None

Returns:

Type Description
Tuple[Tensor, Tensor]

Tuple of (rotated_query, rotated_key) tensors with the same shapes as inputs.
Source code in multimolecule/modules/embeddings/rotary.py
Python
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
def forward(self, q: Tensor, k: Tensor, offset: int = 0, seq_length: int | None = None) -> Tuple[Tensor, Tensor]:
    """
    Apply rotary position embeddings to query and key tensors.

    Args:
        q: Query tensor of shape `(batch_size, num_heads, seq_length, embedding_dim)`
        k: Key tensor of shape `(batch_size, num_heads, seq_length, embedding_dim)`
        offset: Position offset for the start of the sequence (used with past_key_values).
                Defaults to 0.
        seq_length: Full sequence length including offset. If None, uses the sequence length
                from the input tensors. Required when offset > 0.

    Returns:
        Tuple of (rotated_query, rotated_key) tensors with the same shapes as inputs.
    """
    if not self._initialized:
        inv_freq_exponent = (
            torch.arange(0, self.embedding_dim, 2, device=q.device, dtype=self.inv_freq.dtype) / self.embedding_dim
        )
        self.register_buffer("inv_freq", 1.0 / (self.base**inv_freq_exponent), persistent=False)
        self._initialized = True
    if offset > 0 and seq_length is None:
        raise ValueError("seq_length must be provided when offset > 0")

    if isinstance(q, NestedTensor):
        reference = q[..., 0, :, 0]
        positions = torch.cumsum(torch.ones_like(reference, dtype=torch.long), dim=-1) - 1 + offset
        cos, sin = self.cos_sin_tables(positions, dtype=q.dtype)
        return (
            self.apply_rotary(q, cos, sin, interleaved=self.interleaved),
            self.apply_rotary(k, cos, sin, interleaved=self.interleaved),
        )

    if seq_length is None:
        seq_length = k.shape[-2]

    self._update_cos_sin_tables(k, seq_len_dim=-2, seq_length=seq_length)
    return self.apply_rotary_pos_emb(q, offset=offset), self.apply_rotary_pos_emb(k, offset=offset)

apply_rotary_pos_emb

Python
apply_rotary_pos_emb(x: Tensor, offset: int = 0) -> Tensor

Apply rotary position embeddings to a tensor.

Parameters:

Name Type Description Default
x
Tensor

Input tensor of shape (batch_size, num_heads, seq_length, embedding_dim)

 required
offset
int

Position offset for the start of the sequence (used with past_key_values). Defaults to 0.

 0

Returns:

Type Description
Tensor

Rotated tensor with the same shape as input.
Source code in multimolecule/modules/embeddings/rotary.py
Python
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
def apply_rotary_pos_emb(self, x: Tensor, offset: int = 0) -> Tensor:
    """
    Apply rotary position embeddings to a tensor.

    Args:
        x: Input tensor of shape `(batch_size, num_heads, seq_length, embedding_dim)`
        offset: Position offset for the start of the sequence (used with past_key_values).
                Defaults to 0.

    Returns:
        Rotated tensor with the same shape as input.
    """
    if self._cos_cached is None or self._sin_cached is None:
        raise RuntimeError("Cos/sin tables not initialized. Call forward() or _update_cos_sin_tables() first.")

    cos = self._cos_cached[:, :, offset : offset + x.shape[-2], :]
    sin = self._sin_cached[:, :, offset : offset + x.shape[-2], :]
    rotated = (x * cos) + (self.rotate_half(x, interleaved=self.interleaved) * sin)
    return rotated.type_as(x)

cos_sin_tables

Python
cos_sin_tables(position_ids: Tensor, dtype: dtype | None = None) -> Tuple[Tensor, Tensor]

Compute (cos, sin) rotary tables for explicit position_ids.

Unlike :meth:forward, positions come from position_ids rather than an internal arange over the sequence. This is what autoregressive models (KV-cache offsets, generation) and packed NestedTensor inputs (per-sequence positions) require. Works for dense position_ids of shape (batch, seq) and ragged NestedTensor positions alike; the result broadcasts over the head dimension and is not cached (positions vary per call).

Parameters:

Name Type Description Default
position_ids
Tensor

Integer positions, shape (batch, seq) (dense or NestedTensor).

 required
dtype
dtype | None

Optional output dtype for cos/sin. Frequencies are always accumulated in the inv_freq dtype (float32 by default) for numerical stability, then cast.

 None

Returns:

Type Description
Tuple[Tensor, Tensor]

Tuple of (cos, sin) each shaped (batch, 1, seq, embedding_dim).
Source code in multimolecule/modules/embeddings/rotary.py
Python
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
def cos_sin_tables(self, position_ids: Tensor, dtype: torch.dtype | None = None) -> Tuple[Tensor, Tensor]:
    """Compute ``(cos, sin)`` rotary tables for explicit ``position_ids``.

    Unlike :meth:`forward`, positions come from ``position_ids`` rather than an internal
    ``arange`` over the sequence. This is what autoregressive models (KV-cache offsets,
    generation) and packed ``NestedTensor`` inputs (per-sequence positions) require. Works for
    dense ``position_ids`` of shape ``(batch, seq)`` and ragged ``NestedTensor`` positions alike;
    the result broadcasts over the head dimension and is *not* cached (positions vary per call).

    Args:
        position_ids: Integer positions, shape ``(batch, seq)`` (dense or ``NestedTensor``).
        dtype: Optional output dtype for ``cos``/``sin``. Frequencies are always accumulated in the
               ``inv_freq`` dtype (float32 by default) for numerical stability, then cast.

    Returns:
        Tuple of ``(cos, sin)`` each shaped ``(batch, 1, seq, embedding_dim)``.
    """
    if not self._initialized:
        inv_freq_exponent = (
            torch.arange(0, self.embedding_dim, 2, device=position_ids.device, dtype=self.inv_freq.dtype)
            / self.embedding_dim
        )
        self.register_buffer("inv_freq", 1.0 / (self.base**inv_freq_exponent), persistent=False)
        self._initialized = True
    freqs = position_ids.unsqueeze(-1).to(self.inv_freq.dtype) * self.inv_freq / self.scale
    emb = torch.repeat_interleave(freqs, 2, dim=-1) if self.interleaved else torch.cat((freqs, freqs), dim=-1)
    cos = emb.cos().unsqueeze(1)
    sin = emb.sin().unsqueeze(1)
    if dtype is not None:
        cos, sin = cos.to(dtype), sin.to(dtype)
    return cos, sin

apply_rotary staticmethod

Python
apply_rotary(x: Tensor, cos: Tensor, sin: Tensor, interleaved: bool = False) -> Tensor

Apply precomputed cos/sin tables to x (no internal cache).

Companion to :meth:cos_sin_tables for models that precompute the tables once at the model level and apply them inside attention (e.g. autoregressive models passing position_embeddings down).

Source code in multimolecule/modules/embeddings/rotary.py
Python
228
229
230
231
232
233
234
235
@staticmethod
def apply_rotary(x: Tensor, cos: Tensor, sin: Tensor, interleaved: bool = False) -> Tensor:
    """Apply precomputed ``cos``/``sin`` tables to ``x`` (no internal cache).

    Companion to :meth:`cos_sin_tables` for models that precompute the tables once at the model level and
    apply them inside attention (e.g. autoregressive models passing ``position_embeddings`` down).
    """
    return (x * cos) + (RotaryEmbedding.rotate_half(x, interleaved=interleaved) * sin)

SinusoidalEmbedding

Bases: Embedding

Sinusoidal positional embeddings for inputs with any length.

Freezing

The embeddings are frozen and cannot be trained. They will not be saved in the model’s state_dict.

Padding Idx

Padding symbols are ignored if the padding_idx is specified.

Sequence Length

These embeddings get automatically extended in forward if more positions is needed.

Parameters:

Name Type Description Default

num_embeddings

int

The number of embeddings to use.

 required

embedding_dim

int

The dimension of the embeddings.

 required

padding_idx

int | None

The index of the padding symbol.

 None

bias

int

The bias of the embeddings.

 1
Example

embedding = SinusoidalEmbedding(num_embeddings=128, embedding_dim=64) input_ids = torch.arange(28).repeat(4).view(4, -1) input_embeds = torch.randn(4, 28, 64) embeddings = embedding(input_ids) embeddings.shape # no batch dimension if padding_idx is None torch.Size([28, 64]) input_embeds = input_embeds + embeddings input_embeds.shape torch.Size([4, 28, 64]) embedding = SinusoidalEmbedding(num_embeddings=128, embedding_dim=64, padding_idx=0) embeddings = embedding(input_ids) embeddings.shape # batch dimension if padding_idx is not None torch.Size([4, 28, 64]) embedding.state_dict() # no weight in state_dict OrderedDict()

Source code in multimolecule/modules/embeddings/sinusoidal.py
Python
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
@POSITION_EMBEDDINGS.register("sinusoidal")
@POSITION_EMBEDDINGS_HF.register("sinusoidal")
class SinusoidalEmbedding(nn.Embedding):
    r"""
    Sinusoidal positional embeddings for inputs with any length.

    Note: **Freezing**
        The embeddings are frozen and cannot be trained.
        They will not be saved in the model's state_dict.

    Tip: **Padding Idx**
        Padding symbols are ignored if the padding_idx is specified.

    Success: **Sequence Length**
        These embeddings get automatically extended in forward if more positions is needed.

    Args:
        num_embeddings: The number of embeddings to use.
        embedding_dim: The dimension of the embeddings.
        padding_idx: The index of the padding symbol.
        bias: The bias of the embeddings.

    Example:
        >>> embedding = SinusoidalEmbedding(num_embeddings=128, embedding_dim=64)
        >>> input_ids = torch.arange(28).repeat(4).view(4, -1)
        >>> input_embeds = torch.randn(4, 28, 64)
        >>> embeddings = embedding(input_ids)
        >>> embeddings.shape  # no batch dimension if padding_idx is None
        torch.Size([28, 64])
        >>> input_embeds = input_embeds + embeddings
        >>> input_embeds.shape
        torch.Size([4, 28, 64])
        >>> embedding = SinusoidalEmbedding(num_embeddings=128, embedding_dim=64, padding_idx=0)
        >>> embeddings = embedding(input_ids)
        >>> embeddings.shape  # batch dimension if padding_idx is not None
        torch.Size([4, 28, 64])
        >>> embedding.state_dict()  # no weight in state_dict
        OrderedDict()
    """

    _is_hf_initialized = True

    def __init__(
        self,
        num_embeddings: int,
        embedding_dim: int,
        padding_idx: int | None = None,
        bias: int = 1,
        device: torch.device | None = None,
        dtype: torch.dtype = torch.float32,
        **kwargs,
    ):
        weight = self.get_embedding(num_embeddings, embedding_dim, padding_idx, device=device, dtype=dtype)
        super().__init__(num_embeddings, embedding_dim, padding_idx, _weight=weight.detach(), _freeze=True, **kwargs)
        del self.weight
        self.register_buffer("weight", weight, persistent=False)
        self.bias = bias
        self._initialized = False

    def update_weight(self, num_embeddings: int):
        weight = self.get_embedding(
            num_embeddings, self.embedding_dim, self.padding_idx, dtype=self.weight.dtype, device=self.weight.device
        )
        self.register_buffer("weight", weight, persistent=False)

    @staticmethod
    def get_embedding(
        num_embeddings: int,
        embedding_dim: int,
        padding_idx: int | None = None,
        device: torch.device | None = None,
        dtype: torch.dtype = torch.float32,
    ) -> Tensor:
        """
        Build sinusoidal embeddings.

        This matches the implementation in tensor2tensor, but differs slightly from the description in Section 3.5 of
        "Attention Is All You Need".
        """
        if device is None:
            device = get_default_device()
        half_dim = embedding_dim // 2
        emb = torch.exp(torch.arange(half_dim, device=device, dtype=dtype) * -(math.log(10000) / (half_dim - 1)))
        emb = torch.arange(num_embeddings, device=device, dtype=dtype).unsqueeze(1) * emb.unsqueeze(0)
        emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1).view(num_embeddings, -1)
        if embedding_dim % 2 == 1:
            emb = torch.cat([emb, torch.zeros(num_embeddings, 1, dtype=dtype, device=device)], dim=1)
        if padding_idx is not None:
            emb[padding_idx, :] = 0
        return emb.detach()

    @staticmethod
    def get_position_ids(tensor: Tensor, padding_idx: int | None = None):
        """
        Replace non-padding symbols with their position numbers.

        Position numbers begin at padding_idx+1. Padding symbols are ignored.
        """
        # The series of casts and type-conversions here are carefully
        # balanced to both work with ONNX export and XLA. In particular XLA
        # prefers ints, cumsum defaults to output longs, and ONNX doesn't know
        # how to handle the dtype kwarg in cumsum.
        if padding_idx is None:
            return torch.cumsum(tensor.new_ones(tensor.size(1), dtype=torch.long), dim=0) - 1
        mask = tensor.ne(padding_idx).long()
        return torch.cumsum(mask, dim=1, dtype=torch.long) * mask + padding_idx

    def forward(self, input_ids: Tensor) -> Tensor:
        if not self._initialized:
            self.update_weight(self.num_embeddings)
            self._initialized = True
        _, seq_length = input_ids.shape[:2]
        # expand embeddings if needed
        max_position = seq_length + self.bias + 1
        if self.padding_idx is not None:
            max_position += self.padding_idx
        if max_position > self.weight.size(0):
            self.update_weight(max_position)
        # Need to shift the position ids by the padding index
        position_ids = self.get_position_ids(input_ids, self.padding_idx) + self.bias
        return super().forward(position_ids)

get_embedding staticmethod

Python
get_embedding(num_embeddings: int, embedding_dim: int, padding_idx: int | None = None, device: device | None = None, dtype: dtype = float32) -> Tensor

Build sinusoidal embeddings.

This matches the implementation in tensor2tensor, but differs slightly from the description in Section 3.5 of “Attention Is All You Need”.

Source code in multimolecule/modules/embeddings/sinusoidal.py
Python
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
@staticmethod
def get_embedding(
    num_embeddings: int,
    embedding_dim: int,
    padding_idx: int | None = None,
    device: torch.device | None = None,
    dtype: torch.dtype = torch.float32,
) -> Tensor:
    """
    Build sinusoidal embeddings.

    This matches the implementation in tensor2tensor, but differs slightly from the description in Section 3.5 of
    "Attention Is All You Need".
    """
    if device is None:
        device = get_default_device()
    half_dim = embedding_dim // 2
    emb = torch.exp(torch.arange(half_dim, device=device, dtype=dtype) * -(math.log(10000) / (half_dim - 1)))
    emb = torch.arange(num_embeddings, device=device, dtype=dtype).unsqueeze(1) * emb.unsqueeze(0)
    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1).view(num_embeddings, -1)
    if embedding_dim % 2 == 1:
        emb = torch.cat([emb, torch.zeros(num_embeddings, 1, dtype=dtype, device=device)], dim=1)
    if padding_idx is not None:
        emb[padding_idx, :] = 0
    return emb.detach()

get_position_ids staticmethod

Python
get_position_ids(tensor: Tensor, padding_idx: int | None = None)

Replace non-padding symbols with their position numbers.

Position numbers begin at padding_idx+1. Padding symbols are ignored.

Source code in multimolecule/modules/embeddings/sinusoidal.py
Python
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
@staticmethod
def get_position_ids(tensor: Tensor, padding_idx: int | None = None):
    """
    Replace non-padding symbols with their position numbers.

    Position numbers begin at padding_idx+1. Padding symbols are ignored.
    """
    # The series of casts and type-conversions here are carefully
    # balanced to both work with ONNX export and XLA. In particular XLA
    # prefers ints, cumsum defaults to output longs, and ONNX doesn't know
    # how to handle the dtype kwarg in cumsum.
    if padding_idx is None:
        return torch.cumsum(tensor.new_ones(tensor.size(1), dtype=torch.long), dim=0) - 1
    mask = tensor.ne(padding_idx).long()
    return torch.cumsum(mask, dim=1, dtype=torch.long) * mask + padding_idx