Skip to content

layer

Classes

fastvideo.attention.layer.DistributedAttention 

DistributedAttention(num_heads: int, head_size: int, num_kv_heads: int | None = None, softmax_scale: float | None = None, causal: bool = False, supported_attention_backends: tuple[AttentionBackendEnum, ...] | None = None, prefix: str = '', **extra_impl_args)

Bases: Module

Distributed attention layer.

Source code in fastvideo/attention/layer.py 
21
22
23
24
25
26
27
28
29
30
31
32
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
def __init__(self,
             num_heads: int,
             head_size: int,
             num_kv_heads: int | None = None,
             softmax_scale: float | None = None,
             causal: bool = False,
             supported_attention_backends: tuple[AttentionBackendEnum, ...]
             | None = None,
             prefix: str = "",
             **extra_impl_args) -> None:
    super().__init__()
    if softmax_scale is None:
        self.softmax_scale = head_size**-0.5
    else:
        self.softmax_scale = softmax_scale

    if num_kv_heads is None:
        num_kv_heads = num_heads

    dtype = get_compute_dtype()
    attn_backend = get_attn_backend(
        head_size,
        dtype,
        supported_attention_backends=supported_attention_backends)
    impl_cls = attn_backend.get_impl_cls()
    self.attn_impl = impl_cls(num_heads=num_heads,
                              head_size=head_size,
                              causal=causal,
                              softmax_scale=self.softmax_scale,
                              num_kv_heads=num_kv_heads,
                              prefix=f"{prefix}.impl",
                              **extra_impl_args)
    self.num_heads = num_heads
    self.head_size = head_size
    self.num_kv_heads = num_kv_heads
    self.backend = backend_name_to_enum(attn_backend.get_name())
    self.dtype = dtype

Functions

fastvideo.attention.layer.DistributedAttention.forward 
forward(q: Tensor, k: Tensor, v: Tensor, replicated_q: Tensor | None = None, replicated_k: Tensor | None = None, replicated_v: Tensor | None = None, freqs_cis: tuple[Tensor, Tensor] | None = None, attention_mask: Tensor | None = None) -> tuple[Tensor, Tensor | None]

Forward pass for distributed attention.

Parameters:

Name Type Description Default
q Tensor

Query tensor [batch_size, seq_len, num_heads, head_dim]

 required
k Tensor

Key tensor [batch_size, seq_len, num_heads, head_dim]

 required
v Tensor

Value tensor [batch_size, seq_len, num_heads, head_dim]

 required
replicated_q Optional[Tensor]

Replicated query tensor, typically for text tokens

 None
replicated_k Optional[Tensor]

Replicated key tensor

 None
replicated_v Optional[Tensor]

Replicated value tensor

 None
attention_mask Optional[Tensor]

Attention mask [batch_size, seq_len]

 None

Returns:

Type Description
tuple[Tensor, Tensor | None]

Tuple[torch.Tensor, Optional[torch.Tensor]]: A tuple containing: - o (torch.Tensor): Output tensor after attention for the main sequence - replicated_o (Optional[torch.Tensor]): Output tensor for replicated tokens, if provided
Source code in fastvideo/attention/layer.py 
 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
@torch.compiler.disable
def forward(
    self,
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    replicated_q: torch.Tensor | None = None,
    replicated_k: torch.Tensor | None = None,
    replicated_v: torch.Tensor | None = None,
    freqs_cis: tuple[torch.Tensor, torch.Tensor] | None = None,
    attention_mask: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor | None]:
    """Forward pass for distributed attention.

    Args:
        q (torch.Tensor): Query tensor [batch_size, seq_len, num_heads, head_dim]
        k (torch.Tensor): Key tensor [batch_size, seq_len, num_heads, head_dim]
        v (torch.Tensor): Value tensor [batch_size, seq_len, num_heads, head_dim]
        replicated_q (Optional[torch.Tensor]): Replicated query tensor, typically for text tokens
        replicated_k (Optional[torch.Tensor]): Replicated key tensor
        replicated_v (Optional[torch.Tensor]): Replicated value tensor
        attention_mask (Optional[torch.Tensor]): Attention mask [batch_size, seq_len]

    Returns:
        Tuple[torch.Tensor, Optional[torch.Tensor]]: A tuple containing:
            - o (torch.Tensor): Output tensor after attention for the main sequence
            - replicated_o (Optional[torch.Tensor]): Output tensor for replicated tokens, if provided
    """
    # Check input shapes
    assert q.dim() == 4 and k.dim() == 4 and v.dim(
    ) == 4, "Expected 4D tensors"
    batch_size, seq_len, num_heads, head_dim = q.shape
    local_rank = get_sp_parallel_rank()
    world_size = get_sp_world_size()

    forward_context: ForwardContext = get_forward_context()
    ctx_attn_metadata = forward_context.attn_metadata

    # Stack QKV
    qkv = torch.cat([q, k, v],
                    dim=0)  # [3*batch, seq_len, num_heads, head_dim]

    # Redistribute heads across sequence dimension
    qkv = sequence_model_parallel_all_to_all_4D(qkv,
                                                scatter_dim=2,
                                                gather_dim=1)

    # After all-to-all, each rank has the full sequence but only a subset of heads
    # The attention mask should now apply to the full sequence length
    # Since mask is [batch, full_seq_len], it's already in the correct format

    # LOAY TODO, instead of slicing repeatedly maintain an original qkv and rewrite into that
    valid_seq_len = None
    if attention_mask is not None:
        valid_seq_len = (attention_mask[0] == 1).sum().item()
        qkv = qkv[:, :valid_seq_len, :, :]

    if freqs_cis is not None:
        cos, sin = freqs_cis
        qkv[:batch_size * 2] = _apply_rotary_emb(qkv[:batch_size * 2],
                                                 cos,
                                                 sin,
                                                 is_neox_style=False)
    # Apply backend-specific preprocess_qkv
    qkv = self.attn_impl.preprocess_qkv(qkv, ctx_attn_metadata)

    # Concatenate with replicated QKV if provided
    if replicated_q is not None:
        assert replicated_k is not None and replicated_v is not None
        replicated_qkv = torch.cat(
            [replicated_q, replicated_k, replicated_v],
            dim=0)  # [3, seq_len, num_heads, head_dim]
        heads_per_rank = num_heads // world_size
        replicated_qkv = replicated_qkv[:, :, local_rank *
                                        heads_per_rank:(local_rank + 1) *
                                        heads_per_rank]
        qkv = torch.cat([qkv, replicated_qkv], dim=1)

    q, k, v = qkv.chunk(3, dim=0)

    output = self.attn_impl.forward(q, k, v, ctx_attn_metadata)

    # Redistribute back if using sequence parallelism
    replicated_output = None
    if replicated_q is not None:
        split_idx = seq_len * world_size if valid_seq_len is None else valid_seq_len
        replicated_output = output[:, split_idx:]
        output = output[:, :split_idx]
        # TODO: make this asynchronous
        replicated_output = sequence_model_parallel_all_gather(
            replicated_output.contiguous(), dim=2)
    # Apply backend-specific postprocess_output
    output = self.attn_impl.postprocess_output(output, ctx_attn_metadata)

    if attention_mask is not None:
        pad_len = (attention_mask[0] == 0).sum().item()
        output = torch.nn.functional.pad(output, (0, 0, 0, 0, 0, pad_len))

    output = sequence_model_parallel_all_to_all_4D(output,
                                                   scatter_dim=1,
                                                   gather_dim=2)

    return output, replicated_output

fastvideo.attention.layer.DistributedAttention_VSA 

DistributedAttention_VSA(num_heads: int, head_size: int, num_kv_heads: int | None = None, softmax_scale: float | None = None, causal: bool = False, supported_attention_backends: tuple[AttentionBackendEnum, ...] | None = None, prefix: str = '', **extra_impl_args)

Bases: DistributedAttention

Distributed attention layer with VSA support.

Source code in fastvideo/attention/layer.py 
21
22
23
24
25
26
27
28
29
30
31
32
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
def __init__(self,
             num_heads: int,
             head_size: int,
             num_kv_heads: int | None = None,
             softmax_scale: float | None = None,
             causal: bool = False,
             supported_attention_backends: tuple[AttentionBackendEnum, ...]
             | None = None,
             prefix: str = "",
             **extra_impl_args) -> None:
    super().__init__()
    if softmax_scale is None:
        self.softmax_scale = head_size**-0.5
    else:
        self.softmax_scale = softmax_scale

    if num_kv_heads is None:
        num_kv_heads = num_heads

    dtype = get_compute_dtype()
    attn_backend = get_attn_backend(
        head_size,
        dtype,
        supported_attention_backends=supported_attention_backends)
    impl_cls = attn_backend.get_impl_cls()
    self.attn_impl = impl_cls(num_heads=num_heads,
                              head_size=head_size,
                              causal=causal,
                              softmax_scale=self.softmax_scale,
                              num_kv_heads=num_kv_heads,
                              prefix=f"{prefix}.impl",
                              **extra_impl_args)
    self.num_heads = num_heads
    self.head_size = head_size
    self.num_kv_heads = num_kv_heads
    self.backend = backend_name_to_enum(attn_backend.get_name())
    self.dtype = dtype

Functions

fastvideo.attention.layer.DistributedAttention_VSA.forward 
forward(q: Tensor, k: Tensor, v: Tensor, replicated_q: Tensor | None = None, replicated_k: Tensor | None = None, replicated_v: Tensor | None = None, gate_compress: Tensor | None = None, freqs_cis: tuple[Tensor, Tensor] | None = None, attention_mask: Tensor | None = None) -> tuple[Tensor, Tensor | None]

Forward pass for distributed attention.

Parameters:

Name Type Description Default
q Tensor

Query tensor [batch_size, seq_len, num_heads, head_dim]

 required
k Tensor

Key tensor [batch_size, seq_len, num_heads, head_dim]

 required
v Tensor

Value tensor [batch_size, seq_len, num_heads, head_dim]

 required
gate_compress Tensor

Gate compress tensor [batch_size, seq_len, num_heads, head_dim]

 None
replicated_q Optional[Tensor]

Replicated query tensor, typically for text tokens

 None
replicated_k Optional[Tensor]

Replicated key tensor

 None
replicated_v Optional[Tensor]

Replicated value tensor

 None
attention_mask Optional[Tensor]

Attention mask [batch_size, seq_len]

 None

Returns:

Type Description
tuple[Tensor, Tensor | None]

Tuple[torch.Tensor, Optional[torch.Tensor]]: A tuple containing: - o (torch.Tensor): Output tensor after attention for the main sequence - replicated_o (Optional[torch.Tensor]): Output tensor for replicated tokens, if provided
Source code in fastvideo/attention/layer.py 
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
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
@torch.compiler.disable
def forward(
    self,
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    replicated_q: torch.Tensor | None = None,
    replicated_k: torch.Tensor | None = None,
    replicated_v: torch.Tensor | None = None,
    gate_compress: torch.Tensor | None = None,
    freqs_cis: tuple[torch.Tensor, torch.Tensor] | None = None,
    attention_mask: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor | None]:
    """Forward pass for distributed attention.

    Args:
        q (torch.Tensor): Query tensor [batch_size, seq_len, num_heads, head_dim]
        k (torch.Tensor): Key tensor [batch_size, seq_len, num_heads, head_dim]
        v (torch.Tensor): Value tensor [batch_size, seq_len, num_heads, head_dim]
        gate_compress (torch.Tensor): Gate compress tensor [batch_size, seq_len, num_heads, head_dim]
        replicated_q (Optional[torch.Tensor]): Replicated query tensor, typically for text tokens
        replicated_k (Optional[torch.Tensor]): Replicated key tensor
        replicated_v (Optional[torch.Tensor]): Replicated value tensor
        attention_mask (Optional[torch.Tensor]): Attention mask [batch_size, seq_len]

    Returns:
        Tuple[torch.Tensor, Optional[torch.Tensor]]: A tuple containing:
            - o (torch.Tensor): Output tensor after attention for the main sequence
            - replicated_o (Optional[torch.Tensor]): Output tensor for replicated tokens, if provided
    """
    # Check text tokens are not supported for VSA now
    assert replicated_q is None and replicated_k is None and replicated_v is None, "Replicated QKV is not supported for VSA now"
    # Check input shapes
    assert q.dim() == 4 and k.dim() == 4 and v.dim(
    ) == 4, "Expected 4D tensors"

    forward_context: ForwardContext = get_forward_context()
    ctx_attn_metadata = forward_context.attn_metadata

    batch_size, seq_len, num_heads, head_dim = q.shape
    # Stack QKV
    qkvg = torch.cat([q, k, v, gate_compress],
                     dim=0)  # [4*batch, seq_len, num_heads, head_dim]

    # Redistribute heads across sequence dimension
    # Before: [4*batch, shard_seq_len, num_heads, head_dim]
    # After:  [4*batch, full_seq_len, shard_num_heads, head_dim]
    qkvg = sequence_model_parallel_all_to_all_4D(qkvg,
                                                 scatter_dim=2,
                                                 gather_dim=1)

    # After all-to-all, each rank has the full sequence but only a subset of heads
    # The attention mask should now apply to the full sequence length

    if attention_mask is not None:
        valid_seq_len = (attention_mask[0] == 1).sum().item()
        qkvg = qkvg[:, :valid_seq_len, :, :]

    if freqs_cis is not None:
        cos, sin = freqs_cis
        qkvg[:batch_size * 2] = _apply_rotary_emb(qkvg[:batch_size * 2],
                                                  cos,
                                                  sin,
                                                  is_neox_style=False)

    qkvg = self.attn_impl.preprocess_qkv(qkvg, ctx_attn_metadata)

    q, k, v, gate_compress = qkvg.chunk(4, dim=0)
    output = self.attn_impl.forward(
        q, k, v, gate_compress, ctx_attn_metadata)  # type: ignore[call-arg]

    # Redistribute back if using sequence parallelism
    replicated_output = None

    # Apply backend-specific postprocess_output
    output = self.attn_impl.postprocess_output(output, ctx_attn_metadata)

    if attention_mask is not None:
        pad_len = (attention_mask[0] == 0).sum().item()
        output = torch.nn.functional.pad(output, (0, 0, 0, 0, 0, pad_len))

    output = sequence_model_parallel_all_to_all_4D(output,
                                                   scatter_dim=1,
                                                   gather_dim=2)
    return output, replicated_output

fastvideo.attention.layer.LocalAttention 

LocalAttention(num_heads: int, head_size: int, num_kv_heads: int | None = None, softmax_scale: float | None = None, causal: bool = False, supported_attention_backends: tuple[AttentionBackendEnum, ...] | None = None, **extra_impl_args)

Bases: Module

Attention layer.

Source code in fastvideo/attention/layer.py 
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
def __init__(self,
             num_heads: int,
             head_size: int,
             num_kv_heads: int | None = None,
             softmax_scale: float | None = None,
             causal: bool = False,
             supported_attention_backends: tuple[AttentionBackendEnum, ...]
             | None = None,
             **extra_impl_args) -> None:
    super().__init__()
    if softmax_scale is None:
        self.softmax_scale = head_size**-0.5
    else:
        self.softmax_scale = softmax_scale
    if num_kv_heads is None:
        num_kv_heads = num_heads

    dtype = get_compute_dtype()
    attn_backend = get_attn_backend(
        head_size,
        dtype,
        supported_attention_backends=supported_attention_backends)
    impl_cls = attn_backend.get_impl_cls()
    self.attn_impl = impl_cls(num_heads=num_heads,
                              head_size=head_size,
                              softmax_scale=self.softmax_scale,
                              num_kv_heads=num_kv_heads,
                              causal=causal,
                              **extra_impl_args)
    self.num_heads = num_heads
    self.head_size = head_size
    self.num_kv_heads = num_kv_heads
    self.backend = backend_name_to_enum(attn_backend.get_name())
    self.dtype = dtype

Functions

fastvideo.attention.layer.LocalAttention.forward 
forward(q: Tensor, k: Tensor, v: Tensor, freqs_cis: tuple[Tensor, Tensor] | None = None) -> Tensor

Apply local attention between query, key and value tensors.

Parameters:

Name Type Description Default
q Tensor

Query tensor of shape [batch_size, seq_len, num_heads, head_dim]

 required
k Tensor

Key tensor of shape [batch_size, seq_len, num_heads, head_dim]

required
v Tensor

Value tensor of shape [batch_size, seq_len, num_heads, head_dim]

 required

Returns:

Type Description
Tensor

torch.Tensor: Output tensor after local attention
Source code in fastvideo/attention/layer.py 
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
def forward(
    self,
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    freqs_cis: tuple[torch.Tensor, torch.Tensor] | None = None,
) -> torch.Tensor:
    """
    Apply local attention between query, key and value tensors.

    Args:
        q (torch.Tensor): Query tensor of shape [batch_size, seq_len, num_heads, head_dim]
        k (torch.Tensor): Key tensor of shape [batch_size, seq_len, num_heads, head_dim] 
        v (torch.Tensor): Value tensor of shape [batch_size, seq_len, num_heads, head_dim]

    Returns:
        torch.Tensor: Output tensor after local attention
    """
    # Check input shapes
    assert q.dim() == 4 and k.dim() == 4 and v.dim(
    ) == 4, "Expected 4D tensors"

    forward_context: ForwardContext = get_forward_context()
    ctx_attn_metadata = forward_context.attn_metadata

    if freqs_cis is not None:
        cos, sin = freqs_cis
        q = _apply_rotary_emb(q, cos, sin, is_neox_style=False)
        k = _apply_rotary_emb(k, cos, sin, is_neox_style=False)

    output = self.attn_impl.forward(q, k, v, ctx_attn_metadata)
    return output

Functions