File size: 7,946 Bytes
f8d22a5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
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
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
"""Attention backend shim — switchable between Flash Attention 2 and 4.

Exports a single ``flash_attn_varlen_func`` with the FA2 calling convention.
The underlying kernel is selected at runtime via ``set_attn_backend(name)``
(default: ``"flash2"``). The selected kernel is resolved lazily on the first
call so model-config-driven selection (which happens after this module is
imported) takes effect.

Modules that previously did ``from flash_attn import flash_attn_varlen_func``
should import from here instead.

For the FA4 path, calling-convention differences are normalised:

* ``window_size=(-1, -1)`` (FA2 "no window") -> ``(None, None)`` (FA4).
* ``block_table`` -> ``page_table``.
* FA4's optional ``(out, lse)`` tuple return is unwrapped to ``out``.
* ``dropout_p>0`` / ``alibi_slopes`` / ``return_attn_probs`` raise on FA4.
"""

from __future__ import annotations

from typing import Any, Callable

_FA2_ALIASES = {"flash2", "fa2", "flash_attention_2", "flash_attn_2"}
_FA4_ALIASES = {"flash4", "fa4", "flash_attention_4", "flash_attn_4"}
_SDPA_ALIASES = {"sdpa", "torch_sdpa", "scaled_dot_product_attention"}

_BACKEND: str = "flash2"
_RESOLVED_FN: Callable[..., Any] | None = None


def _normalize(name: str) -> str:
    n = name.lower().strip()
    if n in _FA2_ALIASES:
        return "flash2"
    if n in _FA4_ALIASES:
        return "flash4"
    if n in _SDPA_ALIASES:
        return "sdpa"
    raise ValueError(
        f"Unknown attention backend {name!r}; expected one of "
        f"{sorted(_FA2_ALIASES | _FA4_ALIASES | _SDPA_ALIASES)}"
    )


def set_attn_backend(name: str) -> None:
    """Select the flash-attn backend used by ``flash_attn_varlen_func``.

    Safe to call multiple times; clears the cached resolution on change.
    """
    global _BACKEND, _RESOLVED_FN
    new = _normalize(name)
    if new != _BACKEND:
        _RESOLVED_FN = None
    _BACKEND = new



def _resolve_fa2() -> Callable[..., Any]:
    from flash_attn import flash_attn_varlen_func as _fn
    return _fn


def _resolve_fa4() -> Callable[..., Any]:
    from flash_attn.cute import flash_attn_varlen_func as _fa4_fn

    def _fa4_wrapper(
        q,
        k,
        v,
        cu_seqlens_q=None,
        cu_seqlens_k=None,
        max_seqlen_q=None,
        max_seqlen_k=None,
        dropout_p: float = 0.0,
        softmax_scale=None,
        causal: bool = False,
        window_size=(-1, -1),
        softcap: float = 0.0,
        alibi_slopes=None,
        deterministic: bool = False,
        return_attn_probs: bool = False,
        block_table=None,
        **_unused: Any,
    ):
        if dropout_p and dropout_p > 0:
            raise NotImplementedError("FA4 backend does not support dropout_p>0")
        if alibi_slopes is not None:
            raise NotImplementedError("FA4 backend does not support alibi_slopes")
        if return_attn_probs:
            raise NotImplementedError("FA4 backend does not support return_attn_probs")

        win_l, win_r = window_size
        if win_l == -1:
            win_l = None
        if win_r == -1:
            win_r = None

        out = _fa4_fn(
            q,
            k,
            v,
            cu_seqlens_q=cu_seqlens_q,
            cu_seqlens_k=cu_seqlens_k,
            max_seqlen_q=max_seqlen_q,
            max_seqlen_k=max_seqlen_k,
            softmax_scale=softmax_scale,
            causal=causal,
            window_size=(win_l, win_r),
            softcap=softcap,
            deterministic=deterministic,
            page_table=block_table,
            return_lse=False,
        )
        if isinstance(out, tuple):
            out = out[0]
        return out

    return _fa4_wrapper


def _resolve_sdpa() -> Callable[..., Any]:
    """FA2 varlen → per-sequence torch.SDPA fallback.

    Use when flash-attn is unavailable (e.g. CUDA 13 has no prebuilt wheel
    and source build is brittle). Slower than FA2 (one SDPA dispatch per
    sequence), but functionally equivalent for the dense / causal / no-alibi
    paths mageflow actually uses. Window / softcap / alibi / paged-attn /
    return_attn_probs are not supported and will raise.
    """
    import torch
    import torch.nn.functional as F

    def _sdpa_wrapper(
        q,
        k,
        v,
        cu_seqlens_q=None,
        cu_seqlens_k=None,
        max_seqlen_q=None,
        max_seqlen_k=None,
        dropout_p: float = 0.0,
        softmax_scale=None,
        causal: bool = False,
        window_size=(-1, -1),
        softcap: float = 0.0,
        alibi_slopes=None,
        deterministic: bool = False,
        return_attn_probs: bool = False,
        block_table=None,
        **_unused: Any,
    ):
        if dropout_p and dropout_p > 0:
            raise NotImplementedError("SDPA backend does not support dropout_p>0")
        if alibi_slopes is not None:
            raise NotImplementedError("SDPA backend does not support alibi_slopes")
        if return_attn_probs:
            raise NotImplementedError("SDPA backend does not support return_attn_probs")
        if softcap and softcap > 0:
            raise NotImplementedError("SDPA backend does not support softcap")
        if window_size not in ((-1, -1), (None, None), (0, 0)):
            raise NotImplementedError(
                f"SDPA backend does not support sliding window (got {window_size})"
            )
        if block_table is not None:
            raise NotImplementedError("SDPA backend does not support paged attention")
        if cu_seqlens_q is None or cu_seqlens_k is None:
            raise ValueError("SDPA backend requires cu_seqlens_q and cu_seqlens_k")

        # GQA: FA2 broadcasts k/v across query head groups natively; torch SDPA
        # does not (the q vs k head-dim mismatch is the AssertionError "tensor
        # a (32) must match tensor b (8) at non-singleton dimension 1" we'd see
        # otherwise). Repeat k/v along the head dim to match q before the loop.
        n_heads_q = q.shape[1]
        n_heads_kv = k.shape[1]
        if n_heads_q != n_heads_kv:
            if n_heads_q % n_heads_kv != 0:
                raise ValueError(
                    f"SDPA backend GQA expansion requires q heads ({n_heads_q}) "
                    f"to be divisible by k/v heads ({n_heads_kv})"
                )
            repeat = n_heads_q // n_heads_kv
            k = k.repeat_interleave(repeat, dim=1)
            v = v.repeat_interleave(repeat, dim=1)

        # q/k/v: (total_tokens, nheads, head_dim). Dispatch SDPA per sequence,
        # then concat. Python-level loop is fine since nseq is small (one per
        # image in the pack) and image-gen latency is dominated by sampling.
        cu_q = cu_seqlens_q.tolist()
        cu_k = cu_seqlens_k.tolist()
        outs = []
        for qs, qe, ks, ke in zip(cu_q[:-1], cu_q[1:], cu_k[:-1], cu_k[1:]):
            # (s, h, d) → (1, h, s, d)
            q_i = q[qs:qe].transpose(0, 1).unsqueeze(0)
            k_i = k[ks:ke].transpose(0, 1).unsqueeze(0)
            v_i = v[ks:ke].transpose(0, 1).unsqueeze(0)
            out_i = F.scaled_dot_product_attention(
                q_i,
                k_i,
                v_i,
                attn_mask=None,
                dropout_p=0.0,
                is_causal=causal,
                scale=softmax_scale,
            )
            # (1, h, s, d) → (s, h, d)
            outs.append(out_i.squeeze(0).transpose(0, 1))
        return torch.cat(outs, dim=0).contiguous()

    return _sdpa_wrapper


def _resolve() -> Callable[..., Any]:
    global _RESOLVED_FN
    if _RESOLVED_FN is None:
        if _BACKEND == "flash4":
            _RESOLVED_FN = _resolve_fa4()
        elif _BACKEND == "sdpa":
            _RESOLVED_FN = _resolve_sdpa()
        else:
            _RESOLVED_FN = _resolve_fa2()
    return _RESOLVED_FN


def flash_attn_varlen_func(*args, **kwargs):
    return _resolve()(*args, **kwargs)


__all__ = ["flash_attn_varlen_func", "set_attn_backend"]