File size: 5,797 Bytes
5ccb4fd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) 2026 Simulacra Research Inc.
# SPDX-License-Identifier: Apache-2.0


from __future__ import annotations

import jax
import jax.numpy as jnp

from .pallas_mha import BlockSizes, noncausal_bias_mha


def _validate_rectangular_attention_inputs(
    q_local: jax.Array,
    k_global: jax.Array,
    v_global: jax.Array,
    edge_bias_local: jax.Array,
    key_mask: jax.Array,
) -> None:
    if q_local.ndim != 3 or k_global.ndim != 3 or v_global.ndim != 3:
        raise ValueError("q, k, and v must have shapes [sequence, heads, head_dim]")
    q_len, n_heads, head_dim = q_local.shape
    kv_len = k_global.shape[0]
    if q_len < 1 or kv_len < 1:
        raise ValueError("query and key sequence lengths must both be positive")
    if k_global.shape != v_global.shape:
        raise ValueError(
            f"k and v shapes must match; got {k_global.shape} and {v_global.shape}"
        )
    if k_global.shape[1:] != (n_heads, head_dim):
        raise ValueError(
            "q, k, and v must have the same head count and head dimension; "
            f"got {q_local.shape}, {k_global.shape}, and {v_global.shape}"
        )
    if edge_bias_local.shape != (q_len, kv_len, n_heads):
        raise ValueError(
            "edge bias must have shape [local_queries, global_keys, heads]; "
            f"got {edge_bias_local.shape}, expected {(q_len, kv_len, n_heads)}"
        )
    if key_mask.shape != (kv_len,):
        raise ValueError(f"key mask must have shape {(kv_len,)}, got {key_mask.shape}")
    arrays = (q_local, k_global, v_global, edge_bias_local)
    if any(value.dtype != jnp.float32 for value in arrays):
        raise TypeError(
            "large-N rectangular attention is fp32-only; got "
            + ", ".join(str(value.dtype) for value in arrays)
        )


def _dividing_block_size(length: int, requested: int | None) -> int:
    block = min(length, 128 if requested is None else requested)
    if block < 1:
        raise ValueError(f"block size must be positive, got {block}")
    while length % block:
        block //= 2
    return block


def pallas_rectangular_edge_attention(
    q_local: jax.Array,
    k_global: jax.Array,
    v_global: jax.Array,
    edge_bias_local: jax.Array,
    key_mask: jax.Array,
    *,
    block_k: int | None = None,
) -> jax.Array:

    _validate_rectangular_attention_inputs(
        q_local, k_global, v_global, edge_bias_local, key_mask
    )
    sm_scale = float(q_local.shape[-1]) ** -0.5
    q_len = q_local.shape[0]
    kv_len = k_global.shape[0]
    bq = _dividing_block_size(q_len, None)
    bk = _dividing_block_size(kv_len, block_k)
    block_sizes = BlockSizes(block_q=bq, block_k=bk)

    masked_bias = jnp.where(
        key_mask[None, :, None].astype(bool),
        edge_bias_local,
        jnp.asarray(-1.0e30, dtype=jnp.float32),
    )
    return noncausal_bias_mha(
        (q_local * jnp.asarray(sm_scale, dtype=jnp.float32))[None],
        k_global[None],
        v_global[None],
        masked_bias[None],
        block_sizes=block_sizes,
    )[0]


def ring_learned_fwl2_local(
    a_local: jax.Array,
    b_local: jax.Array,
    *,
    axis_name: str,
    axis_size: int,
) -> jax.Array:

    local_rows, global_columns, channels = a_local.shape
    if b_local.shape != (local_rows, global_columns, channels):
        raise ValueError(
            f"local 2-FWL shapes must match; got {a_local.shape}, {b_local.shape}"
        )
    return ring_learned_fwl2_columns_local(
        a_local,
        b_local,
        axis_name=axis_name,
        axis_size=axis_size,
    )


def ring_learned_fwl2_columns_local(
    a_local: jax.Array,
    b_local_columns: jax.Array,
    *,
    axis_name: str,
    axis_size: int,
) -> jax.Array:

    if a_local.ndim != 3 or b_local_columns.ndim != 3:
        raise ValueError(
            "local 2-FWL operands must both be rank three; got "
            f"{a_local.shape} and {b_local_columns.shape}"
        )
    local_rows, global_columns, channels = a_local.shape
    if b_local_columns.shape[0] != local_rows:
        raise ValueError(
            "local A/B row counts must match; got "
            f"{local_rows} and {b_local_columns.shape[0]}"
        )
    if b_local_columns.shape[2] != channels:
        raise ValueError(
            "local A/B channel counts must match; got "
            f"{channels} and {b_local_columns.shape[2]}"
        )
    if a_local.dtype != b_local_columns.dtype:
        raise TypeError(
            "local A/B dtypes must match; got "
            f"{a_local.dtype} and {b_local_columns.dtype}"
        )
    if global_columns != local_rows * axis_size:
        raise ValueError(
            "2-FWL row shards must evenly tile the contracted axis; "
            f"got local_rows={local_rows}, columns={global_columns}, "
            f"axis_size={axis_size}"
        )

    def contribution(b_panel, origin):
        a_panel = jax.lax.dynamic_slice_in_dim(
            a_local, origin * local_rows, local_rows, axis=1
        )
        return jnp.einsum("ikc,kjc->ijc", a_panel, b_panel)

    origin0 = jax.lax.axis_index(axis_name).astype(jnp.int32)
    accumulator0 = contribution(b_local_columns, origin0)
    ring_permutation = [(lane, (lane + 1) % axis_size) for lane in range(axis_size)]

    def ring_step(carry, _):
        b_panel, origin, accumulator = carry
        b_panel = jax.lax.ppermute(b_panel, axis_name=axis_name, perm=ring_permutation)
        origin = (origin - jnp.asarray(1, jnp.int32)) % axis_size
        accumulator = accumulator + contribution(b_panel, origin)
        return (b_panel, origin, accumulator), None

    (_, _, result), _ = jax.lax.scan(
        ring_step,
        (b_local_columns, origin0, accumulator0),
        xs=None,
        length=axis_size - 1,
    )
    return result