File size: 14,102 Bytes
0f775e2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
231
232
233
234
235
236
237
238
239
240
241
242
243
244
"""Spec for `based-backward` — the training-side counterpart of based-forward."""
import pathlib
import sys

sys.path.insert(0, str(pathlib.Path(__file__).resolve().parents[1]))
from spec import TaskSpec

# the chunked Based forward, reused as the differentiable spec (it matches the based-forward reference to
# ~3e-7 in fp32; the chunk length is an internal detail of both)
_FWD = '''
CHUNK_SIZE = 128


def _based_forward(q, k, v, scale=None):
    """Based / Taylor linear attention forward in fp32 — differentiable; autograd through this defines the
    gradients. The exact 2nd-order Taylor attention: explicit feature map across chunks, direct polynomial
    inside a chunk."""
    B, T, H, K = q.shape
    V = v.shape[-1]
    C = CHUNK_SIZE
    if scale is None:
        scale = K ** -0.5
    q, k, v = [x.transpose(1, 2).to(torch.float32) for x in (q, k, v)]
    q = q * scale
    P = 1 + K + K * K
    dev = q.device
    rt2 = 2.0 ** -0.5

    def phi(x):                                                    # [1, x, x (x) x / sqrt(2)]
        n = x.shape[2]
        return torch.cat([torch.ones(B, H, n, 1, device=dev, dtype=torch.float32), x,
                          (x[..., :, None] * x[..., None, :]).reshape(B, H, n, K * K) * rt2], -1)

    S = q.new_zeros(B, H, P, V)                                    # sum_j phi(k_j) v_j^T
    z = q.new_zeros(B, H, P, 1)                                    # sum_j phi(k_j)
    tri = torch.tril(torch.ones(C, C, device=dev, dtype=torch.float32))
    outs = []
    for i in range(0, T, C):
        q_i, k_i, v_i = q[:, :, i:i + C], k[:, :, i:i + C], v[:, :, i:i + C]
        n = q_i.shape[2]
        pq = phi(q_i)
        s = q_i @ k_i.transpose(-1, -2)
        A = (1.0 + s + 0.5 * s * s) * tri[:n, :n]                  # causal 2nd-order Taylor weights
        num = pq @ S + A @ v_i
        den = pq @ z + A.sum(-1, keepdim=True)
        outs.append(num / (den + 1e-6))
        pk = phi(k_i)
        S = S + pk.transpose(-1, -2) @ v_i
        z = z + pk.sum(-2)[..., None]
    return torch.cat(outs, 2).transpose(1, 2).contiguous()


def based_backward(q, k, v, do, scale=None):
    """Based backward — the baseline runs the chunked fp32 forward under autograd."""
    ins = [x.detach().clone().requires_grad_(True) for x in (q, k, v)]
    o = _based_forward(*ins, scale=scale)
    return torch.autograd.grad(o, ins, do.float())
'''

SPEC = TaskSpec(
    name="based-backward",
    title="Write a fast Based (Taylor linear attention) BACKWARD kernel",
    blurb=("The training-side counterpart of Based: softmax replaced by its 2nd-order Taylor expansion, "
           "1 + s + s^2/2, which is exactly a dot product of the feature map [1, x, x (x) x / sqrt(2)] — a "
           "linear attention over 1 + K + K^2 features plus a running normaliser. The backward has to "
           "differentiate through BOTH the huge (P, V) state and the quotient by that normaliser, and the "
           "reference does it by replaying the whole chunked graph under autograd."),
    keywords=["mle", "kernel-generation", "based", "taylor", "backward", "linear-attention", "feature-map"],
    module="based_bwd.py",
    func="based_backward",
    signature="based_backward(q, k, v, do, scale=None)",
    returns_doc="""Based / Taylor linear attention backward.

Args:
    q, k: (B, T, H, K) bfloat16 — queries / keys. K is the small Taylor FEATURE dimension (8 or 16).
    v:    (B, T, H, V) bfloat16 — values. V is the (larger) head dimension.
    do:   (B, T, H, V) bfloat16 — incoming gradient w.r.t. the forward output.
    scale: float or None — query scale; None means K ** -0.5.

Returns:
    (dq, dk, dv) with shapes (B,T,H,K), (B,T,H,K), (B,T,H,V).""",

    reference_imports="import torch",
    reference_src=_FWD,
    make_inputs_src='''
def _mk(B, T, H, K, V, seed):
    gen = torch.Generator(device="cuda").manual_seed(seed)
    q = torch.randn(B, T, H, K, device="cuda", dtype=torch.bfloat16, generator=gen)
    k = torch.randn(B, T, H, K, device="cuda", dtype=torch.bfloat16, generator=gen)
    v = torch.randn(B, T, H, V, device="cuda", dtype=torch.bfloat16, generator=gen)
    do = torch.randn(B, T, H, V, device="cuda", dtype=torch.bfloat16, generator=gen)
    return q, k, v, do
''',
    flops_src='''
def canonical_work(B, T, H, K, V, C=64):
    """FLOPs attributed to one Based BACKWARD, from the SHAPE ALONE (chunk length C = 64).

    The forward's linear form carries a (P, V) state with P = 1 + K + K*K: per token 2*P*V to fold the token
    into the state and 2*P*V to read it back out -> 4*P*V. The intra-chunk part is done in the quadratic
    form: 2*C*K for the C scores of one query and 2*C*V to apply them -> 2*C*(K + V) per token. The backward
    is credited the standard 2x the forward. Building the feature map, the running normaliser and the
    divide are O(K*K + V) per token and are not counted.
    """
    P = 1 + K + K * K
    return 2 * (B * H * T * (4 * P * V + 2 * C * (K + V)))
''',
    flops_formula="2 * ( B*H*T * (4*(1 + K + K*K)*V + 2*C*(K + V)) )        with C = 64   # 2x the forward",

    metric="TFLOP/s",
    compare="tuple",
    tuple_names=("dq", "dk", "dv"),
    tol=2e-2,
    shape_names=("B", "T", "H", "K", "V"),
    grader_shapes=[(4, 8192, 32, 16, 128), (8, 4096, 32, 16, 128), (8, 8192, 16, 16, 128),
                   (6, 8192, 24, 16, 128), (8, 8192, 32, 8, 128)],
    measure_shapes=[(4, 8192, 24, 16, 128), (8, 4096, 16, 16, 128), (4, 8192, 32, 16, 64),
                    (6, 4096, 32, 16, 128), (4, 8192, 32, 8, 128)],
    measure_quick_shapes=[(1, 2048, 16, 16, 128), (2, 2048, 8, 16, 128), (1, 4096, 16, 8, 128)],
    correct_shapes=[(1, 512, 4, 16, 128), (2, 1024, 8, 8, 64), (1, 1024, 6, 16, 64), (2, 256, 4, 16, 128)],

    spec_md="""The forward is causal attention with softmax replaced by its 2nd-order Taylor expansion. Per
batch `b` and head `h`, with `s[i, j] = (scale * q_i) . k_j` and `scale` defaulting to `K ** -0.5`:

```
A[i, j] = 1 + s[i, j] + s[i, j]^2 / 2          for j <= i,   0 otherwise
o_i     = ( sum_j A[i, j] * v_j ) / ( sum_j A[i, j] + 1e-6 )
```

You must return the gradients of that forward with respect to `q, k, v`, given the incoming gradient `do` of
the loss with respect to `o`.

**Why this is not an O(T^2) op.** The polynomial is an exact inner product of the feature map

```
phi(x) = [ 1 ,  x ,  x (x) x / sqrt(2) ]                (length P = 1 + K + K*K)
```

because `phi(q) . phi(k) = 1 + (q.k) + (q.k)^2 / 2`, so the layer is an ordinary linear attention over a
`P`-dimensional feature space with state `S = sum_j phi(k_j) v_j^T` and normaliser `z = sum_j phi(k_j)`. `K`
is the small Taylor **feature** dimension (8 or 16), not the head dimension — `V` is the head dimension —
which is what keeps `P` (73 or 273) manageable.

Two things make the backward its own problem rather than a transposed forward. First, the output is a
**quotient**: `do` has to be pushed through `num/(den + 1e-6)`, which produces a second, `V`-contracted
signal `-o * do / den` that feeds the *normaliser* path `z`, so every one of `dq`, `dk`, `dv` has a
numerator term and a denominator term. Second, `dq` needs the derivative of `phi(q)` — the outer product
`q (x) q` differentiates to something that touches every feature twice — and `dk` needs the same for
`phi(k)` while `k` also appears inside the reverse state.

`/app/reference.py` gives you `_based_forward` — the forward in its chunked form (chunk length 128: the
explicit feature map across chunks, the direct polynomial inside a chunk) — and obtains the gradients by
running it under **autograd**. That is the specification, and it is what torch gives you for free, but it
materialises `phi(q)` and `phi(k)` (`P/K` times bigger than `q` and `k`), every `C x C` weight tile and the
`(P, V)` state at every chunk boundary in HBM.

You may reach the same gradients any way you like: derive and fuse the analytic backward, recompute
intermediates instead of storing them, use a different chunk length, or restructure the reverse scan. Only
the returned numbers are specified.""",

    contract_md="""| arg | shape | dtype | meaning |
|-----|-------|-------|---------|
| `q` | `(B, T, H, K)` | `bfloat16` | queries; `K` is the Taylor **feature** dimension |
| `k` | `(B, T, H, K)` | `bfloat16` | keys |
| `v` | `(B, T, H, V)` | `bfloat16` | values; `V` is the head dimension |
| `do` | `(B, T, H, V)` | `bfloat16` | incoming gradient w.r.t. the forward output `o` |
| `scale` | scalar | `float` or `None` | query scale; `None` means `K ** -0.5` (note: `K`, not `V`) |

**Return** a 3-tuple `(dq, dk, dv)` **in that order**, with shapes `(B,T,H,K)`, `(B,T,H,K)`, `(B,T,H,V)` —
the same shapes as `q`, `k`, `v`. Each may be `bfloat16` or `float32`.

**All three are graded.** Getting two of three right scores **0**.

`scale` multiplies **`q` only**, before the dot product, so the polynomial is evaluated at
`s = (scale * q_i) . k_j` and `dq` carries that factor. The denominator guard is exactly `+ 1e-6`, added
after the row sum, and it is inside the derivative: `d/d(den) [num/(den + 1e-6)] = -num/(den + 1e-6)^2`.

Attention is fully causal and includes the diagonal (`j <= i`), so position `0` has the single weight
`A[0, 0]` and never an empty row. `K != V` in general and `K` is small. There is no initial or final state,
no state gradient, no sliding window and no softmax.

All tensors are CUDA and contiguous. `T` is a multiple of 128. No variable-length packing, no GQA. Treat all
inputs as read-only.""",

    regime_md="""**Shape regime you are graded in** (the exact grader sizes are *not* disclosed): `B` in
4–8, `T` in 4096–8192, `H` in 16–32, `K` in {8, 16}, `V` in {64, 128}. `K` changes `P = 1 + K + K*K` from 73
to 273 — a nearly 4x change in state size and in the work per token — so a kernel that hard-codes one
feature dimension will fail or score badly on the other. `B*H` is 96–256. Write a **general** kernel.""",

    correctness_md="""**All three** gradients must match the reference (evaluated in fp32) within **relative
Frobenius error `2e-2`** at every graded shape, including the timed ones. Getting two of three right scores
**0**.""",

    perf_md="""The reference is already sub-quadratic and already vectorised, so there is no free win from
"removing python loops". What it wastes is memory traffic, and autograd multiplies it: `phi(q)`, `phi(k)`,
every `C x C` tile of `s` and `A`, and the `(P, V)` state at every chunk boundary are all kept alive for the
reverse pass.

The kernel you want never materialises `phi` at all. `phi(x) = [1, x, outer(x, x)/sqrt(2)]`, so the `(P, V)`
state is really a `(1, V)` block, a `(K, V)` block and a `(K, K, V)` block, and the same decomposition
applies to the **reverse** state `dS = sum_{i >= j} phi(q_i) (do_i)^T` that `dk` and `dv` are read out of.
Each block has its own natural update, and the quadratic block — `K` independent `(K, V)` linear-attention
states, one per feature of `k` — is where essentially all the FLOPs are.

Three structural points specific to the backward:

- **The normaliser is a second, cheaper linear attention.** `den` is the same computation with `v` replaced
  by a column of ones, and its adjoint is the same reverse scan with `do` replaced by `-o . do / den`
  (a per-position scalar). Fuse it into the same pass — do not spend a second set of matmuls on it.
- **Inside a chunk, stay in the quadratic form.** Evaluating `1 + s + s^2/2` directly from a `C x C` score
  tile is much cheaper than going through the `P`-dimensional feature map, and its adjoint is just
  `dA -> (1 + s) * dA` folded back into `dq` and `dk` — two ordinary bf16 matmuls plus one elementwise
  polynomial, exactly as in the forward.
- **Recompute, don't store.** `A`, `phi(q)`, `phi(k)` and the running state are all cheap to rebuild from
  `q, k, v` during the reverse scan; at `K = 16, V = 128` the quadratic state block alone is `256 x 128`
  floats, so how you split it across warps/CTAs and whether you keep it in bf16 with fp32 accumulation of
  the increments is the main design decision.""",

    precision_md="""All inputs and outputs are **bfloat16** — this is an LLM-training kernel and that is the
precision it runs at in production. Your kernel is expected to do its matmuls on **bf16 tensor cores with
fp32 accumulation**. **fp8** is acceptable anywhere you can still hold the tolerance.

Both state carries **must be fp32**: the forward `S = sum phi(k_j) v_j^T` / `z = sum phi(k_j)`, and the
reverse `dS = sum phi(q_i) do_i^T`. These are **ungated** sums over the whole prefix (resp. suffix) — there
is no decay to forget old terms — so by the end of the sequence they are reductions over thousands of
contributions while each increment stays `O(1)`. In bf16 the increments stop moving the accumulator well
before the end and the error compounds with `T`. `z` is a **denominator**, so an error there is a relative
error on the whole output row and on its entire gradient, not a damped one.

**Gradient reductions drift first**, and here `dq` and `dk` are the exposed ones: each is a contraction over
the full `P`-dimensional feature axis *and* over `V`, so they want fp32 accumulators throughout. Note also
that `s^2` computed in bf16 doubles its relative error before it is summed — square in fp32.

For calibration, a measured bf16 pipeline for this op (bf16 matmul operands, fp32 accumulation, fp32 state
and normaliser) lands at **3.5e-3 / 3.5e-3 / 2.7e-3** relative error on `dq / dk / dv` at
`(B, T, H, K, V) = (2, 2048, 8, 16, 128)` — roughly 6x inside the `2e-2` gate, and flat in `T`.

For a sense of scale on the other side: truncating the feature map to first order (`A = 1 + s`, i.e. plain
linear attention) misses by **2.9 / 2.6 / 0.75** on `dq / dk / dv` at that shape. The 2nd-order term is the
entire point of the method and is not optional.

Do **not** infer from the reference that fp32 compute is wanted. It runs in fp32 purely to be a stable
numerical *specification*, and its speed has no bearing on your score.""",
).validate()