File size: 14,208 Bytes
0f775e2 3a10d86 0f775e2 3a10d86 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 | """Spec for `comba-decode-step` — one Comba recurrent step (decoupled read key p) over a decode batch."""
import pathlib
import sys
sys.path.insert(0, str(pathlib.Path(__file__).resolve().parents[1]))
from spec import TaskSpec
SPEC = TaskSpec(
name="comba-decode-step",
title="Write a fast Comba decode-step (single-token state update) kernel",
blurb=("This is what a Comba layer runs at every generated token: ONE gated delta-rule recurrent step "
"with a DECOUPLED read key — the state correction is read with p while the write still uses k — "
"applied to every sequence in the decode batch at once. No chunking and no WY transform: it is a "
"pure bandwidth problem over the (K, V) recurrent state, which for a large decode batch is a "
"gigabyte of HBM traffic per layer per token."),
keywords=["mle", "kernel-generation", "comba", "decode", "linear-attention", "delta-rule",
"memory-bound", "gpu"],
module="comba_step.py",
func="comba_decode_step",
signature="comba_decode_step(state, q, k, v, p, beta, g, scale=None)",
returns_doc="""One Comba decode step, batched over B sequences.
Args:
state: (B, H, K, V) float32 — the recurrent state carried in from the previous token.
q: (B, H, K) bfloat16 — this token's query.
k: (B, H, K) bfloat16 — this token's WRITE key (L2-normalised along K).
v: (B, H, V) bfloat16 — this token's value.
p: (B, H, K) bfloat16 — this token's READ key (L2-normalised along K).
beta: (B, H) bfloat16 — delta-rule step size in (0, 1).
g: (B, H) float32 — this token's SCALAR log-decay (<= 0); exp(g) is the gate.
scale: float or None — query scale; None means K ** -0.5.
Returns:
(state_new, o) where
state_new: (B, H, K, V) float32 — the NEW state (functional; `state` is not modified)
o: (B, H, V) bfloat16/float32 — this token's output""",
reference_imports="import torch",
reference_src='''
def comba_decode_step(state, q, k, v, p, beta, g, scale=None):
"""One Comba recurrent step, written out in fp32.
Correct and simple — it is the numerical SPECIFICATION, not a performance target.
"""
K = q.shape[-1]
if scale is None:
scale = K ** -0.5
q, k, v, p, beta, g = [x.float() for x in (q, k, v, p, beta, g)]
S = state.float() * g.exp()[..., None, None] # scalar forget gate on the whole state
u = (v - (S * p.unsqueeze(-1)).sum(-2)) * beta.unsqueeze(-1) # correction, read with p (NOT with k)
S = S + k.unsqueeze(-1) * u.unsqueeze(-2) # rank-1 write, using k
o = (S * (q * scale).unsqueeze(-1)).sum(-2) # readout from the UPDATED state
return S, o
''',
make_inputs_src='''
_WARMUP = 16 # decode steps run from a zero state, so `state` has realistic decode-time magnitudes
def _mk(B, H, K, V, seed):
import torch.nn.functional as F
gen = torch.Generator(device="cuda").manual_seed(seed)
def _token():
q = torch.randn(B, H, K, device="cuda", dtype=torch.bfloat16, generator=gen)
k = F.normalize(torch.randn(B, H, K, device="cuda", generator=gen), dim=-1).to(torch.bfloat16)
v = torch.randn(B, H, V, device="cuda", dtype=torch.bfloat16, generator=gen)
p = F.normalize(torch.randn(B, H, K, device="cuda", generator=gen), dim=-1).to(torch.bfloat16)
beta = torch.rand(B, H, device="cuda", generator=gen).sigmoid().to(torch.bfloat16)
g = (F.logsigmoid(torch.randn(B, H, device="cuda", generator=gen)) / 4.0).to(torch.float32)
return q, k, v, p, beta, g
state = torch.zeros(B, H, K, V, device="cuda", dtype=torch.float32)
for _ in range(_WARMUP):
state = _ref(state, *_token())[0]
return (state, *_token())
''',
flops_src='''
def canonical_work(B, H, K, V):
"""BYTES moved by one Comba decode step, from the SHAPE ALONE.
The state dominates and is unavoidable: (B, H, K, V) fp32 read once and written once. Everything else is
one token's worth of activations per (b, h): q, k and p (bf16, K each), v (bf16, V), beta (bf16) and the
scalar log-gate g (fp32), and the output o (bf16, V). At K = V = 128 the state is 99.1% of this number.
Score = this byte count / your runtime, i.e. achieved HBM bandwidth.
"""
state = 2 * B * H * K * V * 4
token = B * H * (2 * K + 2 * K + 2 * K + 2 * V + 2 + 4 + 2 * V)
return state + token
''',
flops_formula="""bytes = 2 * B*H*K*V * 4 # fp32 state: read once + written once (>= 99% of the traffic)
+ B*H * (6*K + 4*V + 6) # one token in: q, k, p, v, beta, g and one token out: o""",
metric="GB/s",
compare="tuple",
tuple_names=("state_new", "o"),
tol=2e-2,
shape_names=("B", "H", "K", "V"),
grader_shapes=[(1024, 16, 128, 128), (512, 32, 128, 128), (2048, 8, 128, 128),
(1536, 16, 64, 128), (768, 24, 128, 128)],
measure_shapes=[(896, 16, 128, 128), (448, 32, 128, 128), (1792, 8, 128, 128),
(1280, 16, 64, 128), (640, 24, 128, 128)],
measure_quick_shapes=[(256, 16, 128, 128), (128, 32, 128, 128), (512, 16, 64, 128)],
correct_shapes=[(8, 4, 128, 128), (16, 2, 64, 128), (4, 8, 128, 64), (32, 4, 64, 64)],
spec_md="""One step of the Comba recurrence, for every `(b, h)` independently. `S` is the incoming
`state[b, h]` of shape `(K, V)`:
```
S = exp(g) * S # scalar forget gate on the whole state
u = beta * ( v - S^T p ) # correction, read with p (NOT with k)
S_new = S + k u^T # rank-1 write, using k
o = S_new^T (scale * q) # readout, from the UPDATED state
```
`scale` defaults to `K ** -0.5` and multiplies `q` only. Written out elementwise, with `i` indexing `K` and
`j` indexing `V`:
```
S[i, j] <- exp(g) * state[i, j]
u[j] = beta * ( v[j] - sum_i p[i] * S[i, j] )
S_new[i, j] = S[i, j] + k[i] * u[j]
o[j] = sum_i (scale * q[i]) * S_new[i, j]
```
The single thing that distinguishes Comba from a gated delta rule is that the state is **read with `p` and
written with `k`** — in DeltaNet those are the same vector. `p` and `k` are separate `(B, H, K)` inputs and
both must be streamed; using `k` in place of `p` is a different, wrong computation.
The correction `u` is read out of the state **after** the decay and **before** the write, and the readout `o`
uses the state **after** the write. `u` is a reduction over the whole `K` axis and the write then needs `u`,
so the state is needed on both sides of a reduction — see *Where the performance comes from*.
This is the same recurrence as the `comba-forward` task, run for exactly one token. Running this step `T`
times in a loop reproduces that task's output to ~1e-7 relative error.
`/app/reference.py` writes the step out in fp32. That is the exact specification; it is deliberately simple
rather than fast, and its runtime has no bearing on your score.""",
contract_md="""| arg | shape | dtype | meaning |
|-----|-------|-------|---------|
| `state` | `(B, H, K, V)` | `float32` | recurrent state carried in from the previous token |
| `q` | `(B, H, K)` | `bfloat16` | this token's query |
| `k` | `(B, H, K)` | `bfloat16` | **write** key (L2-normalised along `K`) |
| `v` | `(B, H, V)` | `bfloat16` | this token's value |
| `p` | `(B, H, K)` | `bfloat16` | **read** key (L2-normalised along `K`) |
| `beta` | `(B, H)` | `bfloat16` | delta-rule step size, in `(0, 1)` |
| `g` | `(B, H)` | `float32` | **scalar** log-decay per `(b, h)`, `<= 0` (`exp(g)` is the gate) |
| `scale` | scalar | `float` or `None` | query scale; `None` means `K ** -0.5` |
Note the argument order is `(state, q, k, v, p, beta, g)` — `p` comes after `v`, and `g` is last. This matches
the `comba-forward` task's `(q, k, v, p, beta, g)` with `state` prepended.
**Return a 2-tuple `(state_new, o)` in that order:**
| out | shape | dtype |
|-----|-------|-------|
| `state_new` | `(B, H, K, V)` | `float32` |
| `o` | `(B, H, V)` | `bfloat16` or `float32` |
**The update is functional, not in-place.** `state` is an input and must be treated as **read-only**;
`state_new` must be a **new** tensor. The grader calls your function and the reference on the *same* input
tensors, so scribbling on `state` makes the reference disagree with you and you fail the correctness gate.
**`state_new` must be `float32`** — a genuine fp32 tensor, not bf16/fp8 values in a wider container and not a
narrower dtype. Half the graded byte count is the fp32 state write; returning a narrower state is a contract
violation, not an optimisation.
`B` is the decode batch (one recurrent state per in-flight sequence), `H` the number of heads. All tensors are
CUDA and contiguous. `K` and `V` are multiples of 32. No variable-length packing, no GQA, no cache
indirection — the batch is dense.""",
regime_md="""**Shape regime you are graded in** (the exact grader sizes are *not* disclosed): `B`
(concurrent sequences) in 512–2048, `H` in 8–32, `K` in {64, 128}, `V` = 128. `B * H` is in the tens of
thousands, so the state alone is 0.5–1.2 GB and a roofline kernel takes 300–500 us — large enough that
bandwidth, not launch overhead, decides the score. Write a **general** kernel; one tuned to a single shape
will not score well.""",
correctness_md="""**Both** returned tensors must match the reference (evaluated in fp32 as a stable
ground truth) within **relative Frobenius error `2e-2`** at every graded shape, including the timed ones —
the new state as well as the output. The worse of the two is what is checked.""",
perf_md="""This kernel is **memory-bound and nothing else**. Per state element you read 4 bytes, write 4
bytes and do about four flops; arithmetic intensity is well under 1 flop/byte. Your score is achieved HBM
bandwidth and the target is that device's roofline — measure it with a large stream-copy rather than
taking a datasheet number on faith.
The reference is a chain of separate elementwise/reduction ops over the `(B, H, K, V)` state, so it drags the
state through HBM five or six times. A good kernel moves it **exactly twice**: read once, write once.
Getting to exactly twice is the whole problem, because a delta-rule step is not a single streaming pass:
```
u = beta * (v - S^T p) # a reduction over the ENTIRE K axis of the decayed state
S += k u^T # a write that cannot start until that reduction has finished
```
The state is needed on both sides of a reduction. Reading it twice costs you a third of your bandwidth
budget. The fix is to keep it **resident**: the reduction runs down the `K` axis independently for each column
of `V`, so `V` splits cleanly across CTAs. Give each CTA a slab of `V` columns and the full `K` axis — a
`K x V_slab` fp32 tile (128 x 32 = 16 KB, or the whole 128 x 128 = 64 KB state of one head) fits in shared
memory (check `p.shared_memory_per_block_optin`) or even registers. Load the slab once, apply the scalar
decay, reduce down `K` against `p`
for `u`, apply the rank-1 write with `k` and accumulate the readout `o = S_new^T (scale q)` in the same
registers, then store the slab once.
Beyond that it is straight bandwidth engineering:
- `V` is the contiguous axis of `state`, so tile so that every load and store is a fully coalesced 128-bit
(`float4`) access. Vectorised fp32 traffic is the single biggest lever.
- `B * H` is in the tens of thousands, so occupancy is free — but the *tail* is not. Choose the CTA tile so
the grid is close to a whole number of waves, and consider a persistent grid.
- `q`, `k`, `p`, `v`, `beta`, `g` are one token's worth per `(b, h)` — a few KB total. Load them once into
registers/shared and reuse them across every tile of the slab. Comba streams one more `(B, H, K)` vector
than Gated DeltaNet does (`p` as well as `k`), but at `K = 128` that is still under 1% of the traffic.
- Fuse everything into **one** kernel. Decay, reduce, write and readout as four launches is exactly what the
reference does and it is why it is slow.
- `torch.compile` on the reference will fuse some of this and is a fair sanity baseline, but it will not find
the resident-slab schedule and it will not hit the roofline.""",
precision_md="""The activations are **bfloat16** — this is an LLM decode kernel and that is the precision
it runs at in production. The **state is `float32`, in and out**, and all state arithmetic must be done in
**fp32**: the state is the quantity that is accumulated across thousands of decode steps, so rounding it is
not a local error, it compounds token after token. That is also why the graded byte count charges you 4 bytes
each way for it.
Concretely: upcast `q`, `k`, `v`, `p`, `beta` to fp32 as you load them, keep `exp(g)` in fp32, do the
`K`-reduction for `u` in an fp32 accumulator, and write `state_new` back as fp32. There is no tensor-core
matmul anywhere in this kernel — there is nothing to gain from bf16 or fp8 arithmetic, and **narrowing the
state is not a legal trade**: it would halve the write traffic that the score is defined against.
For calibration, a correct fp32-state kernel reproduces the reference to about **1e-7** relative error — this
is not a low-precision kernel, and there is no algebraic rearrangement here that needs a tolerance. The `2e-2`
gate is generous on purpose so that reasonable reduction orders and fused-multiply-add differences never bite;
if your error is anywhere near it, you have a bug or you have dropped part of the spec. For reference, a
kernel that rounds the state to bf16 lands at ~1.5e-3 (inside the gate, but it is still a contract violation),
one that drops the forget gate lands at ~2.0e-1, one that reads the correction with `k` instead of `p` lands
at ~6.4e-2, and one that drops the `- S^T p` correction entirely lands at ~4.5e-2. Every one of those is
outside the gate, and every one of them gets *worse* the longer the model decodes, because the error feeds
straight back into the next step's state.""",
).validate()
|