tools/factory/specs/causal_conv1d_decode.py

"""Spec for `causal-conv1d-decode-step` — the single-token decode form of the SSM/linear-attn short conv."""
import pathlib
import sys

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

SPEC = TaskSpec(
    name="causal-conv1d-decode-step",
    title="Write a fast causal conv1d decode-step kernel",
    blurb=("At every generated token, a Mamba-2 / Qwen3-Next block advances a short depthwise causal "
           "convolution by ONE position: read the rolling window state, produce one output, shift the state. "
           "Batched over hundreds of concurrent sequences this is pure bandwidth on the conv state, and it "
           "runs once per layer per token — a different kernel shape entirely from the prefill convolution."),
    keywords=["mle", "kernel-generation", "conv1d", "mamba", "qwen3-next", "decode", "memory-bound"],
    module="conv1d_step.py",
    func="causal_conv1d_decode_step",
    signature="causal_conv1d_decode_step(x, state, weight, bias)",
    returns_doc="""Causal depthwise conv1d, one decode step.

Args:
    x:      (B, D)      bfloat16 — this token's input, one vector per sequence.
    state:  (B, D, W-1) bfloat16 — the previous W-1 inputs, oldest at index 0.
    weight: (D, W)      bfloat16 — per-channel causal filter taps.
    bias:   (D,)        bfloat16 — per-channel bias, added before the activation.

Returns:
    (y, state_out) where
      y:         (B, D)      — silu(conv(window) + bias) for this position
      state_out: (B, D, W-1) — the window shifted by one, ready for the next token""",

    reference_imports="import torch\nimport torch.nn.functional as F",
    reference_src='''
def causal_conv1d_decode_step(x, state, weight, bias):
    """One decode step of the causal depthwise conv, in fp32.

    Correct and simple — it is the numerical SPECIFICATION, not a performance target.
    """
    # window = [ state (oldest .. newest) | x ]   ->  (B, D, W)
    win = torch.cat([state.float(), x.float().unsqueeze(-1)], dim=-1)
    y = (win * weight.float().unsqueeze(0)).sum(-1) + bias.float().unsqueeze(0)
    return F.silu(y), win[:, :, 1:].to(state.dtype)
''',
    make_inputs_src='''
def _mk(B, D, W, seed):
    gen = torch.Generator(device="cuda").manual_seed(seed)
    x = torch.randn(B, D, device="cuda", dtype=torch.bfloat16, generator=gen)
    state = torch.randn(B, D, W - 1, device="cuda", dtype=torch.bfloat16, generator=gen)
    weight = (torch.randn(D, W, device="cuda", dtype=torch.bfloat16, generator=gen) * 0.5)
    bias = (torch.randn(D, device="cuda", dtype=torch.bfloat16, generator=gen) * 0.1)
    return x, state, weight, bias
''',
    flops_src='''
def canonical_work(B, D, W):
    """BYTES attributed to one decode step, from the SHAPE ALONE.

    Read x and the (W-1)-deep state, write y and the shifted state, all bf16; the (D, W) taps and (D,) bias
    are negligible and stay resident. This is a bandwidth kernel, so the score is achieved bandwidth against
    this fixed byte count.
    """
    return 2 * (B * D * 2) + 2 * (B * D * (W - 1) * 2)
''',

    metric="GB/s",
    compare="tuple",
    tuple_names=("y", "state_out"),
    tol=4e-3,     # MEASURED: 2.3x the 1.73e-3 spread between two independent correct implementations
    shape_names=("B", "D", "W"),
    grader_shapes=[(16384, 8192, 4), (32768, 4096, 4), (8192, 16384, 4),
                   (24576, 8192, 4), (65536, 2048, 4)],
    measure_shapes=[(12288, 8192, 4), (24576, 4096, 4), (8192, 12288, 4),
                    (16384, 6144, 4), (49152, 2048, 4)],
    measure_quick_shapes=[(2048, 4096, 4), (4096, 2048, 4), (1024, 8192, 4)],
    correct_shapes=[(64, 256, 4), (129, 512, 4), (257, 128, 4), (32, 1024, 4)],

    spec_md="""One step of the depthwise causal convolution, for every sequence in the batch at once.

The rolling window for channel `d` of sequence `b` is the previous `W-1` inputs followed by this token's:

```
win[b, d, :] = [ state[b, d, 0], ..., state[b, d, W-2], x[b, d] ]      # oldest first, current last
y[b, d]      = silu( bias[d] + sum over i in [0, W) of win[b, d, i] * weight[d, i] )
state_out[b, d, :] = win[b, d, 1:]                                     # drop the oldest, keep W-1
```

`silu(z) = z * sigmoid(z)`. Channels never mix — this is depthwise. `weight[d, W-1]` multiplies the
**current** input and `weight[d, 0]` the oldest, matching the prefill convolution's tap order.

`state` is the carry between tokens: what this call returns is what the next token's call receives. The
contract is **functional** — return a new `state_out`; do not mutate `state` in place. The grader calls your
function and the reference on the same buffers, so an in-place update corrupts the comparison and fails.

Note this is the DECODE counterpart of the prefill conv: there is no time axis to parallelise over, only the
batch and the channels, and the state traffic dominates everything.

`/app/reference.py` builds the window with a concatenate and reduces it in fp32. That is the exact
specification; it is deliberately simple rather than fast.""",

    contract_md="""| arg | shape | dtype | meaning |
|-----|-------|-------|---------|
| `x` | `(B, D)` | `bfloat16` | this token's input, one vector per sequence |
| `state` | `(B, D, W-1)` | `bfloat16` | previous `W-1` inputs, **oldest at index 0** |
| `weight` | `(D, W)` | `bfloat16` | per-channel taps; `weight[d, W-1]` hits the current input |
| `bias` | `(D,)` | `bfloat16` | per-channel bias, applied **before** the SiLU |

**Return** a 2-tuple `(y, state_out)` **in that order**:

| out | shape | notes |
|-----|-------|-------|
| `y` | `(B, D)` | this position's output; bf16 or fp32 |
| `state_out` | `(B, D, W-1)` | window shifted by one; **must be bf16** |

`W` is always 4. `B` is **not** guaranteed to be a multiple of any tile size — the correctness shapes include
`B = 129` and `B = 257`, so handle the tail. Treat all inputs as read-only; the update is functional.""",

    regime_md="""**Shape regime you are graded in** (the exact grader sizes are *not* disclosed): `B`
(concurrent sequences) in 8192–65536, `D` (channels) in 2048–16384, `W` = 4. Large batches are the point —
this kernel only matters when hundreds of sequences decode together, and that is what makes it a bandwidth
problem rather than a latency one. Write a **general** kernel that handles a ragged `B`.""",

    correctness_md="""**Both** returned tensors must match the reference (evaluated in fp32) within
**relative error `4e-3`** at every graded shape, including the timed ones — the shifted state as well as the
output. A wrong `state_out` corrupts every subsequent token, so it is graded just as hard as `y`; the grader
scores the **worse** of the two.

`state_out` carries no arithmetic at all — it is a bf16 copy of values that were already bf16 — so a correct
kernel reproduces it **bit for bit** and the gate on it is effectively exact. The whole `4e-3` budget exists
for `y`.

That gate is **measured**, and kernels that skip this task's actual work miss it by orders of magnitude: no
SiLU scores **1.04**, taps applied in the reversed causal order **1.28**, an unshifted `state_out` **1.42**,
and even the subtlest variant tried — bias added *after* the SiLU instead of before — scores **0.104**, still
**26x** the tolerance.""",

    perf_md="""There are four multiply-adds and a SiLU per element, so this is entirely a memory problem:
the floor is one read of `x` and `state` and one write of `y` and `state_out`, which is what
`canonical_work` counts.

The obvious implementation reads `state`, concatenates, writes a new `state`, and moves `(W-1)` values per
channel in each direction. Most of that traffic is the *same data being shifted by one slot* — the window
overlaps itself between consecutive steps by `W-2` elements. A good kernel loads the window once into
registers, computes `y`, and writes back only the shifted view, with vectorised 128-bit accesses; `(B, D)`
is fully coalesced along `D`, so one warp per group of channels streams cleanly.

The taps and bias are tiny and reused by every sequence — hold them in registers or shared memory rather
than re-reading them per element. And watch the tail: `B` is ragged, and a branchy epilogue on a kernel this
short costs a visible fraction of the runtime.""",

    precision_md="""Inputs and outputs are **bfloat16**; do the accumulation and the SiLU in **fp32**. With
only four taps the convolution is numerically benign, but evaluate the `sigmoid` in fp32 — a bf16 sigmoid
loses enough precision near zero to show up in the norm.

`state_out` **must be bfloat16**: it is the carry the next token consumes, and returning it wider would both
break the contract and hide the quantisation the real decode loop lives with.

**fp8 is not useful here** and is not expected — this is a bandwidth kernel on bf16 data with a fixed output
dtype.

Do **not** infer from the reference that fp32 storage is wanted; it computes in fp32 purely to be a stable
numerical *specification*.

**Where the tolerance comes from.** `4e-3` is measured, not inherited. A second, independent implementation —
no concatenate, the four taps applied as explicit FMAs in the *reverse* order, the shifted state built by an
overwrite rather than a slice of a concatenated window — differs from this fp32 reference by relative
Frobenius error **1.73e-3** at worst across the correctness shapes, and 1.67e-3 at a full graded shape.
The tolerance is **2.3x** that. That 1.7e-3 is almost entirely the bf16 rounding of `y` itself: with only
four taps the fp32 accumulation contributes essentially nothing, so this number is the floor for *any*
bf16-returning kernel and there is no reduction-order freedom left to spend.""",
).validate()