tools/factory/specs/causal_conv3d_cache_step.py

"""Spec for `causal-conv3d-cache-step` — the streaming (feat_cache) form of the video-VAE causal Conv3d."""
import pathlib
import sys

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

SPEC = TaskSpec(
    name="causal-conv3d-cache-step",
    title="Write a fast streaming causal Conv3d cache-step kernel (video VAE decode)",
    blurb=("A 720p x 129-frame decode does not fit in memory as one tensor, so Wan and HunyuanVideo decode "
           "the clip in short chunks of frames and carry a per-layer feat_cache holding the last two input "
           "frames. Each step conditions on that cache instead of on zero padding, emits the chunk's "
           "outputs, and hands the last two frames forward. Chunks are only one to a few frames deep but "
           "full pixel resolution, so the kernel is short-and-wide: a two-frame halo dominates the traffic "
           "and the arithmetic has almost no temporal reuse to hide behind."),
    keywords=["mle", "kernel-generation", "conv3d", "causal", "streaming", "cache", "video", "vae", "wan",
              "hunyuanvideo", "decode"],
    module="conv3d_cache.py",
    func="causal_conv3d_cache_step",
    signature="causal_conv3d_cache_step(x, cache, weight, bias)",
    returns_doc="""One streaming step of the causal 3x3x3 convolution.

Args:
    x:      (B, Cin, Tc, H, W)     bfloat16 — this chunk's input frames.
    cache:  (B, Cin, 2, H, W)      bfloat16 — the two INPUT frames immediately before this chunk,
                                              oldest at index 0.
    weight: (Cout, Cin, 3, 3, 3)   bfloat16 — dense convolution weights.
    bias:   (Cout,)                bfloat16 — per-output-channel bias.

Returns:
    (y, cache_out) where
      y:         (B, Cout, Tc, H, W) bfloat16 — this chunk's outputs
      cache_out: (B, Cin, 2, H, W)   bfloat16 — the last two frames of [cache | x], for the next step""",

    reference_imports="import torch\nimport torch.nn.functional as F",
    reference_src='''
def causal_conv3d_cache_step(x, cache, weight, bias):
    """Streaming causal Conv3d: the cache REPLACES the two frames of temporal zero padding.

    Correct and simple -- the numerical SPECIFICATION, not a performance target.
    """
    xf = torch.cat([cache.float(), x.float()], dim=2)         # (B, Cin, 2 + Tc, H, W)
    xp = F.pad(xf, (1, 1, 1, 1, 0, 0))                        # spatial pad only; time is supplied by cache
    with torch.backends.cudnn.flags(enabled=True, allow_tf32=False):
        y = F.conv3d(xp, weight.float(), bias.float())        # -> (B, Cout, Tc, H, W)
    cache_out = torch.cat([cache, x], dim=2)[:, :, -2:].contiguous()
    return y.to(torch.bfloat16), cache_out
''',
    make_inputs_src='''
def _mk(B, Cin, Cout, Tc, H, W, seed):
    gen = torch.Generator(device="cuda").manual_seed(seed)
    x = torch.randn(B, Cin, Tc, H, W, device="cuda", dtype=torch.bfloat16, generator=gen)
    cache = torch.randn(B, Cin, 2, H, W, device="cuda", dtype=torch.bfloat16, generator=gen)
    w = (torch.randn(Cout, Cin, 3, 3, 3, device="cuda", dtype=torch.bfloat16, generator=gen)
         * (1.0 / (Cin * 27) ** 0.5))
    b = (torch.randn(Cout, device="cuda", dtype=torch.bfloat16, generator=gen) * 0.1)
    return x, cache, w, b
''',
    flops_src='''
def canonical_work(B, Cin, Cout, Tc, H, W):
    """FLOPs of one streaming chunk, from the SHAPE ALONE.

    B*Cout*Tc*H*W outputs, each a dot product over Cin*27 taps (one multiply, one add). The cache frames
    contribute to the outputs but are not themselves outputs, so they add no work of their own; copying the
    cache forward is bookkeeping and is not counted.
    """
    return 2 * B * Cout * Cin * 27 * Tc * H * W
''',

    metric="TFLOP/s",
    compare="tuple",
    tuple_names=("y", "cache_out"),
    tol=1e-2,
    shape_names=("B", "Cin", "Cout", "Tc", "H", "W"),
    grader_shapes=[(1, 128, 128, 4, 256, 256), (1, 128, 128, 2, 384, 384), (1, 192, 128, 3, 256, 320),
                   (1, 128, 256, 1, 352, 480), (2, 96, 96, 4, 224, 224)],
    measure_shapes=[(1, 128, 128, 3, 256, 288), (1, 128, 128, 2, 320, 384), (1, 160, 128, 3, 256, 320),
                    (1, 128, 224, 1, 352, 480), (2, 96, 96, 3, 224, 256)],
    measure_quick_shapes=[(1, 64, 64, 2, 96, 96), (1, 96, 64, 1, 128, 128), (1, 32, 96, 4, 64, 96)],
    correct_shapes=[(1, 32, 32, 3, 16, 16), (1, 48, 32, 1, 24, 20), (2, 16, 24, 2, 17, 23),
                    (1, 64, 64, 5, 32, 32)],

    spec_md="""One streaming step of the causal 3x3x3 convolution. The chunk is `Tc` frames deep; the two
frames that precede it arrive separately, in `cache`.

```
xfull = concat([cache, x], dim=T)                      # (B, Cin, 2 + Tc, H, W)
xp    = zero_pad(xfull, W: 1 left / 1 right, H: 1 top / 1 bottom)     # NO temporal padding

y[b, co, t, h, w] = bias[co]
                  + sum_{ci, kt, kh, kw} xp[b, ci, t + kt, h + kh, w + kw] * weight[co, ci, kt, kh, kw]
                                                        for t in [0, Tc)

cache_out = xfull[:, :, -2:]                            # the last two frames of [cache | x]
```

So output frame `t` of the chunk reads input frames `t-2, t-1, t` of the *concatenated* stream: for `t = 0`
that is `cache[0]`, `cache[1]`, `x[0]`.

### Why this is the causal conv, not a different one

The full-clip kernel prepends **two zero frames** to the whole video. In streaming decode the clip is cut
into chunks and the zeros are only correct for the *first* chunk; every later chunk must condition on the
real previous frames, which is what `cache` carries. That is precisely the trick that lets a 4x8x8 tokenizer
decode 129 frames of 720p without ever materialising the whole volume — and it means a kernel that
zero-pads instead of reading the cache is wrong on the first two output frames of every chunk. At `Tc = 1`
that is *every* output.

`cache_out` is the input-side carry for the next chunk: the last two frames of `[cache | x]`. When
`Tc >= 2` that is just `x[:, :, -2:]`; when `Tc == 1` it is `[cache[:, :, 1], x[:, :, 0]]`, and `Tc = 1`
does appear in the correctness shapes. The contract is **functional** — return a new tensor, do not mutate
`cache` in place; the grader calls your function and the reference on the same buffers, so an in-place
update corrupts the comparison and fails.

Spatial padding is the ordinary symmetric zero padding of a `3x3` conv, so `H` and `W` are preserved.

`/app/reference.py` concatenates, spatially pads, and calls `F.conv3d` in fp32 with TF32 disabled. That is
the exact specification; it is deliberately simple rather than fast.""",

    contract_md="""| arg | shape | dtype | meaning |
|-----|-------|-------|---------|
| `x` | `(B, Cin, Tc, H, W)` | `bfloat16` | this chunk's input frames, contiguous NCDHW |
| `cache` | `(B, Cin, 2, H, W)` | `bfloat16` | the two input frames before the chunk, **oldest at index 0** |
| `weight` | `(Cout, Cin, 3, 3, 3)` | `bfloat16` | taps ordered `(kt, kh, kw)`; `weight[..., 2, :, :]` hits the current frame |
| `bias` | `(Cout,)` | `bfloat16` | per-output-channel bias |

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

| out | shape | dtype | notes |
|-----|-------|-------|-------|
| `y` | `(B, Cout, Tc, H, W)` | `bfloat16` | contiguous NCDHW, one output per input frame of the chunk |
| `cache_out` | `(B, Cin, 2, H, W)` | `bfloat16` | last two frames of `[cache \\| x]`; **must be bf16** |

`Tc` can be as small as **1** and is never larger than a handful — this is a streaming decode, not a full
clip. `Cin` and `Cout` are independent and need not be multiples of any tile size (correctness shapes
include `Cin = 16`, `Cout = 24`, `H = 17`, `W = 23`). All inputs are **read-only** and the cache update is
**functional**.""",

    regime_md="""**Shape regime you are graded in** (the exact grader sizes are *not* disclosed): `B` in 1–2,
`Cin`/`Cout` in 64–256, `Tc` (chunk depth) in **1–4**, and `H`/`W` in 224–480 — a decoder layer running at
or near pixel resolution while streaming a 720p clip a few frames at a time. This is the short-and-wide
regime: two frames of halo per chunk is a large fraction of the input, so the kernel has far less temporal
reuse than the full-clip convolution and the cache traffic is a first-order cost. Resolution is capped so
the fp32 reference fits in memory. Write a **general** kernel — `Tc = 1` and ragged channels both appear.""",

    correctness_md="""**Both** returned tensors must match the reference (evaluated in fp32) within
**relative Frobenius error `1e-2`** at every graded shape, including the timed ones. A wrong `cache_out`
corrupts every subsequent chunk, so it is graded exactly as hard as `y`.""",

    perf_md="""The chunk is only a few frames deep but full pixel resolution, which changes the balance
completely relative to the full-clip convolution.

**The halo is expensive.** With `Tc = 2`, the two cache frames are as much data as the chunk itself: half of
the input traffic produces no output. There is nothing to do about the bytes, but there is a lot to do about
*not reading them twice* — load a `(2 + BT, BH + 2, BW + 2)` halo once into shared memory and let every
output frame in the tile consume it.

**Fuse the cache copy into the main kernel.** `cache_out` is a slice of data the convolution already has in
registers or SMEM. Emitting it from a second kernel costs an extra full read of two frames of the volume,
which at `H = W = 384` and `Cin = 128` is another 75 MB round trip per layer per step.

**Layout still decides everything.** NCDHW puts the reduction axis `Cin` slowest. Transpose to NDHWC
internally so the `Cin*27` reduction is contiguous and the convolution is an implicit GEMM of shape
`(B*Tc*H*W) x (Cin*27) x Cout`. Never materialise the `27x` im2col matrix.

**Watch the occupancy at `Tc = 1`.** The output volume is `B*Cout*H*W`, so the parallelism is entirely
spatial; a tiling that assumes a deep `T` axis will leave the machine half idle. Conversely, at `Tc = 1`
every one of the 27 taps still has to be applied — three temporal taps against `cache[0]`, `cache[1]` and
`x[0]` respectively — so there is no shortcut, only better scheduling.

Other things that matter: keep the weight tensor (at most 4.5 MB) resident across the tile; vectorise loads
and stores along the contiguous channel axis of your internal layout; and predicate the spatial halo rather
than writing a padded copy of a multi-hundred-megabyte volume.""",

    precision_md="""Inputs and outputs are **bfloat16**; accumulate in **fp32**. The reference convolves in
fp32 with TF32 disabled, so it is a true fp32 result and a faithful bf16 kernel's only error is the final
rounding of `y` plus accumulation order.

The tolerance was **measured** on the same convolution: an independent fp32 implementation (27 shifted
`Cin x Cout` matmuls) differs from this reference by `4.7e-5`, and a bf16 convolution with fp32
accumulation by `2.9e-3`. The gate is `1e-2`, about 3.4x the observed bf16 noise. Zero-padding instead of
reading the cache is *not* noise: it is wrong on the first two output frames of every chunk, which at
`Tc <= 4` is at least half the output and lands orders of magnitude outside the gate.

`cache_out` **must be bfloat16** — it is the carry the next chunk consumes, and widening it would both break
the contract and hide the quantisation the real streaming decoder lives with.

**fp8 is not appropriate here**; the graded dtype is bf16 in and bf16 out.""",
).validate()