Add torch213-cxx11-cu130-aarch64-linux artifact
Browse files- build/torch213-cxx11-cu130-aarch64-linux/__init__.py +414 -0
- build/torch213-cxx11-cu130-aarch64-linux/_ops.py +6 -0
- build/torch213-cxx11-cu130-aarch64-linux/flashrt_spatiotemporal_layout/__init__.py +14 -0
- build/torch213-cxx11-cu130-aarch64-linux/flashrt_spatiotemporal_layout_test.abi3.so +3 -0
- build/torch213-cxx11-cu130-aarch64-linux/metadata.json +32 -0
build/torch213-cxx11-cu130-aarch64-linux/__init__.py
ADDED
|
@@ -0,0 +1,414 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""FlashRT spatiotemporal layout kernels."""
|
| 2 |
+
|
| 3 |
+
from __future__ import annotations
|
| 4 |
+
|
| 5 |
+
import torch
|
| 6 |
+
|
| 7 |
+
from ._ops import add_op_namespace_prefix, ops
|
| 8 |
+
|
| 9 |
+
|
| 10 |
+
def _check_ncdhw(x: torch.Tensor, name: str) -> None:
|
| 11 |
+
if x.dim() != 5:
|
| 12 |
+
raise RuntimeError(f"{name} must have shape (B, C, T, H, W)")
|
| 13 |
+
|
| 14 |
+
|
| 15 |
+
@torch.library.register_fake(add_op_namespace_prefix("ncdhw_to_blc_bf16"))
|
| 16 |
+
def _ncdhw_to_blc_bf16_fake(x: torch.Tensor, out: torch.Tensor) -> None:
|
| 17 |
+
_check_ncdhw(x, "x")
|
| 18 |
+
b, c, t, h, w = x.shape
|
| 19 |
+
if out.shape != (b, t * h * w, c):
|
| 20 |
+
raise RuntimeError("out must have shape (B, T * H * W, C)")
|
| 21 |
+
return None
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
@torch.library.register_fake(add_op_namespace_prefix("patch_im2col_bf16"))
|
| 25 |
+
def _patch_im2col_bf16_fake(x: torch.Tensor, out: torch.Tensor) -> None:
|
| 26 |
+
if x.dim() != 4 or x.shape[1:] != (224, 224, 3):
|
| 27 |
+
raise RuntimeError("x must have shape (num_views, 224, 224, 3)")
|
| 28 |
+
if out.shape != (x.shape[0] * 256, 588):
|
| 29 |
+
raise RuntimeError("out must have shape (num_views * 256, 588)")
|
| 30 |
+
return None
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
@torch.library.register_fake(add_op_namespace_prefix("time_unshuffle2_bf16"))
|
| 34 |
+
def _time_unshuffle2_bf16_fake(x: torch.Tensor, out: torch.Tensor) -> None:
|
| 35 |
+
_check_ncdhw(x, "x")
|
| 36 |
+
b, c2, t, h, w = x.shape
|
| 37 |
+
if c2 % 2 != 0:
|
| 38 |
+
raise RuntimeError("x.shape[1] must be even")
|
| 39 |
+
if out.shape != (b, c2 // 2, 2 * t, h, w):
|
| 40 |
+
raise RuntimeError("out must have shape (B, C / 2, 2 * T, H, W)")
|
| 41 |
+
return None
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
@torch.library.register_fake(add_op_namespace_prefix("add_bias_ncdhw_bf16"))
|
| 45 |
+
def _add_bias_ncdhw_bf16_fake(x: torch.Tensor, bias: torch.Tensor) -> None:
|
| 46 |
+
_check_ncdhw(x, "x")
|
| 47 |
+
if bias.shape != (x.shape[1],):
|
| 48 |
+
raise RuntimeError("bias must have shape (C,)")
|
| 49 |
+
return None
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
@torch.library.register_fake(add_op_namespace_prefix("update_cache2_ncdhw_bf16"))
|
| 53 |
+
def _update_cache2_ncdhw_bf16_fake(cur: torch.Tensor, prev: torch.Tensor, out: torch.Tensor) -> None:
|
| 54 |
+
_check_ncdhw(cur, "cur")
|
| 55 |
+
b, c, _, h, w = cur.shape
|
| 56 |
+
if prev.shape != (b, c, 2, h, w) or out.shape != (b, c, 2, h, w):
|
| 57 |
+
raise RuntimeError("prev and out must have shape (B, C, 2, H, W)")
|
| 58 |
+
return None
|
| 59 |
+
|
| 60 |
+
|
| 61 |
+
@torch.library.register_fake(add_op_namespace_prefix("channel_to_space3d_bf16"))
|
| 62 |
+
def _channel_to_space3d_bf16_fake(
|
| 63 |
+
x: torch.Tensor,
|
| 64 |
+
out_channels: int,
|
| 65 |
+
temporal_factor: int,
|
| 66 |
+
spatial_factor: int,
|
| 67 |
+
repeats: int,
|
| 68 |
+
first_chunk: bool,
|
| 69 |
+
out: torch.Tensor,
|
| 70 |
+
) -> None:
|
| 71 |
+
_check_ncdhw(x, "x")
|
| 72 |
+
out_t = x.shape[2] * temporal_factor - (
|
| 73 |
+
temporal_factor - 1 if first_chunk else 0
|
| 74 |
+
)
|
| 75 |
+
expected = (
|
| 76 |
+
x.shape[0],
|
| 77 |
+
out_channels,
|
| 78 |
+
out_t,
|
| 79 |
+
x.shape[3] * spatial_factor,
|
| 80 |
+
x.shape[4] * spatial_factor,
|
| 81 |
+
)
|
| 82 |
+
if (
|
| 83 |
+
out_channels <= 0
|
| 84 |
+
or temporal_factor <= 0
|
| 85 |
+
or spatial_factor <= 0
|
| 86 |
+
or repeats <= 0
|
| 87 |
+
or x.shape[1] * repeats
|
| 88 |
+
< out_channels * temporal_factor * spatial_factor * spatial_factor
|
| 89 |
+
or out.shape != expected
|
| 90 |
+
):
|
| 91 |
+
raise RuntimeError("channel_to_space3d_bf16 shape contract failed")
|
| 92 |
+
return None
|
| 93 |
+
|
| 94 |
+
|
| 95 |
+
@torch.library.register_fake(
|
| 96 |
+
add_op_namespace_prefix("pack_causal_cache3_nhwc_bf16")
|
| 97 |
+
)
|
| 98 |
+
def _pack_causal_cache3_nhwc_bf16_fake(
|
| 99 |
+
previous: torch.Tensor, current: torch.Tensor, out: torch.Tensor
|
| 100 |
+
) -> None:
|
| 101 |
+
_check_ncdhw(previous, "previous")
|
| 102 |
+
_check_ncdhw(current, "current")
|
| 103 |
+
b, c, t, h, w = current.shape
|
| 104 |
+
if (
|
| 105 |
+
t != 1
|
| 106 |
+
or previous.shape != (b, c, 2, h, w)
|
| 107 |
+
or out.shape != (b, h, w, 3 * c)
|
| 108 |
+
):
|
| 109 |
+
raise RuntimeError("causal cache pack shape contract failed")
|
| 110 |
+
return None
|
| 111 |
+
|
| 112 |
+
|
| 113 |
+
@torch.library.register_fake(add_op_namespace_prefix("avg_pool3d_channels_bf16"))
|
| 114 |
+
def _avg_pool3d_channels_bf16_fake(
|
| 115 |
+
x: torch.Tensor,
|
| 116 |
+
out_channels: int,
|
| 117 |
+
factor_t: int,
|
| 118 |
+
factor_s: int,
|
| 119 |
+
group_size: int,
|
| 120 |
+
out: torch.Tensor,
|
| 121 |
+
) -> None:
|
| 122 |
+
_check_ncdhw(x, "x")
|
| 123 |
+
expected = (
|
| 124 |
+
x.shape[0],
|
| 125 |
+
out_channels,
|
| 126 |
+
(x.shape[2] + factor_t - 1) // factor_t,
|
| 127 |
+
x.shape[3] // factor_s,
|
| 128 |
+
x.shape[4] // factor_s,
|
| 129 |
+
)
|
| 130 |
+
if out.shape != expected:
|
| 131 |
+
raise RuntimeError("out has the wrong pooled NCDHW shape")
|
| 132 |
+
return None
|
| 133 |
+
|
| 134 |
+
|
| 135 |
+
@torch.library.register_fake(add_op_namespace_prefix("ndhwc_to_ncdhw_bf16"))
|
| 136 |
+
def _ndhwc_to_ncdhw_bf16_fake(x: torch.Tensor, out: torch.Tensor) -> None:
|
| 137 |
+
if x.dim() != 5:
|
| 138 |
+
raise RuntimeError("x must have shape (B, T, H, W, C)")
|
| 139 |
+
expected = (x.shape[0], x.shape[4], x.shape[1], x.shape[2], x.shape[3])
|
| 140 |
+
if out.shape != expected:
|
| 141 |
+
raise RuntimeError("out must have shape (B, C, T, H, W)")
|
| 142 |
+
return None
|
| 143 |
+
|
| 144 |
+
|
| 145 |
+
@torch.library.register_fake(add_op_namespace_prefix("ndhwc_to_ncdhw_bias_bf16"))
|
| 146 |
+
def _ndhwc_to_ncdhw_bias_bf16_fake(
|
| 147 |
+
x: torch.Tensor, bias: torch.Tensor, out: torch.Tensor
|
| 148 |
+
) -> None:
|
| 149 |
+
_ndhwc_to_ncdhw_bf16_fake(x, out)
|
| 150 |
+
if bias.shape != (x.shape[4],):
|
| 151 |
+
raise RuntimeError("bias must have shape (C,)")
|
| 152 |
+
return None
|
| 153 |
+
|
| 154 |
+
|
| 155 |
+
@torch.library.register_fake(add_op_namespace_prefix("ndhwc_to_ncdhw_add_bf16"))
|
| 156 |
+
def _ndhwc_to_ncdhw_add_bf16_fake(
|
| 157 |
+
x: torch.Tensor, residual: torch.Tensor, out: torch.Tensor
|
| 158 |
+
) -> None:
|
| 159 |
+
_ndhwc_to_ncdhw_bf16_fake(x, out)
|
| 160 |
+
if residual.shape != out.shape:
|
| 161 |
+
raise RuntimeError("residual must match the NCDHW output shape")
|
| 162 |
+
return None
|
| 163 |
+
|
| 164 |
+
|
| 165 |
+
@torch.library.register_fake(
|
| 166 |
+
add_op_namespace_prefix("ncdhw_quantize_fp8_static_ndhwc_bf16")
|
| 167 |
+
)
|
| 168 |
+
def _ncdhw_quantize_fp8_static_ndhwc_bf16_fake(
|
| 169 |
+
x: torch.Tensor, scale: float, out: torch.Tensor
|
| 170 |
+
) -> None:
|
| 171 |
+
_check_ncdhw(x, "x")
|
| 172 |
+
expected = (x.shape[0], x.shape[2], x.shape[3], x.shape[4], x.shape[1])
|
| 173 |
+
if out.shape != expected:
|
| 174 |
+
raise RuntimeError("out must have shape (B, T, H, W, C)")
|
| 175 |
+
return None
|
| 176 |
+
|
| 177 |
+
|
| 178 |
+
@torch.library.register_fake(
|
| 179 |
+
add_op_namespace_prefix("upsample2x_quantize_fp8_static_nhwc_bf16")
|
| 180 |
+
)
|
| 181 |
+
def _upsample2x_quantize_fp8_static_nhwc_bf16_fake(
|
| 182 |
+
x: torch.Tensor, scale: float, out: torch.Tensor
|
| 183 |
+
) -> None:
|
| 184 |
+
if x.dim() != 4:
|
| 185 |
+
raise RuntimeError("x must have shape (N, C, H, W)")
|
| 186 |
+
expected = (x.shape[0], 2 * x.shape[2], 2 * x.shape[3], x.shape[1])
|
| 187 |
+
if out.shape != expected:
|
| 188 |
+
raise RuntimeError("out must have shape (N, 2H, 2W, C)")
|
| 189 |
+
return None
|
| 190 |
+
|
| 191 |
+
|
| 192 |
+
def ncdhw_to_blc_bf16(x: torch.Tensor, out: torch.Tensor | None = None) -> torch.Tensor:
|
| 193 |
+
"""Convert BF16 NCDHW tensor to BLC where L = T * H * W."""
|
| 194 |
+
|
| 195 |
+
if out is None:
|
| 196 |
+
out = torch.empty((x.shape[0], x.shape[2] * x.shape[3] * x.shape[4], x.shape[1]), device=x.device, dtype=x.dtype)
|
| 197 |
+
ops.ncdhw_to_blc_bf16(x, out)
|
| 198 |
+
return out
|
| 199 |
+
|
| 200 |
+
|
| 201 |
+
def patch_im2col_bf16(x: torch.Tensor, out: torch.Tensor | None = None) -> torch.Tensor:
|
| 202 |
+
"""Convert BF16 NHWC images into flattened 14x14 patch rows."""
|
| 203 |
+
|
| 204 |
+
if out is None:
|
| 205 |
+
out = torch.empty((x.shape[0] * 256, 588), device=x.device, dtype=x.dtype)
|
| 206 |
+
ops.patch_im2col_bf16(x, out)
|
| 207 |
+
return out
|
| 208 |
+
|
| 209 |
+
|
| 210 |
+
def time_unshuffle2_bf16(x: torch.Tensor, out: torch.Tensor | None = None) -> torch.Tensor:
|
| 211 |
+
"""Convert BF16 (B, 2C, T, H, W) into (B, C, 2T, H, W)."""
|
| 212 |
+
|
| 213 |
+
if out is None:
|
| 214 |
+
out = torch.empty((x.shape[0], x.shape[1] // 2, 2 * x.shape[2], x.shape[3], x.shape[4]), device=x.device, dtype=x.dtype)
|
| 215 |
+
ops.time_unshuffle2_bf16(x, out)
|
| 216 |
+
return out
|
| 217 |
+
|
| 218 |
+
|
| 219 |
+
def add_bias_ncdhw_bf16(x: torch.Tensor, bias: torch.Tensor) -> torch.Tensor:
|
| 220 |
+
"""Add channel bias to an NCDHW tensor in place."""
|
| 221 |
+
|
| 222 |
+
ops.add_bias_ncdhw_bf16(x, bias)
|
| 223 |
+
return x
|
| 224 |
+
|
| 225 |
+
|
| 226 |
+
def update_cache2_ncdhw_bf16(cur: torch.Tensor, prev: torch.Tensor, out: torch.Tensor | None = None) -> torch.Tensor:
|
| 227 |
+
"""Update two-frame latent cache from current and previous NCDHW tensors."""
|
| 228 |
+
|
| 229 |
+
if out is None:
|
| 230 |
+
out = torch.empty((cur.shape[0], cur.shape[1], 2, cur.shape[3], cur.shape[4]), device=cur.device, dtype=cur.dtype)
|
| 231 |
+
ops.update_cache2_ncdhw_bf16(cur, prev, out)
|
| 232 |
+
return out
|
| 233 |
+
|
| 234 |
+
|
| 235 |
+
def channel_to_space3d_bf16(
|
| 236 |
+
x: torch.Tensor,
|
| 237 |
+
out_channels: int,
|
| 238 |
+
temporal_factor: int,
|
| 239 |
+
spatial_factor: int,
|
| 240 |
+
repeats: int = 1,
|
| 241 |
+
first_chunk: bool = False,
|
| 242 |
+
out: torch.Tensor | None = None,
|
| 243 |
+
) -> torch.Tensor:
|
| 244 |
+
"""Move expanded channels into temporal/spatial dimensions."""
|
| 245 |
+
|
| 246 |
+
out_t = x.shape[2] * temporal_factor - (
|
| 247 |
+
temporal_factor - 1 if first_chunk else 0
|
| 248 |
+
)
|
| 249 |
+
if out is None:
|
| 250 |
+
out = torch.empty(
|
| 251 |
+
(
|
| 252 |
+
x.shape[0],
|
| 253 |
+
out_channels,
|
| 254 |
+
out_t,
|
| 255 |
+
x.shape[3] * spatial_factor,
|
| 256 |
+
x.shape[4] * spatial_factor,
|
| 257 |
+
),
|
| 258 |
+
device=x.device,
|
| 259 |
+
dtype=x.dtype,
|
| 260 |
+
)
|
| 261 |
+
ops.channel_to_space3d_bf16(
|
| 262 |
+
x,
|
| 263 |
+
int(out_channels),
|
| 264 |
+
int(temporal_factor),
|
| 265 |
+
int(spatial_factor),
|
| 266 |
+
int(repeats),
|
| 267 |
+
bool(first_chunk),
|
| 268 |
+
out,
|
| 269 |
+
)
|
| 270 |
+
return out
|
| 271 |
+
|
| 272 |
+
|
| 273 |
+
def pack_causal_cache3_nhwc_bf16(
|
| 274 |
+
previous: torch.Tensor,
|
| 275 |
+
current: torch.Tensor,
|
| 276 |
+
out: torch.Tensor | None = None,
|
| 277 |
+
) -> torch.Tensor:
|
| 278 |
+
"""Pack two cached and one current NCDHW frame into NHWC with 3C."""
|
| 279 |
+
|
| 280 |
+
if out is None:
|
| 281 |
+
out = torch.empty(
|
| 282 |
+
(
|
| 283 |
+
current.shape[0],
|
| 284 |
+
current.shape[3],
|
| 285 |
+
current.shape[4],
|
| 286 |
+
3 * current.shape[1],
|
| 287 |
+
),
|
| 288 |
+
device=current.device,
|
| 289 |
+
dtype=current.dtype,
|
| 290 |
+
)
|
| 291 |
+
ops.pack_causal_cache3_nhwc_bf16(previous, current, out)
|
| 292 |
+
return out
|
| 293 |
+
|
| 294 |
+
|
| 295 |
+
def avg_pool3d_channels_bf16(
|
| 296 |
+
x: torch.Tensor,
|
| 297 |
+
out_channels: int,
|
| 298 |
+
factor_t: int,
|
| 299 |
+
factor_s: int,
|
| 300 |
+
group_size: int,
|
| 301 |
+
out: torch.Tensor | None = None,
|
| 302 |
+
) -> torch.Tensor:
|
| 303 |
+
"""Pool NCDHW blocks while folding spatiotemporal factors into channels."""
|
| 304 |
+
|
| 305 |
+
if out is None:
|
| 306 |
+
out = torch.empty(
|
| 307 |
+
(
|
| 308 |
+
x.shape[0],
|
| 309 |
+
out_channels,
|
| 310 |
+
(x.shape[2] + factor_t - 1) // factor_t,
|
| 311 |
+
x.shape[3] // factor_s,
|
| 312 |
+
x.shape[4] // factor_s,
|
| 313 |
+
),
|
| 314 |
+
device=x.device,
|
| 315 |
+
dtype=x.dtype,
|
| 316 |
+
)
|
| 317 |
+
ops.avg_pool3d_channels_bf16(
|
| 318 |
+
x, out_channels, factor_t, factor_s, group_size, out
|
| 319 |
+
)
|
| 320 |
+
return out
|
| 321 |
+
|
| 322 |
+
|
| 323 |
+
def ndhwc_to_ncdhw_bf16(
|
| 324 |
+
x: torch.Tensor, out: torch.Tensor | None = None
|
| 325 |
+
) -> torch.Tensor:
|
| 326 |
+
"""Convert contiguous BF16 NDHWC to contiguous BF16 NCDHW."""
|
| 327 |
+
|
| 328 |
+
if out is None:
|
| 329 |
+
out = torch.empty(
|
| 330 |
+
(x.shape[0], x.shape[4], x.shape[1], x.shape[2], x.shape[3]),
|
| 331 |
+
device=x.device,
|
| 332 |
+
dtype=x.dtype,
|
| 333 |
+
)
|
| 334 |
+
ops.ndhwc_to_ncdhw_bf16(x, out)
|
| 335 |
+
return out
|
| 336 |
+
|
| 337 |
+
|
| 338 |
+
def ndhwc_to_ncdhw_bias_bf16(
|
| 339 |
+
x: torch.Tensor, bias: torch.Tensor, out: torch.Tensor | None = None
|
| 340 |
+
) -> torch.Tensor:
|
| 341 |
+
"""Convert NDHWC to NCDHW and add a BF16 channel bias."""
|
| 342 |
+
|
| 343 |
+
if out is None:
|
| 344 |
+
out = torch.empty(
|
| 345 |
+
(x.shape[0], x.shape[4], x.shape[1], x.shape[2], x.shape[3]),
|
| 346 |
+
device=x.device,
|
| 347 |
+
dtype=x.dtype,
|
| 348 |
+
)
|
| 349 |
+
ops.ndhwc_to_ncdhw_bias_bf16(x, bias, out)
|
| 350 |
+
return out
|
| 351 |
+
|
| 352 |
+
|
| 353 |
+
def ndhwc_to_ncdhw_add_bf16(
|
| 354 |
+
x: torch.Tensor,
|
| 355 |
+
residual: torch.Tensor,
|
| 356 |
+
out: torch.Tensor | None = None,
|
| 357 |
+
) -> torch.Tensor:
|
| 358 |
+
"""Convert NDHWC to NCDHW and add a BF16 NCDHW residual."""
|
| 359 |
+
|
| 360 |
+
if out is None:
|
| 361 |
+
out = torch.empty_like(residual)
|
| 362 |
+
ops.ndhwc_to_ncdhw_add_bf16(x, residual, out)
|
| 363 |
+
return out
|
| 364 |
+
|
| 365 |
+
|
| 366 |
+
def ncdhw_quantize_fp8_static_ndhwc_bf16(
|
| 367 |
+
x: torch.Tensor,
|
| 368 |
+
scale: float,
|
| 369 |
+
out: torch.Tensor | None = None,
|
| 370 |
+
) -> torch.Tensor:
|
| 371 |
+
"""Quantize BF16 NCDHW to FP8 E4M3 NDHWC using a static scale."""
|
| 372 |
+
|
| 373 |
+
if out is None:
|
| 374 |
+
out = torch.empty(
|
| 375 |
+
(x.shape[0], x.shape[2], x.shape[3], x.shape[4], x.shape[1]),
|
| 376 |
+
device=x.device,
|
| 377 |
+
dtype=torch.float8_e4m3fn,
|
| 378 |
+
)
|
| 379 |
+
ops.ncdhw_quantize_fp8_static_ndhwc_bf16(x, scale, out)
|
| 380 |
+
return out
|
| 381 |
+
|
| 382 |
+
|
| 383 |
+
def upsample2x_quantize_fp8_static_nhwc_bf16(
|
| 384 |
+
x: torch.Tensor,
|
| 385 |
+
scale: float,
|
| 386 |
+
out: torch.Tensor | None = None,
|
| 387 |
+
) -> torch.Tensor:
|
| 388 |
+
"""Nearest-neighbor 2x upsample NCHW and emit static-scaled FP8 NHWC."""
|
| 389 |
+
|
| 390 |
+
if out is None:
|
| 391 |
+
out = torch.empty(
|
| 392 |
+
(x.shape[0], 2 * x.shape[2], 2 * x.shape[3], x.shape[1]),
|
| 393 |
+
device=x.device,
|
| 394 |
+
dtype=torch.float8_e4m3fn,
|
| 395 |
+
)
|
| 396 |
+
ops.upsample2x_quantize_fp8_static_nhwc_bf16(x, scale, out)
|
| 397 |
+
return out
|
| 398 |
+
|
| 399 |
+
|
| 400 |
+
__all__ = [
|
| 401 |
+
"ncdhw_to_blc_bf16",
|
| 402 |
+
"patch_im2col_bf16",
|
| 403 |
+
"time_unshuffle2_bf16",
|
| 404 |
+
"add_bias_ncdhw_bf16",
|
| 405 |
+
"update_cache2_ncdhw_bf16",
|
| 406 |
+
"channel_to_space3d_bf16",
|
| 407 |
+
"pack_causal_cache3_nhwc_bf16",
|
| 408 |
+
"avg_pool3d_channels_bf16",
|
| 409 |
+
"ndhwc_to_ncdhw_bf16",
|
| 410 |
+
"ndhwc_to_ncdhw_bias_bf16",
|
| 411 |
+
"ndhwc_to_ncdhw_add_bf16",
|
| 412 |
+
"ncdhw_quantize_fp8_static_ndhwc_bf16",
|
| 413 |
+
"upsample2x_quantize_fp8_static_nhwc_bf16",
|
| 414 |
+
]
|
build/torch213-cxx11-cu130-aarch64-linux/_ops.py
ADDED
|
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
from . import flashrt_spatiotemporal_layout_test
|
| 3 |
+
ops = torch.ops.flashrt_spatiotemporal_layout_test
|
| 4 |
+
|
| 5 |
+
def add_op_namespace_prefix(op_name: str):
|
| 6 |
+
return f"flashrt_spatiotemporal_layout_test::{op_name}"
|
build/torch213-cxx11-cu130-aarch64-linux/flashrt_spatiotemporal_layout/__init__.py
ADDED
|
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import ctypes
|
| 2 |
+
import importlib.util
|
| 3 |
+
import sys
|
| 4 |
+
from pathlib import Path
|
| 5 |
+
|
| 6 |
+
def _import_from_path(file_path: Path):
|
| 7 |
+
path_hash = '{:x}'.format(ctypes.c_size_t(hash(file_path.absolute())).value)
|
| 8 |
+
spec = importlib.util.spec_from_file_location(path_hash, file_path)
|
| 9 |
+
module = importlib.util.module_from_spec(spec)
|
| 10 |
+
sys.modules[path_hash] = module
|
| 11 |
+
spec.loader.exec_module(module)
|
| 12 |
+
return module
|
| 13 |
+
|
| 14 |
+
globals().update(vars(_import_from_path(Path(__file__).parent.parent / '__init__.py')))
|
build/torch213-cxx11-cu130-aarch64-linux/flashrt_spatiotemporal_layout_test.abi3.so
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:16f0b72634927e102b0ce7c17a8addd5eaed791906b19d034f2729cf909e895a
|
| 3 |
+
size 461408
|
build/torch213-cxx11-cu130-aarch64-linux/metadata.json
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"name": "flashrt-spatiotemporal-layout",
|
| 3 |
+
"id": "flashrt_spatiotemporal_layout_test",
|
| 4 |
+
"version": 1,
|
| 5 |
+
"license": "Apache-2.0",
|
| 6 |
+
"python-depends": [],
|
| 7 |
+
"backend": {
|
| 8 |
+
"type": "cuda",
|
| 9 |
+
"archs": [
|
| 10 |
+
"11.0"
|
| 11 |
+
]
|
| 12 |
+
},
|
| 13 |
+
"digest": {
|
| 14 |
+
"algorithm": "sha256",
|
| 15 |
+
"files": {
|
| 16 |
+
"__init__.py": "4WRzvPydgNf2NK7ZljUlhO8NWc3qbONgjAzNb7YayKE=",
|
| 17 |
+
"flashrt_spatiotemporal_layout_test.abi3.so": "FvC3JjSSfhArDOfBeord1erteRkGsZ0DTycpz5CeiVo=",
|
| 18 |
+
"_ops.py": "P29ghFUoONBA3NS6cY+kzbwr/RczjL8AdMewXrgZDCA=",
|
| 19 |
+
"flashrt_spatiotemporal_layout/__init__.py": "v6p5XMfQzddhi1fLSAw4HX9CyS0rQsidvu9VsT01xi4="
|
| 20 |
+
}
|
| 21 |
+
},
|
| 22 |
+
"provenance": {
|
| 23 |
+
"kernel": {
|
| 24 |
+
"sha": "456d297",
|
| 25 |
+
"dirty": false
|
| 26 |
+
},
|
| 27 |
+
"validation": {
|
| 28 |
+
"torch": "2.13.0+cu130",
|
| 29 |
+
"cuda": "13.0"
|
| 30 |
+
}
|
| 31 |
+
}
|
| 32 |
+
}
|