Promote latest kernel artifacts to main
Browse files- README.md +148 -6
- benchmarks/RESULTS.md +0 -9
- benchmarks/benchmark.py +60 -0
- build/torch211-cxx11-cu128-x86_64-linux/__init__.py +120 -0
- build/torch211-cxx11-cu128-x86_64-linux/{_adaptive_layernorm_producers_cuda_9b532b4.abi3.so → _adaptive_layernorm_producers_cuda_bb9b4e5.abi3.so} +2 -2
- build/torch211-cxx11-cu128-x86_64-linux/_ops.py +3 -3
- build/torch211-cxx11-cu128-x86_64-linux/metadata.json +22 -2
README.md
CHANGED
|
@@ -1,9 +1,151 @@
|
|
| 1 |
-
#
|
| 2 |
|
| 3 |
-
|
| 4 |
-
|
| 5 |
|
| 6 |
-
|
| 7 |
|
| 8 |
-
|
| 9 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# adaptive-layernorm-producers
|
| 2 |
|
| 3 |
+
FlashRT native CUDA producer kernels for DiT/Wan-style adaptive LayerNorm
|
| 4 |
+
blocks.
|
| 5 |
|
| 6 |
+
This package fuses the producer side of the block:
|
| 7 |
|
| 8 |
+
```text
|
| 9 |
+
LayerNorm / adaptive LayerNorm modulation -> low-precision activation
|
| 10 |
+
```
|
| 11 |
+
|
| 12 |
+
into a single Kernel Hub operator. It is intended for transformer and diffuser
|
| 13 |
+
runtimes where the following GEMM already consumes FP8 or NVFP4 activations.
|
| 14 |
+
Without this fusion, the runtime pays for separate normalization, modulation,
|
| 15 |
+
BF16 materialization, and quantization launches before the GEMM.
|
| 16 |
+
|
| 17 |
+
The source kernels are derived from the production FlashRT runtime:
|
| 18 |
+
|
| 19 |
+
- `official/FlashRT/csrc/quantize/ada_layer_norm_fp8.cu`
|
| 20 |
+
- `official/FlashRT/csrc/kernels/dit_bf16.cu::layer_norm_no_affine_fp8_static_bf16`
|
| 21 |
+
|
| 22 |
+
The full FlashRT model runtime and serving pipeline live upstream at
|
| 23 |
+
[LiangSu8899/FlashRT](https://github.com/LiangSu8899/FlashRT).
|
| 24 |
+
|
| 25 |
+
## Available Functions
|
| 26 |
+
|
| 27 |
+
- `ada_layer_norm_quant_fp8_bf16(x, scale, shift, act_scale, eps=1e-5, out=None)`
|
| 28 |
+
- `ada_layer_norm_quant_fp8_ptok_bf16(x, scale, shift, act_scale, eps=1e-5, out=None)`
|
| 29 |
+
- `ada_layer_norm_quant_fp8_ptok_table_bf16(x, temb, table, act_scale, shift_idx, scale_idx, eps=1e-5, out=None)`
|
| 30 |
+
- `ada_layer_norm_quant_fp8_modfp8_bf16(x, scale_fp8, shift_fp8, scale_deq, shift_deq, act_scale, eps=1e-5, out=None)`
|
| 31 |
+
- `awq_ada_layer_norm_quant_fp8_bf16(x, scale, shift, inv_s, act_scale, eps=1e-5, out=None)`
|
| 32 |
+
- `ada_layer_norm_quant_nvfp4_swizzled_bf16(x, scale, shift, eps=1e-5, packed=None, sf_swizzled=None)`
|
| 33 |
+
- `ada_layer_norm_quant_nvfp4_swizzled_modfp8_bf16(x, scale_fp8, shift_fp8, scale_deq, shift_deq, eps=1e-5, packed=None, sf_swizzled=None)`
|
| 34 |
+
- `layer_norm_no_affine_quant_fp8_static_bf16(x, act_scale, eps=1e-5, out=None)`
|
| 35 |
+
- `layer_norm_no_affine_quant_nvfp4_swizzled_bf16(x, eps=1e-5, packed=None, sf_swizzled=None)`
|
| 36 |
+
- `adaln_modulation6_bf16(adaln_params, layer_modulation, out=None)`
|
| 37 |
+
- `swizzled_sf_size(rows, dim)`
|
| 38 |
+
|
| 39 |
+
## Tensor Contract
|
| 40 |
+
|
| 41 |
+
- `x`: contiguous CUDA `torch.bfloat16`, shape `(rows, dim)`.
|
| 42 |
+
- Broadcast `scale`, `shift`, and `inv_s`: contiguous CUDA
|
| 43 |
+
`torch.bfloat16`, shape `(dim,)`.
|
| 44 |
+
- Per-token `scale` and `shift`: contiguous CUDA `torch.bfloat16`, shape
|
| 45 |
+
`(rows, dim)`.
|
| 46 |
+
- Table form `temb`: contiguous CUDA `torch.bfloat16`, shape
|
| 47 |
+
`(rows, chunks, dim)`; `table`: contiguous CUDA `torch.float32`, shape
|
| 48 |
+
`(chunks, dim)`.
|
| 49 |
+
- `scale_fp8`, `shift_fp8`: contiguous CUDA `torch.float8_e4m3fn`, shape `(dim,)`.
|
| 50 |
+
- `scale_deq`, `shift_deq`, `act_scale`: CUDA `torch.float32` scalar tensors.
|
| 51 |
+
- FP8 outputs use `torch.float8_e4m3fn`, shape `(rows, dim)`.
|
| 52 |
+
- NVFP4 packed output uses `torch.uint8`, shape `(rows, dim // 2)`.
|
| 53 |
+
- NVFP4 scale output uses the FlashRT/CUTLASS 128x4 swizzled uint8 scale
|
| 54 |
+
layout. Allocate with `swizzled_sf_size(rows, dim)`.
|
| 55 |
+
- `dim` must be even for FP8 outputs and divisible by 16 for NVFP4 outputs.
|
| 56 |
+
- The package targets CUDA 12.8+ and Blackwell-class deployment paths.
|
| 57 |
+
- `adaln_modulation6_bf16` takes FP32 `(batch, sequence, 6, dim)` parameters
|
| 58 |
+
plus FP32 `(6, dim)` layer modulation and returns six BF16
|
| 59 |
+
`(batch, sequence, dim)` tensors.
|
| 60 |
+
|
| 61 |
+
## Minimal Usage
|
| 62 |
+
|
| 63 |
+
```python
|
| 64 |
+
from kernels import get_kernel
|
| 65 |
+
import torch
|
| 66 |
+
|
| 67 |
+
ops = get_kernel("flashrt/adaptive-layernorm-producers", version=1, trust_remote_code=True)
|
| 68 |
+
|
| 69 |
+
rows, dim = 2520, 3072
|
| 70 |
+
x = torch.randn((rows, dim), device="cuda", dtype=torch.bfloat16)
|
| 71 |
+
scale = torch.zeros((dim,), device="cuda", dtype=torch.bfloat16)
|
| 72 |
+
shift = torch.zeros((dim,), device="cuda", dtype=torch.bfloat16)
|
| 73 |
+
act_scale = torch.tensor([0.025], device="cuda", dtype=torch.float32)
|
| 74 |
+
|
| 75 |
+
x_fp8 = ops.ada_layer_norm_quant_fp8_bf16(x, scale, shift, act_scale)
|
| 76 |
+
```
|
| 77 |
+
|
| 78 |
+
Per-token modulation without materializing a broadcast buffer:
|
| 79 |
+
|
| 80 |
+
```python
|
| 81 |
+
ptok_scale = torch.zeros_like(x)
|
| 82 |
+
ptok_shift = torch.zeros_like(x)
|
| 83 |
+
x_fp8 = ops.ada_layer_norm_quant_fp8_ptok_bf16(
|
| 84 |
+
x, ptok_scale, ptok_shift, act_scale
|
| 85 |
+
)
|
| 86 |
+
```
|
| 87 |
+
|
| 88 |
+
Fused host modulation table add and chunk selection:
|
| 89 |
+
|
| 90 |
+
```python
|
| 91 |
+
chunks = 6
|
| 92 |
+
temb = torch.zeros((rows, chunks, dim), device="cuda", dtype=torch.bfloat16)
|
| 93 |
+
table = torch.zeros((chunks, dim), device="cuda", dtype=torch.float32)
|
| 94 |
+
x_fp8 = ops.ada_layer_norm_quant_fp8_ptok_table_bf16(
|
| 95 |
+
x, temb, table, act_scale, shift_idx=0, scale_idx=1
|
| 96 |
+
)
|
| 97 |
+
```
|
| 98 |
+
|
| 99 |
+
Static-buffer usage for CUDA Graph capture:
|
| 100 |
+
|
| 101 |
+
```python
|
| 102 |
+
out = torch.empty_like(x, dtype=torch.float8_e4m3fn)
|
| 103 |
+
ops.ada_layer_norm_quant_fp8_bf16(x, scale, shift, act_scale, out=out)
|
| 104 |
+
```
|
| 105 |
+
|
| 106 |
+
NVFP4 producer:
|
| 107 |
+
|
| 108 |
+
```python
|
| 109 |
+
packed, sf = ops.ada_layer_norm_quant_nvfp4_swizzled_bf16(x, scale, shift)
|
| 110 |
+
|
| 111 |
+
# GROOT N1.7 pre-FFN producer on Thor.
|
| 112 |
+
packed_ln, sf_ln = ops.layer_norm_no_affine_quant_nvfp4_swizzled_bf16(x)
|
| 113 |
+
```
|
| 114 |
+
|
| 115 |
+
Six-way DiT modulation:
|
| 116 |
+
|
| 117 |
+
```python
|
| 118 |
+
params = torch.randn((1, rows, 6, dim), device="cuda", dtype=torch.float32)
|
| 119 |
+
modulation = torch.randn((6, dim), device="cuda", dtype=torch.float32)
|
| 120 |
+
outputs = ops.adaln_modulation6_bf16(params, modulation)
|
| 121 |
+
```
|
| 122 |
+
|
| 123 |
+
## Validation
|
| 124 |
+
|
| 125 |
+
```bash
|
| 126 |
+
python adaptive-layernorm-producers/tests/test_adaptive_layernorm_producers.py --backend source --mode full
|
| 127 |
+
python adaptive-layernorm-producers/benchmarks/benchmark.py --backend source --iters 100
|
| 128 |
+
```
|
| 129 |
+
|
| 130 |
+
The validation suite checks exact FP8 output for small producer shapes, exact
|
| 131 |
+
NVFP4 packed/scalefactor output against the native staged BF16-norm plus FP4
|
| 132 |
+
quantizer contract on SM110, and strict FP8
|
| 133 |
+
boundary accounting for long video shapes. The long-shape policy requires
|
| 134 |
+
`p99_abs == 0` and only a tiny count of adjacent FP8-code boundary differences
|
| 135 |
+
against the eager reference.
|
| 136 |
+
|
| 137 |
+
See `VALIDATION.md` and `benchmarks/RESULTS.md` for the current local RTX 5090
|
| 138 |
+
source-build results.
|
| 139 |
+
|
| 140 |
+
The SM110 FP4 producer uses the native GROOT N1.7 contract from FlashRT commit
|
| 141 |
+
`24df793f4fa2d50780aea03b644208c6e0cb4162`. Its scale conversion and FP4 bin
|
| 142 |
+
boundaries intentionally differ from the older SM120 producer; release tests
|
| 143 |
+
compare each architecture against its own production contract.
|
| 144 |
+
|
| 145 |
+
The aarch64 release variant `torch211-cxx11-cu130-aarch64-linux` contains
|
| 146 |
+
native SM87 and SM110a code objects. Its SM110a path was validated on NVIDIA
|
| 147 |
+
Thor from the exact Hub-downloaded artifact over decode, DiT, vision, and
|
| 148 |
+
2K/4K video rows. The two per-token entry points were also compared directly
|
| 149 |
+
with their raw registered ops using A-B-B-A timing; all six wrapper/raw rows
|
| 150 |
+
were bit-exact and the measured ratio was `0.997-1.000`. SM87 runtime
|
| 151 |
+
validation remains pending on Jetson AGX Orin hardware.
|
benchmarks/RESULTS.md
CHANGED
|
@@ -59,12 +59,3 @@ were bit-exact in every row.
|
|
| 59 |
|
| 60 |
The table reports the uploaded SM87+SM110a fat binary on Thor. SM87 is present
|
| 61 |
as native cubin code but is not included in these runtime measurements.
|
| 62 |
-
|
| 63 |
-
### GROOT N1.7 FP4 producer update
|
| 64 |
-
|
| 65 |
-
The additive FP4 producer from FlashRT
|
| 66 |
-
`24df793f4fa2d50780aea03b644208c6e0cb4162` was rebuilt as
|
| 67 |
-
`torch213-cxx11-cu130-aarch64-linux` on the same Thor host. Packed FP4 values
|
| 68 |
-
and SFA bytes were exact against the native staged contract. The Tensor
|
| 69 |
-
wrapper/raw registered-op measurement was `8.3094/8.1932 us` (`1.0142x`) in
|
| 70 |
-
direct mode and `5.3749/5.3797 us` (`0.9991x`) under CUDA Graph.
|
|
|
|
| 59 |
|
| 60 |
The table reports the uploaded SM87+SM110a fat binary on Thor. SM87 is present
|
| 61 |
as native cubin code but is not included in these runtime measurements.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
benchmarks/benchmark.py
CHANGED
|
@@ -14,6 +14,7 @@ import torch
|
|
| 14 |
ROOT = Path(__file__).resolve().parents[2]
|
| 15 |
sys.path.insert(0, str(ROOT / "adaptive-layernorm-producers" / "tests"))
|
| 16 |
from test_adaptive_layernorm_producers import ( # noqa: E402
|
|
|
|
| 17 |
load_source_ops,
|
| 18 |
make_case,
|
| 19 |
quant_fp8,
|
|
@@ -169,6 +170,45 @@ def run_case(ops, name: str, rows: int, dim: int, eps: float, iters: int) -> dic
|
|
| 169 |
}
|
| 170 |
|
| 171 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 172 |
def main() -> None:
|
| 173 |
parser = argparse.ArgumentParser()
|
| 174 |
parser.add_argument("--backend", choices=["source", "installed"], default="source")
|
|
@@ -190,6 +230,13 @@ def main() -> None:
|
|
| 190 |
("wan_video_4k", 4096, 3072),
|
| 191 |
]
|
| 192 |
rows = [run_case(ops, name, r, d, 1e-5, args.iters) for name, r, d in shapes]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 193 |
modulation_rows = []
|
| 194 |
for name, batch, sequence, dim in [
|
| 195 |
("groot_dit", 1, 51, 1536),
|
|
@@ -212,6 +259,19 @@ def main() -> None:
|
|
| 212 |
)
|
| 213 |
lines.append(line)
|
| 214 |
print(line)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 215 |
lines.extend(
|
| 216 |
[
|
| 217 |
"",
|
|
|
|
| 14 |
ROOT = Path(__file__).resolve().parents[2]
|
| 15 |
sys.path.insert(0, str(ROOT / "adaptive-layernorm-producers" / "tests"))
|
| 16 |
from test_adaptive_layernorm_producers import ( # noqa: E402
|
| 17 |
+
_nvfp4_out,
|
| 18 |
load_source_ops,
|
| 19 |
make_case,
|
| 20 |
quant_fp8,
|
|
|
|
| 170 |
}
|
| 171 |
|
| 172 |
|
| 173 |
+
def run_ptok_table(ops, name: str, rows: int, dim: int, eps: float, iters: int):
|
| 174 |
+
chunks = 6
|
| 175 |
+
x = torch.randn((rows, dim), device="cuda", dtype=torch.bfloat16)
|
| 176 |
+
temb = torch.randn(
|
| 177 |
+
(rows, chunks, dim), device="cuda", dtype=torch.bfloat16
|
| 178 |
+
)
|
| 179 |
+
table = torch.randn((chunks, dim), device="cuda", dtype=torch.float32)
|
| 180 |
+
out = torch.empty_like(x)
|
| 181 |
+
packed, sf = _nvfp4_out(x)
|
| 182 |
+
|
| 183 |
+
def fused_bf16():
|
| 184 |
+
return ops.ada_layer_norm_ptok_table_bf16(
|
| 185 |
+
x, temb, table, 0, 1, eps, out=out
|
| 186 |
+
)
|
| 187 |
+
|
| 188 |
+
def fused_nvfp4():
|
| 189 |
+
return ops.ada_layer_norm_quant_nvfp4_swizzled_ptok_table_bf16(
|
| 190 |
+
x, temb, table, 0, 1, eps, packed=packed, sf_swizzled=sf
|
| 191 |
+
)
|
| 192 |
+
|
| 193 |
+
def eager_bf16():
|
| 194 |
+
xf = x.float()
|
| 195 |
+
mean = xf.mean(dim=-1, keepdim=True)
|
| 196 |
+
var = (xf - mean).square().mean(dim=-1, keepdim=True)
|
| 197 |
+
norm = (xf - mean) * torch.rsqrt(var + eps)
|
| 198 |
+
shift = temb[:, 0].float() + table[0]
|
| 199 |
+
scale = temb[:, 1].float() + table[1]
|
| 200 |
+
return (norm * (1.0 + scale) + shift).to(torch.bfloat16)
|
| 201 |
+
|
| 202 |
+
return {
|
| 203 |
+
"shape": name,
|
| 204 |
+
"rows": rows,
|
| 205 |
+
"dim": dim,
|
| 206 |
+
"bf16_us": time_cuda(fused_bf16, iters=iters),
|
| 207 |
+
"eager_us": time_cuda(eager_bf16, iters=iters),
|
| 208 |
+
"nvfp4_us": time_cuda(fused_nvfp4, iters=iters),
|
| 209 |
+
}
|
| 210 |
+
|
| 211 |
+
|
| 212 |
def main() -> None:
|
| 213 |
parser = argparse.ArgumentParser()
|
| 214 |
parser.add_argument("--backend", choices=["source", "installed"], default="source")
|
|
|
|
| 230 |
("wan_video_4k", 4096, 3072),
|
| 231 |
]
|
| 232 |
rows = [run_case(ops, name, r, d, 1e-5, args.iters) for name, r, d in shapes]
|
| 233 |
+
table_rows = [
|
| 234 |
+
run_ptok_table(ops, name, r, d, 1e-5, args.iters)
|
| 235 |
+
for name, r, d in [
|
| 236 |
+
("video_short", 64, 3072),
|
| 237 |
+
("video_long", 2520, 3072),
|
| 238 |
+
]
|
| 239 |
+
]
|
| 240 |
modulation_rows = []
|
| 241 |
for name, batch, sequence, dim in [
|
| 242 |
("groot_dit", 1, 51, 1536),
|
|
|
|
| 259 |
)
|
| 260 |
lines.append(line)
|
| 261 |
print(line)
|
| 262 |
+
lines.extend([
|
| 263 |
+
"",
|
| 264 |
+
"| Per-token table shape | BF16 fused us | BF16 eager us | Speedup | NVFP4 fused us |",
|
| 265 |
+
"|---|---:|---:|---:|---:|",
|
| 266 |
+
])
|
| 267 |
+
for row in table_rows:
|
| 268 |
+
line = (
|
| 269 |
+
f"| {row['shape']} M{row['rows']} D{row['dim']} | "
|
| 270 |
+
f"{row['bf16_us']:.3f} | {row['eager_us']:.3f} | "
|
| 271 |
+
f"{row['eager_us'] / row['bf16_us']:.2f}x | {row['nvfp4_us']:.3f} |"
|
| 272 |
+
)
|
| 273 |
+
lines.append(line)
|
| 274 |
+
print(line)
|
| 275 |
lines.extend(
|
| 276 |
[
|
| 277 |
"",
|
build/torch211-cxx11-cu128-x86_64-linux/__init__.py
CHANGED
|
@@ -105,6 +105,26 @@ def _ada_layer_norm_quant_fp8_ptok_table_bf16_fake(
|
|
| 105 |
return None
|
| 106 |
|
| 107 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 108 |
@torch.library.register_fake(add_op_namespace_prefix("ada_layer_norm_quant_fp8_modfp8_bf16"))
|
| 109 |
def _ada_layer_norm_quant_fp8_modfp8_bf16_fake(
|
| 110 |
x: torch.Tensor,
|
|
@@ -162,6 +182,32 @@ def _ada_layer_norm_quant_nvfp4_swizzled_bf16_fake(
|
|
| 162 |
return None
|
| 163 |
|
| 164 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 165 |
@torch.library.register_fake(add_op_namespace_prefix("ada_layer_norm_quant_nvfp4_swizzled_modfp8_bf16"))
|
| 166 |
def _ada_layer_norm_quant_nvfp4_swizzled_modfp8_bf16_fake(
|
| 167 |
x: torch.Tensor,
|
|
@@ -198,6 +244,22 @@ def _layer_norm_no_affine_quant_fp8_static_bf16_fake(
|
|
| 198 |
return None
|
| 199 |
|
| 200 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 201 |
def ada_layer_norm_quant_fp8_bf16(
|
| 202 |
x: torch.Tensor,
|
| 203 |
scale: torch.Tensor,
|
|
@@ -266,6 +328,25 @@ def ada_layer_norm_quant_fp8_ptok_table_bf16(
|
|
| 266 |
return out
|
| 267 |
|
| 268 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 269 |
def ada_layer_norm_quant_fp8_modfp8_bf16(
|
| 270 |
x: torch.Tensor,
|
| 271 |
scale_fp8: torch.Tensor,
|
|
@@ -319,6 +400,27 @@ def ada_layer_norm_quant_nvfp4_swizzled_bf16(
|
|
| 319 |
return packed, sf_swizzled
|
| 320 |
|
| 321 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 322 |
def ada_layer_norm_quant_nvfp4_swizzled_modfp8_bf16(
|
| 323 |
x: torch.Tensor,
|
| 324 |
scale_fp8: torch.Tensor,
|
|
@@ -353,6 +455,21 @@ def layer_norm_no_affine_quant_fp8_static_bf16(
|
|
| 353 |
return out
|
| 354 |
|
| 355 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 356 |
@torch.library.register_fake(add_op_namespace_prefix("adaln_modulation6_bf16"))
|
| 357 |
def _adaln_modulation6_bf16_fake(
|
| 358 |
adaln_params: torch.Tensor,
|
|
@@ -413,11 +530,14 @@ __all__ = [
|
|
| 413 |
"ada_layer_norm_quant_fp8_bf16",
|
| 414 |
"ada_layer_norm_quant_fp8_ptok_bf16",
|
| 415 |
"ada_layer_norm_quant_fp8_ptok_table_bf16",
|
|
|
|
| 416 |
"ada_layer_norm_quant_fp8_modfp8_bf16",
|
| 417 |
"awq_ada_layer_norm_quant_fp8_bf16",
|
| 418 |
"ada_layer_norm_quant_nvfp4_swizzled_bf16",
|
|
|
|
| 419 |
"ada_layer_norm_quant_nvfp4_swizzled_modfp8_bf16",
|
| 420 |
"layer_norm_no_affine_quant_fp8_static_bf16",
|
|
|
|
| 421 |
"adaln_modulation6_bf16",
|
| 422 |
"swizzled_sf_size",
|
| 423 |
]
|
|
|
|
| 105 |
return None
|
| 106 |
|
| 107 |
|
| 108 |
+
@torch.library.register_fake(add_op_namespace_prefix("ada_layer_norm_ptok_table_bf16"))
|
| 109 |
+
def _ada_layer_norm_ptok_table_bf16_fake(
|
| 110 |
+
x: torch.Tensor,
|
| 111 |
+
temb: torch.Tensor,
|
| 112 |
+
table: torch.Tensor,
|
| 113 |
+
shift_idx: int,
|
| 114 |
+
scale_idx: int,
|
| 115 |
+
eps: float,
|
| 116 |
+
out: torch.Tensor,
|
| 117 |
+
) -> None:
|
| 118 |
+
_check_x(x)
|
| 119 |
+
if temb.dim() != 3 or temb.shape[0] != x.shape[0] or temb.shape[2] != x.shape[1]:
|
| 120 |
+
raise RuntimeError("temb must have shape (seq_len, n_chunks, dim)")
|
| 121 |
+
if table.shape != (temb.shape[1], x.shape[1]):
|
| 122 |
+
raise RuntimeError("table must have shape (n_chunks, dim)")
|
| 123 |
+
if out.shape != x.shape:
|
| 124 |
+
raise RuntimeError("out must have the same shape as x")
|
| 125 |
+
return None
|
| 126 |
+
|
| 127 |
+
|
| 128 |
@torch.library.register_fake(add_op_namespace_prefix("ada_layer_norm_quant_fp8_modfp8_bf16"))
|
| 129 |
def _ada_layer_norm_quant_fp8_modfp8_bf16_fake(
|
| 130 |
x: torch.Tensor,
|
|
|
|
| 182 |
return None
|
| 183 |
|
| 184 |
|
| 185 |
+
@torch.library.register_fake(
|
| 186 |
+
add_op_namespace_prefix("ada_layer_norm_quant_nvfp4_swizzled_ptok_table_bf16")
|
| 187 |
+
)
|
| 188 |
+
def _ada_layer_norm_quant_nvfp4_swizzled_ptok_table_bf16_fake(
|
| 189 |
+
x: torch.Tensor,
|
| 190 |
+
temb: torch.Tensor,
|
| 191 |
+
table: torch.Tensor,
|
| 192 |
+
shift_idx: int,
|
| 193 |
+
scale_idx: int,
|
| 194 |
+
eps: float,
|
| 195 |
+
packed: torch.Tensor,
|
| 196 |
+
sf_swizzled: torch.Tensor,
|
| 197 |
+
) -> None:
|
| 198 |
+
_check_x(x)
|
| 199 |
+
if temb.dim() != 3 or temb.shape[0] != x.shape[0] or temb.shape[2] != x.shape[1]:
|
| 200 |
+
raise RuntimeError("temb must have shape (seq_len, n_chunks, dim)")
|
| 201 |
+
if table.shape != (temb.shape[1], x.shape[1]):
|
| 202 |
+
raise RuntimeError("table must have shape (n_chunks, dim)")
|
| 203 |
+
rows, dim = x.shape
|
| 204 |
+
if packed.shape != (rows, dim // 2):
|
| 205 |
+
raise RuntimeError("packed must have shape (rows, dim // 2)")
|
| 206 |
+
if sf_swizzled.numel() < swizzled_sf_size(rows, dim):
|
| 207 |
+
raise RuntimeError("sf_swizzled is too small")
|
| 208 |
+
return None
|
| 209 |
+
|
| 210 |
+
|
| 211 |
@torch.library.register_fake(add_op_namespace_prefix("ada_layer_norm_quant_nvfp4_swizzled_modfp8_bf16"))
|
| 212 |
def _ada_layer_norm_quant_nvfp4_swizzled_modfp8_bf16_fake(
|
| 213 |
x: torch.Tensor,
|
|
|
|
| 244 |
return None
|
| 245 |
|
| 246 |
|
| 247 |
+
@torch.library.register_fake(add_op_namespace_prefix("layer_norm_no_affine_quant_nvfp4_swizzled_bf16"))
|
| 248 |
+
def _layer_norm_no_affine_quant_nvfp4_swizzled_bf16_fake(
|
| 249 |
+
x: torch.Tensor,
|
| 250 |
+
eps: float,
|
| 251 |
+
packed: torch.Tensor,
|
| 252 |
+
sf_swizzled: torch.Tensor,
|
| 253 |
+
) -> None:
|
| 254 |
+
_check_x(x)
|
| 255 |
+
rows, dim = x.shape
|
| 256 |
+
if packed.shape != (rows, dim // 2):
|
| 257 |
+
raise RuntimeError("packed must have shape (rows, dim // 2)")
|
| 258 |
+
if sf_swizzled.numel() < swizzled_sf_size(rows, dim):
|
| 259 |
+
raise RuntimeError("sf_swizzled is too small")
|
| 260 |
+
return None
|
| 261 |
+
|
| 262 |
+
|
| 263 |
def ada_layer_norm_quant_fp8_bf16(
|
| 264 |
x: torch.Tensor,
|
| 265 |
scale: torch.Tensor,
|
|
|
|
| 328 |
return out
|
| 329 |
|
| 330 |
|
| 331 |
+
def ada_layer_norm_ptok_table_bf16(
|
| 332 |
+
x: torch.Tensor,
|
| 333 |
+
temb: torch.Tensor,
|
| 334 |
+
table: torch.Tensor,
|
| 335 |
+
shift_idx: int,
|
| 336 |
+
scale_idx: int,
|
| 337 |
+
eps: float = 1e-5,
|
| 338 |
+
out: torch.Tensor | None = None,
|
| 339 |
+
) -> torch.Tensor:
|
| 340 |
+
"""Fuse table add, chunk selection and AdaLN into a BF16 output."""
|
| 341 |
+
|
| 342 |
+
if out is None:
|
| 343 |
+
out = torch.empty_like(x)
|
| 344 |
+
ops.ada_layer_norm_ptok_table_bf16(
|
| 345 |
+
x, temb, table, int(shift_idx), int(scale_idx), float(eps), out
|
| 346 |
+
)
|
| 347 |
+
return out
|
| 348 |
+
|
| 349 |
+
|
| 350 |
def ada_layer_norm_quant_fp8_modfp8_bf16(
|
| 351 |
x: torch.Tensor,
|
| 352 |
scale_fp8: torch.Tensor,
|
|
|
|
| 400 |
return packed, sf_swizzled
|
| 401 |
|
| 402 |
|
| 403 |
+
def ada_layer_norm_quant_nvfp4_swizzled_ptok_table_bf16(
|
| 404 |
+
x: torch.Tensor,
|
| 405 |
+
temb: torch.Tensor,
|
| 406 |
+
table: torch.Tensor,
|
| 407 |
+
shift_idx: int,
|
| 408 |
+
scale_idx: int,
|
| 409 |
+
eps: float = 1e-5,
|
| 410 |
+
packed: torch.Tensor | None = None,
|
| 411 |
+
sf_swizzled: torch.Tensor | None = None,
|
| 412 |
+
) -> tuple[torch.Tensor, torch.Tensor]:
|
| 413 |
+
"""Fuse per-token table AdaLN into NVFP4 and CUTLASS scale factors."""
|
| 414 |
+
|
| 415 |
+
if packed is None or sf_swizzled is None:
|
| 416 |
+
packed, sf_swizzled = _alloc_nvfp4(x)
|
| 417 |
+
ops.ada_layer_norm_quant_nvfp4_swizzled_ptok_table_bf16(
|
| 418 |
+
x, temb, table, int(shift_idx), int(scale_idx), float(eps),
|
| 419 |
+
packed, sf_swizzled
|
| 420 |
+
)
|
| 421 |
+
return packed, sf_swizzled
|
| 422 |
+
|
| 423 |
+
|
| 424 |
def ada_layer_norm_quant_nvfp4_swizzled_modfp8_bf16(
|
| 425 |
x: torch.Tensor,
|
| 426 |
scale_fp8: torch.Tensor,
|
|
|
|
| 455 |
return out
|
| 456 |
|
| 457 |
|
| 458 |
+
def layer_norm_no_affine_quant_nvfp4_swizzled_bf16(
|
| 459 |
+
x: torch.Tensor,
|
| 460 |
+
eps: float = 1e-5,
|
| 461 |
+
packed: torch.Tensor | None = None,
|
| 462 |
+
sf_swizzled: torch.Tensor | None = None,
|
| 463 |
+
) -> tuple[torch.Tensor, torch.Tensor]:
|
| 464 |
+
"""SM110 no-affine LayerNorm producing NVFP4 and CUTLASS SFA."""
|
| 465 |
+
if packed is None or sf_swizzled is None:
|
| 466 |
+
packed, sf_swizzled = _alloc_nvfp4(x)
|
| 467 |
+
ops.layer_norm_no_affine_quant_nvfp4_swizzled_bf16(
|
| 468 |
+
x, float(eps), packed, sf_swizzled
|
| 469 |
+
)
|
| 470 |
+
return packed, sf_swizzled
|
| 471 |
+
|
| 472 |
+
|
| 473 |
@torch.library.register_fake(add_op_namespace_prefix("adaln_modulation6_bf16"))
|
| 474 |
def _adaln_modulation6_bf16_fake(
|
| 475 |
adaln_params: torch.Tensor,
|
|
|
|
| 530 |
"ada_layer_norm_quant_fp8_bf16",
|
| 531 |
"ada_layer_norm_quant_fp8_ptok_bf16",
|
| 532 |
"ada_layer_norm_quant_fp8_ptok_table_bf16",
|
| 533 |
+
"ada_layer_norm_ptok_table_bf16",
|
| 534 |
"ada_layer_norm_quant_fp8_modfp8_bf16",
|
| 535 |
"awq_ada_layer_norm_quant_fp8_bf16",
|
| 536 |
"ada_layer_norm_quant_nvfp4_swizzled_bf16",
|
| 537 |
+
"ada_layer_norm_quant_nvfp4_swizzled_ptok_table_bf16",
|
| 538 |
"ada_layer_norm_quant_nvfp4_swizzled_modfp8_bf16",
|
| 539 |
"layer_norm_no_affine_quant_fp8_static_bf16",
|
| 540 |
+
"layer_norm_no_affine_quant_nvfp4_swizzled_bf16",
|
| 541 |
"adaln_modulation6_bf16",
|
| 542 |
"swizzled_sf_size",
|
| 543 |
]
|
build/torch211-cxx11-cu128-x86_64-linux/{_adaptive_layernorm_producers_cuda_9b532b4.abi3.so → _adaptive_layernorm_producers_cuda_bb9b4e5.abi3.so}
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:646d65e6da951010d153d47c3dbe8e78881b0f670f115cd6e1f63bbea239cbfd
|
| 3 |
+
size 2000416
|
build/torch211-cxx11-cu128-x86_64-linux/_ops.py
CHANGED
|
@@ -1,9 +1,9 @@
|
|
| 1 |
import torch
|
| 2 |
-
from . import
|
| 3 |
-
ops = torch.ops.
|
| 4 |
|
| 5 |
def add_op_namespace_prefix(op_name: str):
|
| 6 |
"""
|
| 7 |
Prefix op by namespace.
|
| 8 |
"""
|
| 9 |
-
return f"
|
|
|
|
| 1 |
import torch
|
| 2 |
+
from . import _adaptive_layernorm_producers_cuda_bb9b4e5
|
| 3 |
+
ops = torch.ops._adaptive_layernorm_producers_cuda_bb9b4e5
|
| 4 |
|
| 5 |
def add_op_namespace_prefix(op_name: str):
|
| 6 |
"""
|
| 7 |
Prefix op by namespace.
|
| 8 |
"""
|
| 9 |
+
return f"_adaptive_layernorm_producers_cuda_bb9b4e5::{op_name}"
|
build/torch211-cxx11-cu128-x86_64-linux/metadata.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
{
|
| 2 |
"name": "adaptive-layernorm-producers",
|
| 3 |
-
"id": "
|
| 4 |
"version": 1,
|
| 5 |
"license": "Apache-2.0",
|
| 6 |
"python-depends": [],
|
|
@@ -19,5 +19,25 @@
|
|
| 19 |
"8.9",
|
| 20 |
"9.0"
|
| 21 |
]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 22 |
}
|
| 23 |
-
}
|
|
|
|
| 1 |
{
|
| 2 |
"name": "adaptive-layernorm-producers",
|
| 3 |
+
"id": "_adaptive_layernorm_producers_cuda_bb9b4e5",
|
| 4 |
"version": 1,
|
| 5 |
"license": "Apache-2.0",
|
| 6 |
"python-depends": [],
|
|
|
|
| 19 |
"8.9",
|
| 20 |
"9.0"
|
| 21 |
]
|
| 22 |
+
},
|
| 23 |
+
"digest": {
|
| 24 |
+
"algorithm": "sha256",
|
| 25 |
+
"files": {
|
| 26 |
+
"__init__.py": "+ZwPYHOKPDe9jTo0jlUuIRwuU4XT1gxfK/Bxk+ZLkNI=",
|
| 27 |
+
"_adaptive_layernorm_producers_cuda_bb9b4e5.abi3.so": "ZG1l5tqVEBDRU9R8Pb6OeIgbD2cPEVzW4fY7vqI5y/0=",
|
| 28 |
+
"_ops.py": "M1g8M15LM5VNF7VWdjaC0SGIFEv/Lu08K3trrsZPuSE=",
|
| 29 |
+
"adaptive_layernorm_producers/__init__.py": "DFYPlrhXwYjEqCl/8n0SmWGZV8NFml5DPhMjKfv98GY="
|
| 30 |
+
}
|
| 31 |
+
},
|
| 32 |
+
"provenance": {
|
| 33 |
+
"kernel-builder": {
|
| 34 |
+
"version": "0.17.0-dev0",
|
| 35 |
+
"sha": "b39ca23f1b36383df00b27b3ffe1276cd5dbea85",
|
| 36 |
+
"dirty": false
|
| 37 |
+
},
|
| 38 |
+
"kernel": {
|
| 39 |
+
"sha": "bb9b4e57f62b1e00f4c87262303558a6d8d5c61a",
|
| 40 |
+
"dirty": false
|
| 41 |
+
}
|
| 42 |
}
|
| 43 |
+
}
|