| from __future__ import annotations |
|
|
| import torch |
| import torch.nn as nn |
|
|
| import triton |
| import triton.language as tl |
|
|
|
|
| @triton.jit |
| def _rope_interleaved( |
| out, |
| x, |
| cos, |
| sin, |
| ): |
| head_start = tl.program_id(0) * 8 |
| pos = tl.program_id(1) |
|
|
| offsets = tl.arange(0, 8 * 128) |
| x_offsets = pos * 24 * 128 + head_start * 128 + offsets |
| values = tl.load(x + x_offsets, cache_modifier=".cg") |
| pairs = tl.reshape(values, (8 * 64, 2)) |
| even, odd = tl.split(pairs) |
| even = tl.reshape(even, (8, 64)).to(tl.float32) |
| odd = tl.reshape(odd, (8, 64)).to(tl.float32) |
|
|
| freq = tl.arange(0, 64) |
| c = tl.load(cos + pos * 64 + freq).to(tl.float32) |
| s = tl.load(sin + pos * 64 + freq).to(tl.float32) |
|
|
| out_even = even * c - odd * s |
| out_odd = odd * c + even * s |
| result = tl.reshape(tl.join(out_even, out_odd), (8 * 128,)) |
| tl.store(out + x_offsets, result, cache_modifier=".cs") |
|
|
|
|
| class DiffusionRoPE(nn.Module): |
| def __init__(self, is_neox_style: bool = False) -> None: |
| super().__init__() |
| self.interleaved = not is_neox_style |
|
|
| def forward( |
| self, |
| x: torch.Tensor, |
| cos: torch.Tensor, |
| sin: torch.Tensor, |
| ) -> torch.Tensor: |
| if cos.dim() == 3: |
| cos = cos[0] |
| sin = sin[0] |
|
|
| |
| assert self.interleaved |
| assert x.shape[0] == 1 and x.shape[2:] == (24, 128) |
| assert cos.shape[-1] == 64 |
| output = torch.empty_like(x) |
| grid = (3, x.shape[1]) |
| _rope_interleaved[grid]( |
| output, |
| x, |
| cos, |
| sin, |
| num_warps=2, |
| ) |
| return output |
|
|
