wehub-resource-sync
59a0a3844c
PR Test AMD / cancel-on-close (push) Has been skipped
PR Test NVIDIA ARM / scan (push) Has been skipped
PR Test NVIDIA / cancel-on-close (push) Has been skipped
PR Test AMD / scan (push) Has been skipped
PR Test NVIDIA ARM / cancel-on-close (push) Has been skipped
PR Test NVIDIA / scan (push) Has been skipped
Release Docker Images / build (cu129-torch-2.11.0) (push) Has been skipped
Release Docker Images / build (cu130-torch-2.11.0) (push) Has been skipped
Release PyPI / publish (push) Has been skipped
Scheduler Python Test / test (push) Successful in 27m19s
Docs / build (push) Successful in 28m8s
Scheduler C++ Test / test (push) Successful in 28m19s
Scheduler C++ Test / test-flat (push) Successful in 28m18s
Docs / deploy (push) Has been cancelled
PR Test AMD / finish (push) Has been cancelled
PR Test NVIDIA / finish (push) Has been cancelled
PR Test NVIDIA ARM / finish (push) Has been cancelled
PR Test NVIDIA ARM / ${{ matrix.name }} (${{ matrix.runner }}) (push) Has been cancelled
PR Test AMD / ${{ matrix.name }} (${{ matrix.runner }}) (push) Has been cancelled
PR Test NVIDIA / ${{ matrix.name }} (${{ matrix.runner }}) (push) Has been cancelled
72 lines
2.6 KiB
Python
72 lines
2.6 KiB
Python
# SPDX-License-Identifier: MIT
|
|
# SPDX-FileCopyrightText: Copyright (c) 2026 LightSeek Foundation
|
|
# SPDX-FileCopyrightText: Copyright (c) 2023-2026, Songlin Yang, Yu Zhang, Zhiyuan Li
|
|
#
|
|
# Copyright (c) 2026 LightSeek Foundation
|
|
#
|
|
# Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
# of this software and associated documentation files (the "Software"), to deal
|
|
# in the Software without restriction, including without limitation the rights
|
|
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
# copies of the Software, and to permit persons to whom the Software is
|
|
# furnished to do so, subject to the following conditions:
|
|
#
|
|
# The above copyright notice and this permission notice shall be included in
|
|
# all copies or substantial portions of the Software.
|
|
#
|
|
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
# SOFTWARE.
|
|
|
|
import torch
|
|
import triton
|
|
import triton.language as tl
|
|
|
|
|
|
@triton.jit
|
|
def fused_gdn_gating_kernel(
|
|
g,
|
|
A_log,
|
|
a,
|
|
dt_bias,
|
|
seq_len,
|
|
NUM_HEADS: tl.constexpr,
|
|
beta: tl.constexpr,
|
|
threshold: tl.constexpr,
|
|
BLK_HEADS: tl.constexpr,
|
|
):
|
|
i_b, i_s, i_d = tl.program_id(0), tl.program_id(1), tl.program_id(2)
|
|
head_off = i_d * BLK_HEADS + tl.arange(0, BLK_HEADS)
|
|
off = i_b * seq_len * NUM_HEADS + i_s * NUM_HEADS + head_off
|
|
mask = head_off < NUM_HEADS
|
|
blk_A_log = tl.load(A_log + head_off, mask=mask)
|
|
blk_a = tl.load(a + off, mask=mask)
|
|
blk_bias = tl.load(dt_bias + head_off, mask=mask)
|
|
x = blk_a.to(tl.float32) + blk_bias.to(tl.float32)
|
|
softplus_x = tl.where(
|
|
beta * x <= threshold, (1 / beta) * tl.log(1 + tl.exp(beta * x)), x
|
|
)
|
|
blk_g = -tl.exp(blk_A_log.to(tl.float32)) * softplus_x
|
|
tl.store(g + off, blk_g.to(g.dtype.element_ty), mask=mask)
|
|
|
|
|
|
def fused_gdn_gating(
|
|
A_log: torch.Tensor,
|
|
a: torch.Tensor,
|
|
dt_bias: torch.Tensor,
|
|
beta: float = 1.0,
|
|
threshold: float = 20.0,
|
|
) -> torch.Tensor:
|
|
batch, num_heads = a.shape
|
|
seq_len = 1
|
|
grid = (batch, seq_len, triton.cdiv(num_heads, 8))
|
|
g = torch.empty_like(a, dtype=torch.float32)
|
|
fused_gdn_gating_kernel[grid](
|
|
g, A_log, a, dt_bias, seq_len, num_heads, beta, threshold, 8, num_warps=1
|
|
)
|
|
return g
|