A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 09:30 UTC
[Kernel] add kernel for FATReLU (#9610)
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Jee Jee Li committed
Oct 24, 2024 at 16:18 UTC
295a061fb34ec6fb251abf1dbece5b1bb7dc9006
6 files changed
+78
-8
csrc/activation_kernels.cu
+42
@@ -89,6 +89,48 @@ void gelu_tanh_and_mul(torch::Tensor& out, // [..., d]
89
90
namespace vllm {
91
92
+template <typename T>
93
+__device__ __forceinline__ T fatrelu_kernel(const T& x, const float threshold) {
94
+ const float f = (float)x;
95
+ return (T)(f > threshold ? f : 0.0f);
96
+}
97
+
98
+template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&, const float)>
99
+__global__ void act_and_mul_kernel_with_param(
100
+ scalar_t* __restrict__ out, const scalar_t* __restrict__ input, const int d,
101
+ const float param) {
102
+ const int64_t token_idx = blockIdx.x;
103
+ for (int64_t idx = threadIdx.x; idx < d; idx += blockDim.x) {
104
+ const scalar_t x = VLLM_LDG(&input[token_idx * 2 * d + idx]);
105
+ const scalar_t y = VLLM_LDG(&input[token_idx * 2 * d + d + idx]);
106
+ out[token_idx * d + idx] = ACT_FN(x, param) * y;
107
+ }
108
+}
109
+
110
+} // namespace vllm
111
+
112
+#define LAUNCH_ACTIVATION_GATE_KERNEL_WITH_PARAM(KERNEL, PARAM) \
113
+ int d = input.size(-1) / 2; \
114
+ int64_t num_tokens = input.numel() / input.size(-1); \
115
+ dim3 grid(num_tokens); \
116
+ dim3 block(std::min(d, 1024)); \
117
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(input)); \
118
+ const cudaStream_t stream = at::cuda::getCurrentCUDAStream(); \
119
+ VLLM_DISPATCH_FLOATING_TYPES( \
120
+ input.scalar_type(), "act_and_mul_kernel_with_param", [&] { \
121
+ vllm::act_and_mul_kernel_with_param<scalar_t, KERNEL<scalar_t>> \
122
+ <<<grid, block, 0, stream>>>(out.data_ptr<scalar_t>(), \
123
+ input.data_ptr<scalar_t>(), d, \
124
+ PARAM); \
125
+ });
126
+
127
+void fatrelu_and_mul(torch::Tensor& out, // [..., d],
128
+ torch::Tensor& input, // [..., 2 * d]
129
+ double threshold) {
130
+ LAUNCH_ACTIVATION_GATE_KERNEL_WITH_PARAM(vllm::fatrelu_kernel, threshold);
131
+}
132
+namespace vllm {
133
+
134
// Element-wise activation kernel template.
135
template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&)>
136
__global__ void activation_kernel(
csrc/ops.h
+3
@@ -48,6 +48,9 @@ void gelu_and_mul(torch::Tensor& out, torch::Tensor& input);
48
49
void gelu_tanh_and_mul(torch::Tensor& out, torch::Tensor& input);
50
51
+void fatrelu_and_mul(torch::Tensor& out, torch::Tensor& input,
52
+ double threshold);
53
+
54
void gelu_new(torch::Tensor& out, torch::Tensor& input);
55
56
void gelu_fast(torch::Tensor& out, torch::Tensor& input);
csrc/torch_bindings.cpp
+4
@@ -60,6 +60,10 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
60
ops.def("gelu_tanh_and_mul(Tensor! out, Tensor input) -> ()");
61
ops.impl("gelu_tanh_and_mul", torch::kCUDA, &gelu_tanh_and_mul);
62
63
+ // FATReLU implementation.
64
+ ops.def("fatrelu_and_mul(Tensor! out, Tensor input, float threshold) -> ()");
65
+ ops.impl("fatrelu_and_mul", torch::kCUDA, &fatrelu_and_mul);
66
+
67
// GELU implementation used in GPT-2.
68
ops.def("gelu_new(Tensor! out, Tensor input) -> ()");
69
ops.impl("gelu_new", torch::kCUDA, &gelu_new);
tests/kernels/test_activation.py
+16
-7
@@ -1,12 +1,13 @@
1
+import random
2
from typing import Type
3
4
import pytest
5
import torch
6
7
from tests.kernels.utils import opcheck
7
-from vllm.model_executor.layers.activation import (FastGELU, GeluAndMul,
8
- NewGELU, QuickGELU,
9
- SiluAndMul)
8
+from vllm.model_executor.layers.activation import (FastGELU, FatreluAndMul,
9
+ GeluAndMul, NewGELU,
10
+ QuickGELU, SiluAndMul)
11
from vllm.utils import seed_everything
12
13
from .allclose_default import get_default_atol, get_default_rtol
@@ -20,7 +21,8 @@ CUDA_DEVICES = [
21
]
22
23
23
-@pytest.mark.parametrize("activation", ["silu", "gelu", "gelu_tanh"])
24
+@pytest.mark.parametrize("activation",
25
+ ["silu", "gelu", "gelu_tanh", "fatrelu"])
26
@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
27
@pytest.mark.parametrize("d", D)
28
@pytest.mark.parametrize("dtype", DTYPES)
@@ -47,16 +49,23 @@ def test_act_and_mul(
49
elif activation == "gelu_tanh":
50
layer = GeluAndMul(approximate="tanh")
51
fn = torch.ops._C.gelu_tanh_and_mul
52
+ elif activation == "fatrelu":
53
+ threshold = random.uniform(0, 1)
54
+ layer = FatreluAndMul(threshold)
55
+ fn = torch.ops._C.fatrelu_and_mul
56
out = layer(x)
57
ref_out = layer.forward_native(x)
52
- # The SiLU and GELU implementations are equivalent to the native PyTorch
53
- # implementations, so we can do exact comparison.
58
+ # The SiLU, GELU and FatReLU implementations are equivalent to the native
59
+ # PyTorch implementations, so we can do exact comparison.
60
torch.testing.assert_close(out, ref_out, atol=0.0, rtol=0.0)
61
62
d = x.shape[-1] // 2
63
output_shape = (x.shape[:-1] + (d, ))
64
out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
59
- opcheck(fn, (out, x))
65
+ if activation == "fatrelu":
66
+ opcheck(fn, (out, x, threshold))
67
+ else:
68
+ opcheck(fn, (out, x))
69
70
71
@pytest.mark.parametrize("activation", [(FastGELU, torch.ops._C.gelu_fast),
vllm/_custom_ops.py
+6
@@ -79,6 +79,12 @@ def gelu_tanh_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
79
torch.ops._C.gelu_tanh_and_mul(out, x)
80
81
82
+def fatrelu_and_mul(out: torch.Tensor,
83
+ x: torch.Tensor,
84
+ threshold: float = 0.0) -> None:
85
+ torch.ops._C.fatrelu_and_mul(out, x, threshold)
86
+
87
+
88
def gelu_fast(out: torch.Tensor, x: torch.Tensor) -> None:
89
torch.ops._C.gelu_fast(out, x)
90
vllm/model_executor/layers/activation.py
+7
-1
@@ -39,7 +39,13 @@ class FatreluAndMul(CustomOp):
39
return x1 * x2
40
41
def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
42
- return self.forward_native(x)
42
+ from vllm import _custom_ops as ops
43
+
44
+ d = x.shape[-1] // 2
45
+ output_shape = (x.shape[:-1] + (d, ))
46
+ out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
47
+ ops.fatrelu_and_mul(out, x, self.threshold)
48
+ return out
49
50
51
@CustomOp.register("silu_and_mul")