1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
// text-embeddings-inference (adapted)
#![allow(unused_variables, unused_imports, dead_code)]
use hanzo_ml::{Device, DeviceLocation, Result, Tensor};
use hanzo_nn::Activation as HanzoActivation;
use std::sync::atomic::{AtomicBool, Ordering};
use std::sync::{LazyLock, Mutex, Once};
/// Controller for the CUBLASLT handle and inhibition flag.
pub struct CublasLtController {
handle: Mutex<Option<&'static CublasLtWrapper>>,
inhibit: AtomicBool,
/// The device location where cuBLASLt was initialized. This is important
/// because cuBLASLt handles are device-specific and using them with tensors
/// on different devices will cause CUBLAS errors.
device_location: Mutex<Option<DeviceLocation>>,
}
impl CublasLtController {
/// Set whether to inhibit CUBLASLT usage.
pub fn set_inhibit(&self, value: bool) {
self.inhibit.store(value, Ordering::SeqCst);
}
/// Get the handle if not inhibited.
///
/// Note: We check inhibit BEFORE reading handle_opt to avoid TOCTOU race.
/// If inhibit is checked after reading handle_opt, another thread could
/// call set_inhibit(true) between reading handle_opt and checking inhibit,
/// causing us to return a handle that should have been inhibited.
pub fn get(&self) -> Option<&'static CublasLtWrapper> {
// Check inhibit first to prevent TOCTOU race condition
if self.inhibit.load(Ordering::SeqCst) {
return None;
}
let handle_opt = self.handle.lock().unwrap();
*handle_opt
}
/// Get the handle only if the given device matches the device where cuBLASLt
/// was initialized. This is important for multi-GPU setups where using a
/// cuBLASLt handle with tensors on a different device causes CUBLAS errors.
pub fn get_for_device(&self, device: &Device) -> Option<&'static CublasLtWrapper> {
// Check inhibit first to prevent TOCTOU race condition
if self.inhibit.load(Ordering::SeqCst) {
return None;
}
// Check if the device matches the initialized device
let device_loc = self.device_location.lock().unwrap();
if let Some(init_loc) = *device_loc {
if device.location() != init_loc {
return None;
}
}
let handle_opt = self.handle.lock().unwrap();
*handle_opt
}
}
pub static CUBLASLT_CONTROLLER: LazyLock<CublasLtController> =
LazyLock::new(|| CublasLtController {
handle: Mutex::new(None),
inhibit: AtomicBool::new(false),
device_location: Mutex::new(None),
});
#[cfg(feature = "cuda")]
mod api;
#[cfg(feature = "cuda")]
mod matmul;
#[cfg(test)]
#[cfg(feature = "cuda")]
mod tests;
#[cfg(feature = "cuda")]
pub use api::{fused_batch_matmul, fused_batch_matmul_f8, CublasLt};
pub fn maybe_init_cublas_lt_wrapper(device: Device) {
static INIT: Once = Once::new();
INIT.call_once(|| {
#[cfg(feature = "cuda")]
{
match device {
Device::Cuda(_) => {
let wrapper = Box::new(CublasLtWrapper {
cublaslt: CublasLt::new(&device).unwrap(),
});
let wrapper_ptr = Box::leak(wrapper) as &'static CublasLtWrapper;
// Set the controller handle and store the device location
let mut handle_lock = CUBLASLT_CONTROLLER.handle.lock().unwrap();
*handle_lock = Some(wrapper_ptr);
let mut device_loc = CUBLASLT_CONTROLLER.device_location.lock().unwrap();
*device_loc = Some(device.location());
}
_ => {
let mut handle_lock = CUBLASLT_CONTROLLER.handle.lock().unwrap();
*handle_lock = None;
}
}
}
#[cfg(not(feature = "cuda"))]
{
let mut handle_lock = CUBLASLT_CONTROLLER.handle.lock().unwrap();
*handle_lock = None;
}
});
}
#[derive(Debug, Clone)]
pub struct CublasLtWrapper {
#[cfg(feature = "cuda")]
pub cublaslt: CublasLt,
}
impl CublasLtWrapper {
/// Fused batch matmul + add + Relu/Gelu activation using CublasLt for F8 dtypes.
///
/// # Arguments
///
/// * `a` - Input tensor of size BxMxK
/// * `b` - Input tensor of size BxNxK
/// * `dequant_a_scale` - F32 scalar tensor, used to `a` the out tensor.
/// * `dequant_b_scale` - F32 scalar tensor, used to `b` the out tensor.
/// * `quantize_scale` - F32 scalar tensor, used to requantize.
/// * `out` - Optional Output tensor of size BxNxK. If set and beta != 0, will be added to the end result of A*B before `act`
/// * `alpha` - Optional scaling factor for A*B
/// * `beta` - Optional scaling factor for C
/// * `bias` - Optional bias tensor of size M
/// * `act` - Optional Gelu or Relu activation. If set, will be added to the end result
///
/// The resulting tensor is of shape NxM
#[allow(clippy::too_many_arguments)]
pub fn batch_matmul_f8(
&self,
a: &Tensor,
b: &Tensor,
dequant_a_scale: &Tensor,
dequant_b_scale: &Tensor,
quantize_scale: &Tensor,
out: Option<&Tensor>,
alpha: Option<f32>,
beta: Option<f32>,
bias: Option<&Tensor>,
act: Option<HanzoActivation>,
) -> Result<Tensor> {
#[cfg(feature = "cuda")]
{
let inner_act = act.map(|a| match a {
HanzoActivation::Relu => matmul::Activation::Relu,
HanzoActivation::Gelu => matmul::Activation::Gelu,
_ => unreachable!("Unsupported activation in cublaslt matmul"),
});
let mut result = fused_batch_matmul_f8(
a,
b,
dequant_a_scale,
dequant_b_scale,
quantize_scale,
out,
alpha,
beta,
bias,
inner_act,
self.cublaslt.clone(),
)?;
if Some(HanzoActivation::Swiglu) == act {
result = hanzo_nn::ops::swiglu(&result)?;
}
Ok(result)
}
#[cfg(not(feature = "cuda"))]
{
hanzo_ml::bail!("`cuda` feature is not enabled")
}
}
/// Fused batch matmul + add + Relu/Gelu activation using CublasLt.
///
/// # Arguments
///
/// * `a` - Input tensor of size BxMxK
/// * `b` - Input tensor of size BxNxK
/// * `out` - Optional Output tensor of size BxNxK. If set and beta != 0, will be added to the end result of A*B before `act`
/// * `alpha` - Optional scaling factor for A*B
/// * `beta` - Optional scaling factor for C
/// * `bias` - Optional bias tensor of size M
/// * `act` - Optional Gelu or Relu activation. If set, will be added to the end result
///
/// The resulting tensor is of shape NxM
#[allow(clippy::too_many_arguments)]
pub fn batch_matmul(
&self,
a: &Tensor,
b: &Tensor,
out: Option<&Tensor>,
alpha: Option<f32>,
beta: Option<f32>,
bias: Option<&Tensor>,
act: Option<HanzoActivation>,
) -> Result<Tensor> {
#[cfg(feature = "cuda")]
{
let inner_act = act.map(|a| match a {
HanzoActivation::Relu => matmul::Activation::Relu,
HanzoActivation::Gelu => matmul::Activation::Gelu,
_ => unreachable!("Unsupported activation in cublaslt matmul"),
});
let mut result = fused_batch_matmul(
a,
b,
out,
alpha,
beta,
bias,
inner_act,
self.cublaslt.clone(),
)?;
if Some(HanzoActivation::Swiglu) == act {
result = hanzo_nn::ops::swiglu(&result)?;
}
Ok(result)
}
#[cfg(not(feature = "cuda"))]
{
hanzo_ml::bail!("`cuda` feature is not enabled")
}
}
}