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
|
use super::{GgmlDType, QStorage};
use crate::backend::BackendStorage;
use crate::{DType, MetalDevice, MetalStorage, Result, Shape};
use metal::Buffer;
use std::sync::Arc;
pub struct QMetalStorage {
dtype: GgmlDType,
device: MetalDevice,
buffer: Arc<Buffer>,
}
impl QMetalStorage {
pub fn zeros(device: &MetalDevice, elem_count: usize, dtype: GgmlDType) -> Result<Self> {
let size = elem_count * dtype.type_size() / dtype.block_size();
let buffer = device.allocate_zeros(size)?;
Ok(Self {
buffer,
device: device.clone(),
dtype,
})
}
pub fn dtype(&self) -> GgmlDType {
self.dtype
}
pub fn device(&self) -> &MetalDevice {
&self.device
}
pub fn buffer(&self) -> &Buffer {
&self.buffer
}
pub fn dequantize(&self, elem_count: usize) -> Result<MetalStorage> {
use crate::quantized::k_quants::GgmlType;
let buffer = self.device.new_buffer_managed(self.buffer.length())?;
let command_buffer = self.device.command_buffer()?;
command_buffer.set_label("to_cpu");
let blit = command_buffer.new_blit_command_encoder();
blit.set_label("blit_to_cpu");
blit.copy_from_buffer(&self.buffer, 0, &buffer, 0, self.buffer.length());
blit.end_encoding();
self.device.wait_until_completed()?;
let mut out = vec![0.0; elem_count];
let block_len = elem_count / self.dtype.block_size();
match self.dtype {
GgmlDType::F32 => {
let vec: Vec<f32> = read_to_vec(&buffer, block_len);
f32::to_float(&vec, &mut out)?;
}
GgmlDType::F16 => {
let vec: Vec<half::f16> = read_to_vec(&buffer, block_len);
half::f16::to_float(&vec, &mut out)?;
}
GgmlDType::Q4_0 => {
let vec: Vec<crate::quantized::BlockQ4_0> = read_to_vec(&buffer, block_len);
crate::quantized::BlockQ4_0::to_float(&vec, &mut out)?;
}
GgmlDType::Q4_1 => {
let vec: Vec<crate::quantized::BlockQ4_1> = read_to_vec(&buffer, block_len);
crate::quantized::BlockQ4_1::to_float(&vec, &mut out)?;
}
GgmlDType::Q5_0 => {
let vec: Vec<crate::quantized::BlockQ5_0> = read_to_vec(&buffer, block_len);
crate::quantized::BlockQ5_0::to_float(&vec, &mut out)?;
}
GgmlDType::Q5_1 => {
let vec: Vec<crate::quantized::BlockQ5_1> = read_to_vec(&buffer, block_len);
crate::quantized::BlockQ5_1::to_float(&vec, &mut out)?;
}
GgmlDType::Q8_0 => {
let vec: Vec<crate::quantized::BlockQ8_0> = read_to_vec(&buffer, block_len);
crate::quantized::BlockQ8_0::to_float(&vec, &mut out)?;
}
GgmlDType::Q8_1 => {
let vec: Vec<crate::quantized::BlockQ8_1> = read_to_vec(&buffer, block_len);
crate::quantized::BlockQ8_1::to_float(&vec, &mut out)?;
}
GgmlDType::Q2K => {
let vec: Vec<crate::quantized::BlockQ2K> = read_to_vec(&buffer, block_len);
crate::quantized::BlockQ2K::to_float(&vec, &mut out)?;
}
GgmlDType::Q3K => {
let vec: Vec<crate::quantized::BlockQ3K> = read_to_vec(&buffer, block_len);
crate::quantized::BlockQ3K::to_float(&vec, &mut out)?;
}
GgmlDType::Q4K => {
let vec: Vec<crate::quantized::BlockQ4K> = read_to_vec(&buffer, block_len);
crate::quantized::BlockQ4K::to_float(&vec, &mut out)?;
}
GgmlDType::Q5K => {
let vec: Vec<crate::quantized::BlockQ5K> = read_to_vec(&buffer, block_len);
crate::quantized::BlockQ5K::to_float(&vec, &mut out)?;
}
GgmlDType::Q6K => {
let vec: Vec<crate::quantized::BlockQ6K> = read_to_vec(&buffer, block_len);
crate::quantized::BlockQ6K::to_float(&vec, &mut out)?;
}
GgmlDType::Q8K => {
let vec: Vec<crate::quantized::BlockQ8K> = read_to_vec(&buffer, block_len);
crate::quantized::BlockQ8K::to_float(&vec, &mut out)?;
}
}
let buffer = self.device.new_buffer_with_data(&out)?;
Ok(MetalStorage::new(
buffer,
self.device.clone(),
elem_count,
DType::F32,
))
}
pub fn quantize(&mut self, src: &MetalStorage) -> Result<()> {
// Quantization only happens on CPU for now.
let src = src.to_cpu::<f32>()?;
let elem_count = src.len();
let src = crate::Storage::Cpu(crate::CpuStorage::F32(src));
let mut qcpu_storage = crate::Device::Cpu.qzeros(elem_count, self.dtype)?;
qcpu_storage.quantize(&src)?;
let buffer = self.device.new_buffer_with_data(&qcpu_storage.data()?)?;
self.buffer = buffer;
Ok(())
}
pub fn storage_size_in_bytes(&self) -> usize {
self.buffer.length() as usize
}
pub fn fwd(
&self,
self_shape: &Shape,
storage: &MetalStorage,
layout: &crate::Layout,
) -> Result<(MetalStorage, Shape)> {
use crate::MetalError;
if !layout.is_contiguous() {
crate::bail!("input tensor is not contiguous {layout:?}")
}
let src_shape = layout.shape();
// self is transposed so n is first then k.
if src_shape.rank() < 2 {
crate::bail!("input tensor has only one dimension {layout:?}")
}
let (n, k) = self_shape.dims2()?;
let mut dst_shape = src_shape.dims().to_vec();
// We always use a single batch dimension and stack all the tensors in the batch on the
// second dimension as the implementation in candle-metal-kernels doesn't handle batch
// properly.
let m = match dst_shape.len() {
3 => dst_shape[0] * dst_shape[1],
2 => dst_shape[0],
n => crate::bail!("Invalid rank {n} for quantized matmul metal"),
};
let last_k = dst_shape.pop().unwrap();
if last_k != k {
crate::bail!("input tensor {layout:?} incompatible with {:?}", self_shape)
}
dst_shape.push(n);
let dst_shape = Shape::from(dst_shape);
let device = storage.device().clone();
let dst = device.new_buffer(dst_shape.elem_count(), DType::F32, "qmatmul")?;
let command_buffer = device.command_buffer()?;
// In some cases it would be better to use the mm variant, though it has its drawbacks
// around memory alignemnt.
for batch_id in 0..m {
candle_metal_kernels::call_quantized_matmul_mv_t(
device.device(),
&command_buffer,
device.kernels(),
self.dtype.into(),
(1, 1, n, k),
storage.buffer(),
(layout.start_offset() + batch_id * k) * storage.dtype().size_in_bytes(),
&self.buffer,
batch_id * n * DType::F32.size_in_bytes(),
&dst,
)
.map_err(MetalError::from)?;
}
let dst_storage = crate::MetalStorage::new(dst, device, dst_shape.elem_count(), DType::F32);
Ok((dst_storage, dst_shape))
}
}
pub fn load_quantized<T: super::GgmlType + Send + Sync + 'static>(
device: &MetalDevice,
data: &[T],
) -> Result<QStorage> {
let buffer = device.new_buffer_with_data(data)?;
let device = device.clone();
Ok(QStorage::Metal(QMetalStorage {
dtype: T::DTYPE,
device,
buffer,
}))
}
fn read_to_vec<T: Clone>(buffer: &Buffer, n: usize) -> Vec<T> {
let ptr = buffer.contents() as *const T;
assert!(!ptr.is_null());
let slice = unsafe { std::slice::from_raw_parts(ptr, n) };
slice.to_vec()
}
impl From<GgmlDType> for candle_metal_kernels::GgmlDType {
fn from(value: GgmlDType) -> Self {
match value {
GgmlDType::Q4_0 => candle_metal_kernels::GgmlDType::Q4_0,
GgmlDType::Q4_1 => candle_metal_kernels::GgmlDType::Q4_1,
GgmlDType::Q5_0 => candle_metal_kernels::GgmlDType::Q5_0,
GgmlDType::Q5_1 => candle_metal_kernels::GgmlDType::Q5_1,
GgmlDType::Q8_0 => candle_metal_kernels::GgmlDType::Q8_0,
GgmlDType::Q8_1 => candle_metal_kernels::GgmlDType::Q8_1,
GgmlDType::Q2K => candle_metal_kernels::GgmlDType::Q2K,
GgmlDType::Q3K => candle_metal_kernels::GgmlDType::Q3K,
GgmlDType::Q4K => candle_metal_kernels::GgmlDType::Q4K,
GgmlDType::Q5K => candle_metal_kernels::GgmlDType::Q5K,
GgmlDType::Q6K => candle_metal_kernels::GgmlDType::Q6K,
GgmlDType::Q8K => candle_metal_kernels::GgmlDType::Q8K,
GgmlDType::F16 => candle_metal_kernels::GgmlDType::F16,
GgmlDType::F32 => candle_metal_kernels::GgmlDType::F32,
}
}
}
|
