Spaces:
Running
on
A100
Running
on
A100
File size: 5,547 Bytes
174ae06 |
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 |
# Copyright (c) 2025 NVIDIA CORPORATION.
# Licensed under the MIT license.
# Adapted from https://github.com/NVlabs/VILA/tree/main under the Apache 2.0 license.
# LICENSE is in incl_licenses directory.
import math
import struct
import numpy as np
import torch
import triton
import triton.language as tl
from triton.language.extra.cuda import libdevice
segment_size = 1024**3
def floatExMy_quantize_triton(x, e_bit, m_bit, stochastic):
x_ori_shape = x.shape
x = x.view(-1)
n_elements = x.numel()
if n_elements <= segment_size:
grid = lambda meta: (triton.cdiv(n_elements, meta["BLOCK_SIZE"]),)
y = torch.empty_like(x)
if x.dtype in [torch.bfloat16, torch.float32]:
if stochastic:
noise = x.new(x.shape).uniform_(-0.5, 0.5)
_floatExMy_stochastic_quantize_kernel[grid](x, noise, y, n_elements, e_bit, m_bit)
else:
_floatExMy_quantize_kernel[grid](x, y, n_elements, e_bit, m_bit)
torch.cuda.synchronize()
else:
raise NotImplementedError(f"Other data format {x.dtype} for float quantization triton")
else: # Triton will break when x.numel > 2 * 1024 ** 3
num_segments = n_elements // segment_size + 1
split_size = [segment_size] * (num_segments - 1) + [n_elements - segment_size * (num_segments - 1)]
x_list = x.split(split_size)
y_list = []
del x
for x in x_list:
n_elements = x.numel()
grid = lambda meta: (triton.cdiv(n_elements, meta["BLOCK_SIZE"]),)
y = torch.empty_like(x)
if x.dtype in [torch.bfloat16, torch.float32]:
if stochastic:
noise = x.new(x.shape).uniform_(-0.5, 0.5)
_floatExMy_stochastic_quantize_kernel[grid](x, noise, y, n_elements, e_bit, m_bit)
else:
_floatExMy_quantize_kernel[grid](x, y, n_elements, e_bit, m_bit)
torch.cuda.synchronize()
else:
raise NotImplementedError(f"Other data format {x.dtype} for float quantization triton")
y_list.append(y)
y = torch.concat(y_list)
del y_list
y = y.reshape(x_ori_shape)
return y
@triton.autotune(
configs=[
# triton.Config({'BLOCK_SIZE': 4,}, num_warps=4),
triton.Config(
{
"BLOCK_SIZE": 1024,
},
num_warps=4,
),
triton.Config(
{
"BLOCK_SIZE": 2048,
},
num_warps=4,
),
],
key=["n_elements"],
)
@triton.jit
def _floatExMy_quantize_kernel(
x_ptr,
output_ptr,
n_elements,
e_bit,
m_bit,
BLOCK_SIZE: tl.constexpr,
):
if isinstance(e_bit, tl.constexpr):
ebit = e_bit.value
else:
ebit = e_bit
if isinstance(m_bit, tl.constexpr):
mbit = m_bit.value
else:
mbit = m_bit
pid = tl.program_id(axis=0)
block_start = pid * BLOCK_SIZE
offsets = block_start + tl.arange(0, BLOCK_SIZE)
mask = offsets < n_elements
x = tl.load(x_ptr + offsets, mask=mask)
x = x.to(tl.float32)
sign = 1 - 2 * libdevice.signbit(x)
x_abs = tl.abs(x)
Elow = -tl.exp2((ebit - 1).to(tl.float32)) + 2
Ehigh = tl.exp2((ebit - 1).to(tl.float32))
Mhigh = tl.exp2(mbit.to(tl.float32))
expo = tl.floor(tl.log2(x_abs))
expo = tl.clamp(expo, min=Elow, max=Ehigh)
mant = x_abs / tl.exp2(expo)
mant_int = tl.floor(mant)
mant_frac = mant - mant_int
mant_frac = mant_frac * Mhigh
# mant_frac = mant_frac + noise
mant_frac = libdevice.round(mant_frac)
mant_q = mant_int + mant_frac / Mhigh
y = sign * tl.exp2(expo) * mant_q
y = y.to(x_ptr.dtype.element_ty)
tl.store(output_ptr + offsets, y, mask=mask)
@triton.autotune(
configs=[
# triton.Config({'BLOCK_SIZE': 4,}, num_warps=4),
triton.Config(
{
"BLOCK_SIZE": 1024,
},
num_warps=4,
),
triton.Config(
{
"BLOCK_SIZE": 2048,
},
num_warps=4,
),
],
key=["n_elements"],
)
@triton.jit
def _floatExMy_stochastic_quantize_kernel(
x_ptr,
noise_ptr,
output_ptr,
n_elements,
e_bit,
m_bit,
BLOCK_SIZE: tl.constexpr,
):
if isinstance(e_bit, tl.constexpr):
ebit = e_bit.value
else:
ebit = e_bit
if isinstance(m_bit, tl.constexpr):
mbit = m_bit.value
else:
mbit = m_bit
pid = tl.program_id(axis=0)
block_start = pid * BLOCK_SIZE
offsets = block_start + tl.arange(0, BLOCK_SIZE)
mask = offsets < n_elements
x = tl.load(x_ptr + offsets, mask=mask)
noise = tl.load(noise_ptr + offsets, mask=mask)
x = x.to(tl.float32)
sign = 1 - 2 * libdevice.signbit(x)
x_abs = tl.abs(x)
Elow = -tl.exp2((ebit - 1).to(tl.float32)) + 2
Ehigh = tl.exp2((ebit - 1).to(tl.float32))
Mhigh = tl.exp2(mbit.to(tl.float32))
expo = tl.floor(tl.log2(x_abs))
expo = tl.clamp(expo, min=Elow, max=Ehigh)
mant = x_abs / tl.exp2(expo)
mant_int = tl.floor(mant)
mant_frac = mant - mant_int
mant_frac = mant_frac * Mhigh
mant_frac = mant_frac + noise
mant_frac = libdevice.round(mant_frac)
mant_q = mant_int + mant_frac / Mhigh
y = sign * tl.exp2(expo) * mant_q
y = y.to(x_ptr.dtype.element_ty)
tl.store(output_ptr + offsets, y, mask=mask)
|