File size: 6,472 Bytes
165b25c |
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 |
from typing import TYPE_CHECKING
import torch
# neuron has torch version that doesn't even have impl_abstract
if TYPE_CHECKING:
def register_fake(fn):
return lambda name: fn
else:
try:
from torch.library import register_fake
except ImportError:
from torch.library import impl_abstract as register_fake
try:
from ._ops import ops, add_op_namespace_prefix
except ImportError as e:
# Fallback for local development.
try:
import _quantization
ops = torch.ops._quantization
def add_op_namespace_prefix(op_name: str):
return f"_quantization::{op_name}"
except ImportError:
raise e
from .scalar_type import ScalarType
# fp8 marlin
def fp8_marlin_gemm(
a: torch.Tensor,
b_q_weight: torch.Tensor,
b_scales: torch.Tensor,
workspace: torch.Tensor,
num_bits: int,
size_m: int,
size_n: int,
size_k: int,
) -> torch.Tensor:
return ops.fp8_marlin_gemm(
a, b_q_weight, b_scales, workspace, num_bits, size_m, size_n, size_k
)
# gptq_marlin
def gptq_marlin_gemm(
a: torch.Tensor,
b_q_weight: torch.Tensor,
b_scales: torch.Tensor,
b_zeros: torch.Tensor,
g_idx: torch.Tensor,
perm: torch.Tensor,
workspace: torch.Tensor,
b_q_type: ScalarType,
size_m: int,
size_n: int,
size_k: int,
is_k_full: bool,
has_zp: bool = False,
use_fp32_reduce: bool = False,
is_zp_float: bool = False,
) -> torch.Tensor:
return ops.gptq_marlin_gemm(
a,
b_q_weight,
b_scales,
b_zeros,
g_idx,
perm,
workspace,
b_q_type.id,
size_m,
size_n,
size_k,
is_k_full,
has_zp,
use_fp32_reduce,
is_zp_float,
)
# gptq_marlin
def gptq_marlin_repack(
b_q_weight: torch.Tensor,
perm: torch.Tensor,
size_k: int,
size_n: int,
num_bits: int,
) -> torch.Tensor:
return ops.gptq_marlin_repack(b_q_weight, perm, size_k, size_n, num_bits)
# gptq_marlin
def awq_marlin_repack(
b_q_weight: torch.Tensor, size_k: int, size_n: int, num_bits: int
) -> torch.Tensor:
return ops.awq_marlin_repack(b_q_weight, size_k, size_n, num_bits)
# marlin
def marlin_gemm(
a: torch.Tensor,
b_q_weight: torch.Tensor,
b_scales: torch.Tensor,
workspace: torch.Tensor,
size_m: int,
size_n: int,
size_k: int,
) -> torch.Tensor:
return ops.marlin_gemm(
a, b_q_weight, b_scales, workspace, size_m, size_n, size_k
)
# marlin_24
def gptq_marlin_24_gemm(
a: torch.Tensor,
b_q_weight: torch.Tensor,
b_meta: torch.Tensor,
b_scales: torch.Tensor,
workspace: torch.Tensor,
b_q_type: ScalarType,
size_m: int,
size_n: int,
size_k: int,
) -> torch.Tensor:
return ops.gptq_marlin_24_gemm(
a, b_q_weight, b_meta, b_scales, workspace, b_q_type.id, size_m, size_n, size_k
)
# qqq ops
def marlin_qqq_gemm(
a: torch.Tensor,
b_q_weight: torch.Tensor,
s_tok: torch.Tensor,
s_ch: torch.Tensor,
s_group: torch.Tensor,
workspace: torch.Tensor,
size_m: int,
size_n: int,
size_k: int,
) -> torch.Tensor:
return ops.marlin_qqq_gemm(
a, b_q_weight, s_tok, s_ch, s_group, workspace, size_m, size_n, size_k
)
# Fake ops
if hasattr(ops, "gptq_marlin_24_gemm"):
@register_fake(add_op_namespace_prefix("fp8_marlin_gemm"))
def _fp8_marlin_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
b_scales: torch.Tensor, workspace: torch.Tensor,
num_bits: int, size_m: torch.SymInt,
size_n: torch.SymInt,
size_k: torch.SymInt) -> torch.Tensor:
return torch.empty((size_m, size_n), dtype=a.dtype, device=a.device)
@register_fake(add_op_namespace_prefix("gptq_marlin_24_gemm"))
def _gptq_marlin_24_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
b_meta: torch.Tensor, b_scales: torch.Tensor,
workspace: torch.Tensor,
b_q_type: ScalarType, size_m: torch.SymInt,
size_n: torch.SymInt,
size_k: torch.SymInt) -> torch.Tensor:
return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)
@register_fake(add_op_namespace_prefix("gptq_marlin_gemm"))
def _gptq_marlin_gemm_fake(a: torch.Tensor,
b_q_weight: torch.Tensor,
b_scales: torch.Tensor,
b_zeros: torch.Tensor,
g_idx: torch.Tensor,
perm: torch.Tensor,
workspace: torch.Tensor,
b_q_type: ScalarType,
size_m: torch.SymInt,
size_n: torch.SymInt,
size_k: torch.SymInt,
is_k_full: bool,
has_zp: bool = False,
use_fp32_reduce: bool = False,
is_zp_float: bool = False) -> torch.Tensor:
return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)
@register_fake(add_op_namespace_prefix("marlin_qqq_gemm"))
def _marlin_qqq_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
s_tok: torch.Tensor, s_ch: torch.Tensor,
s_group: torch.Tensor, workspace: torch.Tensor,
size_m: torch.SymInt, size_n: torch.SymInt,
size_k: torch.SymInt) -> torch.Tensor:
return torch.empty((size_m, size_n),
dtype=torch.float16,
device=a.device)
@register_fake(add_op_namespace_prefix("marlin_gemm"))
def _marlin_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
b_scales: torch.Tensor, workspace: torch.Tensor,
size_m: torch.SymInt, size_n: torch.SymInt,
size_k: torch.SymInt) -> torch.Tensor:
return torch.empty((size_m, size_n),
dtype=torch.float16,
device=a.device)
|