pull/4821/head
Tong Li 1 year ago committed by GitHub
parent fb46d05cdf
commit bbbcac26e8
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23

@ -1,19 +1,17 @@
import os
import torch
import re
import torch
from .builder import Builder
from .utils import append_nvcc_threads, get_cuda_cc_flag
from .utils import append_nvcc_threads
class GPTQBuilder(Builder):
class GPTQBuilder(Builder):
NAME = "cu_gptq"
PREBUILT_IMPORT_PATH = "colossalai._C.cu_gptq"
def __init__(self):
super().__init__(name=GPTQBuilder.NAME,
prebuilt_import_path=GPTQBuilder.PREBUILT_IMPORT_PATH)
super().__init__(name=GPTQBuilder.NAME, prebuilt_import_path=GPTQBuilder.PREBUILT_IMPORT_PATH)
def include_dirs(self):
ret = [self.csrc_abs_path("gptq"), self.get_cuda_home_include()]
@ -21,32 +19,38 @@ class GPTQBuilder(Builder):
def sources_files(self):
ret = [
self.csrc_abs_path(fname) for fname in [
'gptq/linear_gptq.cpp',
'gptq/column_remap.cu',
'gptq/cuda_buffers.cu',
'gptq/q4_matmul.cu',
'gptq/q4_matrix.cu'
self.csrc_abs_path(fname)
for fname in [
"gptq/linear_gptq.cpp",
"gptq/column_remap.cu",
"gptq/cuda_buffers.cu",
"gptq/q4_matmul.cu",
"gptq/q4_matrix.cu",
]
]
return ret
def cxx_flags(self):
return ['-O3'] + self.version_dependent_macros
return ["-O3"] + self.version_dependent_macros
def nvcc_flags(self):
extra_cuda_flags = ['-v',
'-std=c++14', '-U__CUDA_NO_HALF_OPERATORS__', '-U__CUDA_NO_HALF_CONVERSIONS__',
'-U__CUDA_NO_HALF2_OPERATORS__', '-DTHRUST_IGNORE_CUB_VERSION_CHECK', "-lcublas", "-std=c++17"
extra_cuda_flags = [
"-v",
"-std=c++14",
"-U__CUDA_NO_HALF_OPERATORS__",
"-U__CUDA_NO_HALF_CONVERSIONS__",
"-U__CUDA_NO_HALF2_OPERATORS__",
"-DTHRUST_IGNORE_CUB_VERSION_CHECK",
"-lcublas",
"-std=c++17",
]
for arch in torch.cuda.get_arch_list():
res = re.search(r'sm_(\d+)', arch)
res = re.search(r"sm_(\d+)", arch)
if res:
arch_cap = res[1]
if int(arch_cap) >= 80:
extra_cuda_flags.extend(['-gencode', f'arch=compute_{arch_cap},code={arch}'])
extra_cuda_flags.extend(["-gencode", f"arch=compute_{arch_cap},code={arch}"])
ret = ['-O3', '--use_fast_math'] + self.version_dependent_macros + extra_cuda_flags
return append_nvcc_threads(ret)
ret = ["-O3", "--use_fast_math"] + self.version_dependent_macros + extra_cuda_flags
return append_nvcc_threads(ret)

Loading…
Cancel
Save