[mypy] Enable type checking for test directory (#5017)
This commit is contained in:
@@ -25,24 +25,25 @@ GPTQ_MARLIN_SUPPORTED_SYM = [True]
|
||||
|
||||
|
||||
# Permutations for Marlin scale shuffling
|
||||
def get_scale_perms(num_bits):
|
||||
scale_perm = []
|
||||
def get_scale_perms(num_bits: int):
|
||||
scale_perm: List[int] = []
|
||||
for i in range(8):
|
||||
scale_perm.extend([i + 8 * j for j in range(8)])
|
||||
scale_perm_single = []
|
||||
scale_perm_single: List[int] = []
|
||||
for i in range(4):
|
||||
scale_perm_single.extend(
|
||||
[2 * i + j for j in [0, 1, 8, 9, 16, 17, 24, 25]])
|
||||
return scale_perm, scale_perm_single
|
||||
|
||||
|
||||
def get_pack_factor(num_bits):
|
||||
def get_pack_factor(num_bits: int):
|
||||
assert (num_bits in GPTQ_MARLIN_SUPPORTED_NUM_BITS
|
||||
), f"Unsupported num_bits = {num_bits}"
|
||||
return 32 // num_bits
|
||||
|
||||
|
||||
def marlin_permute_scales(s, size_k, size_n, group_size, num_bits):
|
||||
def marlin_permute_scales(s: torch.Tensor, size_k: int, size_n: int,
|
||||
group_size: int, num_bits: int):
|
||||
scale_perm, scale_perm_single = get_scale_perms(num_bits)
|
||||
if group_size < size_k and group_size != -1:
|
||||
s = s.reshape((-1, len(scale_perm)))[:, scale_perm]
|
||||
|
||||
@@ -1,4 +1,6 @@
|
||||
"""This file is used for /tests and /benchmarks"""
|
||||
from typing import Dict, List
|
||||
|
||||
import numpy
|
||||
import torch
|
||||
|
||||
@@ -11,10 +13,10 @@ import torch
|
||||
#
|
||||
# As a result of this reordering, the vector loads inside the kernel will get the data as it is needed for tensor-core # noqa: E501
|
||||
# (without the need to use ldmatrix instructions) # noqa: E501
|
||||
def get_perms_24(num_bits):
|
||||
perm_list = []
|
||||
def get_perms_24(num_bits: int):
|
||||
perm_list: List[int] = []
|
||||
for i in range(32):
|
||||
perm1 = []
|
||||
perm1: List[int] = []
|
||||
col = i // 4
|
||||
col_o = col // 2
|
||||
for block in [0, 1]:
|
||||
@@ -39,18 +41,18 @@ def get_perms_24(num_bits):
|
||||
|
||||
perm = perm.reshape((-1, len(interleave)))[:, interleave].ravel()
|
||||
perm = torch.from_numpy(perm)
|
||||
scale_perm = []
|
||||
scale_perm: List[int] = []
|
||||
for i in range(8):
|
||||
scale_perm.extend([i * 8 + j for j in [0, 4, 1, 5, 2, 6, 3, 7]])
|
||||
scale_perm_single = []
|
||||
scale_perm_single: List[int] = []
|
||||
for i in range(8):
|
||||
scale_perm_single.extend([8 * i + j for j in [0, 1, 2, 3, 4, 5, 6, 7]])
|
||||
return perm, scale_perm, scale_perm_single
|
||||
|
||||
|
||||
marlin_24_perm = {}
|
||||
marlin_24_scale_perm = {}
|
||||
marlin_24_scale_perm_single = {}
|
||||
marlin_24_perm: Dict[int, torch.Tensor] = {}
|
||||
marlin_24_scale_perm: Dict[int, List[int]] = {}
|
||||
marlin_24_scale_perm_single: Dict[int, List[int]] = {}
|
||||
for num_bits in [4, 8]:
|
||||
perm_24, scale_perm_24, scale_perm_single_24 = get_perms_24(num_bits)
|
||||
marlin_24_perm[num_bits] = perm_24
|
||||
|
||||
@@ -1,4 +1,6 @@
|
||||
"""This file is used for /tests and /benchmarks"""
|
||||
from typing import Dict, List
|
||||
|
||||
import numpy
|
||||
import torch
|
||||
|
||||
@@ -11,10 +13,10 @@ import torch
|
||||
#
|
||||
# As a result of this reordering, the vector loads inside the kernel will get the data as it is needed for tensor-core # noqa: E501
|
||||
# (without the need to use ldmatrix instructions) # noqa: E501
|
||||
def get_perms(num_bits):
|
||||
perm_list = []
|
||||
def get_perms(num_bits: int):
|
||||
perm_list: List[int] = []
|
||||
for i in range(32):
|
||||
perm1 = []
|
||||
perm1: List[int] = []
|
||||
col = i // 4
|
||||
for block in [0, 1]:
|
||||
for row in [
|
||||
@@ -38,19 +40,19 @@ def get_perms(num_bits):
|
||||
|
||||
perm = perm.reshape((-1, len(interleave)))[:, interleave].ravel()
|
||||
perm = torch.from_numpy(perm)
|
||||
scale_perm = []
|
||||
scale_perm: List[int] = []
|
||||
for i in range(8):
|
||||
scale_perm.extend([i + 8 * j for j in range(8)])
|
||||
scale_perm_single = []
|
||||
scale_perm_single: List[int] = []
|
||||
for i in range(4):
|
||||
scale_perm_single.extend(
|
||||
[2 * i + j for j in [0, 1, 8, 9, 16, 17, 24, 25]])
|
||||
return perm, scale_perm, scale_perm_single
|
||||
|
||||
|
||||
marlin_perm = {}
|
||||
marlin_scale_perm = {}
|
||||
marlin_scale_perm_single = {}
|
||||
marlin_perm: Dict[int, torch.Tensor] = {}
|
||||
marlin_scale_perm: Dict[int, List[int]] = {}
|
||||
marlin_scale_perm_single: Dict[int, List[int]] = {}
|
||||
for num_bits in [4, 8]:
|
||||
perm, scale_perm, scale_perm_single = get_perms(num_bits)
|
||||
marlin_perm[num_bits] = perm
|
||||
|
||||
Reference in New Issue
Block a user