Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

use uint16 for g_idx #934

Closed
wants to merge 3 commits into from
Closed

use uint16 for g_idx #934

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1acdb93
use uint16 for g_idx
Qubitium Dec 19, 2024
aa5d6eb
cleanup
Qubitium Dec 19, 2024
7c6e1d7
update
Qubitium Dec 20, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
30 changes: 15 additions & 15 deletions gptqmodel/nn_modules/qlinear/torch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -79,7 +79,7 @@ def __init__(
)
self.register_buffer(
"g_idx",
torch.tensor([i // self.group_size for i in range(infeatures)], dtype=torch.int32),
torch.tensor([i // self.group_size for i in range(infeatures)], dtype=torch.uint16),
)
if bias:
self.register_buffer("bias", torch.zeros((outfeatures), dtype=weight_dtype))
Expand Down Expand Up @@ -107,7 +107,7 @@ def post_init(self):
self.outfeatures // 32 * self.bits
)
self.scales.resize_((math.ceil(self.padded_infeatures / self.group_size), self.outfeatures), )
self.g_idx = torch.tensor([i // self.group_size for i in range(self.padded_infeatures)], dtype=torch.int32,
self.g_idx = torch.tensor([i // self.group_size for i in range(self.padded_infeatures)], dtype=torch.uint16,
device=self.g_idx.device)

def pack(self, linear, scales, zeros, g_idx=None):
Expand Down Expand Up @@ -174,7 +174,7 @@ def pack(self, linear, scales, zeros, g_idx=None):
i = 0
col = 0
while col < qzeros.shape[1]:
if self.bits in [2, 4, 8]:
if self.bits == 4 or self.bits == 8 or self.bits == 2:
for j in range(i, i + (32 // self.bits)):
qzeros[:, col] |= zeros[:, j] << (self.bits * (j - i))
i += 32 // self.bits
Expand Down Expand Up @@ -213,7 +213,7 @@ def forward(self, x: torch.Tensor):
if self.wf.device != self.qzeros.device:
self.wf = self.wf.to(self.qzeros.device)

if self.bits in [2, 4, 8]:
if self.bits == 4 or self.bits in [8, 2]:
zeros = torch.bitwise_right_shift(
torch.unsqueeze(self.qzeros, 2).expand(-1, -1, 32 // self.bits),
self.wf.unsqueeze(0),
Expand Down Expand Up @@ -251,23 +251,23 @@ def forward(self, x: torch.Tensor):

weight = weight.reshape(weight.shape[0] * weight.shape[1], weight.shape[2])

num_itr = self.g_idx.shape[0] // x.shape[-1]
if num_itr == 1:
weights = self.scales[self.g_idx.long()] * (weight - zeros[self.g_idx.long()])
count = self.g_idx.shape[0] // x.shape[-1]
if count == 1:
g_idx_int32 = self.g_idx.int()
weights = self.scales[g_idx_int32] * (weight - zeros[g_idx_int32])
else:
num_dim = self.g_idx.shape[0] // num_itr
num_dim = self.g_idx.shape[0] // count
weights = []
for i in range(num_itr):
for i in range(count):
scale_i = self.scales[:, i * num_dim : (i + 1) * num_dim]
weight_i = weight[:, i * num_dim : (i + 1) * num_dim]
zeros_i = zeros[:, i * num_dim : (i + 1) * num_dim]
g_idx_i = self.g_idx[i * num_dim : (i + 1) * num_dim]
weights.append(scale_i[g_idx_i.long()] * (weight_i - zeros_i[g_idx_i.long()]))
g_idx_i = self.g_idx[i * num_dim : (i + 1) * num_dim].int()
weights.append(scale_i[g_idx_i] * (weight_i - zeros_i[g_idx_i]))
weights = torch.cat(weights, dim=1)
out = torch.matmul(x, weights)
out = out.to(x_dtype)
out = out.reshape(out_shape)
out = out + self.bias if self.bias is not None else out
out = torch.matmul(x, weights).reshape(out_shape)
if self.bias is not None:
out = out + self.bias
return out


Expand Down
Loading