Skip to content

Optimize bitspack.py with torch.arange and bitwise operators #190

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

Merged
hzhangxyz merged 1 commit into from
Mar 5, 2026
Merged

Optimize bitspack.py with torch.arange and bitwise operators #190

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
28 changes: 11 additions & 17 deletions qmp/utility/bitspack.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -41,16 +41,14 @@ def pack_int(tensor: torch.Tensor, size: int) -> torch.Tensor:
# Calculate the number of groups
num_bytes = last_dim // elements_per_byte

# Reshape the tensor to (..., num_bytes, elements_per_bytes)
# Reshape the tensor to (..., num_bytes, elements_per_byte)
shape.append(num_bytes)
shape.append(elements_per_byte)
tensor = tensor.view(shape)

# Define the weights tensor
weights = torch.tensor([1 << i for i in range(0, 8, size)], device=tensor.device, dtype=torch.uint8)

# Calculate the byte value for each group
packed = torch.sum(tensor * weights, dim=-1, dtype=torch.uint8)
# Define the shifts and calculate the byte value for each group
shifts = torch.arange(0, 8, size, device=tensor.device, dtype=torch.uint8)
packed = torch.sum(tensor << shifts, dim=-1, dtype=torch.uint8)

assert packed.dtype == torch.uint8
return packed
Expand Down Expand Up @@ -78,21 +76,17 @@ def unpack_int(tensor: torch.Tensor, size: int, last_dim: int) -> torch.Tensor:
assert tensor.dtype == torch.uint8
assert size in [1, 2, 4, 8]

# Define the weights tensor and shifts tensor
weights = (
torch.tensor([1 << i for i in range(8)], device=tensor.device, dtype=torch.uint8)
.view([8 // size, size])
.sum(dim=-1, dtype=torch.uint8)
)
shifts = torch.tensor(list(range(0, 8, size)), device=tensor.device, dtype=torch.uint8)
# Define the shifts and mask
shifts = torch.arange(0, 8, size, device=tensor.device, dtype=torch.uint8)
mask = torch.tensor((1 << size) - 1, device=tensor.device, dtype=torch.uint8)

# Calculate the unpacked values
result = torch.bitwise_right_shift(torch.bitwise_and(tensor.unsqueeze(-1), weights), shifts)
unpacked = torch.bitwise_and(torch.bitwise_right_shift(tensor.unsqueeze(-1), shifts), mask)

# Reshape the result to the original shape
shape = list(result.shape)
# Reshape the unpacked tensor to the original shape
shape = list(unpacked.shape)
shape.append(shape.pop() * shape.pop())
unpacked = result.view(shape)[..., :last_dim]
unpacked = unpacked.view(shape)[..., :last_dim]

assert unpacked.dtype == torch.uint8
return unpacked