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

Jump to bottom

Add support for symmetric quantization #124

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Add support for symmetric quantization #124

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
6de66ec
add support for symmetric quantization during compression
julian-q Sep 16, 2023
d67c1b7
clean up symmetric / asymmetric compression
julian-q Sep 16, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
57 changes: 36 additions & 21 deletions flexgen/compression.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,16 +33,16 @@ def allocate(self, shape, dtype, comp_config, pin_memory=None, name=None):
"""Allocate a compressed TorchTensor. Round up the shape to group boundary."""
assert comp_config.num_bits == 4 and dtype == np.float16

group_size, group_dim = comp_config.group_size, comp_config.group_dim
group_size, group_dim, symmetric = comp_config.group_size, comp_config.group_dim, comp_config.symmetric

# Round up
num_groups = (shape[group_dim] + group_size - 1) // group_size
data_shape = (
shape[:group_dim] + (num_groups * (group_size // 2),) + shape[group_dim+1:])
scale_shape = (
shape[:group_dim] + (num_groups, 2) + shape[group_dim+1:])
shape[:group_dim] + (num_groups, 1 if symmetric else 2) + shape[group_dim+1:])

data = self.base_device.allocate(data_shape, np.uint8, pin_memory=pin_memory)
data = self.base_device.allocate(data_shape, np.int8 if symmetric else np.uint8, pin_memory=pin_memory)
scale = self.base_device.allocate(scale_shape, np.float16, pin_memory=pin_memory)

return TorchTensor(shape, np_dtype_to_torch_dtype[dtype],
Expand Down Expand Up @@ -89,7 +89,7 @@ def compress(self, tensor, comp_config):
group_size, num_bits, group_dim, symmetric = (
comp_config.group_size, comp_config.num_bits,
comp_config.group_dim, comp_config.symmetric)
assert num_bits == 4 and group_size % 2 == 0 and not symmetric
assert num_bits == 4 and group_size % 2 == 0

if tensor.device.type == "cpu" and tensor.dtype == torch.float16:
tensor = tensor.float()
Expand All @@ -110,14 +110,21 @@ def compress(self, tensor, comp_config):
data = tensor.view(new_shape)

# Quantize
B = 2 ** num_bits - 1
mn = torch.min(data, dim=group_dim + 1, keepdim=True)[0]
mx = torch.max(data, dim=group_dim + 1, keepdim=True)[0]

scale = B / (mx - mn)
data = data - mn
data.mul_(scale)
data = data.clamp_(0, B).round_().to(torch.uint8)
if symmetric:
B = 2 ** (num_bits - 1) - 1
mn = None
mx = torch.max(data.abs(), dim=group_dim + 1, keepdim=True)[0]
scale = B / mx
data = data * scale
data = data.clamp_(-(B + 1), B).round_().to(torch.int8)
else:
B = 2 ** num_bits - 1
mn = torch.min(data, dim=group_dim + 1, keepdim=True)[0]
mx = torch.max(data, dim=group_dim + 1, keepdim=True)[0]
scale = B / (mx - mn)
data = data - mn
data.mul_(scale)
data = data.clamp_(0, B).round_().to(torch.uint8)

# Pack
left_indices = (
Expand All @@ -127,15 +134,19 @@ def compress(self, tensor, comp_config):
tuple(slice(0, x) for x in data.shape[:group_dim+1]) +
(slice(1, data.shape[group_dim+1], 2),))
data = torch.bitwise_or(
data[left_indices].bitwise_left_shift(4), data[right_indices])
data[left_indices].bitwise_left_shift(4),
data[right_indices].bitwise_and(0xF))

# Reshape
data_shape = (
shape[:group_dim] + (num_groups * (group_size // 2),) + shape[group_dim+1:])
scale_shape = (
shape[:group_dim] + (num_groups, 2) + shape[group_dim+1:])
data = data.view(data_shape)
scale = torch.cat([scale, mn], dim=group_dim+1).view(scale_shape)
scale_shape = (
shape[:group_dim] + (num_groups, 1 if symmetric else 2) + shape[group_dim+1:])
if symmetric:
scale = scale.view(scale_shape)
else:
scale = torch.cat([scale, mn], dim=group_dim+1).view(scale_shape)

data = TorchTensor.create_from_torch(data, self.base_device)
scale = TorchTensor.create_from_torch(scale, self.base_device)
Expand Down Expand Up @@ -182,12 +193,16 @@ def decompress(self, tensor):
tuple(slice(0, x) for x in data.shape[:group_dim+1]) +
(slice(1, data.shape[group_dim+1], 2),))
data[left_indices] = packed.bitwise_right_shift(4)
data[right_indices] = packed.bitwise_and(0xF)
data[right_indices] = packed.bitwise_and(0xF).bitwise_left_shift(4).bitwise_right_shift(4)

# Dequantize
scale, mn = scale.data.split(1, dim=group_dim + 1)
data.div_(scale)
data.add_(mn)
if symmetric:
scale = scale.data
data = data / scale
else:
scale, mn = scale.data.split(1, dim=group_dim + 1)
data.div_(scale)
data.add_(mn)

# Reshape
unpad_len = (group_size - tensor.shape[group_dim] % group_size) % group_size
Expand Down Expand Up @@ -281,7 +296,7 @@ def compress(tensor, config):
B = 2 ** (num_bits - 1) - 1
scale = B / torch.max(data.abs(), dim=group_dim + 1, keepdim=True)[0]
data = data * scale
data = data.clamp_(-B, B).round_().to(torch.int8)
data = data.clamp_(-(B + 1), B).round_().to(torch.int8)
return data, scale, original_shape
else:
B = 2 ** num_bits - 1
Expand Down
2 changes: 1 addition & 1 deletion flexgen/flex_opt.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1205,7 +1205,7 @@ def run_flexgen(args):
group_dim=0, symmetric=False),
args.compress_cache,
CompressionConfig(num_bits=4, group_size=64,
group_dim=2, symmetric=False))
group_dim=2, symmetric=True))
assert not (args.compress_cache and args.attn_sparsity < 1.0), "Not implemented"

opt_config = get_opt_config(args.model)
Expand Down