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

Fix cuda hardcode for inference woq #5565

Merged
merged 3 commits into from
Jun 5, 2024
Merged

Fix cuda hardcode for inference woq #5565

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
90dc71b
fix cuda hardcode
Liangliang-Ma May 24, 2024
c9c7ba5
update rest cuda hardcode
Liangliang-Ma May 27, 2024
9c58fc9
Merge branch 'master' into xpu_quant
loadams May 28, 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
18 changes: 9 additions & 9 deletions deepspeed/inference/quantization/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,14 +14,14 @@

device = get_accelerator().device_name() if get_accelerator().is_available() else 'cpu'

quantizer_cuda_module = None
quantizer_module = None


def get_quantizer_cuda_module():
global quantizer_cuda_module
if quantizer_cuda_module is None:
quantizer_cuda_module = deepspeed.ops.op_builder.QuantizerBuilder().load()
return quantizer_cuda_module
def get_quantizer_module():
global quantizer_module
if quantizer_module is None:
quantizer_module = deepspeed.ops.op_builder.QuantizerBuilder().load()
return quantizer_module


def tensor_clamp(tensor: Tensor, min, max) -> Tensor:
Expand Down Expand Up @@ -107,19 +107,19 @@ def dequantize(self, tensor: Tensor, quant_scale: Tensor, quant_min: Tensor) ->
if self.config['group_size'] % 8 == 0 and \
(self.config['num_bits'] == 4 or self.config['num_bits'] == 8) and \
self.config['group_dim'] == len(tensor.shape) - 1 and \
self.dtype == torch.float16 and device == 'cuda':
self.dtype == torch.float16 and device == get_accelerator().device_name():
loadams marked this conversation as resolved.
Show resolved Hide resolved

last_dimension_size = self.config['group_size']
if self.config['num_bits'] == 4:
last_dimension_size = last_dimension_size // 2
quantized_tensor = get_quantizer_cuda_module().dequantize_int4_to_half_experimental(
quantized_tensor = get_quantizer_module().dequantize_int4_to_half_experimental(
tensor.reshape(-1, last_dimension_size), quant_scale, quant_min,
tensor.numel() // last_dimension_size, self.config['group_size'])
shape = list(tensor.shape)
shape[-1] = shape[-1] * 2
elif self.config['num_bits'] == 8:
# last_dimension_size = last_dimension_size // 2
quantized_tensor = get_quantizer_cuda_module().dequantize_int8_to_half_experimental(
quantized_tensor = get_quantizer_module().dequantize_int8_to_half_experimental(
tensor.reshape(-1, last_dimension_size), quant_scale, quant_min,
tensor.numel() // last_dimension_size, self.config['group_size'])
shape = list(tensor.shape)
Expand Down
Loading