Skip to content

resolve huge memory usage of MX data type #859

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

Closed
wants to merge 1 commit into from
+8 −2
Closed

resolve huge memory usage of MX data type #859

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
10 changes: 8 additions & 2 deletions auto_round/data_type/mxfp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -77,6 +77,12 @@ def quant_element(tensor, ebits, mbits, max_norm, mantissa_rounding="even"):
return tensor


def quant_element_ste(tensor, ebits, mbits, max_norm, mantissa_rounding="even"):
with torch.no_grad():
tensor_q = quant_element(tensor, ebits, mbits, max_norm, mantissa_rounding)
return (tensor_q - tensor).detach() + tensor


def quant_mx(tensor, bits=4, group_size=-1, v=0, max_scale=1.0, mantissa_rounding="even", data_type="mx_fp", **kwargs):
"""Quantize the given tensor using the specified parameters.

Expand Down Expand Up @@ -120,7 +126,7 @@ def quant_mx(tensor, bits=4, group_size=-1, v=0, max_scale=1.0, mantissa_roundin
scale = torch.pow(2, shared_exp)
tensor = tensor / scale + v
tensor = torch.clamp(tensor, min=-max_norm, max=max_norm)
tensor = quant_element(tensor, ebits, mbits, max_norm, mantissa_rounding)
tensor = quant_element_ste(tensor, ebits, mbits, max_norm, mantissa_rounding)

tensor = tensor * scale
tensor = revert_tensor_by_pad(tensor, orig_shape=orig_shape, pad_len=pad_len)
Expand Down Expand Up @@ -171,7 +177,7 @@ def quant_mx_rceil(
scale = torch.pow(2, shared_exp)
tensor = tensor / scale + v
tensor = torch.clamp(tensor, min=-max_norm, max=max_norm)
tensor = quant_element(tensor, ebits, mbits, max_norm, mantissa_rounding)
tensor = quant_element_ste(tensor, ebits, mbits, max_norm, mantissa_rounding)

tensor = tensor * scale
tensor = revert_tensor_by_pad(tensor, orig_shape=orig_shape, pad_len=pad_len)
Expand Down