Skip to content

Assert quant_min/quant_max in embedding4bit #7410

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
merged 2 commits into from
Jan 28, 2025
Merged

Assert quant_min/quant_max in embedding4bit #7410

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
d0ca255
init
metascroy Dec 20, 2024
a007d75
format fix
metascroy 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
6 changes: 3 additions & 3 deletions examples/models/llama/source_transformation/quantize.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -729,18 +729,18 @@ def __init__(
def forward(self, indices: torch.Tensor) -> torch.Tensor:
if not self.packed: # 8bit
return torch.ops.quantized_decomposed.embedding_byte.dtype(
self.weight, self.scales, None, 0, 0, indices, dtype=self.dtype
self.weight, self.scales, None, -128, 127, indices, dtype=self.dtype
)
else: # packed
if self.bitwidth == 2:
return torch.ops.quantized_decomposed.embedding_2bit.dtype(
self.weight, self.scales, None, 0, 0, indices, dtype=self.dtype
self.weight, self.scales, None, -2, 1, indices, dtype=self.dtype
)

# Remaining case (always return to make pyre happy)
assert self.bitwidth == 4
return torch.ops.quantized_decomposed.embedding_4bit.dtype(
self.weight, self.scales, None, 0, 0, indices, dtype=self.dtype
self.weight, self.scales, None, -8, 7, indices, dtype=self.dtype
)


Expand Down
28 changes: 28 additions & 0 deletions exir/passes/_quant_patterns_and_replacements.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -202,6 +202,13 @@ def embedding_2bit(
weight_quant_max: int,
indices: torch.Tensor,
) -> torch.Tensor:
assert (
weight_quant_min == -2
), "embedding_2bit in ExecuTorch expects weight_quant_min == -2"
assert (
weight_quant_max == 1
), "embedding_2bit in ExecuTorch expects weight_quant_max == 1"

embedding_weight_checks(weight, weight_scales, weight_zero_points)
group_size = (4 * weight.size(1)) // (
weight_scales.size(1) if weight_scales.dim() == 2 else 1
Expand Down Expand Up @@ -257,6 +264,13 @@ def embedding_2bit_dtype(
indices: torch.Tensor,
dtype: Optional[torch.dtype],
) -> torch.Tensor:
assert (
weight_quant_min == -2
), "embedding_2bit_dtype in ExecuTorch expects weight_quant_min == -2"
assert (
weight_quant_max == 1
), "embedding_2bit_dtype in ExecuTorch expects weight_quant_max == 1"

embedding_weight_checks(weight, weight_scales, weight_zero_points)
group_size = (4 * weight.size(1)) // (
weight_scales.size(1) if weight_scales.dim() == 2 else 1
Expand Down Expand Up @@ -334,6 +348,13 @@ def embedding_4bit(
weight_quant_max: int,
indices: torch.Tensor,
) -> torch.Tensor:
assert (
weight_quant_min == -8
), "embedding_4bit in ExecuTorch expects weight_quant_min == -8"
assert (
weight_quant_max == 7
), "embedding_4bit in ExecuTorch expects weight_quant_max == 7"

embedding_weight_checks(weight, weight_scales, weight_zero_points)
group_size = (2 * weight.size(1)) // (
weight_scales.size(1) if weight_scales.dim() == 2 else 1
Expand Down Expand Up @@ -387,6 +408,13 @@ def embedding_4bit_dtype(
indices: torch.Tensor,
dtype: Optional[torch.dtype],
) -> torch.Tensor:
assert (
weight_quant_min == -8
), "embedding_4bit_dtype in ExecuTorch expects weight_quant_min == -8"
assert (
weight_quant_max == 7
), "embedding_4bit_dtype in ExecuTorch expects weight_quant_max == 7"

embedding_weight_checks(weight, weight_scales, weight_zero_points)
group_size = (2 * weight.size(1)) // (
weight_scales.size(1) if weight_scales.dim() == 2 else 1
Expand Down
Loading