Skip to content

[Transforms] Update examples for R4 and transform_block_size option #1870

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
brian-dellabetta merged 5 commits into from
Sep 30, 2025
Merged

[Transforms] Update examples for R4 and transform_block_size option #1870

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
083e687
update transforms examples with new options
brian-dellabetta Sep 24, 2025
fb73c0d
switch block_size back to head_dim
brian-dellabetta Sep 26, 2025
6077d6b
examples comment update
brian-dellabetta Sep 26, 2025
075ac5f
Merge branch 'main' into bdellabe/transforms-updates
brian-dellabetta Sep 26, 2025
1ed7409
Merge branch 'main' into bdellabe/transforms-updates
brian-dellabetta Sep 30, 2025
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
4 changes: 3 additions & 1 deletion examples/transform/quip_example.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,7 +21,9 @@
# * apply quip transforms to model in order to make quantization easier
# * quantize the weights to 4 bit with a group size 128
recipe = [
QuIPModifier(rotations=["v", "u"], transform_type="random-hadamard"),
QuIPModifier(
rotations=["v", "u"], transform_block_size=128, transform_type="random-hadamard"
),
QuantizationModifier(targets="Linear", scheme="W4A16", ignore=["lm_head"]),
]

Expand Down
11 changes: 6 additions & 5 deletions examples/transform/spinquant_example.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,14 +11,15 @@
model = AutoModelForCausalLM.from_pretrained(MODEL_ID, torch_dtype="auto")
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

# NOTE: currently only fused rotations (R1 & R2) are available
# Learned rotations and online rotations (R3 & R4) will be added
# in a future release.
# NOTE: currently only rotations R1, R2, and R4 are available
# R3 and learned R1/R2 rotations will be added in a future release.
# Configure the quantization algorithm to run.
# * apply spinquant transforms to model to reduce quantization loss
# * quantize the weights to 4 bit with group size 128
recipe = [
SpinQuantModifier(rotations=["R1", "R2"], transform_type="hadamard"),
SpinQuantModifier(
rotations=["R1", "R2", "R4"], transform_block_size=64, transform_type="hadamard"
),
QuantizationModifier(targets="Linear", scheme="W4A16", ignore=["lm_head"]),
]

Expand All @@ -37,6 +38,6 @@
print("==========================================\n\n")

# Save to disk compressed.
SAVE_DIR = MODEL_ID.split("/")[1] + "-spinquantR1R2-w4a16"
SAVE_DIR = MODEL_ID.split("/")[1] + "-spinquantR1R2R4-w4a16"
model.save_pretrained(SAVE_DIR, save_compressed=True)
tokenizer.save_pretrained(SAVE_DIR)
4 changes: 2 additions & 2 deletions src/llmcompressor/modifiers/transform/quip/base.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -135,7 +135,7 @@ def _create_config(self) -> TransformConfig:
def _create_v_scheme(self) -> TransformScheme:
return TransformScheme(
type=self.transform_type,
block_size=self.transform_block_size,
head_dim=self.transform_block_size,
apply=[
TransformArgs(
targets=self.targets,
Expand All @@ -157,7 +157,7 @@ def _create_v_scheme(self) -> TransformScheme:
def _create_u_scheme(self) -> TransformScheme:
return TransformScheme(
type=self.transform_type,
block_size=self.transform_block_size,
head_dim=self.transform_block_size,
apply=[
TransformArgs(
targets=self.targets,
Expand Down
6 changes: 3 additions & 3 deletions src/llmcompressor/modifiers/transform/spinquant/base.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -193,7 +193,7 @@ def _create_r1_scheme(self) -> TransformScheme:
randomize=self.randomize,
requires_grad=self.learnable,
precision=self.precision,
block_size=self.transform_block_size,
head_dim=self.transform_block_size,
apply=[
TransformArgs(
targets=[
Expand Down Expand Up @@ -240,7 +240,7 @@ def _create_r2_scheme(self, model: PreTrainedModel) -> TransformScheme:
randomize=self.randomize,
requires_grad=self.learnable,
precision=self.precision,
block_size=head_dim,
head_dim=head_dim,
apply=[
TransformArgs(targets=[self.mappings.attn_v], location="weight_output"),
TransformArgs(
Expand All @@ -262,7 +262,7 @@ def _create_r4_scheme(self) -> TransformScheme:
randomize=self.randomize,
requires_grad=self.learnable,
precision=self.precision,
block_size=self.transform_block_size,
head_dim=self.transform_block_size,
apply=[
TransformArgs(
targets=[*self.mappings.mlp_out],
Expand Down