Skip to content

[Transforms] Use get_head_dim util #1918

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

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
+6 −22
Open

[Transforms] Use get_head_dim util #1918

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
28 changes: 6 additions & 22 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 @@ -9,7 +9,7 @@
TransformScheme,
apply_transform_config,
)
from compressed_tensors.utils import TorchDtype
from compressed_tensors.utils import TorchDtype, get_head_dim
from pydantic import Field, ValidationInfo, field_validator
from transformers import PreTrainedModel

Expand Down Expand Up @@ -126,16 +126,17 @@ def on_initialize(self, state: State, **kwargs) -> bool:

self.mappings = infer_mapping_from_model(state.model)
self.norm_mappings = infer_norm_mapping_from_model(state.model)
head_dim = get_head_dim(state.model.config)

config_groups = {}
if SpinquantRotation.R1 in self.rotations:
config_groups["R1"] = self._create_r1_scheme()

if SpinquantRotation.R2 in self.rotations:
config_groups["R2"] = self._create_r2_scheme(state.model)
config_groups["R2"] = self._create_r2_scheme(head_dim)

if SpinquantRotation.R3 in self.rotations:
config_groups["R3"] = self._create_r3_scheme()
config_groups["R3"] = self._create_r3_scheme(head_dim)

if SpinquantRotation.R4 in self.rotations:
config_groups["R4"] = self._create_r4_scheme()
Expand Down Expand Up @@ -217,24 +218,7 @@ def _create_r1_scheme(self) -> TransformScheme:
],
)

def _create_r2_scheme(self, model: PreTrainedModel) -> TransformScheme:
config = model.config

if hasattr(config, "head_dim"):
head_dim = config.head_dim
elif hasattr(config, "hidden_size") and hasattr(config, "num_attention_heads"):
head_dim = config.hidden_size // config.num_attention_heads
else:
raise NotImplementedError()

if self.transform_block_size:
if head_dim % self.transform_block_size != 0:
raise ValueError(
f"transform_block_size {self.transform_block_size} must be set "
f"such that model's head_dim {head_dim} is evenly divisible by it"
)
head_dim = self.transform_block_size

def _create_r2_scheme(self, head_dim: int) -> TransformScheme:
return TransformScheme(
type=self.transform_type,
randomize=self.randomize,
Expand All @@ -251,7 +235,7 @@ def _create_r2_scheme(self, model: PreTrainedModel) -> TransformScheme:
],
)

def _create_r3_scheme(self) -> TransformScheme:
def _create_r3_scheme(self, head_dim: int) -> TransformScheme:
raise NotImplementedError(
"SpinQuant R3 rotations will be added in a future release"
)
Expand Down
Loading