Skip to content

Perform scale preshuffling in addition to weight preshuffling #2538

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 2 commits into
base: main
Choose a base branch
from
Open

Perform scale preshuffling in addition to weight preshuffling #2538

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
e855143
Preshuffle scales
KyleHerndon Oct 16, 2025
b3980ae
Add missing file
KyleHerndon Oct 16, 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
7 changes: 5 additions & 2 deletions sharktank/sharktank/examples/export_paged_llm_v1.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -280,9 +280,12 @@ def main():
llama_config.hp.rope_interleave_emb = False

# Override matmul_kernel if the weights were shuffled
if dataset.properties.get("use_shuffled_kernel", False):
shuffle_version = dataset.properties.get("use_shuffled_kernel", False)
if shuffle_version:
kernel_selection = f"sharktank.asm.shuffled;{llama_config.matmul_kernel}"
logger.debug(f"Using preshuffle kernel variant: {kernel_selection}")
logger.debug(
f"Using preshuffle kernel variant: {kernel_selection} (version={shuffle_version})"
)
llama_config.matmul_kernel = kernel_selection

hp = llama_config.hp
Expand Down
8 changes: 8 additions & 0 deletions sharktank/sharktank/kernels/assembly_binaries.py
View file
Open in desktop

Large diffs are not rendered by default.

100 changes: 89 additions & 11 deletions sharktank/sharktank/kernels/gemm_fp4_asm.py
View file
Open in desktop

Large diffs are not rendered by default.

11 changes: 8 additions & 3 deletions sharktank/sharktank/models/llama/tools/import_quark_dataset.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,7 +30,7 @@
_optional_int_prop,
_int_prop,
)
from sharktank.kernels.gemm_fp4_asm import shuffle_weight
from sharktank.kernels.gemm_fp4_asm import shuffle_weight, shuffle_scale


def _load_json(p: Path):
Expand Down Expand Up @@ -133,9 +133,10 @@ def create_fp4_block_tensor(

expected_shape = list(original_shape[:-1]) + [num_blocks, packed_block_size]

# Apply weight shuffling during preprocessing to avoid runtime shuffling (if enabled)
# Apply weight and scale shuffling during preprocessing to avoid runtime shuffling (if enabled)
if apply_shuffle:
weight_tensor = shuffle_weight(weight_tensor, layout=(16, 16))
scale_tensor = shuffle_scale(scale_tensor)
weight_tensor = weight_tensor.view(*expected_shape)

layout = BlockScaledFp4Layout(
Expand Down Expand Up @@ -514,7 +515,11 @@ def main(argv):
updated_properties = convert_hf_hparams_to_gguf(ds.properties)

# Store shuffle configuration for kernel selection
updated_properties["use_shuffled_kernel"] = args.apply_shuffle
# Version tracking: True (v1, weights only), "v2" (weights + scales), False (no preshuffle)
if args.apply_shuffle:
updated_properties["use_shuffled_kernel"] = "v2" # weights + scales preshuffled
else:
updated_properties["use_shuffled_kernel"] = False

head_count = (updated_properties["llama.attention.head_count"],)

Expand Down
Loading