Skip to content

Compress the embedding table to SFP at load time #993

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
Mikyx-1 wants to merge 4 commits into
base: dev
Choose a base branch
from
Open

Compress the embedding table to SFP at load time #993

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
1e2fe4d
Compress the embedding table to SFP at load time
Mikyx-1 Aug 23, 2026
4b49a8f
Use Highway scalar conversions in weights test
Mikyx-1 Aug 28, 2026
a075d4e
Vectorize the embedding max-absolute scan
Mikyx-1 Aug 28, 2026
f3407af
Use Highway namespace alias in weights
Mikyx-1 Aug 28, 2026
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
22 changes: 21 additions & 1 deletion BUILD.bazel
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -349,7 +349,10 @@ cc_library(
cc_library(
name = "weights",
srcs = ["gemma/weights.cc"],
hdrs = ["gemma/weights.h"],
hdrs = [
"gemma/weights.h",
"gemma/weights_internal.h",
],
deps = [
":configs",
":gemma_args",
Expand Down Expand Up @@ -379,6 +382,23 @@ cc_test(
],
)

cc_test(
name = "weights_test",
srcs = ["gemma/weights_test.cc"],
deps = [
":gemma_args",
":mat",
":threading_context",
":weights",
"//compression:compress",
"//compression:types",
"//io",
"//io:blob_store",
"@googletest//:gtest_main", # buildcleaner: keep
"@highway//:hwy",
],
)

# For building all tests in one command, so we can test several.
test_suite(
name = "ops_tests",
Expand Down
2 changes: 2 additions & 0 deletions CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -206,6 +206,7 @@ set(SOURCES
gemma/vit.h
gemma/weights.cc
gemma/weights.h
gemma/weights_internal.h
io/blob_store.cc
io/blob_store.h
io/fields.cc
Expand Down Expand Up @@ -383,6 +384,7 @@ set(GEMMA_TEST_FILES
deepseek/deepseek_test.cc
gemma/gemma_args_test.cc
gemma/tensor_info_test.cc
gemma/weights_test.cc
io/blob_store_test.cc
io/fields_test.cc
ops/bench_matmul.cc
Expand Down
8 changes: 8 additions & 0 deletions gemma/gemma_args.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,6 +53,7 @@ struct LoaderArgs : public ArgsBase<LoaderArgs> {
Path weights; // weights file location
Tristate map;
Tristate to_bf16;
Tristate sfp_embedding;
Tristate wrapping;

template <class Visitor>
Expand All @@ -65,6 +66,13 @@ struct LoaderArgs : public ArgsBase<LoaderArgs> {
"Enable memory-mapping? -1 = auto, 0 = no, 1 = yes.");
visitor(to_bf16, "to_bf16", Tristate::kDefault,
"Convert weights to bf16? -1 = auto, 0 = no, 1 = yes.");
visitor(sfp_embedding, "sfp_embedding", Tristate::kDefault,
"Compress the embedding table to SFP while loading? This halves\n"
" its footprint and, for models with tied input/output\n"
" embeddings, the bandwidth of the per-token logits MatMul,\n"
" in exchange for 8-bit precision on that tensor. Enabling this\n"
" disables automatic mapping; explicit --map=1 still wins.\n"
" -1 = auto (currently off), 0 = no, 1 = yes.");
visitor(wrapping, "wrapping", Tristate::kDefault,
"Enable prompt wrapping? Specify 0 for pre-2025 format PT models.");
}
Expand Down
Loading
Loading