Skip to content

mla: split latent_scale identity/dynamic compile-cache entries #52

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

Closed
MadeBy561 wants to merge 1 commit into from
Closed

mla: split latent_scale identity/dynamic compile-cache entries #52

Show file tree
Hide file tree
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
9 changes: 5 additions & 4 deletions b12x/attention/mla/kernel.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2201,10 +2201,11 @@ def _sparse_mla_decode_grid_flat_launch(
)
compile_spec = KernelCompileSpec.from_fields(
"attention.mla.sm120.decode",
# v17 adds the runtime FP8-RoPE record format specialization. The
# explicit cache key does not hash source, so this ABI bump rejects
# cubins compiled for the 432-byte BF16-RoPE record.
17,
# v18: latent_scale == 1.0 is folded out at trace time, producing a
# cubin without the outer-scale multiply; the identity and dynamic
# variants must not share a cache entry.
18,
key_field("latent_scale_identity", int(float(latent_scale) == 1.0)),
*spec_fields,
)
if per_token_len:
Expand Down
8 changes: 5 additions & 3 deletions b12x/attention/mla/prefill_mg.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3766,9 +3766,11 @@ def _sparse_mla_prefill_mg_flat_launch(
)
compile_spec = KernelCompileSpec.from_fields(
"attention.mla.sm120.prefill_mg",
# v3 adds the FP8-RoPE record specialization/read path. The explicit
# cache key does not hash source, so do not reuse v2 KLD-only cubins.
3,
# v4: latent_scale == 1.0 is folded out at trace time, producing a
# cubin without the outer-scale multiply; the identity and dynamic
# variants must not share a cache entry.
4,
key_field("latent_scale_identity", int(float(latent_scale) == 1.0)),
*spec_fields,
)
entry = kernel.call_dual if has_extra else kernel
Expand Down