Skip to content

fix(dsv4): zero inactive prefill MoE tails #744

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
high-cloud wants to merge 1 commit into
base: main
Choose a base branch
from
+33 −0
Open

fix(dsv4): zero inactive prefill MoE tails #744

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
33 changes: 33 additions & 0 deletions models/deepseek/v4/prefill_fwd.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -127,6 +127,34 @@
# the final RMSNorm gamma — mirrors decode_fwd.
HC_HEAD_NAMES = ["hc_head_fn", "hc_head_scale", "hc_head_base"]
FINAL_NORM_NAMES = ["final_norm_w"]
TAIL_D_CHUNK = 512
TAIL_D_BLOCKS = D // TAIL_D_CHUNK
assert D % TAIL_D_CHUNK == 0


@pl.jit.inline
def _zero_hc_tail(
x_hc: pl.Tensor[[T, HC_MULT, D], pl.BF16],
num_tokens: pl.Scalar[pl.INT32],
) -> pl.Tensor[[T, HC_MULT, D], pl.BF16]:
active_tokens = pl.cast(num_tokens, pl.INDEX)
if active_tokens < 0:
active_tokens = pl.cast(0, pl.INDEX)
if active_tokens > T:
active_tokens = pl.cast(T, pl.INDEX)

for block in pl.spmd(T * HC_MULT, name_hint="prefill_zero_hc_tail"):
if block == 0:
x_hc[0:1, 0:1, 0:16] = x_hc[0:1, 0:1, 0:16]

@gemini-code-assist gemini-code-assist Bot Jul 10, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

This no-op assignment appears to be a compiler workaround, likely to act as a keep-alive for x_hc and ensure it's materialized before the loop. For improved maintainability, please add a comment explaining its purpose.

Suggested change
x_hc[0:1, 0:1, 0:16] = x_hc[0:1, 0:1, 0:16]
# Keep-alive for x_hc to ensure it's materialized before use in the loop.
x_hc[0:1, 0:1, 0:16] = x_hc[0:1, 0:1, 0:16]

t = block // HC_MULT
h = block % HC_MULT
if t >= active_tokens:
for db in pl.pipeline(TAIL_D_BLOCKS, stage=2):
d0 = db * TAIL_D_CHUNK
x_hc[t:t + 1, h:h + 1, d0:d0 + TAIL_D_CHUNK] = pl.full(
[1, 1, TAIL_D_CHUNK], dtype=pl.BF16, value=0.0
)
return x_hc

# Per-FWD-layer stacked weights (sliced by the FWD layer index 0..42).
FWD_LAYER_STACKED_NAMES = [
Expand Down Expand Up @@ -358,6 +386,7 @@ def prefill_fwd(
attn_sink_l0, wo_a_l0, wo_b_l0, wo_b_scale_l0,
x_attn0, nt,
)
x_attn0 = _zero_hc_tail(x_attn0, nt)
with pl.scope():
moe(
x_attn0,
Expand Down Expand Up @@ -420,6 +449,7 @@ def prefill_fwd(
attn_sink_l1, wo_a_l1, wo_b_l1, wo_b_scale_l1,
x_attn1, nt,
)
x_attn1 = _zero_hc_tail(x_attn1, nt)
with pl.scope():
moe(
x_attn1,
Expand Down Expand Up @@ -517,6 +547,7 @@ def prefill_fwd(
attn_sink_csa, wo_a_csa, wo_b_csa, wo_b_scale_csa,
x_attn_csa, nt,
)
x_attn_csa = _zero_hc_tail(x_attn_csa, nt)
with pl.scope():
moe(
x_attn_csa,
Expand Down Expand Up @@ -589,6 +620,7 @@ def prefill_fwd(
attn_sink_hca, wo_a_hca, wo_b_hca, wo_b_scale_hca,
x_attn_hca, nt,
)
x_attn_hca = _zero_hc_tail(x_attn_hca, nt)
with pl.scope():
moe(
x_attn_hca,
Expand Down Expand Up @@ -682,6 +714,7 @@ def prefill_fwd(
attn_sink_last, wo_a_last, wo_b_last, wo_b_scale_last,
x_attn_last, nt,
)
x_attn_last = _zero_hc_tail(x_attn_last, nt)
with pl.scope():
moe(
x_attn_last,
Expand Down
Loading