Skip to content

fix scaling factor for DeepSeek #521

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

Merged
merged 2 commits into from
Aug 21, 2025
Merged

fix scaling factor for DeepSeek #521

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
005f6e7
fix scaling factor for DeepSseek
bzgoogle Aug 20, 2025
06312be
move scale to __post_init__
bzgoogle Aug 21, 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
12 changes: 10 additions & 2 deletions tpu_commons/models/jax/common/attention/deepseek_v3_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,4 @@
import math
from dataclasses import dataclass
from typing import Any, Tuple

Expand Down Expand Up @@ -63,6 +64,7 @@ class MLA(nnx.Module):
attention_chunk_size: int | None = None
rope_input_ordering: str = "split"
quant: Any | None = None
rope_mscale_all_dim: float = 1.0

def __post_init__(self):
self.N = self.num_attention_heads
Expand All @@ -73,6 +75,13 @@ def __post_init__(self):

assert self.N == self.K, "N and K must be equal for MLA"

if self.rope_scaling["factor"] <= 1.0:
yarn_mscale = 1.0
else:
yarn_mscale = 0.1 * self.rope_mscale_all_dim * math.log(
self.rope_scaling["factor"]) + 1.0
self.scale = self.qk_head_dim**-0.5 * yarn_mscale**2

self.rope = DeepseekScalingRotaryEmbedding(
self.qk_rope_head_dim,
self.rope_theta,
Expand Down Expand Up @@ -180,7 +189,6 @@ def __call__(self,
# Concatenate the nope and rope queries.
q_TNH = jnp.concatenate([q_nope_TNH, q_rope_TNH], axis=-1)
# Multiple the query by scaling factor
q_TNH = q_TNH * self.qk_head_dim**-0.5
q_TNH = nnx.with_sharding_constraint(q_TNH, self.query_tnh)

with jax.named_scope("kv_proj"):
Expand Down Expand Up @@ -293,7 +301,7 @@ def attention(
def _ragged_paged_attention(*args):
return ragged_paged_attention(
*args,
sm_scale=q_TNH.shape[-1]**-0.5,
sm_scale=self.scale,
)

output_TNH, kv_cache = jax.jit(
Expand Down