Skip to content

Separate q,k,v projection in sliding attention layers #3376

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
patrickji2014 merged 3 commits into from
Aug 13, 2026
Merged

Separate q,k,v projection in sliding attention layers #3376

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
c0693df
separate q,k,v projection in sliding attention layers
patrickji2014 Aug 11, 2026
451b533
add comment for tracking qkv_proj removal
patrickji2014 Aug 12, 2026
56278c7
Merge branch 'main' into patrickji.qkv_weight_unmerge
patrickji2014 Aug 12, 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
83 changes: 36 additions & 47 deletions tpu_inference/models/jax/gemma4.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,8 +33,7 @@
from tpu_inference.layers.jax import JaxModule
from tpu_inference.layers.jax.embed import JaxEmbed
from tpu_inference.layers.jax.linear import (JaxEinsum, JaxLinear, JaxLmHead,
JaxMergedColumnParallelLinear,
JaxQKVParallelLinear)
JaxMergedColumnParallelLinear)
from tpu_inference.layers.jax.moe.moe import JaxRoutedExperts
from tpu_inference.layers.jax.norm import JaxRmsNorm
from tpu_inference.layers.jax.pp_utils import PPMissingLayer, make_layers
Expand Down Expand Up @@ -309,23 +308,36 @@ def __init__(self,
None) if _shard_kv_on_k else (None, None, "model")
_kv_bias_spec = ("model", None) if _shard_kv_on_k else (None, "model")

self.q_proj = JaxEinsum(
"TD,DNH->TNH",
(self.hidden_size, self.num_heads, self.head_dim),
bias_shape=(self.num_heads,
self.head_dim) if config.attention_bias else None,
param_dtype=dtype,
kernel_init=nnx.with_partitioning(init_fn, (None, "model", None)),
bias_init=nnx.with_partitioning(init_fn, ("model", None))
if config.attention_bias else None,
rngs=rng,
quant_config=quant_config,
prefix=prefix + ".q_proj",
)
self.k_proj = JaxEinsum(
"TD,DKH->TKH",
(self.hidden_size, self.num_kv_heads, self.head_dim),
bias_shape=(self.num_kv_heads,
self.head_dim) if config.attention_bias else None,
param_dtype=dtype,
kernel_init=nnx.with_partitioning(init_fn, _kv_kernel_spec),
bias_init=nnx.with_partitioning(init_fn, _kv_bias_spec)
if config.attention_bias else None,
rngs=rng,
quant_config=quant_config,
prefix=prefix + ".k_proj",
)
if use_k_eq_v: # TODO: Add QKV fusion logic for k == v case.
self.qkv_proj = None
self.q_proj = JaxEinsum(
"TD,DNH->TNH",
(self.hidden_size, self.num_heads, self.head_dim),
bias_shape=(self.num_heads,
self.head_dim) if config.attention_bias else None,
param_dtype=dtype,
kernel_init=nnx.with_partitioning(init_fn,
(None, "model", None)),
bias_init=nnx.with_partitioning(init_fn, ("model", None))
if config.attention_bias else None,
rngs=rng,
quant_config=quant_config,
prefix=prefix + ".q_proj",
)
self.k_proj = JaxEinsum(
self.v_proj = None
else:
self.v_proj = JaxEinsum(
"TD,DKH->TKH",
(self.hidden_size, self.num_kv_heads, self.head_dim),
bias_shape=(self.num_kv_heads,
Expand All @@ -336,24 +348,8 @@ def __init__(self,
if config.attention_bias else None,
rngs=rng,
quant_config=quant_config,
prefix=prefix + ".k_proj",
prefix=prefix + ".v_proj",
)
self.v_proj = None
else:
self.qkv_proj = JaxQKVParallelLinear(
hidden_size=self.hidden_size,
num_heads=self.num_heads,
num_kv_heads=self.num_kv_heads,
head_dim=self.head_dim,
use_bias=config.attention_bias,
dtype=dtype,
rngs=rng,
quant_config=quant_config,
prefix=prefix,
)
self.q_proj = None
self.k_proj = None
self.v_proj = None

self.q_norm = JaxRmsNorm(
self.head_dim,
Expand Down Expand Up @@ -428,13 +424,10 @@ def __call__(
attention_metadata: AttentionMetadata,
) -> Tuple[jax.Array, jax.Array]:
md = attention_metadata
if self.qkv_proj is not None:
q, k, v = self.qkv_proj(x)
else:
k = self.k_proj(x)
v = k
# q: (T, N, H)
q = self.q_proj(x)
k = self.k_proj(x)
v = self.v_proj(x) if self.v_proj is not None else k
# q: (T, N, H)
q = self.q_proj(x)
# Q norm (always applied)
q = self.q_norm(q)

Expand Down Expand Up @@ -1008,12 +1001,8 @@ def __call__(


class Gemma4ForCausalLM(JaxModule, LoadableWithIterator):
# qkv_proj packing is removed in PR 3376 for performance gain
packed_modules_mapping = {
"qkv_proj": [
Comment thread
patrickji2014 marked this conversation as resolved.
"q_proj",
"k_proj",
"v_proj",
],
"gate_up_proj": [
"gate_proj",
"up_proj",
Expand Down
Loading