Skip to content

[Feature] Support MLA dummy load & optimize MLA value padding memory #7775

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
Jiang-Jia-Jun merged 1 commit into from
May 11, 2026
+19 −1
Merged

[Feature] Support MLA dummy load & optimize MLA value padding memory #7775

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
17 changes: 17 additions & 0 deletions fastdeploy/model_executor/layers/linear.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1010,6 +1010,23 @@ def __init__(
# Override weight keys to use the combined kv_b_proj
self.weight_key = f"{prefix}.weight" # e.g., "kv_b_proj.weight"

if self.fd_config.load_config.load_choices == "dummy":
# Create K projection weight
self.k_b_proj_weight = self.create_parameter(
shape=[self.num_heads_per_partition, qk_nope_head_dim, kv_lora_rank],
dtype=self.weight_dtype,
is_bias=False,
default_initializer=paddle.nn.initializer.Constant(0),
)

# Create V projection weight
self.v_b_proj_weight = self.create_parameter(
shape=[self.num_heads_per_partition, kv_lora_rank, v_head_dim],
dtype=self.weight_dtype,
is_bias=False,
default_initializer=paddle.nn.initializer.Constant(0),
)

def process_weights_after_loading(self):
if self.fd_config.load_config.dynamic_load_weight:
return
Expand Down
3 changes: 2 additions & 1 deletion fastdeploy/model_executor/models/deepseek_v3.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -420,7 +420,8 @@ def forward(
key = paddle.empty([full_k_pe.shape[0], self.num_attention_heads_tp, self.qk_head_dim], dtype=query.dtype)
key[..., : self.qk_nope_head_dim] = key_nope
key[..., self.qk_nope_head_dim :] = full_k_pe.unsqueeze(1)
value = paddle.nn.functional.pad(value, [0, self.qk_head_dim - self.v_head_dim], value=0)
if self.qk_head_dim - self.v_head_dim != 0:
value = paddle.nn.functional.pad(value, [0, self.qk_head_dim - self.v_head_dim], value=0)

fmha_out = self.mla_attn(
q=query,
Expand Down
Loading