Skip to content

[Models] Optimize DeepSeek V3 MLA projection layout #8083

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
chang-wenbin wants to merge 1 commit into
base: develop
Choose a base branch
from
+37 −31
Open

[Models] Optimize DeepSeek V3 MLA projection layout #8083

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
24 changes: 24 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 @@ -1129,6 +1129,30 @@ def forward_k_b(self, x: paddle.Tensor) -> paddle.Tensor:
out = paddle.bmm(x, self.k_b_proj_weight)
return out

def forward_k_b_thd(self, x: paddle.Tensor) -> paddle.Tensor:
"""
Forward K_b projection for token-head-dim layout.

Args:
x: Input tensor with shape [tokens, heads, qk_nope_head_dim]

Returns:
K_b projection output with shape [tokens, heads, kv_lora_rank]
"""
return paddle.einsum("thd,hdr->thr", x, self.k_b_proj_weight)

def forward_v_b_htr(self, x: paddle.Tensor) -> paddle.Tensor:
"""
Forward V_b projection for token-head-rank layout.

Args:
x: Input tensor with shape [tokens, heads, kv_lora_rank]

Returns:
V_b projection output with shape [tokens, heads, v_head_dim]
"""
return paddle.einsum("thr,hrv->thv", x, self.v_b_proj_weight)

def forward_v_b(self, x: paddle.Tensor) -> paddle.Tensor:
"""
Forward pass for V_b projection using bmm
Expand Down
44 changes: 13 additions & 31 deletions fastdeploy/model_executor/models/deepseek_v3.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -454,8 +454,7 @@ def forward_swa_static(
key_pe: paddle.Tensor,
):
"""MLA static attention with sliding window indexer."""
q_nope_out = self.kv_b_proj_bmm(query_nope.transpose([1, 0, 2]), proj_type="k").transpose([1, 0, 2])

q_nope_out = self.kv_b_proj_bmm.forward_k_b_thd(query_nope)
q_input = paddle.concat([q_nope_out, query_pe], axis=-1)
q_input.reshape_(
[
Expand Down Expand Up @@ -488,13 +487,10 @@ def forward_swa_static(
attn_softmax_scale=self.attn_softmax_scale,
)

fmqa_out = fmqa_out.reshape_([-1, self.num_attention_heads_tp, self.kv_lora_rank]).transpose([1, 0, 2])
fmqa_out = fmqa_out.reshape_([-1, self.num_attention_heads_tp, self.kv_lora_rank])

return (
self.kv_b_proj_bmm(fmqa_out, proj_type="v")
.transpose([1, 0, 2])
.reshape_([-1, self.num_attention_heads_tp * self.v_head_dim])
)
fmqa_out = self.kv_b_proj_bmm.forward_v_b_htr(fmqa_out)
return fmqa_out.reshape_([-1, self.num_attention_heads_tp * self.v_head_dim])

def forward(
self,
Expand Down Expand Up @@ -578,8 +574,6 @@ def forward(
)
key[..., : self.qk_nope_head_dim] = key_nope
key[..., self.qk_nope_head_dim :] = full_k_pe.unsqueeze(1)
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 All @@ -591,8 +585,6 @@ def forward(
forward_meta=forward_meta,
)

fmha_out.reshape_([-1, self.num_attention_heads_tp, self.qk_head_dim])
fmha_out = fmha_out[:, :, : self.v_head_dim]
fmha_out.reshape_([-1, self.num_attention_heads_tp * self.v_head_dim])
attn_out = fmha_out

Expand Down Expand Up @@ -627,7 +619,7 @@ def forward(
query_nope = decoder_query_nope.reshape([0, -1, self.qk_nope_head_dim])
query_pe = decoder_query_pe.reshape([0, -1, self.qk_rope_head_dim])

q_nope_out = self.kv_b_proj_bmm(query_nope.transpose([1, 0, 2]), proj_type="k").transpose([1, 0, 2])
q_nope_out = self.kv_b_proj_bmm.forward_k_b_thd(query_nope)

q_input = paddle.concat([q_nope_out, query_pe], axis=-1)
q_input.reshape_(
Expand All @@ -647,13 +639,9 @@ def forward(
forward_meta=forward_meta,
)

fmqa_out = fmqa_out.reshape_([-1, self.num_attention_heads_tp, self.kv_lora_rank]).transpose([1, 0, 2])

fmqa_out = (
self.kv_b_proj_bmm(fmqa_out, proj_type="v")
.transpose([1, 0, 2])
.reshape_([-1, self.num_attention_heads_tp * self.v_head_dim])
)
fmqa_out = fmqa_out.reshape_([-1, self.num_attention_heads_tp, self.kv_lora_rank])
fmqa_out = self.kv_b_proj_bmm.forward_v_b_htr(fmqa_out)
fmqa_out = fmqa_out.reshape_([-1, self.num_attention_heads_tp * self.v_head_dim])

if int(os.getenv("USE_FLASH_MLA", "0")) == 0 and self.prop.major == 9:
pass
Expand Down Expand Up @@ -1036,8 +1024,8 @@ def forward(
query_nope, query_pe = query.split([self.qk_nope_head_dim, self.qk_rope_head_dim], axis=-1)

query_pe, key_pe = self.rotary_emb(forward_meta.position_ids, query_pe, key_pe)
q_nope_out = self.kv_b_proj_bmm(query_nope.transpose([1, 0, 2]).contiguous(), proj_type="k")
q_input = paddle.concat([q_nope_out.transpose([1, 0, 2]).contiguous(), query_pe], axis=-1)
q_nope_out = self.kv_b_proj_bmm.forward_k_b_thd(query_nope)
q_input = paddle.concat([q_nope_out, query_pe], axis=-1)

compressed_kv = self.kv_a_layernorm(compressed_kv)[0]
# kv = paddle.concat([compressed_kv, key_pe.squeeze(1)], axis=-1)
Expand All @@ -1053,15 +1041,9 @@ def forward(
forward_meta=forward_meta,
)

fmha_out = fmha_out.reshape_([-1, self.num_attention_heads_tp, self.kv_lora_rank]).transpose([1, 0, 2])
fmha_out = (
self.kv_b_proj_bmm(
fmha_out,
proj_type="v",
)
.transpose([1, 0, 2])
.reshape_([-1, self.num_attention_heads_tp * self.v_head_dim])
)
fmha_out = fmha_out.reshape_([-1, self.num_attention_heads_tp, self.kv_lora_rank])
fmha_out = self.kv_b_proj_bmm.forward_v_b_htr(fmha_out)
fmha_out = fmha_out.reshape_([-1, self.num_attention_heads_tp * self.v_head_dim])

output = self.o_proj(fmha_out)

Expand Down
Loading