Skip to content

remove contiguous before rope #20

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
blzheng merged 1 commit into from
Jan 7, 2026
Merged

remove contiguous before rope #20

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
8 changes: 4 additions & 4 deletions python/sglang/srt/layers/attention/vision.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -671,9 +671,9 @@ def forward(
q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)

# [b, s, embed_dim] --> [b * s, head, head_size]
q = q.reshape(bsz * s, head, -1).contiguous()
k = k.reshape(bsz * s, kv_head, -1).contiguous()
v = v.reshape(bsz * s, kv_head, -1).contiguous()
q = q.reshape(bsz * s, head, -1)
k = k.reshape(bsz * s, kv_head, -1)
v = v.reshape(bsz * s, kv_head, -1)
else:
# [b, s, embed_dim] --> [s, b, embed_dim]
x = rearrange(x, "b s ... -> s b ...")
Expand All @@ -692,7 +692,7 @@ def forward(

# [s, b, head, head_size] --> [b, s, head, head_size]
q, k, v = [
rearrange(x, "s b ... -> b s ...").contiguous() for x in (q, k, v)
rearrange(x, "s b ... -> b s ...") for x in (q, k, v)
]

if position_embeddings is not None:
Expand Down
4 changes: 2 additions & 2 deletions sgl-kernel/csrc/cpu/rope.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -622,8 +622,8 @@ std::tuple<at::Tensor, at::Tensor> rotary_embedding_cpu(
std::tuple<at::Tensor, at::Tensor>
apply_rotary_pos_emb_cpu(at::Tensor& query, at::Tensor& key, at::Tensor& cos, at::Tensor& sin) {
RECORD_FUNCTION("sgl-kernel::apply_rotary_pos_emb_cpu", std::vector<c10::IValue>({query, key}));
CHECK_INPUT(query);
CHECK_INPUT(key);
CHECK_LAST_DIM_CONTIGUOUS_INPUT(query);
CHECK_LAST_DIM_CONTIGUOUS_INPUT(key);
CHECK_INPUT(cos);
CHECK_INPUT(sin);
CHECK_DIM(3, query);
Expand Down
10 changes: 5 additions & 5 deletions test/srt/cpu/test_rope.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -253,15 +253,15 @@ def test_apply_rotary_pos_emb(self):
num_tokens = 1024
num_heads = 8
head_size = 72
query = torch.randn(num_tokens, num_heads, head_size).to(torch.bfloat16)
key = torch.randn(num_tokens, num_heads, head_size).to(torch.bfloat16)
qkv = torch.randn(num_tokens, num_heads * head_size * 3).to(torch.bfloat16)
query, key, _ = qkv.split([num_heads * head_size, num_heads * head_size, num_heads * head_size], dim=-1)
query = query.view(num_tokens, num_heads, head_size)
key = key.view(num_tokens, num_heads, head_size)
cos = torch.rand(num_tokens, head_size).to(torch.float32)
sin = torch.rand(num_tokens, head_size).to(torch.float32)
query_clone = query.clone()
key_clone = key.clone()
q_out_ref, k_out_ref = apply_rotary_pos_emb_native(query, key, cos, sin)
q_out_sgl, k_out_sgl = torch.ops.sgl_kernel.apply_rotary_pos_emb_cpu(
query_clone, key_clone, cos, sin
query, key, cos, sin
)
torch.testing.assert_close(q_out_ref, q_out_sgl, atol=1e-2, rtol=1e-2)
torch.testing.assert_close(k_out_ref, k_out_sgl, atol=1e-2, rtol=1e-2)
Expand Down
Loading