Skip to content

More fixes to Wave decode #12

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
Hardcode84 merged 5 commits into from
May 13, 2025
Merged

More fixes to Wave decode #12

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
d151bb5
view truncate
Hardcode84 May 13, 2025
68e61bc
shapes
Hardcode84 May 13, 2025
9d6d8d8
shapes
Hardcode84 May 13, 2025
a9d2d32
fix
Hardcode84 May 13, 2025
5822bff
remove assert
Hardcode84 May 13, 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
6 changes: 4 additions & 2 deletions python/sglang/srt/layers/attention/wave_backend.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -326,12 +326,14 @@ def init_cuda_graph_state(
self, max_bs: int, kv_indices_buf: Optional[torch.Tensor] = None
):
self.cuda_graph_attn_logits = torch.zeros(
(max_bs, self.num_head, self.max_kv_splits, self.v_head_dim),
(self.max_kv_splits, max_bs, self.v_head_dim, self.num_head),
# (max_bs, self.num_head, self.max_kv_splits, self.v_head_dim),
dtype=torch.float32,
device=self.device,
)
self.cuda_graph_attn_lse = torch.zeros(
(max_bs, self.num_head, self.max_kv_splits),
(self.max_kv_splits, max_bs, self.num_head),
# (max_bs, self.num_head, self.max_kv_splits),
dtype=torch.float32,
device=self.device,
)
Expand Down
11 changes: 8 additions & 3 deletions python/sglang/srt/layers/attention/wave_ops/decode_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -643,6 +643,11 @@ def decode_attention_fwd_grouped(
_decode_softmax_reducev_fwd(attn_logits, q, o, v_buffer, b_seq_len, num_kv_splits)


def view_trunc(tensor, shape):
size = math.prod(shape)
return tensor.view(-1)[:size].view(shape)


def decode_attention_wave(
q,
k_buffer,
Expand Down Expand Up @@ -673,8 +678,8 @@ def decode_attention_wave(
seq_len,
)

k_buffer = k_buffer.view(num_seqs, seq_len, num_kv_heads, head_size)
v_buffer = v_buffer.view(num_seqs, seq_len, num_kv_heads, head_size_kv)
k_buffer = view_trunc(k_buffer, (num_seqs, seq_len, num_kv_heads, head_size))
v_buffer = view_trunc(v_buffer, (num_seqs, seq_len, num_kv_heads, head_size_kv))

# Get the kernels (either compile or load from cache).
if mha:
Expand Down Expand Up @@ -765,7 +770,7 @@ def decode_attention_fwd(
sm_scale,
logit_cap=0.0,
):
assert max_kv_splits == attn_logits.shape[2]
# assert max_kv_splits == attn_logits.shape[2]
decode_attention_wave(
q,
k_buffer,
Expand Down
Loading