Skip to content

Fix attention kernel of GQA use case #78

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
merged 1 commit into from
May 10, 2024
Merged

Fix attention kernel of GQA use case #78

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
16 changes: 8 additions & 8 deletions jetstream_pt/layers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -138,14 +138,14 @@ def __call__(self, xq, xk, xv, mask, cache):
"""
Args:
xq: torch.Tensor of (batch size, num_heads, seqlen, head_dim)
xk: torch.Tensor of (batch size, num_heads, seqlen, head_dim)
xv: torch.Tensor of (batch size, num_heads, seqlen, head_dim)
xk: torch.Tensor of (batch size, num_kv_heads, seqlen, head_dim)
xv: torch.Tensor of (batch size, num_kv_heads, seqlen, head_dim)
mask: mask with 0 and -inf, or None
cache: CacheManagerInterface object
"""
bsz, num_heads, seqlen, head_dim = xq.shape
_, _, _, kv_head_dim = xk.shape
n_rep = head_dim // kv_head_dim
_, num_kv_heads, _, kv_head_dim = xk.shape
n_rep = num_heads // num_kv_heads
if seqlen == 1:
xq = torch.broadcast_to(xq, (xq.shape[0], xq.shape[1], 2, xq.shape[3]))

Expand Down Expand Up @@ -191,14 +191,14 @@ def __call__(self, xq, xk, xv, mask, cache):
"""
Args:
xq: torch.Tensor of (batch size, num_heads, seqlen, head_dim)
xk: torch.Tensor of (batch size, num_heads, seqlen, head_dim)
xv: torch.Tensor of (batch size, num_heads, seqlen, head_dim)
xk: torch.Tensor of (batch size, num_kv_heads, seqlen, head_dim)
xv: torch.Tensor of (batch size, num_kv_heads, seqlen, head_dim)
mask: mask with 0 and -inf, or None
cache: CacheManagerInterface object
"""
bsz, num_heads, seqlen, head_dim = xq.shape
_, _, _, kv_head_dim = xk.shape
n_rep = head_dim // kv_head_dim
_, num_kv_heads, _, kv_head_dim = xk.shape
n_rep = num_heads // num_kv_heads
if seqlen == 1:
xq = torch.broadcast_to(xq, (xq.shape[0], xq.shape[1], 2, xq.shape[3]))

Expand Down