Skip to content

Adds attention mask with model.forward(..., cache=None). #89

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
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Adds attention mask with model.forward(..., cache=None). #89

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: 11 additions & 6 deletions mistral/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,6 +14,7 @@
from mistral.moe import MoeArgs, MoeLayer

from xformers.ops.fmha import memory_efficient_attention
from xformers.ops.fmha.attn_bias import BlockDiagonalCausalMask


@dataclass
Expand Down Expand Up @@ -80,6 +81,7 @@ def forward(
x: torch.Tensor,
freqs_cis: torch.Tensor,
cache: Optional[CacheView],
mask: Optional[torch.Tensor]=None,
) -> torch.Tensor:
seqlen_sum, _ = x.shape

Expand Down Expand Up @@ -109,9 +111,9 @@ def forward(

# xformers requires (B=1, S, H, D)
xq, key, val = xq[None, ...], key[None, ...], val[None, ...]
output = memory_efficient_attention(
xq, key, val, None if cache is None else cache.mask
)
if mask is None and cache is not None:
mask = cache.mask
output = memory_efficient_attention(xq, key, val, mask)

return self.wo(output.view(seqlen_sum, self.n_heads * self.head_dim))

Expand Down Expand Up @@ -163,9 +165,10 @@ def __init__(self, args: ModelArgs):
self.feed_forward = FeedForward(args=args)

def forward(
self, x: torch.Tensor, freqs_cis: torch.Tensor, cache: Optional[CacheView]
self, x: torch.Tensor, freqs_cis: torch.Tensor, cache: Optional[CacheView],
mask: Optional[torch.Tensor]=None,
) -> torch.Tensor:
r = self.attention.forward(self.attention_norm(x), freqs_cis, cache)
r = self.attention.forward(self.attention_norm(x), freqs_cis, cache, mask)
h = x + r
r = self.feed_forward.forward(self.ffn_norm(h))
out = h + r
Expand Down Expand Up @@ -251,8 +254,10 @@ def forward_partial(
assert sum(seqlens) == num_toks, (sum(seqlens), num_toks)
if cache is not None:
input_metadata = cache.get_input_metadata(seqlens)
mask = None
else:
input_metadata = SimpleInputMetadata.from_seqlens(seqlens, self.device)
mask = BlockDiagonalCausalMask.from_seqlens(seqlens).make_local_attention(max(seqlens))

if self.pipeline_rank == 0:
assert self.tok_embeddings is not None
Expand All @@ -271,7 +276,7 @@ def forward_partial(
cache_view = cache.get_view(local_layer_id, input_metadata)
else:
cache_view = None
h = layer(h, freqs_cis, cache_view)
h = layer(h, freqs_cis, cache_view, mask)

if cache is not None:
cache.update_seqlens(seqlens)
Expand Down