Skip to content

Use correct values to update encoder KV Cache for streaming models #15323

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

Closed
MahmoudAshraf97 wants to merge 3 commits into from
Closed

Use correct values to update encoder KV Cache for streaming models #15323

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
07f9734
use correct update indices
MahmoudAshraf97 Jan 27, 2026
6b0ee60
Merge branch 'main' into fix_cache
KunalDhawan Feb 7, 2026
7850301
Apply isort and black reformatting
KunalDhawan Feb 7, 2026
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
2 changes: 2 additions & 0 deletions nemo/collections/asr/modules/conformer_encoder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1026,8 +1026,10 @@ def setup_streaming_params(
if hasattr(m, "_max_cache_len"):
if isinstance(m, MultiHeadAttention):
m.cache_drop_size = streaming_cfg.cache_drop_size
m.valid_query_length = streaming_cfg.valid_out_len
if isinstance(m, CausalConv1D):
m.cache_drop_size = streaming_cfg.cache_drop_size
m.valid_query_length = streaming_cfg.valid_out_len

self.streaming_cfg = streaming_cfg

Expand Down
11 changes: 5 additions & 6 deletions nemo/collections/asr/parts/submodules/causal_convs.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,7 +62,8 @@ def __init__(
)

def forward(
self, x,
self,
x,
):
x = F.pad(x, pad=(self._left_padding, self._right_padding, self._left_padding, self._right_padding))
x = super().forward(x)
Expand Down Expand Up @@ -96,6 +97,7 @@ def __init__(
dtype=None,
) -> None:
self.cache_drop_size = None
self.valid_query_length = None
if padding is None:
self._left_padding = kernel_size - 1
self._right_padding = stride - 1
Expand Down Expand Up @@ -134,11 +136,8 @@ def update_cache(self, x, cache=None):
else:
new_x = F.pad(x, pad=(0, self._right_padding))
new_x = torch.cat([cache, new_x], dim=-1)
if self.cache_drop_size > 0:
next_cache = new_x[:, :, : -self.cache_drop_size]
else:
next_cache = new_x
next_cache = next_cache[:, :, -cache.size(-1) :]
cache_keep_size = self.valid_query_length - self.cache_drop_size
next_cache = torch.cat([cache, x[:, :, :cache_keep_size]], dim=-1)[:, :, -cache.size(-1) :]
return new_x, next_cache

def forward(self, x, cache=None):
Expand Down
3 changes: 2 additions & 1 deletion nemo/collections/asr/parts/submodules/multi_head_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -86,6 +86,7 @@ def __init__(
self.use_pytorch_sdpa_backends = use_pytorch_sdpa_backends

self.cache_drop_size = None
self.valid_query_length = None
self.use_bias = use_bias
self.dropout_rate = dropout_rate
assert n_feat % n_head == 0
Expand Down Expand Up @@ -204,7 +205,7 @@ def forward(self, query, key, value, mask, pos_emb=None, cache=None):
def update_cache(self, key, value, query, cache):
if cache is not None:
key = value = torch.cat([cache, key], dim=1)
q_keep_size = query.shape[1] - self.cache_drop_size
q_keep_size = self.valid_query_length - self.cache_drop_size
cache = torch.cat([cache[:, q_keep_size:, :], query[:, :q_keep_size, :]], dim=1)
return key, value, query, cache

Expand Down
Loading