Warn on small Mamba chunked prefill size (#30938)
Co-authored-by: Daniel Afrimi <dafrimi@aws-dfw-cs-001-login-01.cm.cluster>
This commit is contained in:
co-authored by
Daniel Afrimi
parent
0c29c8fece
commit
24da0e51b6
@@ -5250,6 +5250,19 @@ class ServerArgs:
|
|||||||
assert view.mamba_track_interval % view.page_size == 0
|
assert view.mamba_track_interval % view.page_size == 0
|
||||||
assert self.mamba_cache_chunk_size is not None
|
assert self.mamba_cache_chunk_size is not None
|
||||||
|
|
||||||
|
if (
|
||||||
|
view.chunked_prefill_size is not None
|
||||||
|
and 0 < view.chunked_prefill_size < self.mamba_cache_chunk_size
|
||||||
|
):
|
||||||
|
logger.warning(
|
||||||
|
"Mamba radix extra-buffer is enabled with chunked_prefill_size=%s "
|
||||||
|
"smaller than mamba_cache_chunk_size=%s. This can make "
|
||||||
|
"mamba_track_mask false for unfinished chunked-prefill handoff "
|
||||||
|
"and skip Mamba state checkpoints.",
|
||||||
|
view.chunked_prefill_size,
|
||||||
|
self.mamba_cache_chunk_size,
|
||||||
|
)
|
||||||
|
|
||||||
def _handle_mamba_radix_cache(self, model_arch: str):
|
def _handle_mamba_radix_cache(self, model_arch: str):
|
||||||
# Resolution moved to the resolution pipeline (arg_groups/overrides.py:
|
# Resolution moved to the resolution pipeline (arg_groups/overrides.py:
|
||||||
# _mamba_radix_cache_resolution), invoked here at each legacy call
|
# _mamba_radix_cache_resolution), invoked here at each legacy call
|
||||||
|
|||||||
Reference in New Issue
Block a user