[Fix] Land the decode mamba checkpoint depth on the tree page under DCP (#35412)
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> Co-authored-by: Ke Bao <ispobaoke@gmail.com>
This commit is contained in:
co-authored by
Claude Opus 5
Ke Bao
parent
308bc1228b
commit
eac91ac362
@@ -1490,6 +1490,14 @@ def mamba_checkpoint_grid(tree_page: int) -> int:
|
||||
return math.lcm(mamba_cache_chunk_size(), tree_page)
|
||||
|
||||
|
||||
def mamba_track_grid(tree_page: int) -> int:
|
||||
"""The same granularity for a decode-donated checkpoint, which additionally
|
||||
has to land on the requested ``mamba_track_interval``."""
|
||||
return math.lcm(
|
||||
mamba_checkpoint_grid(tree_page), get_exec().mamba.mamba_track_interval
|
||||
)
|
||||
|
||||
|
||||
def max_speculative_num_draft_tokens() -> int | None:
|
||||
"""The largest draft-token count speculative decoding may use.
|
||||
|
||||
|
||||
Reference in New Issue
Block a user