Fix extra_buffer_lazy guard bypass (#31942)

This commit is contained in:
Ke Bao
2026-07-21 20:41:35 +08:00
committed by GitHub
parent e4eea7ce2f
commit 0a06cc5317
+5 -7
View File
@@ -4822,14 +4822,12 @@ class ServerArgs:
"extra_buffer_lazy unsupported under PD disaggregation; use " "extra_buffer_lazy unsupported under PD disaggregation; use "
"--mamba-radix-cache-strategy extra_buffer." "--mamba-radix-cache-strategy extra_buffer."
) )
if view.speculative_num_draft_tokens is not None: algo = (view.speculative_algorithm or "").upper()
if view.mamba_radix_cache_strategy == "extra_buffer_lazy": assert algo not in ("DFLASH", "DSPARK"), (
# The dflash family's verify bypasses prepare_mamba_track_for_verify. f"extra_buffer_lazy unsupported with {view.speculative_algorithm}; "
assert view.speculative_algorithm not in ("DFLASH", "DSPARK"), ( "use --mamba-radix-cache-strategy extra_buffer."
f"extra_buffer_lazy unsupported with "
f"{view.speculative_algorithm}; use "
"--mamba-radix-cache-strategy extra_buffer."
) )
if view.speculative_num_draft_tokens is not None:
assert view.mamba_track_interval >= view.speculative_num_draft_tokens assert view.mamba_track_interval >= view.speculative_num_draft_tokens
if view.page_size is not None: if view.page_size is not None:
assert view.mamba_track_interval % view.page_size == 0 assert view.mamba_track_interval % view.page_size == 0