Add bfloat16 KV cache validation for HiSparse (#22505)
This commit is contained in:
@@ -6565,6 +6565,12 @@ class ServerArgs:
|
|||||||
f"Please use --nsa-{label}-backend=flashmla_sparse or omit it."
|
f"Please use --nsa-{label}-backend=flashmla_sparse or omit it."
|
||||||
)
|
)
|
||||||
|
|
||||||
|
if self.kv_cache_dtype != "bfloat16":
|
||||||
|
raise ValueError(
|
||||||
|
f"HiSparse requires bfloat16 KV cache, but got --kv-cache-dtype={self.kv_cache_dtype}. "
|
||||||
|
f"Please use --kv-cache-dtype=bfloat16."
|
||||||
|
)
|
||||||
|
|
||||||
assert (
|
assert (
|
||||||
self.schedule_conservativeness >= 0
|
self.schedule_conservativeness >= 0
|
||||||
), "schedule_conservativeness must be non-negative"
|
), "schedule_conservativeness must be non-negative"
|
||||||
|
|||||||
Reference in New Issue
Block a user