Add bfloat16 KV cache validation for HiSparse (#22505)

This commit is contained in:
Mohammad Miadh Angkad
2026-04-13 12:41:42 +08:00
committed by GitHub
parent fae0a2fc3c
commit 4dbd59850b
+6
View File
@@ -6565,6 +6565,12 @@ class ServerArgs:
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 (
self.schedule_conservativeness >= 0
), "schedule_conservativeness must be non-negative"