[WIP] Before refactor the nanovllm sparse policy.

2026-01-19 22:34:44 +08:00
parent b5da802dff
commit b97b0b96a0
8 changed files with 475 additions and 837 deletions
--- a/nanovllm/kvcache/sparse/init.py
+++ b/nanovllm/kvcache/sparse/init.py
@@ -61,8 +61,6 @@ def create_sparse_policy(policy_type: SparsePolicyType, **kwargs) -> SparsePolic
            block_size=kwargs.get("block_size", 128),
            samples_per_chunk=kwargs.get("samples_per_chunk", 128),
            threshold=kwargs.get("threshold", 0.9),
-            use_triton=kwargs.get("use_triton", True),
-            stride=kwargs.get("stride", 8),
        )

    else: