[WIP] need refactor.

2026-01-22 22:20:34 +08:00
parent 69b779e252
commit 5fb0f67295
11 changed files with 514 additions and 548 deletions
--- a/nanovllm/config.py
+++ b/nanovllm/config.py
@@ -62,6 +62,7 @@ class Config:
    xattn_keep_sink: bool = False  # Always keep first block (sink tokens)
    xattn_keep_recent: bool = False  # Always keep recent diagonal blocks
    xattn_norm: float = 1.0  # Normalization factor for attention scores
+    xattn_use_bsa: bool = True  # Use Block Sparse Attention library (requires installation)

    def __post_init__(self):
        assert os.path.isdir(self.model)