✨ feat: add configurable stride and chunk_size for XAttention BSA
- Add sparse_chunk_size config option (default: 16384) - Pass stride, chunk_size, use_triton through factory function - Add --sparse-stride CLI option to test_ruler.py Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This commit is contained in:
@@ -61,6 +61,9 @@ 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),
|
||||
stride=kwargs.get("stride", 8),
|
||||
chunk_size=kwargs.get("chunk_size", 16384),
|
||||
use_triton=kwargs.get("use_triton", True),
|
||||
)
|
||||
|
||||
else:
|
||||
|
||||
Reference in New Issue
Block a user