zijie-tian/nano-vllm
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[WIP] Before refactor the compute)_chunked_prefill.

Browse Source
This commit is contained in:
Zijie Tian
2026-01-23 03:36:12 +08:00
parent edc006463b
commit ca32ea6f93
7 changed files with 914 additions and 114 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
nanovllm/config.py
View File

@@ -48,7 +48,7 @@ class Config:
# XAttention BSA specific parameters
sparse_block_size: int = 128 # Block size for BSA (tokens per block)
sparse_samples_per_chunk: int = 128 # Samples per chunk for estimation
sparse_threshold: float = 0.9 # Cumulative attention threshold (0-1)
sparse_threshold: float = 0.95 # Cumulative attention threshold (tau in XAttention)
sparse_use_triton: bool = True # Use Triton kernels for estimation
sparse_stride: int = 8 # Stride for Q/K downsampling