xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-09-09 08:27:03 +08:00
Code Issues Packages Projects Releases Wiki Activity

some docstring updates

Browse Source 
Signed-off-by: Vladimir Anisimov <vanisimov@nvidia.com>
This commit is contained in:
Vladimir Anisimov 2025-12-24 02:09:52 -08:00
parent b0be6298c4
commit 2817110aa3
1 changed files with 3 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/v1/core/kv_cache_utils.py
View File

@ -947,7 +947,7 @@ def _find_best_group_size(
overhead_threshold: float = 0.10) -> int: overhead_threshold: float = 0.10) -> int:
""" """
Find the optimal group size that minimizes padding memory, preferring Find the optimal group size that minimizes padding memory, preferring
larger group sizes (fewer tensors). larger group sizes.
For each layer type, padding = (group_size - count % group_size) % group_size For each layer type, padding = (group_size - count % group_size) % group_size
weighted by that layer's max_memory_usage_bytes. Different layer types weighted by that layer's max_memory_usage_bytes. Different layer types
@ -1000,7 +1000,7 @@ def _find_best_group_size(
def find_best_in_range(start: int, end: int) -> int: def find_best_in_range(start: int, end: int) -> int:
"""Find best group size in [start, end] range. """Find best group size in [start, end] range.
Prefers larger group sizes (fewer tensors) when padding is equal. Prefers larger group sizes when padding is equal.
Key: (padding_memory, -group_size) so larger group_size wins ties. Key: (padding_memory, -group_size) so larger group_size wins ties.
""" """
return min(range(start, end + 1), return min(range(start, end + 1),
@ -1110,7 +1110,7 @@ def _get_kv_cache_groups_uniform_page_size(
# (full.0, full.1), (sw.0, sw.2), (sw.1, padding). # (full.0, full.1), (sw.0, sw.2), (sw.1, padding).
# Find optimal group_size by trying all options and choosing the one with # Find optimal group_size by trying all options and choosing the one with
# minimal padding (weighted by layer memory size). Prefers larger group sizes # minimal padding (weighted by layer memory size). Prefers larger group sizes
# (fewer tensors) and enforces group_size >= 3 unless overhead exceeds 20%. # and enforces group_size >= 3 unless overhead exceeds the threshold.
group_size = _find_best_group_size(same_type_layers, vllm_config) group_size = _find_best_group_size(same_type_layers, vllm_config)
grouped_layers = [] grouped_layers = []
for layers in same_type_layers.values(): for layers in same_type_layers.values():