Fix: Dynamic RoPE Cache Expansion to Prevent Position-ID Out-of-Bounds in EAGLE + Long-Sequence Workloads (#10788)
This commit is contained in:
@@ -222,6 +222,11 @@ class Envs:
|
||||
SGLANG_TRITON_PREFILL_TRUNCATION_ALIGN_SIZE = EnvInt(4096)
|
||||
SGLANG_TRITON_DECODE_SPLIT_TILE_SIZE = EnvInt(256)
|
||||
|
||||
# RoPE cache configuration
|
||||
SGLANG_SPEC_EXPANSION_SAFETY_FACTOR = EnvInt(2)
|
||||
SGLANG_ROPE_CACHE_SAFETY_MARGIN = EnvInt(256)
|
||||
SGLANG_ROPE_CACHE_ALIGN = EnvInt(128)
|
||||
|
||||
# Overlap Spec V2
|
||||
SGLANG_ENABLE_OVERLAP_PLAN_STREAM = EnvBool(False)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user