Fix: Safe RoPE Cache Expansion to Prevent Position-ID Out-of-Bounds in EAGLE + Long-Sequence Workloads (#11871)

2025-11-25 01:19:06 +08:00
parent a146f833f1
commit 98b38de3f2
4 changed files with 102 additions and 0 deletions
@@ -277,6 +277,11 @@ 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_SPEC_V2 = EnvBool(False)
    SGLANG_ENABLE_OVERLAP_PLAN_STREAM = EnvBool(False)