[V32/GLM5] Control the threshold of applying dense attention with an environ (#20062)

2026-03-09 14:36:10 -07:00
parent d39ed074cf
commit be63f982b7
6 changed files with 32 additions and 59 deletions
@@ -378,8 +378,7 @@ class Envs:
    SGLANG_NSA_FUSE_TOPK = EnvBool(True)
    SGLANG_NSA_ENABLE_MTP_PRECOMPUTE_METADATA = EnvBool(True)
    SGLANG_USE_FUSED_METADATA_COPY = EnvBool(True)
-    SGLANG_VERIFY_FUSED_METADATA_COPY = EnvBool(False)
-    SGLANG_NSA_FORCE_MLA = EnvBool(False)
+    SGLANG_NSA_PREFILL_DENSE_ATTN_KV_LEN_THRESHOLD = EnvInt(2048)

    # sgl-kernel
    SGLANG_SKIP_SGL_KERNEL_VERSION_CHECK = EnvBool(False)