[Feature] add feature mla_ag_after_qlora for dsv3.2 (#19428)

Co-authored-by: JiaruiChang5268 <changjiarui1@huawei.com>
2026-03-02 20:00:31 +08:00
parent 3f36f27eae
commit b3718982a1
5 changed files with 101 additions and 9 deletions
@@ -312,6 +312,8 @@ class Envs:
    SGLANG_NPU_FORWARD_NATIVE_GELUTANH = EnvBool(False)
    # Forward native implementation for gemma rms norm for model Skywork-Reward-Gemma-2-27B-v0.2
    SGLANG_NPU_FORWARD_NATIVE_GEMMA_RMS_NORM = EnvBool(False)
+    # Delay all-gather after qlora for better performance for Deepseek v3.2
+    SGLANG_USE_AG_AFTER_QLORA = EnvBool(False)

    # Quantization
    SGLANG_INT4_WEIGHT = EnvBool(False)