update Q/K/V_SCALE_CONSTANT

vllm/envs.py

@@ -118,9 +118,9 @@ if TYPE_CHECKING:
     VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
     VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
     VLLM_DISABLE_COMPILE_CACHE: bool = False
-    Q_SCALE_CONSTANT: int = 200
-    K_SCALE_CONSTANT: int = 200
-    V_SCALE_CONSTANT: int = 100
+    Q_SCALE_CONSTANT: int = 10
+    K_SCALE_CONSTANT: int = 10
+    V_SCALE_CONSTANT: int = 10
     VLLM_SERVER_DEV_MODE: bool = False
     VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: int = 128
     VLLM_MLA_DISABLE: bool = False
@@ -1049,13 +1049,13 @@ environment_variables: dict[str, Callable[[], Any]] = {
 
     # Divisor for dynamic query scale factor calculation for FP8 KV Cache
     "Q_SCALE_CONSTANT":
-    lambda: int(os.getenv("Q_SCALE_CONSTANT", "200")),
+    lambda: int(os.getenv("Q_SCALE_CONSTANT", "10")),
     # Divisor for dynamic key scale factor calculation for FP8 KV Cache
     "K_SCALE_CONSTANT":
-    lambda: int(os.getenv("K_SCALE_CONSTANT", "200")),
+    lambda: int(os.getenv("K_SCALE_CONSTANT", "10")),
     # Divisor for dynamic value scale factor calculation for FP8 KV Cache
     "V_SCALE_CONSTANT":
-    lambda: int(os.getenv("V_SCALE_CONSTANT", "100")),
+    lambda: int(os.getenv("V_SCALE_CONSTANT", "10")),
 
     # If set, enable multiprocessing in LLM for the V1 code path.
     "VLLM_ENABLE_V1_MULTIPROCESSING":