[FP8][Kernel] Dynamic kv cache scaling factors computation (#11906)

Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>
Co-authored-by: Micah Williamson <micah.williamson@amd.com>

vllm/envs.py

@@ -73,6 +73,8 @@ if TYPE_CHECKING:
     VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
     VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
     VLLM_DISABLE_COMPILE_CACHE: bool = False
+    K_SCALE_CONSTANT: int = 200
+    V_SCALE_CONSTANT: int = 100
     VLLM_SERVER_DEV_MODE: bool = False
     VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: int = 128
 
@@ -474,6 +476,13 @@ environment_variables: Dict[str, Callable[[], Any]] = {
     "VLLM_USE_V1":
     lambda: bool(int(os.getenv("VLLM_USE_V1", "0"))),
 
+    # Divisor for dynamic key scale factor calculation for FP8 KV Cache
+    "K_SCALE_CONSTANT":
+    lambda: int(os.getenv("K_SCALE_CONSTANT", "200")),
+
+    # Divisor for dynamic value scale factor calculation for FP8 KV Cache
+    "V_SCALE_CONSTANT":
+    lambda: int(os.getenv("V_SCALE_CONSTANT", "100")),
     # If set, enable multiprocessing in LLM for the V1 code path.
     "VLLM_ENABLE_V1_MULTIPROCESSING":
     lambda: bool(int(os.getenv("VLLM_ENABLE_V1_MULTIPROCESSING", "1"))),