[Attention] Flash Attention 3 - fp8 (#14570)

Signed-off-by: Mickael Seznec <mickael@mistral.ai>

vllm/envs.py

@@ -78,6 +78,7 @@ 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
     VLLM_SERVER_DEV_MODE: bool = False
@@ -524,13 +525,17 @@ environment_variables: dict[str, Callable[[], Any]] = {
     # Pad the fp8 weights to 256 bytes for ROCm
     "VLLM_ROCM_FP8_PADDING":
     lambda: bool(int(os.getenv("VLLM_ROCM_FP8_PADDING", "1"))),
+
+    # Divisor for dynamic query scale factor calculation for FP8 KV Cache
+    "Q_SCALE_CONSTANT":
+    lambda: int(os.getenv("Q_SCALE_CONSTANT", "200")),
     # 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"))),