[Misc] Load FP8 kv-cache scaling factors from checkpoints (#4893)

The 2nd PR for #4532.

This PR supports loading FP8 kv-cache scaling factors from a FP8 checkpoint (with .kv_scale parameter).

vllm/engine/arg_utils.py

@@ -191,12 +191,11 @@ class EngineArgs:
         parser.add_argument(
             '--kv-cache-dtype',
             type=str,
-            choices=['auto', 'fp8'],
+            choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
             default=EngineArgs.kv_cache_dtype,
             help='Data type for kv cache storage. If "auto", will use model '
-            'data type. FP8_E5M2 (without scaling) is only supported on cuda '
-            'version greater than 11.8. On ROCm (AMD GPU), FP8_E4M3 is instead '
-            'supported for common inference criteria.')
+            'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
+            'ROCm (AMD GPU) supports fp8 (=fp8_e4m3)')
         parser.add_argument(
             '--quantization-param-path',
             type=nullable_str,