[Quantization][FP8] Add support for FP8 models with input_scale for output projection and QK quantization (#15734)

Signed-off-by: Randall Smith <Randall.Smith@amd.com>
Signed-off-by: Luka Govedič <lgovedic@redhat.com>
Co-authored-by: Luka Govedič <lgovedic@redhat.com>

vllm/engine/arg_utils.py

@@ -1368,6 +1368,23 @@ class EngineArgs:
                                recommend_to_remove=False)
             return False
 
+        if current_platform.is_rocm():
+            from vllm.model_executor.layers.quantization.fp8 import Fp8Config
+            load_config = self.create_load_config()
+            quantization_config = VllmConfig.get_quantization_config(
+                model_config, load_config)
+            if isinstance(quantization_config, Fp8Config):
+                _raise_or_fallback(feature_name="fp8 for ROCm",
+                                   recommend_to_remove=False)
+                return False
+            from vllm.model_executor.layers.quantization.quark.quark import (
+                QuarkConfig)
+
+            if isinstance(quantization_config, QuarkConfig
+                          ) and quantization_config.has_fp8_layer_weights():
+                _raise_or_fallback(feature_name="Quark fp8 for ROCm",
+                                   recommend_to_remove=False)
+
         # No Fp8 KV cache so far.
         if self.kv_cache_dtype != "auto":
             fp8_attention = self.kv_cache_dtype.startswith("fp8")