[KV Sharing] Raise error if using eagle with fast prefill (#24350)

Signed-off-by: Yong Hoon Shin <yhshin@meta.com>

vllm/config/__init__.py

@@ -3665,6 +3665,24 @@ class VllmConfig:
                 " Disabling `torch.compile`.")
             self.compilation_config.level = CompilationLevel.NO_COMPILATION
 
+        if self.cache_config.kv_sharing_fast_prefill:
+            if not envs.VLLM_USE_V1:
+                raise NotImplementedError(
+                    "Fast prefill optimization for KV sharing is not supported "
+                    "in V0 currently.")
+
+            if self.speculative_config is not None and \
+                self.speculative_config.use_eagle():
+                raise NotImplementedError(
+                    "Fast prefill optimization for KV sharing is not "
+                    "compatible with EAGLE as EAGLE requires correct logits "
+                    "for all tokens while fast prefill gives incorrect logits "
+                    "for prompt tokens.")
+
+            logger.warning_once(
+                "--kv-sharing-fast-prefill requires changes on model side for "
+                "correctness and to realize prefill savings. ")
+
         if ((not envs.VLLM_USE_V1) and self.lora_config is not None
                 and self.compilation_config.level
                 != CompilationLevel.NO_COMPILATION):

vllm/config/cache.py

@@ -145,19 +145,12 @@ class CacheConfig:
 
         self._verify_cache_dtype()
         self._verify_prefix_caching()
-        self._verify_kv_sharing_fast_prefill()
 
     def metrics_info(self):
         # convert cache_config to dict(key: str, value: str) for prometheus
         # metrics info
         return {key: str(value) for key, value in self.__dict__.items()}
 
-    def _verify_kv_sharing_fast_prefill(self) -> None:
-        if self.kv_sharing_fast_prefill and not envs.VLLM_USE_V1:
-            raise NotImplementedError(
-                "Fast prefill optimization for KV sharing is not supported "
-                "in V0 currently.")
-
     @model_validator(mode='after')
     def _verify_args(self) -> Self:
         if self.cpu_offload_gb < 0: