fix: add ENFORCE_EAGER env var to skip torch.compile on ROCm

torch._dynamo.exc.Unsupported crashes EngineCore during graph tracing
of LlamaDecoderLayer on gfx1151.  ENFORCE_EAGER=true bypasses
torch.compile and CUDA graph capture entirely.

ray_serve/serve_llm.py

@@ -54,6 +54,7 @@ class LLMDeployment:
         self.enable_chunked_prefill = os.environ.get("ENABLE_CHUNKED_PREFILL", "true").lower() == "true"
         self.num_speculative_tokens = int(os.environ.get("NUM_SPECULATIVE_TOKENS", "0"))
         self.ngram_prompt_lookup_max = int(os.environ.get("NGRAM_PROMPT_LOOKUP_MAX", "0"))
+        self.enforce_eager = os.environ.get("ENFORCE_EAGER", "false").lower() == "true"
 
         engine_kwargs: dict[str, Any] = {
             "model": self.model_id,
@@ -62,6 +63,7 @@ class LLMDeployment:
             "trust_remote_code": True,
             "enable_prefix_caching": self.enable_prefix_caching,
             "enable_chunked_prefill": self.enable_chunked_prefill,
+            "enforce_eager": self.enforce_eager,
         }
 
         # n-gram speculative decoding (no draft model needed)
@@ -75,6 +77,7 @@ class LLMDeployment:
 
         print(f"Prefix caching: {self.enable_prefix_caching}")
         print(f"Chunked prefill: {self.enable_chunked_prefill}")
+        print(f"Enforce eager (no torch.compile): {self.enforce_eager}")
 
         engine_args = AsyncEngineArgs(**engine_kwargs)
         self.engine = AsyncLLMEngine.from_engine_args(engine_args)
@@ -102,6 +105,7 @@ class LLMDeployment:
                     "enable_chunked_prefill": str(self.enable_chunked_prefill),
                     "num_speculative_tokens": str(self.num_speculative_tokens),
                     "ngram_prompt_lookup_max": str(self.ngram_prompt_lookup_max),
+                    "enforce_eager": str(self.enforce_eager),
                 }
             )
         except ImportError: