Add llm_methods metadata to multimethod PTE export

extension/llm/export/config/llm_config.py

@@ -310,11 +310,15 @@ class MethodConfig:
             torch.export tracing. Controls which graph path is captured, e.g.
             prefill vs decode, or for YOCO, where all layers run for decode
             but not prefill. When unset, uses the model's default input length.
+        phase: Optional inference phase tag ("prefill" or "decode"). When set,
+            the method name is recorded in llm_methods metadata so the runtime
+            can pick the correct method for each inference phase.
     """
 
     method_name: str
     lora_config: Optional[LoraConfig] = None
     export_seq_len: Optional[int] = None
+    phase: Optional[str] = None
 
 
 @dataclass

extension/llm/runner/constants.h

@@ -19,6 +19,10 @@ inline constexpr auto kVocabSize = "get_vocab_size";
 inline constexpr auto kUseKVCache = "use_kv_cache";
 inline constexpr auto kUseSDPAWithKVCache = "use_sdpa_with_kv_cache";
 
+// LLM multimethod phase metadata (e.g. YOCO prefill/decode)
+inline constexpr auto kLlmMethodsPrefill = "llm_methods_prefill";
+inline constexpr auto kLlmMethodsDecode = "llm_methods_decode";
+
 // Multimodal method name conventions
 inline constexpr auto kVisionEncoderMethod = "vision_encoder";
 inline constexpr auto kAudioEncoderMethod = "audio_encoder";