fix(hi3): fail fast when the AR KV cache cannot be built

unirl/models/hunyuan_image3/ar.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import sys
 from dataclasses import dataclass
 from dataclasses import field as dc_field
 from typing import Any, Dict, List, Optional, Tuple
@@ -99,8 +100,13 @@ def init_state(
         batch_size = int(input_ids.shape[0])
 
         prompt_len = int(input_ids.shape[1])
-        past_kv_initial = self._build_kv_cache(
-            transformer, batch_size=batch_size, max_cache_len=prompt_len + int(max_new_tokens)
+        cache_cls = sys.modules[type(transformer).__module__].HunyuanStaticCache
+        past_kv_initial = cache_cls(
+            config=transformer.config,
+            batch_size=batch_size,
+            max_cache_len=prompt_len + int(max_new_tokens),
+            dtype=torch.bfloat16,
+            dynamic=True,
         )
 
         cond_vit = conditions.cond_vit
@@ -214,29 +220,6 @@ def step(
 
         return token_id, log_prob, state
 
-    @staticmethod
-    def _build_kv_cache(transformer, *, batch_size: int, max_cache_len: int):
-        """Pre-build a ``HunyuanStaticCache`` for the AR loop."""
-        import sys as _sys
-
-        upstream_mod = _sys.modules.get(type(transformer).__module__)
-        cache_cls = getattr(upstream_mod, "HunyuanStaticCache", None)
-        if cache_cls is None:
-            return None
-        config = getattr(transformer, "config", None)
-        if config is None:
-            return None
-        try:
-            return cache_cls(
-                config=config,
-                batch_size=batch_size,
-                max_cache_len=max_cache_len,
-                dtype=torch.bfloat16,
-                dynamic=True,
-            )
-        except Exception:  # noqa: BLE001 -- fall back to HF default cache
-            return None
-
 
 class HunyuanImage3ARStage(ARStage[HunyuanImage3ARConditions]):
     """Rollout-level AR stage: ``HunyuanImage3ARConditions → TextSegment``."""