relay working with qwen2.5;

relay anounced on node ready

packages/node/meshnet_node/model_backend.py

@@ -308,7 +308,11 @@ class TorchModelShard:
         self._norm = _final_norm(self.model) if self.is_tail else None
         self._lm_head = getattr(self.model, "lm_head", None) if self.is_tail else None
         # Per-session KV/recurrent-state cache for this shard's layer range.
-        self.supports_kv_cache = True
+        # Hybrid/linear-attention models such as Qwen3.6 can dispatch Triton
+        # recurrent-cache kernels when use_cache=True. Those kernels cannot
+        # consume CPU tensors ("Pointer argument cannot be accessed from Triton"),
+        # so CPU shards intentionally stay on the stateless prefill path.
+        self.supports_kv_cache = self.device.type != "cpu"
         self.kv_sessions = SessionCacheStore(
             max_sessions=int(os.environ.get("MESHNET_KV_MAX_SESSIONS", "8")),
             ttl_seconds=float(os.environ.get("MESHNET_KV_TTL_SECONDS", "600")),
@@ -612,9 +616,13 @@ class TorchModelShard:
                 hidden_states, attention_mask, position_ids,
                 start_layer=start_layer, cache=cache, past_len=0,
             )
-        except TypeError as exc:
+        except Exception as exc:
+            if not _cache_unsupported_for_shard(exc):
+                raise
             # Layers reject cache kwargs (exotic architecture) — disable caching
-            # for this backend and stay on the stateless path.
+            # for this backend and stay on the stateless path. Some hybrid
+            # CPU paths also accept cache kwargs but fail at runtime inside
+            # Triton-only kernels; treat those as cache-unsupported too.
             self.supports_kv_cache = False
             print(f"  [node] kv cache unsupported by {self.model_id}: {exc}", flush=True)
             return self._run_layers(
@@ -1146,3 +1154,13 @@ def _looks_like_oom(exc: BaseException) -> bool:
             return True
         current = current.__cause__ or current.__context__
     return False
+
+
+def _cache_unsupported_for_shard(exc: BaseException) -> bool:
+    """True when a layer failure means session cache is unsupported, not fatal."""
+    text = str(exc).lower()
+    return (
+        isinstance(exc, TypeError)
+        or "pointer argument cannot be accessed from triton" in text
+        or ("triton" in text and "cpu tensor" in text)
+    )