update

Author: gufengc

src/parallax/models/qwen3_5.py

@@ -38,9 +38,7 @@ def __call__(
         if prefix_lens is not None and context_lengths is not None:
             input_lengths = context_lengths - prefix_lens
 
-        if target_len == 1 or (
-            prefix_lens is not None and bool(mx.any(prefix_lens > 0))
-        ):
+        if target_len == 1 or (prefix_lens is not None and bool(mx.any(prefix_lens > 0))):
             conv_state, state = cache.read_states(state_slot_mapping)
         else:
             conv_state = mx.zeros(

src/parallax/models/qwen3_next.py

@@ -15,8 +15,8 @@
 
 from parallax.server.cache.base import BaseCache
 from parallax.utils.prefix_cache_utils import compute_attention_with_prefix_cache
-from parallax_utils.logging_config import get_logger
 from parallax_extensions.ops import paged_attention_v1, reshape_and_cache
+from parallax_utils.logging_config import get_logger
 
 logger = get_logger(__name__)
 
@@ -167,9 +167,7 @@ def __call__(
         if prefix_lens is not None and context_lengths is not None:
             input_lengths = context_lengths - prefix_lens
 
-        if target_len == 1 or (
-            prefix_lens is not None and bool(mx.any(prefix_lens > 0))
-        ):
+        if target_len == 1 or (prefix_lens is not None and bool(mx.any(prefix_lens > 0))):
             conv_state, state1 = cache.read_states(state_slot_mapping)
         else:
             conv_state = mx.zeros(

src/parallax/server/cache/linear_cache.py

@@ -66,17 +66,13 @@ def get_indexer_cache(self) -> Optional[mx.array]:
     def zero_slot(self, slot_idx: int):
         """Reset a request slot to an empty recurrent state."""
         if self.conv_state_cache is not None:
-            self.conv_state_cache[0, slot_idx] = mx.zeros_like(
-                self.conv_state_cache[0, slot_idx]
-            )
+            self.conv_state_cache[0, slot_idx] = mx.zeros_like(self.conv_state_cache[0, slot_idx])
         if self.linear_state_cache is not None:
             self.linear_state_cache[0, slot_idx] = mx.zeros_like(
                 self.linear_state_cache[0, slot_idx]
             )
 
-    def snapshot_slot(
-        self, slot_idx: int
-    ) -> Tuple[Optional[mx.array], Optional[mx.array]]:
+    def snapshot_slot(self, slot_idx: int) -> Tuple[Optional[mx.array], Optional[mx.array]]:
         """Copy the recurrent state currently stored in a request slot."""
         conv_state = None
         linear_state = None