Clean up server_args, triton cache manager (#8332)

python/sglang/srt/model_executor/forward_batch_info.py

@@ -74,8 +74,6 @@ class ForwardMode(IntEnum):
     MIXED = auto()
     # No sequence to forward. For data parallel attention, some workers will be IDLE if no sequence are allocated.
     IDLE = auto()
-    # Split Prefill for PD multiplexing
-    SPLIT_PREFILL = auto()
 
     # Used in speculative decoding: verify a batch in the target model.
     TARGET_VERIFY = auto()
@@ -86,6 +84,9 @@ class ForwardMode(IntEnum):
     # It is now used for triggering the sampling_info_done event for the first prefill batch.
     DUMMY_FIRST = auto()
 
+    # Split Prefill for PD multiplexing
+    SPLIT_PREFILL = auto()
+
     def is_prefill(self):
         return self.is_extend()
 
@@ -103,12 +104,12 @@ class ForwardMode(IntEnum):
     def is_mixed(self):
         return self == ForwardMode.MIXED
 
-    def is_split_prefill(self):
-        return self == ForwardMode.SPLIT_PREFILL
-
     def is_idle(self):
         return self == ForwardMode.IDLE
 
+    def is_decode_or_idle(self):
+        return self == ForwardMode.DECODE or self == ForwardMode.IDLE
+
     def is_target_verify(self):
         return self == ForwardMode.TARGET_VERIFY
 
@@ -132,8 +133,8 @@ class ForwardMode(IntEnum):
     def is_dummy_first(self):
         return self == ForwardMode.DUMMY_FIRST
 
-    def is_decode_or_idle(self):
-        return self == ForwardMode.DECODE or self == ForwardMode.IDLE
+    def is_split_prefill(self):
+        return self == ForwardMode.SPLIT_PREFILL
 
 
 @total_ordering

python/sglang/srt/model_executor/model_runner.py

@@ -109,7 +109,6 @@ from sglang.srt.utils import (
     get_bool_env_var,
     get_cpu_ids_by_node,
     init_custom_process_group,
-    is_cuda,
     is_fa3_default_architecture,
     is_flashinfer_available,
     is_hip,