fix: avoid unused patch state globals

Author: tengqm

vllm_fl/dispatch/backends/vendor/gcu/impl/flash_attn_backend.py

@@ -38,15 +38,12 @@
 
 logger = logging.getLogger(__name__)
 
-_patched = False
-
 _FA_UTILS = "vllm.v1.attention.backends.fa_utils"
 _FLASH_ATTN = "vllm.v1.attention.backends.flash_attn"
 
 
 def apply_flash_attn_backend_gcu_patch() -> None:
-    global _patched
-    if _patched:
+    if getattr(sys.modules.get(_FA_UTILS), "_gcu_flash_attn_patched", False):
         return
 
     try:
@@ -72,6 +69,7 @@ def apply_flash_attn_backend_gcu_patch() -> None:
     fa_utils.get_scheduler_metadata = get_scheduler_metadata
     fa_utils.reshape_and_cache_flash = reshape_and_cache_flash
     fa_utils._GCU_FLASH_ATTN_AVAILABLE = True
+    fa_utils._gcu_flash_attn_patched = True
     fa_utils.is_flash_attn_varlen_func_available = lambda: True
 
     # If flash_attn.py already imported (its gate ran False and it skipped the
@@ -84,7 +82,6 @@ def apply_flash_attn_backend_gcu_patch() -> None:
         flash_attn_mod.flash_attn_supports_sinks = fa_utils.flash_attn_supports_sinks
         flash_attn_mod.is_flash_attn_varlen_func_available = lambda: True
 
-    _patched = True
     logger.info(
         "GCU: enabled native FLASH_ATTN backend "
         "(vendor flash_attn_varlen_func + flag_gems reshape_and_cache_flash)"

vllm_fl/dispatch/backends/vendor/gcu/impl/slot_mapping.py

@@ -40,8 +40,6 @@
 
 logger = logging.getLogger(__name__)
 
-_patched = False
-
 
 def _compute_slot_mapping_int32(self, num_reqs, query_start_loc, positions):
     from vllm.v1.attention.backends.utils import PAD_SLOT_ID
@@ -105,11 +103,10 @@ def _compute_slot_mapping_int32(self, num_reqs, query_start_loc, positions):
 
 def apply_slot_mapping_gcu_patch() -> None:
     """Replace BlockTable.compute_slot_mapping with the on-device int32 version."""
-    global _patched
-    if _patched:
-        return
     from vllm.v1.worker.block_table import BlockTable
 
+    if BlockTable.compute_slot_mapping is _compute_slot_mapping_int32:
+        return
+
     BlockTable.compute_slot_mapping = _compute_slot_mapping_int32
-    _patched = True
     logger.info("GCU: patched BlockTable.compute_slot_mapping (on-device int32)")