switch INT8 support to official flagos version

Author: cyx11111

vllm_fl/dispatch/backends/vendor/sunrise/patch.py

@@ -44,6 +44,32 @@ def apply_sunrise_patches():
     patch_ptpu_trunc_normal_init()
     patch_minicpmo_resampler_device()
     patch_moe_force_config()
+    patch_native_moe_ops()
+    patch_native_int8_routing()
+
+
+def patch_native_moe_ops() -> None:
+    """Give vLLM's fused-MoE path implementations that exist on PTPU.
+
+    Deferred rather than import-time: the shims resolve through the FL dispatch
+    registry, which is only populated once the backends have registered. Must
+    precede ``patch_native_int8_routing``, which routes W8A8 MoE onto them.
+    """
+    from .patches import patch_moe_native_ops
+
+    patch_moe_native_ops.apply_patch()
+
+
+def patch_native_int8_routing() -> None:
+    """Re-assert the INT8 routing that ``register_oot_ops`` can overwrite.
+
+    ``register_oot_ops`` installs ``install_fl_w8a8_moe_selector`` and clones the
+    CUDA linear kernels into the OOT slot before it calls us, so sunrise has to
+    re-run the order-sensitive INT8 patches afterwards to be the last writer.
+    """
+    from .patches import patch_int8_native
+
+    patch_int8_native.install_late_patches()
 
 
 def patch_moe_force_config() -> None:

vllm_fl/dispatch/backends/vendor/sunrise/patches/__init__.py

@@ -13,7 +13,9 @@
   optional decode profiler, torch profiler PrivateUse1 bridge.
 * **Deferred** (``patch.apply_sunrise_patches``): CUDA stream shims, FlagCX
   comm/collectives, cudagraph, sampler/penalties, distributed runtime, OOT
-  layer registration, and other worker-init hooks.
+  layer registration, the ``_moe_C`` dispatch shims, and other worker-init
+  hooks. INT8 also re-asserts its MoE routing there, because
+  ``register_oot_ops`` installs the generic FL W8A8 selector in between.
 """
 
 from . import patch_fla_ops