Reply to delete code

Author: luoyc123

vllm_fl/__init__.py

@@ -150,6 +150,7 @@ def register_router():
 
 def register_model():
     """Register FL-specific models not yet upstream."""
+    from vllm import ModelRegistry
     _register_flagcx_connector()
 
     # Register OOT quant kernels so kernel selection can find them
@@ -176,7 +177,6 @@ def register_model():
     except Exception as e:
         logger.error(f"Register DeepseekV4 model error: {str(e)}")
 
-
     # Register DeepseekV4 model
     try:
         ModelRegistry.register_model(

vllm_fl/ops/custom_ops.py

@@ -23,6 +23,7 @@
 OOT_OPS = {
     "silu_and_mul": (SiluAndMulFL, "SiluAndMul"),  # noqa F405
     "gelu_and_mul": (GeluAndMulFL, "GeluAndMul"),  # noqa F405
+    "silu_and_mul_with_clamp": (SiluAndMulWithClampFL, "SiluAndMulWithClamp"),  # noqa F405
     "rms_norm": (RMSNormFL, "RMSNorm"),  # noqa F405
     "rotary_embedding": (RotaryEmbeddingFL, "RotaryEmbedding"),  # noqa F405
     "fused_moe": (FusedMoEFL, "FusedMoE"),  # noqa F405

vllm_fl/ops/fused_moe/fused_moe_utils.py

@@ -368,8 +368,11 @@ def apply(
         # Fast path (no LoRA): single fused call via patched fused_experts_impl.
         if self._lora_context is None:
             from vllm_fl.ops.fused_moe.fused_moe import fused_experts_impl as _fused_experts_impl
+        # Fast path (no LoRA, NVIDIA only): single fused FlagGems call.
+        if self._lora_context is None and current_platform.is_cuda():
+            import flag_gems
 
-            output.copy_(_fused_experts_impl(
+            output.copy_(flag_gems.fused_experts_impl(
                 hidden_states,
                 w1,
                 w2,