feat: P1 — add eager warmup_fused_swiglu_compilation for SharedExpert (1-group)

single_shot_inference.py

@@ -1042,6 +1042,15 @@ def main():
         se.set_fused_swiglu(True)
         # EAGERLY process shared expert weights
         se._ensure_initialized()
+        # P1: Eagerly warmup fused SwiGLU compilation for SE (1-group)
+        if se._fused_swiglu:
+            from dsv4.ops.gemm_runner import warmup_fused_swiglu_compilation
+            K_packed = H // 2
+            N_packed_l1 = (2 * cfg.get("moe_intermediate_size", 3072)) // 2  # gate+up
+            warmup_fused_swiglu_compilation(
+                1, K_packed, N_packed_l1, dev,
+                swiglu_limit=cfg.get("swiglu_limit", 10.0),
+            )
         # Fix activation global scales — _ensure_initialized sets gsa from l1_gs (which is 1.0)
         # FIX: Same runtime gsa for SharedExpert
         se._use_runtime_gsa = True