diff --git a/vllm_xpu_kernels/fused_moe_interface.py b/vllm_xpu_kernels/fused_moe_interface.py
index 2e3ab4a..327562f 100755
--- a/vllm_xpu_kernels/fused_moe_interface.py
+++ b/vllm_xpu_kernels/fused_moe_interface.py
@@ -64,6 +64,10 @@ ONEDNN_SIDECAR_ALLOW_GRAPH_CAPTURE_ENV = (
     "VLLM_XPU_MOE_ONEDNN_SIDECAR_ALLOW_GRAPH_CAPTURE")
 W8A8_MIDDLE_LAYERLET_ENV = "VLLM_XPU_MOE_W8A8_MIDDLE_LAYERLET"
 W8A8_FULL_LAYERLET_ENV = "VLLM_XPU_MOE_W8A8_FULL_LAYERLET"
+W8A8_FULL_LAYERLET_GATE_TRACE_ENV = (
+    "VLLM_XPU_MOE_W8A8_FULL_LAYERLET_GATE_TRACE_FILE")
+W8A8_FULL_LAYERLET_GATE_TRACE_MAX_LINES_ENV = (
+    "VLLM_XPU_MOE_W8A8_FULL_LAYERLET_GATE_TRACE_MAX_LINES")
 
 _live_abi_lock = threading.Lock()
 _live_abi_writes = 0
@@ -76,6 +80,8 @@ _sidecar_probe_writes = 0
 _sidecar_reference_parity_calls = 0
 _sidecar_probe_disabled = False
 _sidecar_probe_warned = False
+_full_layerlet_gate_lock = threading.Lock()
+_full_layerlet_gate_writes = 0
 
 
 try:
@@ -938,6 +944,29 @@ def _maybe_log_sidecar_probe(record):
         return
 
 
+def _maybe_log_full_layerlet_gate(record):
+    global _full_layerlet_gate_writes
+
+    path = os.environ.get(W8A8_FULL_LAYERLET_GATE_TRACE_ENV)
+    if not path:
+        return
+    max_lines = max(0, _int_env(W8A8_FULL_LAYERLET_GATE_TRACE_MAX_LINES_ENV,
+                                256))
+    if max_lines == 0:
+        return
+    try:
+        line = json.dumps(record, sort_keys=True)
+        with _full_layerlet_gate_lock:
+            if _full_layerlet_gate_writes >= max_lines:
+                return
+            os.makedirs(os.path.dirname(path) or ".", exist_ok=True)
+            with open(path, "a", encoding="utf-8") as handle:
+                handle.write(line + "\n")
+            _full_layerlet_gate_writes += 1
+    except Exception:
+        return
+
+
 def _maybe_probe_onednn_sidecar(*,
                                 hidden_states,
                                 w13,
@@ -2148,6 +2177,9 @@ def xpu_fused_moe(hidden_states,
         unpermuted_row_to_permuted_row = int8_scratch[
             "unpermuted_row_to_permuted_row"]
 
+    full_layerlet_fused_offset_gate = _should_use_int8_fused_prologue_offset(
+        diagnostic_context)
+    full_layerlet_op_gate = _should_use_w8a8_full_layerlet()
     use_w8a8_full_layerlet = (
         is_int8 and activation == "silu" and num_rows == 1
         and expert_map is None and ep_size == 1 and int8_scratch is not None
@@ -2158,8 +2190,50 @@ def xpu_fused_moe(hidden_states,
         and int8_scratch.get("gemm2_a") is not None
         and int8_scratch.get("gemm2_a_scales") is not None
         and int8_scratch.get("gemm2_output") is not None
-        and _should_use_int8_fused_prologue_offset(diagnostic_context)
-        and _should_use_w8a8_full_layerlet())
+        and full_layerlet_fused_offset_gate
+        and full_layerlet_op_gate)
+    if os.environ.get(W8A8_FULL_LAYERLET_GATE_TRACE_ENV):
+        scratch_keys = []
+        if isinstance(int8_scratch, dict):
+            scratch_keys = sorted(str(key) for key in int8_scratch.keys())
+        _maybe_log_full_layerlet_gate({
+            "event": "w8a8_full_layerlet_gate",
+            "use_w8a8_full_layerlet": bool(use_w8a8_full_layerlet),
+            "is_int8": bool(is_int8),
+            "activation": str(activation),
+            "num_rows": int(num_rows),
+            "hidden_size": int(hidden_size),
+            "n_experts_per_token": int(n_experts_per_token),
+            "expert_map_is_none": expert_map is None,
+            "ep_size": int(ep_size),
+            "has_int8_scratch": int8_scratch is not None,
+            "scratch_keys": scratch_keys,
+            "has_prologue_workspace": (
+                int8_scratch is not None
+                and int8_scratch.get("prologue_workspace") is not None),
+            "has_w8a8_offsets": (
+                int8_scratch is not None
+                and int8_scratch.get("w8a8_offsets") is not None),
+            "has_gemm1_a": (
+                int8_scratch is not None
+                and int8_scratch.get("gemm1_a") is not None),
+            "has_gemm1_a_scales": (
+                int8_scratch is not None
+                and int8_scratch.get("gemm1_a_scales") is not None),
+            "has_gemm2_a": (
+                int8_scratch is not None
+                and int8_scratch.get("gemm2_a") is not None),
+            "has_gemm2_a_scales": (
+                int8_scratch is not None
+                and int8_scratch.get("gemm2_a_scales") is not None),
+            "has_gemm2_output": (
+                int8_scratch is not None
+                and int8_scratch.get("gemm2_output") is not None),
+            "fused_offset_gate": bool(full_layerlet_fused_offset_gate),
+            "full_layerlet_op_gate": bool(full_layerlet_op_gate),
+            "stream_capture_active": _stream_capture_active(),
+            "diagnostic_context": str(diagnostic_context),
+        })
     if use_w8a8_full_layerlet:
         with timed_region("xpu_moe.w8a8_full_layerlet"):
             _w8a8_full_layerlet_out(