[Graph Optimization][CI] Add ERNIE45T 21B sot test (#5538)

.github/workflows/_base_test.yml

@@ -206,6 +206,13 @@ jobs:
           check_service 90
           python -m pytest -sv test_max_waiting_time.py || TEST_EXIT_CODE=1
 
+          curl -X POST http://0.0.0.0:${FLASK_PORT}/switch \
+            -H "Content-Type: application/json" \
+            -d "{\"--model\": \"/MODELDATA/ernie-4_5-21b-a3b-bf16-paddle\", \"--config\": \"ernie45t_21b_sot.yaml\", \"--enable-logprob\": \"False\"}"
+          check_service 360
+          export TEMPLATE=TOKEN_NORMAL
+          python -m pytest -sv test_seed_usage.py -k "not test_seed_stream" || TEST_EXIT_CODE=1
+
           curl -X POST http://0.0.0.0:${FLASK_PORT}/switch \
             -H "Content-Type: application/json" \
             -d "{\"--model\": \"/MODELDATA/ernie-4_5-21b-a3b-bf16-paddle\", \"--config\": \"ernie45t_21b_cinn.yaml\", \"--enable-logprob\": \"False\"}"

fastdeploy/model_executor/layers/attention/append_attn_backend.py

@@ -244,8 +244,6 @@ class AppendAttentionBackend(AttentionBackend):
                 # 128 is qwen3
                 # 32 is glm
                 assert forward_meta.rotary_embs.shape[4] in [128, 32]
-            else:
-                assert forward_meta.rotary_embs.shape == [2, 1, self.max_seq_len, 1, 64]
 
         if self.pd_disaggregation_mode == "per_query":
             metadata.kv_signal_data_list[layer.layer_id] = init_signal_layerwise(

tests/ce/deploy/ernie45t_21b_sot.yaml

@@ -0,0 +1,8 @@
+max_model_len: 32768
+max_num_seqs: 128
+tensor_parallel_size: 1
+quantization: wint4
+graph_optimization_config:
+  graph_opt_level: 1
+  sot_warmup_sizes: [2,16,32,64]
+  use_cudagraph: True