[Graph Optimization][BugFix][CI] Fix 0size bug && add unitest (#5495)

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

fastdeploy/model_executor/forward_meta.py

@@ -149,6 +149,9 @@ class ForwardMeta:
     moe_num_chunk: int = 1
     max_moe_num_chunk: int = 1
 
+    # for zero size
+    is_zero_size: bool = False
+
     def clear_caches(self):
         """Safely clean up the caches"""
         if self.caches:

fastdeploy/model_executor/layers/embeddings.py

@@ -288,7 +288,7 @@ class VocabParallelEmbedding(nn.Layer):
             h2d_copy(param[:, : shard_weight.shape[1]], shard_weight)
             param[:, shard_weight.shape[1] :].fill_(0)
 
-    def forward(self, ids_remove_padding=None) -> paddle.Tensor:
+    def forward(self, ids_remove_padding, forward_meta=None) -> paddle.Tensor:
         """
         Defines the forward computation of the layer.
 
@@ -299,7 +299,7 @@ class VocabParallelEmbedding(nn.Layer):
         Returns:
             Tensor: Embedded tensor representation of the input IDs.
         """
-        if ids_remove_padding.shape[0] == 0:
+        if forward_meta is not None and forward_meta.is_zero_size:
             return paddle.empty([0, self.embedding_dim], dtype=self.embeddings.weight.dtype)
         if self.column_cut:
             input_embedings = self.embeddings(ids_remove_padding)

fastdeploy/model_executor/models/ernie4_5_moe.py

@@ -448,7 +448,7 @@ class Ernie4_5_Model(nn.Layer):
         ids_remove_padding: paddle.Tensor,
         forward_meta: ForwardMeta,
     ):
-        hidden_states = self.embed_tokens(ids_remove_padding=ids_remove_padding)
+        hidden_states = self.embed_tokens(ids_remove_padding=ids_remove_padding, forward_meta=forward_meta)
 
         if current_platform.is_iluvatar() and forward_meta.attn_backend.mixed:
             hidden_states = forward_meta.attn_backend.transpose(hidden_states)

fastdeploy/worker/gpu_model_runner.py

@@ -1541,6 +1541,9 @@ class GPUModelRunner(ModelRunnerBase):
         for attn_backend in self.attn_backends:
             attn_backend.init_attention_metadata(self.forward_meta)
 
+        # for zero size
+        self.forward_meta.is_zero_size = self.forward_meta.ids_remove_padding.shape[0] == 0
+
     def initialize_kv_cache(self, profile: bool = False) -> None:
         """
         Initialize kv cache

tests/ce/deploy/ernie45t_21b_cinn.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: 2
+  sot_warmup_sizes: [2,16,32,64]
+  use_cudagraph: True

tests/e2e/test_EB_VL_Lite_sot_serving.py

@@ -91,7 +91,7 @@ def setup_and_run_server():
         "--reasoning-parser",
         "ernie-45-vl",
         "--graph-optimization-config",
-        '{"graph_opt_level": 1, "use_cudagraph": true, "full_cuda_graph": false}',
+        '{"graph_opt_level": 2, "use_cudagraph": true, "full_cuda_graph": false}',
     ]
 
     # Start subprocess in new process group