Update Unit Test for PaddleOCR-VL (#4802)

* fix paddleocr prefix cache bug

* add test for paddleocr_vl

* disable prefix-caching in ocr

* add test for paddleocr_vl

* Fix top_p for rejection sampling

* add test for ocr processor; fix top_p for rejection sampling

* add test for ocr processor; fix top_p for rejection sampling

* add test for ocr processor; fix top_p for rejection sampling

* add test for ocr processor; fix top_p for rejection sampling

* add test for ocr processor; fix top_p for rejection sampling

---------

Co-authored-by: ming1753 <ideaminghp@163.com>
Co-authored-by: ming1753 <61511741+ming1753@users.noreply.github.com>

fastdeploy/input/ernie4_5_vl_processor/ernie4_5_vl_processor.py

@@ -26,6 +26,8 @@ from fastdeploy.utils import data_processor_logger
 
 from .process import DataProcessor
 
+_SAMPLING_EPS = 1e-5
+
 
 class Ernie4_5_VLProcessor(Ernie4_5Processor):
     """The processor class for ERNIE MoE VL models."""
@@ -268,6 +270,9 @@ class Ernie4_5_VLProcessor(Ernie4_5Processor):
             request["reasoning_max_tokens"] = max(int(request["max_tokens"] * 0.8), 1)
         data_processor_logger.info(f"Processed request {request}")
 
+        if request.get("top_p") is not None and request.get("top_p") < _SAMPLING_EPS:
+            request["top_p"] = _SAMPLING_EPS
+
         return request
 
     def append_completion_tokens(self, multimodal_inputs, completion_token_ids):

fastdeploy/input/paddleocr_vl_processor/paddleocr_vl_processor.py

@@ -22,6 +22,8 @@ from fastdeploy.utils import data_processor_logger
 
 from .process import DataProcessor
 
+_SAMPLING_EPS = 1e-5
+
 
 class PaddleOCRVLProcessor(TextProcessor):
     """
@@ -61,7 +63,6 @@ class PaddleOCRVLProcessor(TextProcessor):
             tool_parser_obj: Tool parser instance
         """
         super().__init__(model_name_or_path, reasoning_parser_obj, tool_parser_obj)
-
         data_processor_logger.info(f"model_name_or_path: {model_name_or_path}")
         processor_kwargs = self._parse_processor_kwargs(mm_processor_kwargs)
         self.processor = DataProcessor(
@@ -252,6 +253,9 @@ class PaddleOCRVLProcessor(TextProcessor):
         if request.get("max_tokens") is None:
             request["max_tokens"] = max(1, max_model_len - len(request["prompt_token_ids"]))  # Ensure at least 1 token
 
+        if request.get("top_p") is not None and request.get("top_p") < _SAMPLING_EPS:
+            request["top_p"] = _SAMPLING_EPS
+
         return request
 
     def append_generated_tokens(self, multimodal_inputs, generated_token_ids):

tests/e2e/test_paddleocr_vl_serving.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import json
 import os
 import shutil
 import signal
@@ -99,7 +98,6 @@ def setup_and_run_server():
         model_path = "./PaddleOCR-VL-0.9B"
 
     log_path = "server.log"
-    limit_mm_str = json.dumps({"image": 100, "video": 100})
 
     cmd = [
         sys.executable,
@@ -109,8 +107,6 @@ def setup_and_run_server():
         model_path,
         "--port",
         str(FD_API_PORT),
-        "--tensor-parallel-size",
-        "2",
         "--engine-worker-queue-port",
         str(FD_ENGINE_QUEUE_PORT),
         "--metrics-port",
@@ -119,18 +115,13 @@ def setup_and_run_server():
         str(FD_CACHE_QUEUE_PORT),
         "--enable-mm",
         "--max-model-len",
-        "32768",
+        "16384",
         "--max-num-batched-tokens",
-        "384",
+        "16384",
         "--max-num-seqs",
         "128",
-        "--limit-mm-per-prompt",
-        limit_mm_str,
-        "--enable-chunked-prefill",
-        "--kv-cache-ratio",
-        "0.71",
-        "--quantization",
-        "wint4",
+        "--gpu-memory-utilization",
+        "0.9",
         "--graph-optimization-config",
         '{"graph_opt_level":0, "use_cudagraph":true}',
     ]