[cherry-pick] add qwen2vl video infer benchmark (#952)

deploy/qwen2_vl/README.md

@@ -36,31 +36,31 @@ python setup_cuda.py install
 
 ### 3.1. 文本&单张图像输入高性能推理
 ```bash
-python deploy/qwen2_vl/single_image_infer.py \
+CUDA_VISIBLE_DEVICES=0 python deploy/qwen2_vl/single_image_infer.py \
     --model_name_or_path Qwen/Qwen2-VL-2B-Instruct \
     --dtype bfloat16 \
     --benchmark True \
 ```
 
-- 在 NVIDIA A100-SXM4-80GB 上测试的单图端到端速度性能如下：
+- 在 NVIDIA A800-80GB 上测试的单图端到端速度性能如下：
 
 | model                  | Paddle Inference|    PyTorch   | Paddle 动态图 |
 | ---------------------- | --------------- | ------------ | ------------ |
-| Qwen2-VL-2B-Instruct   |      1.44 s     |     2.35 s   |    5.215 s   |
-| Qwen2-VL-7B-Instruct   |      1.73 s     |      4.4s    |    6.339 s   |
+| Qwen2-VL-2B-Instruct   |      1.053 s     |     2.086 s   |   5.766 s   |
+| Qwen2-VL-7B-Instruct   |      2.293 s     |     3.132 s   |   6.221 s   |
 
 
 ### 3.2. 文本&视频输入高性能推理
 ```bash
-python deploy/qwen2_vl/video_infer.py \
-    --model_name_or_path Qwen/Qwen2-VL-7B-Instruct \
+CUDA_VISIBLE_DEVICES=0 python deploy/qwen2_vl/video_infer.py \
+    --model_name_or_path Qwen/Qwen2-VL-2B-Instruct \
     --dtype bfloat16 \
-    --benchmark 1
+    --benchmark True
 ```
 
-- 在 NVIDIA A100-SXM4-80GB 上测试的单图端到端速度性能如下：
+- 在 NVIDIA A800-80GB 上测试的单视频端到端速度性能如下：
 
 | model                  | Paddle Inference|    PyTorch   | Paddle 动态图 |
 | ---------------------- | --------------- | ------------ | ------------ |
-| Qwen2-VL-2B-Instruct   |      1.503 s    |       -      |    47.922 s  |
-| Qwen2-VL-7B-Instruct   |      2.715 s    |       -      |    33.597 s  |
+| Qwen2-VL-2B-Instruct   |      2.890 s    |     3.143 s  |    6.183 s  |
+| Qwen2-VL-7B-Instruct   |      2.534 s    |     2.715 s  |    5.721 s  |

deploy/qwen2_vl/single_image_infer.py

@@ -267,11 +267,15 @@ def run_model():
             duringtime = endtime - starttime
             duringtime = duringtime.seconds * 1000 + duringtime.microseconds / 1000.0
             sumtime += duringtime
-            print(f"Single {predictor_args.model_name_or_path} end to end time : ", duringtime, "ms")
+            print(f"Single Image Inference: {predictor_args.model_name_or_path} end-to-end time : ", duringtime, "ms")
             inference_global_mem = paddle.device.cuda.memory_reserved() / (1024**3)
             print(f"Inference used CUDA memory : {inference_global_mem:.3f} GiB")
 
-    print(f"Single {predictor_args.model_name_or_path} ave end to end time : ", sumtime / repeat_times, "ms")
+    print(
+        f"Single Image Inference: {predictor_args.model_name_or_path} average end-to-end time : ",
+        sumtime / repeat_times,
+        "ms",
+    )
 
 else:
     generated_text = run_model()

deploy/qwen2_vl/video_infer.py

@@ -267,11 +267,15 @@ def run_model():
             duringtime = endtime - starttime
             duringtime = duringtime.seconds * 1000 + duringtime.microseconds / 1000.0
             sumtime += duringtime
-            print(f"Single {predictor_args.model_name_or_path} end to end time : ", duringtime, "ms")
+            print(f"Single Video Inference: {predictor_args.model_name_or_path} end to end time : ", duringtime, "ms")
             inference_global_mem = paddle.device.cuda.memory_reserved() / (1024**3)
             print(f"Inference used CUDA memory : {inference_global_mem:.3f} GiB")
 
-    print(f"Single {predictor_args.model_name_or_path} ave end to end time : ", sumtime / repeat_times, "ms")
+    print(
+        f"Single Video Inference: {predictor_args.model_name_or_path} average end-to-end time : ",
+        sumtime / repeat_times,
+        "ms",
+    )
 
 else:
     generated_text = run_model()

paddlemix/examples/qwen2_vl/README.md

@@ -42,17 +42,17 @@ python -m pip install paddlenlp==3.0.0b3
 
 ### a. 单图预测
 ```bash
-python paddlemix/examples/qwen2_vl/single_image_infer.py
+CUDA_VISIBLE_DEVICES=0 python paddlemix/examples/qwen2_vl/single_image_infer.py
 ```
 
 ### b. 多图预测
 ```bash
-python paddlemix/examples/qwen2_vl/multi_image_infer.py
+CUDA_VISIBLE_DEVICES=0 python paddlemix/examples/qwen2_vl/multi_image_infer.py
 ```
 
 ### c. 视频预测
 ```bash
-python paddlemix/examples/qwen2_vl/video_infer.py
+CUDA_VISIBLE_DEVICES=0 python paddlemix/examples/qwen2_vl/video_infer.py
 ```
 
 ## 4 模型微调
@@ -106,19 +106,19 @@ sh paddlemix/examples/qwen2_vl/shell/basline_7b_lora_bs32_1e8.sh
 同按步骤3中的模型推理预测，只需将`paddlemix/examples/qwen2_vl/single_image_infer.py`中的`--model_path`参数修改为微调后的模型路径即可。
 
 ```bash
-python paddlemix/examples/qwen2_vl/single_image_infer.py
+CUDA_VISIBLE_DEVICES=0 python paddlemix/examples/qwen2_vl/single_image_infer.py
 ```
 
 ### 5 高性能推理优化
 
 [Paddle高性能推理优化后](../../../deploy/qwen2_vl/)，测试结果如下：
 
-- 在 NVIDIA A100-SXM4-80GB 上测试的单图端到端速度性能如下：
+- 在 NVIDIA A800-80GB 上测试的单图端到端速度性能如下：
 
 | model                  | Paddle Inference|    PyTorch   | Paddle 动态图 |
 | ---------------------- | --------------- | ------------ | ------------ |
-| Qwen2-VL-2B-Instruct   |      1.44 s     |     2.35 s   |    5.215 s   |
-| Qwen2-VL-7B-Instruct   |      1.73 s     |      4.4s    |    6.339 s   |
+| Qwen2-VL-2B-Instruct   |      1.053 s     |     2.086 s   |   5.766 s   |
+| Qwen2-VL-7B-Instruct   |      2.293 s     |     3.132 s   |   6.221 s   |
 
 
 

paddlemix/examples/qwen2_vl/single_image_infer.py

@@ -94,10 +94,10 @@ def main(args):
             if i > 10:
                 total += time.time() - start
         print("s/it: ", total / 10)
-        print(f"\nGPU memory_allocated: {paddle.device.cuda.memory_allocated() / 1024 ** 3:.2f} GB")
-        print(f"\nGPU max_memory_allocated: {paddle.device.cuda.max_memory_allocated() / 1024 ** 3:.2f} GB")
-        print(f"\nGPU memory_reserved: {paddle.device.cuda.memory_reserved() / 1024 ** 3:.2f} GB")
-        print(f"\nGPU max_memory_reserved: {paddle.device.cuda.max_memory_reserved() / 1024 ** 3:.2f} GB")
+        print(f"GPU memory_allocated: {paddle.device.cuda.memory_allocated() / 1024 ** 3:.2f} GB")
+        print(f"GPU max_memory_allocated: {paddle.device.cuda.max_memory_allocated() / 1024 ** 3:.2f} GB")
+        print(f"GPU memory_reserved: {paddle.device.cuda.memory_reserved() / 1024 ** 3:.2f} GB")
+        print(f"GPU max_memory_reserved: {paddle.device.cuda.max_memory_reserved() / 1024 ** 3:.2f} GB")
         print("output_text:\n", output_text)
 
     else:

paddlemix/examples/qwen2_vl/video_infer.py

@@ -12,53 +12,114 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import argparse
+
+import paddle
+
 from paddlemix.models.qwen2_vl import MIXQwen2Tokenizer
 from paddlemix.models.qwen2_vl.modeling_qwen2_vl import Qwen2VLForConditionalGeneration
 from paddlemix.processors.qwen2_vl_processing import (
     Qwen2VLImageProcessor,
     Qwen2VLProcessor,
     process_vision_info,
 )
+from paddlemix.utils.log import logger
 
-MODEL_NAME = "Qwen/Qwen2-VL-7B-Instruct"
-model = Qwen2VLForConditionalGeneration.from_pretrained(MODEL_NAME, dtype="bfloat16")
-
-image_processor = Qwen2VLImageProcessor()
-tokenizer = MIXQwen2Tokenizer.from_pretrained(MODEL_NAME)
-min_pixels = 256 * 28 * 28  # 200704
-max_pixels = 1280 * 28 * 28  # 1003520
-processor = Qwen2VLProcessor(image_processor, tokenizer, min_pixels=min_pixels, max_pixels=max_pixels)
-
-# Messages containing a video and a text query
-messages = [
-    {
-        "role": "user",
-        "content": [
-            {
-                "type": "video",
-                "video": "paddlemix/demo_images/red-panda.mp4",
-                "max_pixels": 360 * 420,
-                "fps": 1.0,
-            },
-            {"type": "text", "text": "Describe this video."},
-        ],
-    }
-]
-
-image_inputs, video_inputs = process_vision_info(messages)
-question = "Describe this video."
-video_pad_token = "<|vision_start|><|video_pad|><|vision_end|>"
-text = f"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n{video_pad_token}{question}<|im_end|>\n<|im_start|>assistant\n"
-
-inputs = processor(
-    text=[text],
-    images=image_inputs,
-    videos=video_inputs,
-    padding=True,
-    return_tensors="pd",
-)
-# Inference: Generation of the output
-generated_ids = model.generate(**inputs, max_new_tokens=128)  # already trimmed in paddle
-# print("generated_ids:\n", generated_ids)
-output_text = processor.batch_decode(generated_ids[0], skip_special_tokens=True, clean_up_tokenization_spaces=False)
-print("output_text:\n", output_text[0])
+
+def main(args):
+    paddle.seed(seed=0)
+    compute_dtype = "float16" if args.fp16 else "bfloat16"
+    if "npu" in paddle.get_device():
+        is_bfloat16_supported = True
+    else:
+        is_bfloat16_supported = paddle.amp.is_bfloat16_supported()
+    if compute_dtype == "bfloat16" and not is_bfloat16_supported:
+        logger.warning("bfloat16 is not supported on your device,change to float32")
+        compute_dtype = "float32"
+
+    model = Qwen2VLForConditionalGeneration.from_pretrained(args.model_path, dtype="bfloat16")
+
+    image_processor = Qwen2VLImageProcessor()
+    tokenizer = MIXQwen2Tokenizer.from_pretrained(args.model_path)
+    min_pixels = 256 * 28 * 28  # 200704
+    max_pixels = 1280 * 28 * 28  # 1003520
+    processor = Qwen2VLProcessor(image_processor, tokenizer, min_pixels=min_pixels, max_pixels=max_pixels)
+
+    # Messages containing a video and a text query
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "video",
+                    "video": f"{args.video_file}",
+                    "max_pixels": 360 * 420,
+                    "fps": 1.0,
+                },
+                {"type": "text", "text": f"{args.question}"},
+            ],
+        }
+    ]
+
+    # Preparation for inference
+    image_inputs, video_inputs = process_vision_info(messages)
+
+    question = messages[0]["content"][1]["text"]
+    video_pad_token = "<|vision_start|><|video_pad|><|vision_end|>"
+    text = f"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n{video_pad_token}{question}<|im_end|>\n<|im_start|>assistant\n"
+    text = [text]
+
+    inputs = processor(
+        text=text,
+        images=image_inputs,
+        videos=video_inputs,
+        padding=True,
+        return_tensors="pd",
+    )
+
+    if args.benchmark:
+        import time
+
+        start = 0.0
+        total = 0.0
+        for i in range(20):
+            if i > 10:
+                start = time.time()
+            with paddle.no_grad():
+                generated_ids = model.generate(
+                    **inputs, max_new_tokens=args.max_new_tokens, temperature=args.temperature
+                )  # already trimmed in paddle
+                output_text = processor.batch_decode(
+                    generated_ids[0], skip_special_tokens=True, clean_up_tokenization_spaces=False
+                )
+            if i > 10:
+                total += time.time() - start
+        print("s/it: ", total / 10)
+        print(f"GPU memory_allocated: {paddle.device.cuda.memory_allocated() / 1024 ** 3:.2f} GB")
+        print(f"GPU max_memory_allocated: {paddle.device.cuda.max_memory_allocated() / 1024 ** 3:.2f} GB")
+        print(f"GPU memory_reserved: {paddle.device.cuda.memory_reserved() / 1024 ** 3:.2f} GB")
+        print(f"GPU max_memory_reserved: {paddle.device.cuda.max_memory_reserved() / 1024 ** 3:.2f} GB")
+        print("output_text:\n", output_text)
+
+    else:
+        # Inference: Generation of the output
+        generated_ids = model.generate(
+            **inputs, max_new_tokens=args.max_new_tokens, temperature=args.temperature
+        )  # already trimmed in paddle
+        output_text = processor.batch_decode(
+            generated_ids[0], skip_special_tokens=True, clean_up_tokenization_spaces=False
+        )
+        print("output_text:\n", output_text)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model_path", type=str, default="Qwen/Qwen2-VL-2B-Instruct")
+    parser.add_argument("--question", type=str, default="Describe this video.")
+    parser.add_argument("--video_file", type=str, default="paddlemix/demo_images/red-panda.mp4")
+    parser.add_argument("--temperature", type=float, default=0.01)
+    parser.add_argument("--max_new_tokens", type=int, default=128)
+    parser.add_argument("--fp16", action="store_true")
+    parser.add_argument("--benchmark", action="store_true")
+    args = parser.parse_args()
+    main(args)