1from __future__ import annotations
2
3import os
4import daft
5from daft import col
6from daft.aihc.common.udf import aihc_udf
7from daft.aihc.functions.multimodal.qwen_vl_video_understanding_vllm import QwenVLVideoUnderstandingVLLM
8
9if __name__ == "__main__":
10    if os.getenv("DAFT_RUNNER", "native") == "ray":
11        import ray
12        ray.init(dashboard_host="0.0.0.0", ignore_reinit_error=True)
13        daft.set_runner_ray()
14    daft.set_execution_config(actor_udf_ready_timeout=6000, min_cpu_per_task=0)
15
16    samples = {
17        "video_path": [
18            f"https://{bucket}.bj.bcebos.com/sample.mp4"
19        ],
20        "prompt": ["请给出视频的详细描述。"]
21    }
22
23    video_src_type = "video_url"
24    model_path = os.getenv("MODEL_PATH", "/opt/aihc/models")
25    model_name = os.getenv("MODEL_NAME", "Qwen/Qwen2.5-VL-7B-Instruct")
26    dtype = "bfloat16"
27    default_prompt = None
28    max_caption_length = 256
29    min_pixels = 320 * 160
30    max_pixels = 320 * 160
31    fps = 1
32    batch_size = 2
33    seed = 42
34    max_model_len = 12800
35    max_num_seqs = 128
36    enable_prefix_caching = True
37    gpu_memory_utilization = 0.95
38    enforce_eager = True
39
40    ds = daft.from_pydict(samples)
41    ds = ds.with_column(
42        "caption",
43        aihc_udf(
44            QwenVLVideoUnderstandingVLLM,
45            construct_args={
46                "video_src_type": video_src_type,
47                "model_path": model_path,
48                "model_name": model_name,
49                "dtype": dtype,
50                "prompt": default_prompt,
51                "max_caption_length": max_caption_length,
52                "min_pixels": min_pixels,
53                "max_pixels": max_pixels,
54                "fps": fps,
55                "batch_size": batch_size,
56                "seed": seed,
57                "max_model_len": max_model_len,
58                "max_num_seqs": max_num_seqs,
59                "enable_prefix_caching": enable_prefix_caching,
60                "gpu_memory_utilization": gpu_memory_utilization,
61                "enforce_eager": enforce_eager,
62            },
63            num_gpus=1,
64            batch_size=batch_size,
65            concurrency=1,
66        )(col("video_path"), col("prompt")),
67    )
68
69    ds.show()
70#╭────────────────────────────────┬────────────────────────┬─────────────────────────────────────────────────────────────╮                                                                                      
71#│ video_path                     ┆ prompt                 ┆ caption                                                     │
72#│ ---                            ┆ ---                    ┆ ---                                                         │
73#│ String                         ┆ String                 ┆ String                                                      │
74#╞════════════════════════════════╪════════════════════════╪═════════════════════════════════════════════════════════════╡
75#│ https://{bucket}.bj.bcebos.com/┆ 请给出视频的详细描述。   ┆ 这段视频展示了一对男女在室内对话的场景。背景中可以看到一些…      │
76#╰────────────────────────────────┴────────────────────────┴─────────────────────────────────────────────────────────────╯