add base test (#1044)

This pull request introduces new configuration files, adds a flexible
inference script, and implements tools for extracting and analyzing
controller metrics related to latency and GPU utilization. It also
enhances the controller service to record and export detailed metrics
for improved monitoring and analysis.

**Key changes:**

Configuration and Inference:

* Added two new model configuration files for different inference
settings: `wan22_moe_i2v_baseline.json` (4-step inference) and
`wan22_moe_i2v_baseline_50steps.json` (50-step inference), supporting
the Wan2.2 I2V model with parallelization options.
[[1]](diffhunk://#diff-54e707f816456395c7a8b8a1d469798796ccd0fe68b1f3baa996311d4db4df3fR1-R48)
[[2]](diffhunk://#diff-7e5b85d9092b5921ea7744d9b62c0b8756183336eb41af8a7818e87e7557d2e4R1-R36)
* Introduced a new, comprehensive `infer.py` script for running
inference with various LightX2V models, supporting a wide range of tasks
and flexible command-line arguments.

Controller Metrics and Monitoring:

* Enhanced the controller (`controller.py`) to record detailed
monitoring samples, track controller uptime, and save all metrics
(including requests and periodic samples) to a configurable JSON file on
shutdown. This enables more thorough post-run analysis.
[[1]](diffhunk://#diff-cd2bb47b50681fc168f658410a3c5239855d829c498e4f0b40b06a370816ba84R69-R76)
[[2]](diffhunk://#diff-cd2bb47b50681fc168f658410a3c5239855d829c498e4f0b40b06a370816ba84R881-R890)
[[3]](diffhunk://#diff-cd2bb47b50681fc168f658410a3c5239855d829c498e4f0b40b06a370816ba84R1081-R1099)
[[4]](diffhunk://#diff-cd2bb47b50681fc168f658410a3c5239855d829c498e4f0b40b06a370816ba84R1931-R1932)
[[5]](diffhunk://#diff-cd2bb47b50681fc168f658410a3c5239855d829c498e4f0b40b06a370816ba84R2181-R2185)

Metrics Extraction Utilities:

* Added `extract_base_latency.py` to extract and export baseline request
latency data from metrics JSON to CSV, including finish time and e2e
latency.
* Added `extract_base_utilization.py` to process and export average GPU
utilization and memory occupancy rates from controller monitoring
samples to CSV for further analysis.

Author: zhtshr

configs/disagg/baseline/wan22_moe_i2v_baseline.json

@@ -0,0 +1,48 @@
+{
+    "infer_steps": 4,
+    "in_dim": 36,
+    "dim": 5120,
+    "ffn_dim": 13824,
+    "freq_dim": 256,
+    "num_heads": 40,
+    "num_layers": 40,
+    "out_dim": 16,
+    "eps": 1e-06,
+    "model_type": "i2v",
+    "target_video_length": 81,
+    "text_len": 512,
+    "target_height": 480,
+    "target_width": 832,
+    "self_attn_1_type": "sage_attn2",
+    "cross_attn_1_type": "sage_attn2",
+    "cross_attn_2_type": "sage_attn2",
+    "sample_guide_scale": [
+        3.5,
+        3.5
+    ],
+    "sample_shift": 5.0,
+    "enable_cfg": false,
+    "cpu_offload": true,
+    "offload_granularity": "block",
+    "t5_cpu_offload": false,
+    "vae_cpu_offload": false,
+    "fps": 16,
+    "use_image_encoder": false,
+    "boundary_step_index": 2,
+    "denoising_step_list": [
+        1000,
+        750,
+        500,
+        250
+    ],
+    "dit_quantized": true,
+    "dit_quant_scheme": "int8-q8f",
+    "high_noise_quantized_ckpt": "/root/zht/LightX2V/models/lightx2v/Wan2.2-Distill-Models/wan2.2_i2v_A14b_high_noise_int8_lightx2v_4step.safetensors",
+    "low_noise_quantized_ckpt": "/root/zht/LightX2V/models/lightx2v/Wan2.2-Distill-Models/wan2.2_i2v_A14b_low_noise_int8_lightx2v_4step.safetensors",
+    "high_noise_original_ckpt": "/root/zht/LightX2V/models/lightx2v/Wan2.2-Distill-Models/wan2.2_i2v_A14b_high_noise_int8_lightx2v_4step.safetensors",
+    "low_noise_original_ckpt": "/root/zht/LightX2V/models/lightx2v/Wan2.2-Distill-Models/wan2.2_i2v_A14b_low_noise_int8_lightx2v_4step.safetensors",
+    "parallel": {
+        "seq_p_size": 4,
+        "seq_p_attn_type": "ulysses"
+    }
+}

configs/disagg/baseline/wan22_moe_i2v_baseline_50steps.json

@@ -0,0 +1,36 @@
+{
+    "model_path": "/root/zht/LightX2V/models/Wan-AI/Wan2.2-I2V-A14B",
+    "infer_steps": 50,
+    "in_dim": 36,
+    "dim": 5120,
+    "ffn_dim": 13824,
+    "freq_dim": 256,
+    "num_heads": 40,
+    "num_layers": 40,
+    "out_dim": 16,
+    "eps": 1e-06,
+    "model_type": "i2v",
+    "target_video_length": 81,
+    "text_len": 512,
+    "target_height": 480,
+    "target_width": 832,
+    "self_attn_1_type": "sage_attn2",
+    "cross_attn_1_type": "sage_attn2",
+    "cross_attn_2_type": "sage_attn2",
+    "sample_guide_scale": [
+        3.5,
+        3.5
+    ],
+    "sample_shift": 5.0,
+    "enable_cfg": false,
+    "cpu_offload": true,
+    "offload_granularity": "block",
+    "t5_cpu_offload": false,
+    "vae_cpu_offload": false,
+    "fps": 16,
+    "use_image_encoder": false,
+    "parallel": {
+        "seq_p_size": 8,
+        "seq_p_attn_type": "ulysses"
+    }
+}

lightx2v/disagg/examples/infer.py

@@ -0,0 +1,245 @@
+import argparse
+import os
+
+import torch
+import torch.distributed as dist
+from loguru import logger
+
+from lightx2v.common.ops import *
+from lightx2v.models.runners.bagel.bagel_runner import BagelRunner  # noqa: F401
+from lightx2v.models.runners.flux2.flux2_runner import Flux2DevRunner, Flux2KleinRunner  # noqa: F401
+from lightx2v.models.runners.hunyuan_video.hunyuan_video_15_distill_runner import HunyuanVideo15DistillRunner  # noqa: F401
+from lightx2v.models.runners.hunyuan_video.hunyuan_video_15_runner import HunyuanVideo15Runner  # noqa: F401
+from lightx2v.models.runners.longcat_image.longcat_image_runner import LongCatImageRunner  # noqa: F401
+from lightx2v.models.runners.ltx2.ltx2_runner import LTX2Runner  # noqa: F401
+from lightx2v.models.runners.neopp.neopp_runner import NeoppRunner  # noqa: F401
+from lightx2v.models.runners.qwen_image.qwen_image_runner import QwenImageRunner  # noqa: F401
+from lightx2v.models.runners.seedvr.seedvr_runner import SeedVRRunner  # noqa: F401
+from lightx2v.models.runners.wan.wan_animate_runner import WanAnimateRunner  # noqa: F401
+from lightx2v.models.runners.wan.wan_audio_runner import Wan22AudioRunner, WanAudioRunner  # noqa: F401
+from lightx2v.models.runners.wan.wan_distill_runner import WanDistillRunner  # noqa: F401
+from lightx2v.models.runners.wan.wan_matrix_game2_runner import WanSFMtxg2Runner  # noqa: F401
+from lightx2v.models.runners.wan.wan_matrix_game3_runner import WanMatrixGame3Runner  # noqa: F401
+from lightx2v.models.runners.wan.wan_runner import Wan22MoeRunner, WanRunner  # noqa: F401
+from lightx2v.models.runners.wan.wan_sf_runner import WanSFRunner  # noqa: F401
+from lightx2v.models.runners.wan.wan_vace_runner import Wan22MoeVaceRunner, WanVaceRunner  # noqa: F401
+from lightx2v.models.runners.worldplay.worldplay_ar_runner import WorldPlayARRunner  # noqa: F401
+from lightx2v.models.runners.worldplay.worldplay_bi_runner import WorldPlayBIRunner  # noqa: F401
+from lightx2v.models.runners.worldplay.worldplay_distill_runner import WorldPlayDistillRunner  # noqa: F401
+from lightx2v.models.runners.z_image.z_image_runner import ZImageRunner  # noqa: F401
+from lightx2v.utils.envs import *
+from lightx2v.utils.input_info import init_empty_input_info, update_input_info_from_dict
+from lightx2v.utils.profiler import *
+from lightx2v.utils.registry_factory import RUNNER_REGISTER
+from lightx2v.utils.set_config import print_config, set_config, set_parallel_config
+from lightx2v.utils.utils import seed_all, validate_config_paths
+from lightx2v_platform.registry_factory import PLATFORM_DEVICE_REGISTER
+
+try:
+    from lightx2v.models.runners.worldmirror.worldmirror_runner import WorldMirrorRunner  # noqa: F401
+except Exception as exc:  # pragma: no cover - optional dependency guard
+    logger.warning("WorldMirrorRunner import skipped: {}", exc)
+
+
+def init_runner(config):
+    torch.set_grad_enabled(False)
+    runner = RUNNER_REGISTER[config["model_cls"]](config)
+    runner.init_modules()
+    return runner
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--seed", type=int, default=42, help="The seed for random generator")
+    parser.add_argument(
+        "--model_cls",
+        type=str,
+        required=True,
+        choices=[
+            "wan2.1",
+            "wan2.1_distill",
+            "wan2.1_mean_flow_distill",
+            "wan2.1_vace",
+            "wan2.1_sf",
+            "wan2.1_sf_mtxg2",
+            "seko_talk",
+            "wan2.2_moe",
+            "lingbot_world",
+            "wan2.2",
+            "wan2.2_matrix_game3",
+            "wan2.2_moe_audio",
+            "wan2.2_audio",
+            "wan2.2_moe_distill",
+            "wan2.2_moe_vace",
+            "qwen_image",
+            "longcat_image",
+            "wan2.2_animate",
+            "hunyuan_video_1.5",
+            "hunyuan_video_1.5_distill",
+            "worldplay_distill",
+            "worldplay_ar",
+            "worldplay_bi",
+            "z_image",
+            "flux2_klein",
+            "flux2_dev",
+            "ltx2",
+            "bagel",
+            "seedvr2",
+            "neopp",
+            "lingbot_world_fast",
+            "worldmirror",
+        ],
+        default="wan2.1",
+    )
+
+    parser.add_argument("--task", type=str, choices=["t2v", "i2v", "t2i", "i2i", "flf2v", "vace", "animate", "s2v", "rs2v", "t2av", "i2av", "ltx2_s2v", "sr", "recon"], default="t2v")
+    parser.add_argument("--support_tasks", type=str, nargs="+", default=[], help="Set supported tasks for the model")
+    parser.add_argument("--model_path", type=str, required=True)
+    parser.add_argument("--sf_model_path", type=str, required=False)
+    parser.add_argument("--config_json", type=str, required=True)
+    parser.add_argument("--use_prompt_enhancer", action="store_true")
+
+    parser.add_argument("--prompt", type=str, default="", help="The input prompt for text-to-video generation")
+    parser.add_argument("--negative_prompt", type=str, default="")
+
+    parser.add_argument(
+        "--image_path",
+        type=str,
+        default="",
+        help="The path to input image file(s) for image-to-video (i2v) or image-to-audio-video (i2av) task. Multiple paths should be comma-separated. Example: 'path1.jpg,path2.jpg'",
+    )
+    parser.add_argument("--last_frame_path", type=str, default="", help="The path to last frame file for first-last-frame-to-video (flf2v) task")
+    parser.add_argument(
+        "--audio_path",
+        type=str,
+        default="",
+        help="Input audio path: Wan s2v / rs2v, or required for LTX-2 task ltx2_s2v.",
+    )
+    parser.add_argument("--image_strength", type=str, default="1.0", help="i2av: single float, or comma-separated floats (one per image, or one value broadcast). Example: 1.0 or 1.0,0.85,0.9")
+    parser.add_argument(
+        "--image_frame_idx", type=str, default="", help="i2av: comma-separated pixel frame indices (one per image). Omit or empty to evenly space frames in [0, num_frames-1]. Example: 0,40,80"
+    )
+    # [Warning] For vace task, need refactor.
+    parser.add_argument(
+        "--src_ref_images",
+        type=str,
+        default=None,
+        help="The file list of the source reference images. Separated by ','. Default None.",
+    )
+    parser.add_argument(
+        "--src_video",
+        type=str,
+        default=None,
+        help="The file of the source video. Default None.",
+    )
+    parser.add_argument(
+        "--src_mask",
+        type=str,
+        default=None,
+        help="The file of the source mask. Default None.",
+    )
+    parser.add_argument(
+        "--src_pose_path",
+        type=str,
+        default=None,
+        help="The file of the source pose. Default None.",
+    )
+    parser.add_argument(
+        "--src_face_path",
+        type=str,
+        default=None,
+        help="The file of the source face. Default None.",
+    )
+    parser.add_argument(
+        "--src_bg_path",
+        type=str,
+        default=None,
+        help="The file of the source background. Default None.",
+    )
+    parser.add_argument(
+        "--src_mask_path",
+        type=str,
+        default=None,
+        help="The file of the source mask. Default None.",
+    )
+    parser.add_argument(
+        "--pose",
+        type=str,
+        default=None,
+        help="Pose string (e.g., 'w-3, right-0.5') or JSON file path for WorldPlay models.",
+    )
+    parser.add_argument(
+        "--action_path",
+        type=str,
+        default=None,
+        help="Directory path for lingbot camera/action control files (poses.npy, intrinsics.npy, optional action.npy).",
+    )
+    parser.add_argument(
+        "--action_ckpt",
+        type=str,
+        default=None,
+        help="Path to action model checkpoint for WorldPlay models.",
+    )
+    # WorldMirror (3D reconstruction) specific
+    parser.add_argument("--input_path", type=str, default=None, help="(worldmirror/recon) Path to a directory of images, a video file, or a single image.")
+    parser.add_argument("--strict_output_path", type=str, default=None, help="(worldmirror/recon) If set, write outputs directly here instead of under save_result_path/<subdir>/<timestamp>/.")
+    parser.add_argument("--prior_cam_path", type=str, default=None, help="(worldmirror/recon) Optional camera prior JSON (extrinsics + intrinsics).")
+    parser.add_argument("--prior_depth_path", type=str, default=None, help="(worldmirror/recon) Optional depth prior directory (one .npy/.png per image).")
+    parser.add_argument("--subfolder", type=str, default=None, help="(worldmirror/recon) Subfolder inside model_path containing weights. Overrides config.")
+    parser.add_argument("--disable_heads", type=str, nargs="*", default=None, help="(worldmirror/recon) Heads to disable: any of camera depth normal points gs.")
+    parser.add_argument("--enable_bf16", action="store_true", default=False, help="(worldmirror/recon) Run the WorldMirror model in bf16.")
+    parser.add_argument("--save_rendered", action="store_true", default=False, help="(worldmirror/recon) Render an interpolated fly-through video from Gaussian splats.")
+    parser.add_argument("--render_interp_per_pair", type=int, default=None, help="(worldmirror/recon) Interpolated frames per camera pair for --save_rendered.")
+    parser.add_argument("--render_depth", action="store_true", default=False, help="(worldmirror/recon) Also render a depth video with --save_rendered.")
+    parser.add_argument("--wm_config_path", type=str, default=None, help="(worldmirror/recon) Optional training YAML (pair with --wm_ckpt_path).")
+    parser.add_argument("--wm_ckpt_path", type=str, default=None, help="(worldmirror/recon) Optional .ckpt/.safetensors (pair with --wm_config_path).")
+
+    parser.add_argument("--save_result_path", type=str, default=None, help="The path to save video path/file")
+    parser.add_argument("--return_result_tensor", action="store_true", help="Whether to return result tensor. (Useful for comfyui)")
+    parser.add_argument("--target_shape", type=int, nargs="+", default=[], help="Set return video or image shape")
+    parser.add_argument("--target_video_length", type=int, default=81, help="The target video length for each generated clip")
+    parser.add_argument("--aspect_ratio", type=str, default="")
+    parser.add_argument("--video_path", type=str, default=None, help="input video path(for sr/v2v task)")
+    parser.add_argument("--sr_ratio", type=float, default=2.0, help="super resolution ratio for sr task")
+    parser.add_argument(
+        "--num_iterations",
+        type=int,
+        default=None,
+        help="Override the number of Matrix-Game-3 generation segments. Final video length follows 57 + 40 * (num_iterations - 1).",
+    )
+
+    args = parser.parse_args()
+    # validate_task_arguments(args)
+
+    seed_all(args.seed)
+
+    # set config
+    config = set_config(args)
+    # init input_info
+    input_info = init_empty_input_info(args.task, args.support_tasks)
+
+    if config["parallel"]:
+        platform_device = PLATFORM_DEVICE_REGISTER.get(os.getenv("PLATFORM", "cuda"), None)
+        platform_device.init_parallel_env()
+        set_parallel_config(config)
+
+    print_config(config)
+
+    validate_config_paths(config)
+
+    with ProfilingContext4DebugL1("Total Cost"):
+        # init runner
+        runner = init_runner(config)
+        # start to infer
+        data = args.__dict__
+        update_input_info_from_dict(input_info, data)
+        runner.run_pipeline(input_info)
+
+    # Clean up distributed process group
+    if dist.is_initialized():
+        dist.destroy_process_group()
+        logger.info("Distributed process group cleaned up")
+
+
+if __name__ == "__main__":
+    main()