Add SingleStream text2video loadgen integration

text_to_video/wan2.2-t2v-14b/README.md

@@ -28,6 +28,7 @@ cd inference/text_to_video
 - `download_model.py` - Model download
 - `run_inference.py` - Video generation  
 - `run_evaluation.py` - VBench evaluation
+- `run_mlperf.py` - Run mlperf loadgen using Wan2.2 T2V-A14B-Diffusers model
 - `launch.sh` - Docker launcher
 - `data/vbench_prompts.txt` - Input prompts
 - `data/fixed_latent.pt` - Fixed latent tensor for deterministic generation

text_to_video/wan2.2-t2v-14b/run_mlperf.py

@@ -0,0 +1,346 @@
+
+from diffusers import WanPipeline, AutoencoderKLWan
+import argparse
+import yaml
+import json
+import logging
+import os
+import torch
+import array
+import numpy as np
+import mlperf_loadgen as lg
+from pathlib import Path
+
+SCENARIO_MAP = {
+    "SingleStream": lg.TestScenario.SingleStream,
+    "MultiStream": lg.TestScenario.MultiStream,
+    "Server": lg.TestScenario.Server,
+    "Offline": lg.TestScenario.Offline,
+}
+
+NANO_SEC = 1e9
+MILLI_SEC = 1000
+
+
+def setup_logging(rank):
+    """Setup logging configuration for data parallel (all ranks log)."""
+    logging.basicConfig(
+        level=logging.INFO,
+        format=f'[Rank {rank}] %(asctime)s - %(levelname)s - %(message)s',
+        datefmt='%Y-%m-%d %H:%M:%S'
+    )
+
+
+def load_config(config_path):
+    """Load configuration from YAML file."""
+    with open(config_path, 'r') as f:
+        config = yaml.safe_load(f)
+    return config
+
+
+def load_prompts(dataset_path):
+    """Load prompts from dataset file."""
+    with open(dataset_path, 'r') as f:
+        prompts = [line.strip() for line in f if line.strip()]
+    return prompts
+
+
+class Model:
+    def __init__(
+        self, model_path, device, config, prompts, fixed_latent=None, rank=0
+    ):
+        self.device = device
+        self.rank = rank
+        self.height = config['height']
+        self.width = config['width']
+        self.num_frames = config['num_frames']
+        self.fps = config['fps']
+        self.guidance_scale = config['guidance_scale']
+        self.guidance_scale_2 = config['guidance_scale_2']
+        self.boundary_ratio = config['boundary_ratio']
+        self.negative_prompt = config['negative_prompt'].strip()
+        self.sample_steps = config['sample_steps']
+        self.base_seed = config['seed']
+        self.vae = AutoencoderKLWan.from_pretrained(
+            model_path,
+            subfolder="vae",
+            torch_dtype=torch.float32
+        )
+        self.pipe = WanPipeline.from_pretrained(
+            model_path,
+            boundary_ratio=self.boundary_ratio,
+            vae=self.vae,
+            torch_dtype=torch.bfloat16
+        )
+        self.pipe.to(self.device)
+        self.prompts = prompts
+        self.fixed_latent = fixed_latent
+
+    def issue_queries(self, query_samples):
+        if self.rank == 0:
+            idx = [q.index for q in query_samples]
+            query_ids = [q.id for q in query_samples]
+            response = []
+            for i, q in zip(idx, query_ids):
+                pipeline_kwargs = {
+                    "prompt": self.prompts[i],
+                    "negative_prompt": self.negative_prompt,
+                    "height": self.height,
+                    "width": self.width,
+                    "num_frames": self.num_frames,
+                    "guidance_scale": self.guidance_scale,
+                    "guidance_scale_2": self.guidance_scale_2,
+                    "num_inference_steps": self.sample_steps,
+                    "generator": torch.Generator(device=self.device).manual_seed(self.base_seed),
+                }
+                if self.fixed_latent is not None:
+                    pipeline_kwargs["latents"] = self.fixed_latent
+                output = self.pipe(**pipeline_kwargs).frames[0]
+                response_array = array.array(
+                    "B", output.cpu().detach().numpy().tobytes()
+                )
+                bi = response_array.buffer_info()
+                response.append(lg.QuerySampleResponse(q, bi[0], bi[1]))
+            lg.QuerySamplesComplete(response)
+
+    def flush_queries(self):
+        pass
+
+
+class DebugModel:
+    def __init__(
+        self, model_path, device, config, prompts, fixed_latent=None, rank=0
+    ):
+        self.prompts = prompts
+
+    def issue_queries(self, query_samples):
+        idx = [q.index for q in query_samples]
+        query_ids = [q.id for q in query_samples]
+        response = []
+        for i, q in zip(idx, query_ids):
+            print(i, self.prompts[i])
+            output = self.prompts[i]
+            response_array = array.array(
+                "B", output.encode("utf-8")
+            )
+            bi = response_array.buffer_info()
+            response.append(lg.QuerySampleResponse(q, bi[0], bi[1]))
+        lg.QuerySamplesComplete(response)
+
+    def flush_queries(self):
+        pass
+
+
+def load_query_samples(sample_list):
+    pass
+
+
+def unload_query_samples(sample_list):
+    pass
+
+
+def get_args():
+    parser = argparse.ArgumentParser(
+        description="Batch T2V inference with Wan2.2-Diffusers")
+    # Model Arguments
+    parser.add_argument(
+        "--model-path",
+        type=str,
+        default="./models/Wan2.2-T2V-A14B-Diffusers",
+        help="Path to model checkpoint directory (default: ./models/Wan2.2-T2V-A14B-Diffusers)"
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        default="./data/vbench_prompts.txt",
+        help="Path to dataset file (text prompts, one per line) (default: ./data/prompts.txt)"
+    )
+    parser.add_argument(
+        "--output-dir",
+        type=str,
+        default="./output",
+        help="Directory to save generated videos (default: ./data/outputs)"
+    )
+    parser.add_argument(
+        "--config",
+        type=str,
+        default="./inference_config.yaml",
+        help="Path to inference configuration file (default: ./inference_config.yaml)"
+    )
+    parser.add_argument(
+        "--num-iterations",
+        type=int,
+        default=1,
+        help="Number of generation iterations per prompt (default: 1)"
+    )
+    parser.add_argument(
+        "--num-prompts",
+        type=int,
+        default=-1,
+        help="Process only first N prompts (for testing, default: all)"
+    )
+    parser.add_argument(
+        "--fixed-latent",
+        type=str,
+        default="./data/fixed_latent.pt",
+        help="Path to fixed latent .pt file for deterministic generation (default: data/fixed_latent.pt)"
+    )
+    # MLPerf loadgen arguments
+    parser.add_argument(
+        "--scenario",
+        default="SingleStream",
+        help="mlperf benchmark scenario, one of " +
+        str(list(SCENARIO_MAP.keys())),
+    )
+    parser.add_argument(
+        "--user_conf",
+        default="user.conf",
+        help="user config for user LoadGen settings such as target QPS",
+    )
+    parser.add_argument(
+        "--audit_conf", default="audit.config", help="config for LoadGen audit settings"
+    )
+    parser.add_argument(
+        "--performance-sample-count",
+        type=int,
+        help="performance sample count",
+        default=5000,
+    )
+    parser.add_argument(
+        "--accuracy",
+        action="store_true",
+        help="enable accuracy pass"
+    )
+    # Dont overwrite these for official submission
+    parser.add_argument("--count", type=int, help="dataset items to use")
+    parser.add_argument("--time", type=int, help="time to scan in seconds")
+    parser.add_argument("--qps", type=int, help="target qps")
+    parser.add_argument("--debug", action="store_true", help="debug")
+    parser.add_argument(
+        "--samples-per-query",
+        default=8,
+        type=int,
+        help="mlperf multi-stream samples per query",
+    )
+    parser.add_argument(
+        "--max-latency", type=float, help="mlperf max latency in pct tile"
+    )
+
+    return parser.parse_args()
+
+
+def run_mlperf(args, config):
+    # Load dataset
+    dataset = load_prompts(args.dataset)
+
+    # Load model parameters
+    # Parallelism parameters
+    world_size = int(os.environ.get("WORLD_SIZE", 1))
+    rank = int(os.environ.get("RANK", 0))
+    local_rank = int(os.environ.get("LOCAL_RANK", 0))
+
+    torch.cuda.set_device(local_rank)
+    device = torch.device(f"cuda:{local_rank}")
+    setup_logging(rank)
+
+    # Generation parameters from config
+
+    output_dir = Path(args.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+    output_dir_lg = str(args.output_dir)
+
+    fixed_latent = None
+    if args.fixed_latent:
+        fixed_latent = torch.load(args.fixed_latent)
+        logging.info(
+            f"Loaded fixed latent from {args.fixed_latent} with shape: {fixed_latent.shape}")
+        logging.info(f"This latent will be reused for all generations")
+    else:
+        logging.info("No fixed latent provided - using random initial latents")
+
+    # Loading model
+    model = Model(args.model_path, device, config, dataset, fixed_latent, rank)
+    # model = DebugModel(args.model_path, device, config, dataset, fixed_latent, rank)
+    logging.info("Model loaded successfully!")
+
+    # Prepare loadgen for run
+    if rank == 0:
+        log_output_settings = lg.LogOutputSettings()
+        log_output_settings.outdir = output_dir_lg
+        log_output_settings.copy_summary_to_stdout = False
+
+        log_settings = lg.LogSettings()
+        log_settings.enable_trace = args.debug
+        log_settings.log_output = log_output_settings
+
+        user_conf = os.path.abspath(args.user_conf)
+        settings = lg.TestSettings()
+        settings.FromConfig(user_conf, "wan2.2-t2v-14b", args.scenario)
+
+        audit_config = os.path.abspath(args.audit_conf)
+        if os.path.exists(audit_config):
+            settings.FromConfig(audit_config, "wan2.2-t2v-14b", args.scenario)
+        settings.scenario = SCENARIO_MAP[args.scenario]
+
+        settings.mode = lg.TestMode.PerformanceOnly
+        if args.accuracy:
+            settings.mode = lg.TestMode.AccuracyOnly
+
+        if args.time:
+            # override the time we want to run
+            settings.min_duration_ms = args.time * MILLI_SEC
+            settings.max_duration_ms = args.time * MILLI_SEC
+        if args.qps:
+            qps = float(args.qps)
+            settings.server_target_qps = qps
+            settings.offline_expected_qps = qps
+
+        count_override = False
+        count = args.count
+        if count:
+            count_override = True
+
+        if args.count:
+            settings.min_query_count = count
+            settings.max_query_count = count
+        count = len(dataset)
+
+        if args.samples_per_query:
+            settings.multi_stream_samples_per_query = args.samples_per_query
+        if args.max_latency:
+            settings.server_target_latency_ns = int(
+                args.max_latency * NANO_SEC)
+            settings.multi_stream_expected_latency_ns = int(
+                args.max_latency * NANO_SEC)
+
+        performance_sample_count = (
+            args.performance_sample_count
+            if args.performance_sample_count
+            else min(count, 500)
+        )
+
+        sut = lg.ConstructSUT(model.issue_queries, model.flush_queries)
+        qsl = lg.ConstructQSL(
+            count, performance_sample_count, load_query_samples, unload_query_samples
+        )
+
+        lg.StartTestWithLogSettings(
+            sut, qsl, settings, log_settings, audit_config)
+        if args.accuracy:
+            # TODO: output accuracy
+            final_results = {}
+            with open("results.json", "w") as f:
+                json.dump(final_results, f, sort_keys=True, indent=4)
+
+        lg.DestroyQSL(qsl)
+        lg.DestroySUT(sut)
+
+
+def main():
+    args = get_args()
+    config = load_config(args.config)
+    run_mlperf(args, config)
+
+
+if __name__ == "__main__":
+    main()

text_to_video/wan2.2-t2v-14b/user.conf

@@ -0,0 +1,4 @@
+# The format of this config file is 'key = value'.
+# The key has the format 'model.scenario.key'. Value is mostly int64_t.
+# Model maybe '*' as wildcard. In that case the value applies to all models.
+# All times are in milli seconds