Add files using upload-large-folder tool

FlowCache/.gitignore

@@ -0,0 +1,8 @@
+__pycache__
+*.pyc
+*.log
+*.pt
+*.mp4
+ckpt
+downloads
+*.whl

FlowCache/FlowCache4MAGI-1-dev-V1/README.md

@@ -0,0 +1,71 @@
+# FLOW CACHING FOR AUTOREGRESSIVE VIDEO GENERATION
+
+This repository provides the official implementation of **FlowCache** on **MAGI-1** model, a caching-based acceleration method for autoregressive video generation models.
+
+
+## 🚀 Installation
+
+Please follow the installation instructions provided in the [MAGI-1](https://github.com/SandAI-org/MAGI-1), as this implementation is built on top of MAGI-1.
+
+---
+
+## ▶️ Usage
+
+### 1. Single Video Generation
+
+Run accelerated generation using FlowCache:
+
+```bash
+# FlowCache for text-to-video generation
+bash scripts/single_run/flowcache_t2v.sh
+
+# FlowCache for video-to-video generation
+bash scripts/single_run/flowcache_v2v.sh
+
+# Baseline acceleration method (TeaCache) for text-to-video
+bash scripts/single_run/teacache_t2v.sh
+
+# Baseline acceleration method (TeaCache) for video-to-video
+bash scripts/single_run/teacache_v2v.sh
+```
+
+### 2. Benchmark Sampling
+
+Generate videos for evaluation on standard benchmarks:
+
+```bash
+# VBench
+bash scripts/sample/flowcache_vbench.sh
+bash scripts/sample/teacache_vbench.sh
+
+# PhysicsIQ
+bash scripts/sample/flowcache_physicsiq.sh
+bash scripts/sample/teacache_physicsiq.sh
+```
+
+### 3. Quality Evaluation
+
+Compute perceptual and structural similarity metrics between original and accelerated generations:
+
+```bash
+bash scripts/metric.sh
+```
+
+---
+
+## ⚙️ Key Parameters
+
+| Parameter | Description |
+|----------|-------------|
+| `rel_l1_thresh` | Relative L1 distance threshold for cache reuse decision |
+| ` warmup_steps` | Number of denoising steps where reuse is disabled |
+| `total_cache_chunk_nums` (`B_total`) | Total number of cache chunks maintained |
+| `compress_strategy` | Granularity for selecting important KV caches: `token`, `frame`, or `chunk` |
+| `query_granularity` | Granularity for importance scoring: `token`, `frame`, or `chunk` |
+| `mix_lambda` | Weight balancing importance and redundancy (default: `0.07`) |
+| `mode` | Generation mode: `t2v` (text-to-video), `i2v` (image-to-video), or `v2v` (video-to-video) |
+| `prompt` | Input prompt for conditional generation |
+| `output_path` | Path to save generated videos |
+| `config_file` | Path to MAGI-1 model configuration |
+
+---

FlowCache/FlowCache4MAGI-1-dev-V1/requirements.txt

@@ -0,0 +1,21 @@
+accelerate==0.32.1
+beautifulsoup4==4.13.4
+debugpy==1.8.14
+diffusers==0.29.2
+einops>=0.6.0
+ffmpeg-python
+# flash-attn==2.4.2
+flashinfer-python==0.2.0.post2 --extra-index-url https://flashinfer.ai/whl/cu124/torch2.4/
+ftfy==6.2.0
+gpustat==1.1.1
+imageio==2.34.0
+imageio[ffmpeg]
+matplotlib==3.10.1
+numpy==1.26.4
+protobuf==5.28.3
+rich==14.0.0
+sentencepiece==0.2.0
+timm==1.0.15
+torchdiffeq==0.2.4
+transformers==4.42.3
+tqdm

FlowCache/FlowCache4MAGI-1-dev-V1/sample_video.py

@@ -0,0 +1,403 @@
+# Copyright 2024 MAGI Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import re
+import sys
+import argparse
+import csv
+import subprocess
+from pathlib import Path
+
+import multiprocessing as mp
+
+# Constants
+DEFAULT_BASE_PORT = 29510
+PHYSICSIQ_FPS = 24
+
+
+def load_yaml_config(yaml_path: str) -> dict:
+    """Load configuration from YAML file."""
+    import yaml
+
+    with open(yaml_path, "r") as f:
+        return yaml.safe_load(f)
+
+
+def apply_slice(items: list, start: int | None, end: int | None) -> list:
+    """Apply start/end slice to a list with bounds checking."""
+    if start is None and end is None:
+        return items
+
+    slice_start = max(0, start if start is not None else 0)
+    slice_end = min(end if end is not None else len(items), len(items))
+    slice_end = max(slice_start, slice_end)
+
+    return items[slice_start:slice_end]
+
+
+def configure_teacache(transport, config: dict) -> None:
+    """Configure TeaCache reuse strategy on SampleTransport."""
+    from inference.pipeline.teacache import (
+        teacache_forward_velocity,
+        teacache_integrate_velocity,
+    )
+
+    transport.rel_l1_thresh = config["rel_l1_thresh"]
+    transport.accumulated_rel_l1_distance = 0
+    transport.previous_modulated_input = None
+    transport.previous_residual = None
+    transport.cnt = 0
+    transport.forward_velocity = teacache_forward_velocity
+    transport.integrate_velocity = teacache_integrate_velocity
+    transport.reuse_times = 0
+    transport.warmup_steps = config["warmup_steps"]
+    transport.previous_output = None
+    transport.log = config.get("log", False)
+
+
+def configure_kv_cache(transport, config: dict) -> None:
+    """Configure KV cache compression if enabled."""
+    if not config.get("compress_kv_cache", False):
+        transport.compress_kv_cache = False
+        return
+
+    print("KV cache compression is enabled.")
+    transport.compress_kv_cache = True
+
+    assert config.get("total_cache_chunk_nums") is not None
+
+    compression_config = {
+        "method_config": {
+            "compress_strategy": config["compress_strategy"],
+            "mix_lambda": config["mix_lambda"],
+            "query_granularity": config["query_granularity"],
+            "score_weighting_method": config.get("score_weighting_method"),
+            "power": config.get("power", 3),
+        },
+    }
+
+    from inference.pipeline.kvcompress import replace_magi
+
+    replace_magi(compression_config)
+
+
+def configure_flowcache(transport, config: dict) -> None:
+    """Configure FlowCache reuse strategy on SampleTransport."""
+    from inference.pipeline.flowcache import (
+        flowcache_forward_velocity,
+        flowcache_integrate_velocity,
+    )
+
+    configure_kv_cache(transport, config)
+
+    transport.rel_l1_thresh = config["rel_l1_thresh"]
+    transport.chunk_accumulated_rel_l1 = 0
+    transport.previous_modulated_input = None
+    transport.previous_residual = None
+    transport.cnt = 0
+    transport.forward_velocity = flowcache_forward_velocity
+    transport.integrate_velocity = flowcache_integrate_velocity
+    transport.reuse_times = 0
+    transport.warmup_steps = config["warmup_steps"]
+    transport.previous_output = None
+    transport.discard_nearly_clean_chunk = config.get("discard_nearly_clean_chunk", False)
+    transport.chunk_accumulated_rel_l1 = None
+    transport.prev_chunk_features = None
+    transport.chunk_reuse_flags = None
+    transport.total_cache_chunk_nums = config.get("total_cache_chunk_nums")
+    transport.log = config.get("log", False)
+
+
+def configure_reuse_strategy(config: dict) -> None:
+    """Configure the appropriate reuse strategy on SampleTransport."""
+    from inference.pipeline.video_generate import SampleTransport
+
+    strategy = config["reuse_strategy"]
+
+    if strategy == "original":
+        return
+    if strategy == "all":
+        configure_teacache(SampleTransport, config)
+    elif strategy == "chunkwise":
+        configure_flowcache(SampleTransport, config)
+    else:
+        raise ValueError(f"Unknown reuse strategy: {strategy}")
+
+
+def setup_environment(gpu_id: int) -> None:
+    """Set up environment variables for a GPU worker process."""
+    os.environ["CUDA_VISIBLE_DEVICES"] = str(gpu_id)
+    os.environ["WORLD_SIZE"] = "1"
+    os.environ["RANK"] = "0"
+    os.environ["LOCAL_RANK"] = "0"
+    os.environ["MASTER_ADDR"] = "localhost"
+    os.environ["MASTER_PORT"] = str(DEFAULT_BASE_PORT + gpu_id)
+
+    # Enable pdb terminal debugging
+    sys.stdin = open(0)
+
+
+def filter_existing_samples(samples: list, config: dict) -> list:
+    """Filter out samples whose output files already exist."""
+    if config["benchmark"] == "vbench":
+        return [
+            sample
+            for sample in samples
+            if not os.path.exists(os.path.abspath(os.path.join(config["save_path"], f"{sample}-0.mp4")))
+        ]
+    else:  # physicsiq
+        return [
+            sample for sample in samples if not os.path.exists(sample["output_path"])
+        ]
+
+
+def assign_samples_to_gpu(
+    samples: list, gpu_id: int, rank: int, num_gpus: int
+) -> list:
+    """Divide samples across GPUs and return the subset for this GPU."""
+    samples_per_gpu = (len(samples) + num_gpus - 1) // num_gpus
+    start_idx = rank * samples_per_gpu
+    end_idx = min(start_idx + samples_per_gpu, len(samples))
+    return samples[start_idx:end_idx]
+
+
+def process_vbench_sample(pipeline, prompt: str, config: dict, gpu_id: int) -> None:
+    """Process a single vbench text-to-video sample."""
+    output_path = os.path.abspath(os.path.join(config["save_path"], f"{prompt}-0.mp4"))
+
+    if os.path.exists(output_path):
+        print(f"[SKIP GPU {gpu_id}] Already exists: {output_path}")
+        return
+
+    print(f"[GPU {gpu_id}] Generating T2V: '{prompt}' -> {output_path}")
+    pipeline.run_text_to_video(prompt=prompt, output_path=output_path)
+    print(f"[DONE GPU {gpu_id}] Saved: {output_path}")
+
+
+def process_physicsiq_sample(pipeline, sample: dict, gpu_id: int) -> None:
+    """Process a single PhysicsIQ video-to-video sample."""
+    prompt = sample["description"]
+    prefix_video_path = sample["prefix_video_path"]
+    output_path = sample["output_path"]
+
+    if not os.path.exists(prefix_video_path):
+        print(f"[WARN GPU {gpu_id}] Conditioning video not found: {prefix_video_path}")
+        return
+
+    if os.path.exists(output_path):
+        print(f"[SKIP GPU {gpu_id}] Already exists: {output_path}")
+        return
+
+    print(f"[GPU {gpu_id}] Generating V2V: '{prompt}'")
+    print(f"   Input:  {prefix_video_path}")
+    print(f"   Output: {output_path}")
+
+    pipeline.run_video_to_video(
+        prompt=prompt,
+        prefix_video_path=prefix_video_path,
+        output_path=output_path,
+    )
+    print(f"[DONE GPU {gpu_id}] Saved: {output_path}")
+
+
+def worker_process(gpu_id: int, rank: int, config: dict, all_samples: list) -> None:
+    """Independent worker running on each GPU."""
+    setup_environment(gpu_id)
+    configure_reuse_strategy(config)
+
+    try:
+        magi_root = subprocess.check_output(
+            ["git", "rev-parse", "--show-toplevel"]
+        ).decode().strip()
+        os.environ["MAGI_ROOT"] = magi_root
+        os.environ["PYTHONPATH"] = f"{magi_root}:{os.environ.get('PYTHONPATH', '')}"
+    except Exception as e:
+        print(f"[GPU {gpu_id}] Failed to set MAGI_ROOT: {e}")
+        return
+
+    filtered_samples = filter_existing_samples(all_samples, config)
+
+    if not filtered_samples:
+        print(f"[GPU {gpu_id}] No samples need to be generated.")
+        return
+
+    print(f"Processing {len(filtered_samples)} samples.")
+
+    my_samples = assign_samples_to_gpu(
+        filtered_samples, gpu_id, rank, config["num_gpus"]
+    )
+
+    if not my_samples:
+        print(f"[GPU {gpu_id}] No samples assigned.")
+        return
+
+    print(f"[GPU {gpu_id}] Assigned {len(my_samples)} samples")
+
+    from inference.pipeline.entry import MagiPipeline
+
+    print(f"[GPU {gpu_id}] Loading model...")
+    pipeline = MagiPipeline(config["config_file"])
+    print(f"[GPU {gpu_id}] Model loaded.")
+
+    process_func = (
+        process_vbench_sample if config["benchmark"] == "vbench" else process_physicsiq_sample
+    )
+
+    for sample in my_samples:
+        process_func(pipeline, sample, config, gpu_id)
+
+    print(f"[GPU {gpu_id}] Completed.")
+
+
+def build_conditioning_video_path(
+    data_root: str, vid_id: str, scenario: str, fps: int
+) -> str:
+    """Construct the path to the conditioning video file."""
+    conditioning_dir = os.path.join(
+        data_root, "physics-IQ-benchmark", "split-videos", "conditioning", f"{fps}FPS"
+    )
+    match_suffix = re.search(r"_(.*)", scenario)
+    suffix = match_suffix.group(1) if match_suffix else ""
+    filename = f"{vid_id}_conditioning-videos_{fps}FPS_{suffix}"
+    return os.path.join(conditioning_dir, filename)
+
+
+def load_physicsiq_samples(config: dict) -> list[dict]:
+    """Load sample list from PhysicsIQ dataset."""
+    data_root = config["physicsiq_data_dir"]
+    descriptions_csv = os.path.join(data_root, "descriptions", "descriptions.csv")
+    output_dir = config["save_path"]
+
+    if not os.path.exists(descriptions_csv):
+        raise FileNotFoundError(f"descriptions.csv not found at {descriptions_csv}")
+
+    os.makedirs(output_dir, exist_ok=True)
+
+    samples = []
+    with open(descriptions_csv, mode="r") as f:
+        reader = csv.DictReader(f)
+        for row in reader:
+            scenario = row["scenario"].strip()
+            match_id = re.match(r"^(\d+)_", scenario)
+
+            if not match_id:
+                print(f"Cannot extract ID from scenario: {scenario}")
+                continue
+
+            vid_id = match_id.group(1).zfill(4)
+            description = row["description"]
+            generated_video_name = row["generated_video_name"]
+            prefix_video_path = build_conditioning_video_path(
+                data_root, vid_id, scenario, PHYSICSIQ_FPS
+            )
+            output_path = os.path.join(output_dir, generated_video_name)
+
+            os.makedirs(os.path.dirname(output_path), exist_ok=True)
+
+            samples.append({
+                "vid_id": vid_id,
+                "scenario": scenario,
+                "description": description,
+                "generated_video_name": generated_video_name,
+                "prefix_video_path": prefix_video_path,
+                "output_path": output_path,
+            })
+
+    # PhysicsIQ samples are duplicated; take only the first half
+    unique_count = len(samples) // 2
+    samples = samples[:unique_count]
+
+    print(f"Loaded {unique_count} PhysicsIQ samples.")
+
+    return apply_slice(samples, config.get("start"), config.get("end"))
+
+
+def load_vbench_samples(config: dict) -> list[str]:
+    """Load prompt list from vbench dimension file."""
+    prompt_dir = config["vbench_prompt_dir"]
+    dimension = config.get("dimension")
+
+    if not dimension:
+        raise ValueError("For vbench, 'dimension' must be specified in config")
+
+    prompt_file = os.path.join(prompt_dir, f"{dimension}.txt")
+
+    if not os.path.exists(prompt_file):
+        raise FileNotFoundError(f"Prompt file not found: {prompt_file}")
+
+    with open(prompt_file, "r") as f:
+        prompts = [line.strip() for line in f if line.strip()]
+
+    return apply_slice(prompts, config.get("start"), config.get("end"))
+
+
+def setup_save_path(config: dict) -> None:
+    """Configure the output save path based on benchmark type."""
+    base_path = config["base_save_path"]
+
+    if config["benchmark"] == "vbench":
+        dimension = config.get("dimension")
+        videos_dir = os.path.join(base_path, "videos", dimension) if dimension else None
+        config["save_path"] = videos_dir if videos_dir else os.path.join(base_path, "videos")
+    elif config["benchmark"] == "physicsiq":
+        config["save_path"] = os.path.join(base_path, "videos")
+
+    os.makedirs(config["save_path"], exist_ok=True)
+
+
+def main() -> None:
+    """Entry point for video sampling script."""
+    parser = argparse.ArgumentParser(
+        description="Video sampling script using YAML configuration"
+    )
+    parser.add_argument("yaml_config", type=str, help="Path to YAML configuration file")
+    args = parser.parse_args()
+
+    config = load_yaml_config(args.yaml_config)
+    print(f"Loaded configuration from: {args.yaml_config}")
+
+    setup_save_path(config)
+
+    gpu_ids = list(map(int, config["gpus"].split(",")))
+    config["num_gpus"] = len(gpu_ids)
+
+    benchmark = config["benchmark"]
+    if benchmark == "vbench":
+        all_samples = load_vbench_samples(config)
+    elif benchmark == "physicsiq":
+        data_root = config["physicsiq_data_dir"]
+        if not os.path.exists(data_root):
+            raise FileNotFoundError(f"Data directory not found: {data_root}")
+        all_samples = load_physicsiq_samples(config)
+    else:
+        raise ValueError(f"Invalid benchmark: {benchmark}")
+
+    print(f"Total samples: {len(all_samples)}")
+    print(f"GPUs: {gpu_ids}")
+    print(f"Output: {config['save_path']}")
+    print(f"Config: {config['config_file']}")
+
+    processes = []
+    for rank, gpu_id in enumerate(gpu_ids):
+        p = mp.Process(target=worker_process, args=(gpu_id, rank, config, all_samples))
+        p.start()
+        processes.append(p)
+
+    for p in processes:
+        p.join()
+
+
+if __name__ == "__main__":
+    main()

FlowCache/FlowCache4MAGI-1-dev-V2/README.md

@@ -0,0 +1,71 @@
+# FLOW CACHING FOR AUTOREGRESSIVE VIDEO GENERATION
+
+This repository provides the official implementation of **FlowCache** on **MAGI-1** model, a caching-based acceleration method for autoregressive video generation models.
+
+
+## 🚀 Installation
+
+Please follow the installation instructions provided in the [MAGI-1](https://github.com/SandAI-org/MAGI-1), as this implementation is built on top of MAGI-1.
+
+---
+
+## ▶️ Usage
+
+### 1. Single Video Generation
+
+Run accelerated generation using FlowCache:
+
+```bash
+# FlowCache for text-to-video generation
+bash scripts/single_run/flowcache_t2v.sh
+
+# FlowCache for video-to-video generation
+bash scripts/single_run/flowcache_v2v.sh
+
+# Baseline acceleration method (TeaCache) for text-to-video
+bash scripts/single_run/teacache_t2v.sh
+
+# Baseline acceleration method (TeaCache) for video-to-video
+bash scripts/single_run/teacache_v2v.sh
+```
+
+### 2. Benchmark Sampling
+
+Generate videos for evaluation on standard benchmarks:
+
+```bash
+# VBench
+bash scripts/sample/flowcache_vbench.sh
+bash scripts/sample/teacache_vbench.sh
+
+# PhysicsIQ
+bash scripts/sample/flowcache_physicsiq.sh
+bash scripts/sample/teacache_physicsiq.sh
+```
+
+### 3. Quality Evaluation
+
+Compute perceptual and structural similarity metrics between original and accelerated generations:
+
+```bash
+bash scripts/metric.sh
+```
+
+---
+
+## ⚙️ Key Parameters
+
+| Parameter | Description |
+|----------|-------------|
+| `rel_l1_thresh` | Relative L1 distance threshold for cache reuse decision |
+| ` warmup_steps` | Number of denoising steps where reuse is disabled |
+| `total_cache_chunk_nums` (`B_total`) | Total number of cache chunks maintained |
+| `compress_strategy` | Granularity for selecting important KV caches: `token`, `frame`, or `chunk` |
+| `query_granularity` | Granularity for importance scoring: `token`, `frame`, or `chunk` |
+| `mix_lambda` | Weight balancing importance and redundancy (default: `0.07`) |
+| `mode` | Generation mode: `t2v` (text-to-video), `i2v` (image-to-video), or `v2v` (video-to-video) |
+| `prompt` | Input prompt for conditional generation |
+| `output_path` | Path to save generated videos |
+| `config_file` | Path to MAGI-1 model configuration |
+
+---

FlowCache/FlowCache4MAGI-1-dev-V2/requirements.txt

@@ -0,0 +1,21 @@
+accelerate==0.32.1
+beautifulsoup4==4.13.4
+debugpy==1.8.14
+diffusers==0.29.2
+einops>=0.6.0
+ffmpeg-python
+# flash-attn==2.4.2
+flashinfer-python==0.2.0.post2 --extra-index-url https://flashinfer.ai/whl/cu124/torch2.4/
+ftfy==6.2.0
+gpustat==1.1.1
+imageio==2.34.0
+imageio[ffmpeg]
+matplotlib==3.10.1
+numpy==1.26.4
+protobuf==5.28.3
+rich==14.0.0
+sentencepiece==0.2.0
+timm==1.0.15
+torchdiffeq==0.2.4
+transformers==4.42.3
+tqdm

FlowCache/FlowCache4MAGI-1-dev-V2/sample_video.py

@@ -0,0 +1,429 @@
+# Copyright 2024 MAGI Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import re
+import sys
+import argparse
+import csv
+import subprocess
+from pathlib import Path
+
+import multiprocessing as mp
+
+# Constants
+DEFAULT_BASE_PORT = 29510
+PHYSICSIQ_FPS = 24
+
+
+def resolve_gpu_ids(gpus_config) -> list[int]:
+    """Resolve explicit GPU IDs or auto-detect all currently visible GPUs."""
+    if isinstance(gpus_config, int):
+        return [gpus_config]
+
+    gpus_text = str(gpus_config).strip()
+    if not gpus_text:
+        raise ValueError("'gpus' must not be empty")
+
+    if gpus_text.lower() not in {"all", "auto"}:
+        return [int(item.strip()) for item in gpus_text.split(",") if item.strip()]
+
+    visible_devices = os.environ.get("CUDA_VISIBLE_DEVICES")
+    if visible_devices:
+        visible = [item.strip() for item in visible_devices.split(",") if item.strip()]
+        if visible and all(item.isdigit() for item in visible):
+            return [int(item) for item in visible]
+
+    try:
+        output = subprocess.check_output(
+            ["nvidia-smi", "--query-gpu=index", "--format=csv,noheader,nounits"],
+            text=True,
+            timeout=10,
+        )
+        gpu_ids = [int(line.strip()) for line in output.splitlines() if line.strip()]
+        if gpu_ids:
+            return gpu_ids
+    except Exception:
+        pass
+
+    try:
+        import torch
+
+        count = torch.cuda.device_count()
+        if count > 0:
+            return list(range(count))
+    except Exception:
+        pass
+
+    raise RuntimeError("No CUDA GPUs detected for gpus: all")
+
+
+def load_yaml_config(yaml_path: str) -> dict:
+    """Load configuration from YAML file."""
+    import yaml
+
+    with open(yaml_path, "r") as f:
+        return yaml.safe_load(f)
+
+
+def apply_slice(items: list, start: int | None, end: int | None) -> list:
+    """Apply start/end slice to a list with bounds checking."""
+    if start is None and end is None:
+        return items
+
+    slice_start = max(0, start if start is not None else 0)
+    slice_end = min(end if end is not None else len(items), len(items))
+    slice_end = max(slice_start, slice_end)
+
+    return items[slice_start:slice_end]
+
+
+def configure_teacache(transport, config: dict) -> None:
+    """Configure TeaCache reuse strategy on SampleTransport."""
+    from inference.pipeline.teacache import setup_teacache
+
+    setup_teacache(
+        rel_l1_thresh=config["rel_l1_thresh"],
+        warmup_steps=config["warmup_steps"],
+        log=config.get("log", False),
+    )
+
+
+def configure_kv_cache(transport, config: dict) -> None:
+    """Configure KV cache compression if enabled."""
+    if not config.get("compress_kv_cache", False):
+        transport.compress_kv_cache = False
+        return
+
+    print("KV cache compression is enabled.")
+    transport.compress_kv_cache = True
+
+    assert config.get("total_cache_chunk_nums") is not None
+
+    compression_config = {
+        "method_config": {
+            "compress_strategy": config["compress_strategy"],
+            "mix_lambda": config["mix_lambda"],
+            "query_granularity": config["query_granularity"],
+            "score_weighting_method": config.get("score_weighting_method") or "no_weight",
+            "power": config.get("power", 3),
+        },
+    }
+
+    from inference.pipeline.kvcompress import replace_magi
+
+    replace_magi(compression_config)
+
+
+def configure_flowcache(transport, config: dict) -> None:
+    """Configure FlowCache reuse strategy on SampleTransport."""
+    from inference.pipeline.flowcache import setup_flowcache
+
+    configure_kv_cache(transport, config)
+
+    setup_flowcache(
+        rel_l1_thresh=config["rel_l1_thresh"],
+        warmup_steps=config["warmup_steps"],
+        discard_nearly_clean_chunk=config.get("discard_nearly_clean_chunk", False),
+        log=config.get("log", False),
+        total_cache_chunk_nums=config.get("total_cache_chunk_nums", 5),
+        compress_kv_cache=config.get("compress_kv_cache", False),
+    )
+
+
+def configure_reuse_strategy(config: dict) -> None:
+    """Configure the appropriate reuse strategy on SampleTransport."""
+    from inference.pipeline.video_generate import SampleTransport
+
+    strategy = config["reuse_strategy"]
+
+    if strategy == "original":
+        return
+    if strategy == "all":
+        configure_teacache(SampleTransport, config)
+    elif strategy == "chunkwise":
+        configure_flowcache(SampleTransport, config)
+    else:
+        raise ValueError(f"Unknown reuse strategy: {strategy}")
+
+
+def setup_environment(gpu_id: int) -> None:
+    """Set up environment variables for a GPU worker process."""
+    os.environ["CUDA_VISIBLE_DEVICES"] = str(gpu_id)
+    os.environ["WORLD_SIZE"] = "1"
+    os.environ["RANK"] = "0"
+    os.environ["LOCAL_RANK"] = "0"
+    os.environ["MASTER_ADDR"] = "localhost"
+    os.environ["MASTER_PORT"] = str(DEFAULT_BASE_PORT + gpu_id)
+
+    # Enable pdb terminal debugging
+    sys.stdin = open(0)
+
+
+def filter_existing_samples(samples: list, config: dict) -> list:
+    """Filter out samples whose output files already exist."""
+    if config["benchmark"] == "vbench":
+        return [
+            sample
+            for sample in samples
+            if not os.path.exists(os.path.abspath(os.path.join(config["save_path"], f"{sample}-0.mp4")))
+        ]
+    else:  # physicsiq
+        return [
+            sample for sample in samples if not os.path.exists(sample["output_path"])
+        ]
+
+
+def assign_samples_to_gpu(
+    samples: list, gpu_id: int, rank: int, num_gpus: int
+) -> list:
+    """Divide samples across GPUs and return the subset for this GPU."""
+    samples_per_gpu = (len(samples) + num_gpus - 1) // num_gpus
+    start_idx = rank * samples_per_gpu
+    end_idx = min(start_idx + samples_per_gpu, len(samples))
+    return samples[start_idx:end_idx]
+
+
+def process_vbench_sample(pipeline, prompt: str, config: dict, gpu_id: int) -> None:
+    """Process a single vbench text-to-video sample."""
+    output_path = os.path.abspath(os.path.join(config["save_path"], f"{prompt}-0.mp4"))
+
+    if os.path.exists(output_path):
+        print(f"[SKIP GPU {gpu_id}] Already exists: {output_path}")
+        return
+
+    print(f"[GPU {gpu_id}] Generating T2V: '{prompt}' -> {output_path}")
+    pipeline.run_text_to_video(prompt=prompt, output_path=output_path)
+    print(f"[DONE GPU {gpu_id}] Saved: {output_path}")
+
+
+def process_physicsiq_sample(pipeline, sample: dict, gpu_id: int) -> None:
+    """Process a single PhysicsIQ video-to-video sample."""
+    prompt = sample["description"]
+    prefix_video_path = sample["prefix_video_path"]
+    output_path = sample["output_path"]
+
+    if not os.path.exists(prefix_video_path):
+        print(f"[WARN GPU {gpu_id}] Conditioning video not found: {prefix_video_path}")
+        return
+
+    if os.path.exists(output_path):
+        print(f"[SKIP GPU {gpu_id}] Already exists: {output_path}")
+        return
+
+    print(f"[GPU {gpu_id}] Generating V2V: '{prompt}'")
+    print(f"   Input:  {prefix_video_path}")
+    print(f"   Output: {output_path}")
+
+    pipeline.run_video_to_video(
+        prompt=prompt,
+        prefix_video_path=prefix_video_path,
+        output_path=output_path,
+    )
+    print(f"[DONE GPU {gpu_id}] Saved: {output_path}")
+
+
+def worker_process(gpu_id: int, rank: int, config: dict, all_samples: list) -> None:
+    """Independent worker running on each GPU."""
+    setup_environment(gpu_id)
+    configure_reuse_strategy(config)
+
+    try:
+        magi_root = subprocess.check_output(
+            ["git", "rev-parse", "--show-toplevel"]
+        ).decode().strip()
+        os.environ["MAGI_ROOT"] = magi_root
+        os.environ["PYTHONPATH"] = f"{magi_root}:{os.environ.get('PYTHONPATH', '')}"
+    except Exception as e:
+        print(f"[GPU {gpu_id}] Failed to set MAGI_ROOT: {e}")
+        return
+
+    filtered_samples = filter_existing_samples(all_samples, config)
+
+    if not filtered_samples:
+        print(f"[GPU {gpu_id}] No samples need to be generated.")
+        return
+
+    print(f"Processing {len(filtered_samples)} samples.")
+
+    my_samples = assign_samples_to_gpu(
+        filtered_samples, gpu_id, rank, config["num_gpus"]
+    )
+
+    if not my_samples:
+        print(f"[GPU {gpu_id}] No samples assigned.")
+        return
+
+    print(f"[GPU {gpu_id}] Assigned {len(my_samples)} samples")
+
+    from inference.pipeline.entry import MagiPipeline
+
+    print(f"[GPU {gpu_id}] Loading model...")
+    pipeline = MagiPipeline(config["config_file"])
+    print(f"[GPU {gpu_id}] Model loaded.")
+
+    process_func = (
+        process_vbench_sample if config["benchmark"] == "vbench" else process_physicsiq_sample
+    )
+
+    for sample in my_samples:
+        process_func(pipeline, sample, config, gpu_id)
+
+    print(f"[GPU {gpu_id}] Completed.")
+
+
+def build_conditioning_video_path(
+    data_root: str, vid_id: str, scenario: str, fps: int
+) -> str:
+    """Construct the path to the conditioning video file."""
+    conditioning_dir = os.path.join(
+        data_root, "physics-IQ-benchmark", "split-videos", "conditioning", f"{fps}FPS"
+    )
+    match_suffix = re.search(r"_(.*)", scenario)
+    suffix = match_suffix.group(1) if match_suffix else ""
+    filename = f"{vid_id}_conditioning-videos_{fps}FPS_{suffix}"
+    return os.path.join(conditioning_dir, filename)
+
+
+def load_physicsiq_samples(config: dict) -> list[dict]:
+    """Load sample list from PhysicsIQ dataset."""
+    data_root = config["physicsiq_data_dir"]
+    descriptions_csv = os.path.join(data_root, "descriptions", "descriptions.csv")
+    output_dir = config["save_path"]
+
+    if not os.path.exists(descriptions_csv):
+        raise FileNotFoundError(f"descriptions.csv not found at {descriptions_csv}")
+
+    os.makedirs(output_dir, exist_ok=True)
+
+    samples = []
+    with open(descriptions_csv, mode="r") as f:
+        reader = csv.DictReader(f)
+        for row in reader:
+            scenario = row["scenario"].strip()
+            match_id = re.match(r"^(\d+)_", scenario)
+
+            if not match_id:
+                print(f"Cannot extract ID from scenario: {scenario}")
+                continue
+
+            vid_id = match_id.group(1).zfill(4)
+            description = row["description"]
+            generated_video_name = row["generated_video_name"]
+            prefix_video_path = build_conditioning_video_path(
+                data_root, vid_id, scenario, PHYSICSIQ_FPS
+            )
+            output_path = os.path.join(output_dir, generated_video_name)
+
+            os.makedirs(os.path.dirname(output_path), exist_ok=True)
+
+            samples.append({
+                "vid_id": vid_id,
+                "scenario": scenario,
+                "description": description,
+                "generated_video_name": generated_video_name,
+                "prefix_video_path": prefix_video_path,
+                "output_path": output_path,
+            })
+
+    # PhysicsIQ samples are duplicated; take only the first half
+    unique_count = len(samples) // 2
+    samples = samples[:unique_count]
+
+    print(f"Loaded {unique_count} PhysicsIQ samples.")
+
+    return apply_slice(samples, config.get("start"), config.get("end"))
+
+
+def load_vbench_samples(config: dict) -> list[str]:
+    """Load prompt list from vbench dimension file."""
+    prompt_dir = config["vbench_prompt_dir"]
+    dimension = config.get("dimension")
+
+    if not dimension:
+        raise ValueError("For vbench, 'dimension' must be specified in config")
+
+    prompt_file = os.path.join(prompt_dir, f"{dimension}.txt")
+
+    if not os.path.exists(prompt_file):
+        raise FileNotFoundError(f"Prompt file not found: {prompt_file}")
+
+    with open(prompt_file, "r") as f:
+        prompts = [line.strip() for line in f if line.strip()]
+
+    return apply_slice(prompts, config.get("start"), config.get("end"))
+
+
+def setup_save_path(config: dict) -> None:
+    """Configure the output save path based on benchmark type."""
+    base_path = config["base_save_path"]
+
+    if config["benchmark"] == "vbench":
+        dimension = config.get("dimension")
+        videos_dir = os.path.join(base_path, "videos", dimension) if dimension else None
+        config["save_path"] = videos_dir if videos_dir else os.path.join(base_path, "videos")
+    elif config["benchmark"] == "physicsiq":
+        config["save_path"] = os.path.join(base_path, "videos")
+
+    os.makedirs(config["save_path"], exist_ok=True)
+
+
+def main() -> None:
+    """Entry point for video sampling script."""
+    parser = argparse.ArgumentParser(
+        description="Video sampling script using YAML configuration"
+    )
+    parser.add_argument("yaml_config", type=str, help="Path to YAML configuration file")
+    args = parser.parse_args()
+
+    config = load_yaml_config(args.yaml_config)
+    print(f"Loaded configuration from: {args.yaml_config}")
+
+    setup_save_path(config)
+
+    gpu_ids = resolve_gpu_ids(config["gpus"])
+    config["num_gpus"] = len(gpu_ids)
+
+    benchmark = config["benchmark"]
+    if benchmark == "vbench":
+        all_samples = load_vbench_samples(config)
+    elif benchmark == "physicsiq":
+        data_root = config["physicsiq_data_dir"]
+        if not os.path.exists(data_root):
+            raise FileNotFoundError(f"Data directory not found: {data_root}")
+        all_samples = load_physicsiq_samples(config)
+    else:
+        raise ValueError(f"Invalid benchmark: {benchmark}")
+
+    print(f"Total samples: {len(all_samples)}")
+    print(f"GPUs: {gpu_ids}")
+    print(f"Output: {config['save_path']}")
+    print(f"Config: {config['config_file']}")
+
+    processes = []
+    for rank, gpu_id in enumerate(gpu_ids):
+        p = mp.Process(target=worker_process, args=(gpu_id, rank, config, all_samples))
+        p.start()
+        processes.append(p)
+
+    for p in processes:
+        p.join()
+
+    failed = [p.exitcode for p in processes if p.exitcode != 0]
+    if failed:
+        raise RuntimeError(f"{len(failed)} worker process(es) failed with exit codes: {failed}")
+
+
+if __name__ == "__main__":
+    main()

FlowCache/FlowCache4MAGI-1-dev3-motion/README.md

@@ -0,0 +1,71 @@
+# FLOW CACHING FOR AUTOREGRESSIVE VIDEO GENERATION
+
+This repository provides the official implementation of **FlowCache** on **MAGI-1** model, a caching-based acceleration method for autoregressive video generation models.
+
+
+## 🚀 Installation
+
+Please follow the installation instructions provided in the [MAGI-1](https://github.com/SandAI-org/MAGI-1), as this implementation is built on top of MAGI-1.
+
+---
+
+## ▶️ Usage
+
+### 1. Single Video Generation
+
+Run accelerated generation using FlowCache:
+
+```bash
+# FlowCache for text-to-video generation
+bash scripts/single_run/flowcache_t2v.sh
+
+# FlowCache for video-to-video generation
+bash scripts/single_run/flowcache_v2v.sh
+
+# Baseline acceleration method (TeaCache) for text-to-video
+bash scripts/single_run/teacache_t2v.sh
+
+# Baseline acceleration method (TeaCache) for video-to-video
+bash scripts/single_run/teacache_v2v.sh
+```
+
+### 2. Benchmark Sampling
+
+Generate videos for evaluation on standard benchmarks:
+
+```bash
+# VBench
+bash scripts/sample/flowcache_vbench.sh
+bash scripts/sample/teacache_vbench.sh
+
+# PhysicsIQ
+bash scripts/sample/flowcache_physicsiq.sh
+bash scripts/sample/teacache_physicsiq.sh
+```
+
+### 3. Quality Evaluation
+
+Compute perceptual and structural similarity metrics between original and accelerated generations:
+
+```bash
+bash scripts/metric.sh
+```
+
+---
+
+## ⚙️ Key Parameters
+
+| Parameter | Description |
+|----------|-------------|
+| `rel_l1_thresh` | Relative L1 distance threshold for cache reuse decision |
+| ` warmup_steps` | Number of denoising steps where reuse is disabled |
+| `total_cache_chunk_nums` (`B_total`) | Total number of cache chunks maintained |
+| `compress_strategy` | Granularity for selecting important KV caches: `token`, `frame`, or `chunk` |
+| `query_granularity` | Granularity for importance scoring: `token`, `frame`, or `chunk` |
+| `mix_lambda` | Weight balancing importance and redundancy (default: `0.07`) |
+| `mode` | Generation mode: `t2v` (text-to-video), `i2v` (image-to-video), or `v2v` (video-to-video) |
+| `prompt` | Input prompt for conditional generation |
+| `output_path` | Path to save generated videos |
+| `config_file` | Path to MAGI-1 model configuration |
+
+---

FlowCache/FlowCache4MAGI-1-dev3-motion/README_MOTIONCACHE.md

@@ -0,0 +1,92 @@
+# MotionCache on MAGI-1 (dev3-motion)
+
+基于 [FlowCache4MAGI-1](../FlowCache4MAGI-1) 的 **MotionCache** 复现分支，对应论文：
+
+> Xu et al., 2026 — *Motion-Aware Caching for Efficient Autoregressive Video Generation*
+
+官方代码仓库 [ywlq/MotionCache](https://github.com/ywlq/MotionCache) 尚未发布实现，本目录依据论文方法在 MAGI-1 推理框架上完成首版复现。
+
+## 与 FlowCache 的核心差异
+
+| 维度 | FlowCache | MotionCache |
+|------|-----------|-------------|
+| 粒度 | Chunk 级全有或全无 | Phase 2 为 Token（latent 帧×空间）级 |
+| 跳过策略 | 相对 L1 累积阈值 | 运动重要性加权累积 |
+| 调度 | 单一 chunk-wise 策略 | 两阶段 coarse-to-fine |
+
+### 算法概要
+
+1. **全局 Warm-up（m 步）**：前 `warmup_steps` 步禁止 cache reuse
+2. **Phase 1（K 步）**：chunk-wise 二值决策，与 FlowCache 相同
+3. **Phase 2**：基于帧间 latent 差计算运动重要性 `M`，经 soft-mapping 得到权重 `W ∈ [α, 1]`
+4. **Token 累积**：`A[p] += W[p] · Δ_chunk`，当 `A[p] > τ` 时该 token 触发 DiT 计算
+5. **Integrate**：active token 正常积分；inactive token 复用缓存 residual
+
+### MAGI-1 默认超参（论文 Appendix C）
+
+| 参数 | 值 | 说明 |
+|------|-----|------|
+| `alpha` | 0.5 | 静态区域权重下限 |
+| `phase1_steps` (K) | 9 | chunk-wise 阶段持续步数 |
+| `warmup_steps` (m) | 5 | 全局禁止 reuse 的步数 |
+| `rel_l1_thresh` (τ) | 0.015 (slow) / 0.025 (fast) | token 累积阈值 |
+
+## 快速运行
+
+```bash
+cd FlowCache4MAGI-1-dev3-motion
+
+# MotionCache-slow（论文 Table 1 配置）
+bash scripts/single_run/motioncache_t2v.sh
+
+# MotionCache-fast
+MOTIONCACHE_CONFIG=yaml_config/single_run/motioncache_config_fast.yaml \
+  bash scripts/single_run/motioncache_t2v.sh
+```
+
+需先按 FlowCache4MAGI-1 说明安装依赖并下载 MAGI-1 权重（`downloads/` 目录可通过软链接指向原项目）。
+
+## 代码结构
+
+```
+inference/pipeline/
+├── motioncache.py              # 入口与 forward/integrate monkey-patch
+└── cache/
+    ├── motioncache.py          # MotionWiseCache 核心逻辑
+    └── sparse_utils.py         # Phase 2 gather/scatter 与 sparse meta_args
+
+inference/model/dit/dit_module.py   # sparse KV cache 写入 + flash_attn 稀疏 q 分支
+inference/common/dataclass.py       # ModelMetaArgs.sparse_active_indices
+
+yaml_config/single_run/
+├── motioncache_config.yaml     # slow 配置
+└── motioncache_config_fast.yaml
+```
+
+## 当前实现说明（严格对齐论文 §5.3 Phase 2）
+
+- **Phase 1（前 K 个 denoise step / chunk）**：与 FlowCache 完全一致的 chunk-wise 策略（已验证 PSNR=∞）
+- **Phase 2 入口**：每个 chunk 的第 K 步强制全量计算，建立 token-wise 阶段的 residual 基准
+- **Phase 2 运动感知累积**：latent 帧间差 → soft-mapping 权重 W → 累积器 A；`A > τ` 的 token 为 active
+- **Phase 2 稀疏 forward（论文 5.3）**：
+  - 全部 inactive → 跳过 DiT forward，复用 chunk residual
+  - 部分 active → **gather active patch tokens → compact DiT forward → scatter 回完整序列**
+  - KV cache 仅更新 active 位置（`_compresskv_adjust_key_and_value` sparse 分支）
+  - integrate 时 active token 正常积分，inactive token 复用 `previous_residual`
+- **运动 proxy**：latent 空间帧间 L1 差（论文 Eq. 9-10）
+- **跨 chunk 连续性**：chunk 首帧与上一 chunk 末帧比较
+
+### 验证结果（`a woman dancing.`，240×720×720，vs FlowCache baseline）
+
+| 运行 | PSNR | 说明 |
+|------|------|------|
+| phase1only | ∞ | 与 FlowCache 逐像素一致 |
+| sparse2（论文对齐 sparse forward） | **20.44 dB** | 无黑帧，reuse_rate≈15.6% |
+| final（整 chunk fallback） | 20.48 dB | 画质等价，推理慢 ~14% |
+
+sparse2 日志示例：`active_ratio=1.51%` 时仍走 gather/scatter 稀疏路径；`active_ratio=0%` 时 `skip_forward=True`。
+
+## 参考
+
+- FlowCache: chunk-wise cache + KV compression
+- MotionCache 论文预期 MAGI-1 加速：slow 1.64×，fast 2.07×（Table 1）

FlowCache/FlowCache4MAGI-1-dev3-motion/requirements.txt

@@ -0,0 +1,21 @@
+accelerate==0.32.1
+beautifulsoup4==4.13.4
+debugpy==1.8.14
+diffusers==0.29.2
+einops>=0.6.0
+ffmpeg-python
+# flash-attn==2.4.2
+flashinfer-python==0.2.0.post2 --extra-index-url https://flashinfer.ai/whl/cu124/torch2.4/
+ftfy==6.2.0
+gpustat==1.1.1
+imageio==2.34.0
+imageio[ffmpeg]
+matplotlib==3.10.1
+numpy==1.26.4
+protobuf==5.28.3
+rich==14.0.0
+sentencepiece==0.2.0
+timm==1.0.15
+torchdiffeq==0.2.4
+transformers==4.42.3
+tqdm

FlowCache/FlowCache4MAGI-1-dev3-motion/sample_video.py

@@ -0,0 +1,429 @@
+# Copyright 2024 MAGI Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import re
+import sys
+import argparse
+import csv
+import subprocess
+from pathlib import Path
+
+import multiprocessing as mp
+
+# Constants
+DEFAULT_BASE_PORT = 29510
+PHYSICSIQ_FPS = 24
+
+
+def resolve_gpu_ids(gpus_config) -> list[int]:
+    """Resolve explicit GPU IDs or auto-detect all currently visible GPUs."""
+    if isinstance(gpus_config, int):
+        return [gpus_config]
+
+    gpus_text = str(gpus_config).strip()
+    if not gpus_text:
+        raise ValueError("'gpus' must not be empty")
+
+    if gpus_text.lower() not in {"all", "auto"}:
+        return [int(item.strip()) for item in gpus_text.split(",") if item.strip()]
+
+    visible_devices = os.environ.get("CUDA_VISIBLE_DEVICES")
+    if visible_devices:
+        visible = [item.strip() for item in visible_devices.split(",") if item.strip()]
+        if visible and all(item.isdigit() for item in visible):
+            return [int(item) for item in visible]
+
+    try:
+        output = subprocess.check_output(
+            ["nvidia-smi", "--query-gpu=index", "--format=csv,noheader,nounits"],
+            text=True,
+            timeout=10,
+        )
+        gpu_ids = [int(line.strip()) for line in output.splitlines() if line.strip()]
+        if gpu_ids:
+            return gpu_ids
+    except Exception:
+        pass
+
+    try:
+        import torch
+
+        count = torch.cuda.device_count()
+        if count > 0:
+            return list(range(count))
+    except Exception:
+        pass
+
+    raise RuntimeError("No CUDA GPUs detected for gpus: all")
+
+
+def load_yaml_config(yaml_path: str) -> dict:
+    """Load configuration from YAML file."""
+    import yaml
+
+    with open(yaml_path, "r") as f:
+        return yaml.safe_load(f)
+
+
+def apply_slice(items: list, start: int | None, end: int | None) -> list:
+    """Apply start/end slice to a list with bounds checking."""
+    if start is None and end is None:
+        return items
+
+    slice_start = max(0, start if start is not None else 0)
+    slice_end = min(end if end is not None else len(items), len(items))
+    slice_end = max(slice_start, slice_end)
+
+    return items[slice_start:slice_end]
+
+
+def configure_teacache(transport, config: dict) -> None:
+    """Configure TeaCache reuse strategy on SampleTransport."""
+    from inference.pipeline.teacache import setup_teacache
+
+    setup_teacache(
+        rel_l1_thresh=config["rel_l1_thresh"],
+        warmup_steps=config["warmup_steps"],
+        log=config.get("log", False),
+    )
+
+
+def configure_kv_cache(transport, config: dict) -> None:
+    """Configure KV cache compression if enabled."""
+    if not config.get("compress_kv_cache", False):
+        transport.compress_kv_cache = False
+        return
+
+    print("KV cache compression is enabled.")
+    transport.compress_kv_cache = True
+
+    assert config.get("total_cache_chunk_nums") is not None
+
+    compression_config = {
+        "method_config": {
+            "compress_strategy": config["compress_strategy"],
+            "mix_lambda": config["mix_lambda"],
+            "query_granularity": config["query_granularity"],
+            "score_weighting_method": config.get("score_weighting_method") or "no_weight",
+            "power": config.get("power", 3),
+        },
+    }
+
+    from inference.pipeline.kvcompress import replace_magi
+
+    replace_magi(compression_config)
+
+
+def configure_flowcache(transport, config: dict) -> None:
+    """Configure FlowCache reuse strategy on SampleTransport."""
+    from inference.pipeline.flowcache import setup_flowcache
+
+    configure_kv_cache(transport, config)
+
+    setup_flowcache(
+        rel_l1_thresh=config["rel_l1_thresh"],
+        warmup_steps=config["warmup_steps"],
+        discard_nearly_clean_chunk=config.get("discard_nearly_clean_chunk", False),
+        log=config.get("log", False),
+        total_cache_chunk_nums=config.get("total_cache_chunk_nums", 5),
+        compress_kv_cache=config.get("compress_kv_cache", False),
+    )
+
+
+def configure_reuse_strategy(config: dict) -> None:
+    """Configure the appropriate reuse strategy on SampleTransport."""
+    from inference.pipeline.video_generate import SampleTransport
+
+    strategy = config["reuse_strategy"]
+
+    if strategy == "original":
+        return
+    if strategy == "all":
+        configure_teacache(SampleTransport, config)
+    elif strategy == "chunkwise":
+        configure_flowcache(SampleTransport, config)
+    else:
+        raise ValueError(f"Unknown reuse strategy: {strategy}")
+
+
+def setup_environment(gpu_id: int) -> None:
+    """Set up environment variables for a GPU worker process."""
+    os.environ["CUDA_VISIBLE_DEVICES"] = str(gpu_id)
+    os.environ["WORLD_SIZE"] = "1"
+    os.environ["RANK"] = "0"
+    os.environ["LOCAL_RANK"] = "0"
+    os.environ["MASTER_ADDR"] = "localhost"
+    os.environ["MASTER_PORT"] = str(DEFAULT_BASE_PORT + gpu_id)
+
+    # Enable pdb terminal debugging
+    sys.stdin = open(0)
+
+
+def filter_existing_samples(samples: list, config: dict) -> list:
+    """Filter out samples whose output files already exist."""
+    if config["benchmark"] == "vbench":
+        return [
+            sample
+            for sample in samples
+            if not os.path.exists(os.path.abspath(os.path.join(config["save_path"], f"{sample}-0.mp4")))
+        ]
+    else:  # physicsiq
+        return [
+            sample for sample in samples if not os.path.exists(sample["output_path"])
+        ]
+
+
+def assign_samples_to_gpu(
+    samples: list, gpu_id: int, rank: int, num_gpus: int
+) -> list:
+    """Divide samples across GPUs and return the subset for this GPU."""
+    samples_per_gpu = (len(samples) + num_gpus - 1) // num_gpus
+    start_idx = rank * samples_per_gpu
+    end_idx = min(start_idx + samples_per_gpu, len(samples))
+    return samples[start_idx:end_idx]
+
+
+def process_vbench_sample(pipeline, prompt: str, config: dict, gpu_id: int) -> None:
+    """Process a single vbench text-to-video sample."""
+    output_path = os.path.abspath(os.path.join(config["save_path"], f"{prompt}-0.mp4"))
+
+    if os.path.exists(output_path):
+        print(f"[SKIP GPU {gpu_id}] Already exists: {output_path}")
+        return
+
+    print(f"[GPU {gpu_id}] Generating T2V: '{prompt}' -> {output_path}")
+    pipeline.run_text_to_video(prompt=prompt, output_path=output_path)
+    print(f"[DONE GPU {gpu_id}] Saved: {output_path}")
+
+
+def process_physicsiq_sample(pipeline, sample: dict, gpu_id: int) -> None:
+    """Process a single PhysicsIQ video-to-video sample."""
+    prompt = sample["description"]
+    prefix_video_path = sample["prefix_video_path"]
+    output_path = sample["output_path"]
+
+    if not os.path.exists(prefix_video_path):
+        print(f"[WARN GPU {gpu_id}] Conditioning video not found: {prefix_video_path}")
+        return
+
+    if os.path.exists(output_path):
+        print(f"[SKIP GPU {gpu_id}] Already exists: {output_path}")
+        return
+
+    print(f"[GPU {gpu_id}] Generating V2V: '{prompt}'")
+    print(f"   Input:  {prefix_video_path}")
+    print(f"   Output: {output_path}")
+
+    pipeline.run_video_to_video(
+        prompt=prompt,
+        prefix_video_path=prefix_video_path,
+        output_path=output_path,
+    )
+    print(f"[DONE GPU {gpu_id}] Saved: {output_path}")
+
+
+def worker_process(gpu_id: int, rank: int, config: dict, all_samples: list) -> None:
+    """Independent worker running on each GPU."""
+    setup_environment(gpu_id)
+    configure_reuse_strategy(config)
+
+    try:
+        magi_root = subprocess.check_output(
+            ["git", "rev-parse", "--show-toplevel"]
+        ).decode().strip()
+        os.environ["MAGI_ROOT"] = magi_root
+        os.environ["PYTHONPATH"] = f"{magi_root}:{os.environ.get('PYTHONPATH', '')}"
+    except Exception as e:
+        print(f"[GPU {gpu_id}] Failed to set MAGI_ROOT: {e}")
+        return
+
+    filtered_samples = filter_existing_samples(all_samples, config)
+
+    if not filtered_samples:
+        print(f"[GPU {gpu_id}] No samples need to be generated.")
+        return
+
+    print(f"Processing {len(filtered_samples)} samples.")
+
+    my_samples = assign_samples_to_gpu(
+        filtered_samples, gpu_id, rank, config["num_gpus"]
+    )
+
+    if not my_samples:
+        print(f"[GPU {gpu_id}] No samples assigned.")
+        return
+
+    print(f"[GPU {gpu_id}] Assigned {len(my_samples)} samples")
+
+    from inference.pipeline.entry import MagiPipeline
+
+    print(f"[GPU {gpu_id}] Loading model...")
+    pipeline = MagiPipeline(config["config_file"])
+    print(f"[GPU {gpu_id}] Model loaded.")
+
+    process_func = (
+        process_vbench_sample if config["benchmark"] == "vbench" else process_physicsiq_sample
+    )
+
+    for sample in my_samples:
+        process_func(pipeline, sample, config, gpu_id)
+
+    print(f"[GPU {gpu_id}] Completed.")
+
+
+def build_conditioning_video_path(
+    data_root: str, vid_id: str, scenario: str, fps: int
+) -> str:
+    """Construct the path to the conditioning video file."""
+    conditioning_dir = os.path.join(
+        data_root, "physics-IQ-benchmark", "split-videos", "conditioning", f"{fps}FPS"
+    )
+    match_suffix = re.search(r"_(.*)", scenario)
+    suffix = match_suffix.group(1) if match_suffix else ""
+    filename = f"{vid_id}_conditioning-videos_{fps}FPS_{suffix}"
+    return os.path.join(conditioning_dir, filename)
+
+
+def load_physicsiq_samples(config: dict) -> list[dict]:
+    """Load sample list from PhysicsIQ dataset."""
+    data_root = config["physicsiq_data_dir"]
+    descriptions_csv = os.path.join(data_root, "descriptions", "descriptions.csv")
+    output_dir = config["save_path"]
+
+    if not os.path.exists(descriptions_csv):
+        raise FileNotFoundError(f"descriptions.csv not found at {descriptions_csv}")
+
+    os.makedirs(output_dir, exist_ok=True)
+
+    samples = []
+    with open(descriptions_csv, mode="r") as f:
+        reader = csv.DictReader(f)
+        for row in reader:
+            scenario = row["scenario"].strip()
+            match_id = re.match(r"^(\d+)_", scenario)
+
+            if not match_id:
+                print(f"Cannot extract ID from scenario: {scenario}")
+                continue
+
+            vid_id = match_id.group(1).zfill(4)
+            description = row["description"]
+            generated_video_name = row["generated_video_name"]
+            prefix_video_path = build_conditioning_video_path(
+                data_root, vid_id, scenario, PHYSICSIQ_FPS
+            )
+            output_path = os.path.join(output_dir, generated_video_name)
+
+            os.makedirs(os.path.dirname(output_path), exist_ok=True)
+
+            samples.append({
+                "vid_id": vid_id,
+                "scenario": scenario,
+                "description": description,
+                "generated_video_name": generated_video_name,
+                "prefix_video_path": prefix_video_path,
+                "output_path": output_path,
+            })
+
+    # PhysicsIQ samples are duplicated; take only the first half
+    unique_count = len(samples) // 2
+    samples = samples[:unique_count]
+
+    print(f"Loaded {unique_count} PhysicsIQ samples.")
+
+    return apply_slice(samples, config.get("start"), config.get("end"))
+
+
+def load_vbench_samples(config: dict) -> list[str]:
+    """Load prompt list from vbench dimension file."""
+    prompt_dir = config["vbench_prompt_dir"]
+    dimension = config.get("dimension")
+
+    if not dimension:
+        raise ValueError("For vbench, 'dimension' must be specified in config")
+
+    prompt_file = os.path.join(prompt_dir, f"{dimension}.txt")
+
+    if not os.path.exists(prompt_file):
+        raise FileNotFoundError(f"Prompt file not found: {prompt_file}")
+
+    with open(prompt_file, "r") as f:
+        prompts = [line.strip() for line in f if line.strip()]
+
+    return apply_slice(prompts, config.get("start"), config.get("end"))
+
+
+def setup_save_path(config: dict) -> None:
+    """Configure the output save path based on benchmark type."""
+    base_path = config["base_save_path"]
+
+    if config["benchmark"] == "vbench":
+        dimension = config.get("dimension")
+        videos_dir = os.path.join(base_path, "videos", dimension) if dimension else None
+        config["save_path"] = videos_dir if videos_dir else os.path.join(base_path, "videos")
+    elif config["benchmark"] == "physicsiq":
+        config["save_path"] = os.path.join(base_path, "videos")
+
+    os.makedirs(config["save_path"], exist_ok=True)
+
+
+def main() -> None:
+    """Entry point for video sampling script."""
+    parser = argparse.ArgumentParser(
+        description="Video sampling script using YAML configuration"
+    )
+    parser.add_argument("yaml_config", type=str, help="Path to YAML configuration file")
+    args = parser.parse_args()
+
+    config = load_yaml_config(args.yaml_config)
+    print(f"Loaded configuration from: {args.yaml_config}")
+
+    setup_save_path(config)
+
+    gpu_ids = resolve_gpu_ids(config["gpus"])
+    config["num_gpus"] = len(gpu_ids)
+
+    benchmark = config["benchmark"]
+    if benchmark == "vbench":
+        all_samples = load_vbench_samples(config)
+    elif benchmark == "physicsiq":
+        data_root = config["physicsiq_data_dir"]
+        if not os.path.exists(data_root):
+            raise FileNotFoundError(f"Data directory not found: {data_root}")
+        all_samples = load_physicsiq_samples(config)
+    else:
+        raise ValueError(f"Invalid benchmark: {benchmark}")
+
+    print(f"Total samples: {len(all_samples)}")
+    print(f"GPUs: {gpu_ids}")
+    print(f"Output: {config['save_path']}")
+    print(f"Config: {config['config_file']}")
+
+    processes = []
+    for rank, gpu_id in enumerate(gpu_ids):
+        p = mp.Process(target=worker_process, args=(gpu_id, rank, config, all_samples))
+        p.start()
+        processes.append(p)
+
+    for p in processes:
+        p.join()
+
+    failed = [p.exitcode for p in processes if p.exitcode != 0]
+    if failed:
+        raise RuntimeError(f"{len(failed)} worker process(es) failed with exit codes: {failed}")
+
+
+if __name__ == "__main__":
+    main()

FlowCache/FlowCache4MAGI-1-dev4-detail/README.md

@@ -0,0 +1,71 @@
+# FLOW CACHING FOR AUTOREGRESSIVE VIDEO GENERATION
+
+This repository provides the official implementation of **FlowCache** on **MAGI-1** model, a caching-based acceleration method for autoregressive video generation models.
+
+
+## 🚀 Installation
+
+Please follow the installation instructions provided in the [MAGI-1](https://github.com/SandAI-org/MAGI-1), as this implementation is built on top of MAGI-1.
+
+---
+
+## ▶️ Usage
+
+### 1. Single Video Generation
+
+Run accelerated generation using FlowCache:
+
+```bash
+# FlowCache for text-to-video generation
+bash scripts/single_run/flowcache_t2v.sh
+
+# FlowCache for video-to-video generation
+bash scripts/single_run/flowcache_v2v.sh
+
+# Baseline acceleration method (TeaCache) for text-to-video
+bash scripts/single_run/teacache_t2v.sh
+
+# Baseline acceleration method (TeaCache) for video-to-video
+bash scripts/single_run/teacache_v2v.sh
+```
+
+### 2. Benchmark Sampling
+
+Generate videos for evaluation on standard benchmarks:
+
+```bash
+# VBench
+bash scripts/sample/flowcache_vbench.sh
+bash scripts/sample/teacache_vbench.sh
+
+# PhysicsIQ
+bash scripts/sample/flowcache_physicsiq.sh
+bash scripts/sample/teacache_physicsiq.sh
+```
+
+### 3. Quality Evaluation
+
+Compute perceptual and structural similarity metrics between original and accelerated generations:
+
+```bash
+bash scripts/metric.sh
+```
+
+---
+
+## ⚙️ Key Parameters
+
+| Parameter | Description |
+|----------|-------------|
+| `rel_l1_thresh` | Relative L1 distance threshold for cache reuse decision |
+| ` warmup_steps` | Number of denoising steps where reuse is disabled |
+| `total_cache_chunk_nums` (`B_total`) | Total number of cache chunks maintained |
+| `compress_strategy` | Granularity for selecting important KV caches: `token`, `frame`, or `chunk` |
+| `query_granularity` | Granularity for importance scoring: `token`, `frame`, or `chunk` |
+| `mix_lambda` | Weight balancing importance and redundancy (default: `0.07`) |
+| `mode` | Generation mode: `t2v` (text-to-video), `i2v` (image-to-video), or `v2v` (video-to-video) |
+| `prompt` | Input prompt for conditional generation |
+| `output_path` | Path to save generated videos |
+| `config_file` | Path to MAGI-1 model configuration |
+
+---

FlowCache/FlowCache4MAGI-1-dev4-detail/README_MOTIONCACHE.md

@@ -0,0 +1,92 @@
+# MotionCache on MAGI-1 (dev3-motion)
+
+基于 [FlowCache4MAGI-1](../FlowCache4MAGI-1) 的 **MotionCache** 复现分支，对应论文：
+
+> Xu et al., 2026 — *Motion-Aware Caching for Efficient Autoregressive Video Generation*
+
+官方代码仓库 [ywlq/MotionCache](https://github.com/ywlq/MotionCache) 尚未发布实现，本目录依据论文方法在 MAGI-1 推理框架上完成首版复现。
+
+## 与 FlowCache 的核心差异
+
+| 维度 | FlowCache | MotionCache |
+|------|-----------|-------------|
+| 粒度 | Chunk 级全有或全无 | Phase 2 为 Token（latent 帧×空间）级 |
+| 跳过策略 | 相对 L1 累积阈值 | 运动重要性加权累积 |
+| 调度 | 单一 chunk-wise 策略 | 两阶段 coarse-to-fine |
+
+### 算法概要
+
+1. **全局 Warm-up（m 步）**：前 `warmup_steps` 步禁止 cache reuse
+2. **Phase 1（K 步）**：chunk-wise 二值决策，与 FlowCache 相同
+3. **Phase 2**：基于帧间 latent 差计算运动重要性 `M`，经 soft-mapping 得到权重 `W ∈ [α, 1]`
+4. **Token 累积**：`A[p] += W[p] · Δ_chunk`，当 `A[p] > τ` 时该 token 触发 DiT 计算
+5. **Integrate**：active token 正常积分；inactive token 复用缓存 residual
+
+### MAGI-1 默认超参（论文 Appendix C）
+
+| 参数 | 值 | 说明 |
+|------|-----|------|
+| `alpha` | 0.5 | 静态区域权重下限 |
+| `phase1_steps` (K) | 9 | chunk-wise 阶段持续步数 |
+| `warmup_steps` (m) | 5 | 全局禁止 reuse 的步数 |
+| `rel_l1_thresh` (τ) | 0.015 (slow) / 0.025 (fast) | token 累积阈值 |
+
+## 快速运行
+
+```bash
+cd FlowCache4MAGI-1-dev3-motion
+
+# MotionCache-slow（论文 Table 1 配置）
+bash scripts/single_run/motioncache_t2v.sh
+
+# MotionCache-fast
+MOTIONCACHE_CONFIG=yaml_config/single_run/motioncache_config_fast.yaml \
+  bash scripts/single_run/motioncache_t2v.sh
+```
+
+需先按 FlowCache4MAGI-1 说明安装依赖并下载 MAGI-1 权重（`downloads/` 目录可通过软链接指向原项目）。
+
+## 代码结构
+
+```
+inference/pipeline/
+├── motioncache.py              # 入口与 forward/integrate monkey-patch
+└── cache/
+    ├── motioncache.py          # MotionWiseCache 核心逻辑
+    └── sparse_utils.py         # Phase 2 gather/scatter 与 sparse meta_args
+
+inference/model/dit/dit_module.py   # sparse KV cache 写入 + flash_attn 稀疏 q 分支
+inference/common/dataclass.py       # ModelMetaArgs.sparse_active_indices
+
+yaml_config/single_run/
+├── motioncache_config.yaml     # slow 配置
+└── motioncache_config_fast.yaml
+```
+
+## 当前实现说明（严格对齐论文 §5.3 Phase 2）
+
+- **Phase 1（前 K 个 denoise step / chunk）**：与 FlowCache 完全一致的 chunk-wise 策略（已验证 PSNR=∞）
+- **Phase 2 入口**：每个 chunk 的第 K 步强制全量计算，建立 token-wise 阶段的 residual 基准
+- **Phase 2 运动感知累积**：latent 帧间差 → soft-mapping 权重 W → 累积器 A；`A > τ` 的 token 为 active
+- **Phase 2 稀疏 forward（论文 5.3）**：
+  - 全部 inactive → 跳过 DiT forward，复用 chunk residual
+  - 部分 active → **gather active patch tokens → compact DiT forward → scatter 回完整序列**
+  - KV cache 仅更新 active 位置（`_compresskv_adjust_key_and_value` sparse 分支）
+  - integrate 时 active token 正常积分，inactive token 复用 `previous_residual`
+- **运动 proxy**：latent 空间帧间 L1 差（论文 Eq. 9-10）
+- **跨 chunk 连续性**：chunk 首帧与上一 chunk 末帧比较
+
+### 验证结果（`a woman dancing.`，240×720×720，vs FlowCache baseline）
+
+| 运行 | PSNR | 说明 |
+|------|------|------|
+| phase1only | ∞ | 与 FlowCache 逐像素一致 |
+| sparse2（论文对齐 sparse forward） | **20.44 dB** | 无黑帧，reuse_rate≈15.6% |
+| final（整 chunk fallback） | 20.48 dB | 画质等价，推理慢 ~14% |
+
+sparse2 日志示例：`active_ratio=1.51%` 时仍走 gather/scatter 稀疏路径；`active_ratio=0%` 时 `skip_forward=True`。
+
+## 参考
+
+- FlowCache: chunk-wise cache + KV compression
+- MotionCache 论文预期 MAGI-1 加速：slow 1.64×，fast 2.07×（Table 1）

FlowCache/FlowCache4MAGI-1-dev4-detail/README_MOTIONDETAIL.md

@@ -0,0 +1,71 @@
+# MotionDetailCache on MAGI-1 (dev4-detail)
+
+基于 [FlowCache4MAGI-1-dev3-motion](../FlowCache4MAGI-1-dev3-motion) 的扩展分支，在 MotionCache 的 **motion 度量** 之外，增加 **spatial detail 度量**，Phase 2 active token 由两者共同决定。
+
+## 与 dev3 (MotionCache) 的差异
+
+| 维度 | dev3 MotionCache | dev4 MotionDetailCache |
+|------|------------------|------------------------|
+| Phase 2 权重 | 仅 motion 权重 W_m | motion W_m + detail W_d 融合 |
+| Detail proxy | — | latent 局部 k×k 空间方差 |
+| 适用场景 | 运动区域优先计算 | 静止纹理/边缘/语义细节也优先保留 |
+
+## Detail 度量
+
+对每个 latent 像素 `(t, h, w)`：
+
+1. 通道聚合幅度：`mag = mean(|x|, dim=C)` → `[N, T, H, W]`
+2. 局部空间方差：在 `detail_window_size×detail_window_size` 窗口内计算 `Var(mag)`
+3. Soft-mapping：帧内 min-max 归一化 → `W_d ∈ [detail_alpha, 1]`
+
+高方差区域对应边缘、纹理、物体边界等 heterogeneous 邻域。
+
+## Motion + Detail 融合
+
+```
+W_combined = combine(W_motion, W_detail)
+A[p] += W_combined[p] · Δ_chunk
+active[p] = A[p] > τ
+```
+
+`weight_combine_mode`（默认 `max`）：
+
+| 模式 | 公式 | 特点 |
+|------|------|------|
+| `max` | max(W_m, W_d) | 运动或细节任一显著即可加速累积（推荐） |
+| `product` | W_m × W_d | 两者同时高才快速激活，更激进 reuse |
+| `blend` | (1-λ)W_m + λW_d | 线性权衡，λ=`detail_lambda` |
+
+## 默认超参
+
+继承 dev3 的 motion 参数，并新增：
+
+| 参数 | 默认值 | 说明 |
+|------|--------|------|
+| `detail_alpha` | 0.5 | detail 权重下限 |
+| `detail_window_size` | 3 | 局部方差窗口（奇数） |
+| `detail_lambda` | 0.5 | blend 模式下的 detail 权重 |
+| `weight_combine_mode` | max | 融合方式 |
+
+## 快速运行
+
+```bash
+cd FlowCache4MAGI-1-dev4-detail
+conda activate magi
+export CUDA_VISIBLE_DEVICES=0
+export MASTER_ADDR=localhost
+bash scripts/single_run/motiondetail_t2v.sh
+```
+
+## 代码结构
+
+```
+inference/pipeline/cache/
+├── motioncache.py          # dev3 基类 MotionWiseCache
+└── motiondetailcache.py    # dev4 MotionDetailCache
+
+yaml_config/single_run/motiondetail_config.yaml
+scripts/single_run/motiondetail_t2v.sh
+```
+
+其余 sparse forward、KV cache、integrate 逻辑与 dev3 完全一致。

FlowCache/FlowCache4MAGI-1-dev4-detail/requirements.txt

@@ -0,0 +1,21 @@
+accelerate==0.32.1
+beautifulsoup4==4.13.4
+debugpy==1.8.14
+diffusers==0.29.2
+einops>=0.6.0
+ffmpeg-python
+# flash-attn==2.4.2
+flashinfer-python==0.2.0.post2 --extra-index-url https://flashinfer.ai/whl/cu124/torch2.4/
+ftfy==6.2.0
+gpustat==1.1.1
+imageio==2.34.0
+imageio[ffmpeg]
+matplotlib==3.10.1
+numpy==1.26.4
+protobuf==5.28.3
+rich==14.0.0
+sentencepiece==0.2.0
+timm==1.0.15
+torchdiffeq==0.2.4
+transformers==4.42.3
+tqdm

FlowCache/FlowCache4MAGI-1-dev4-detail/sample_video.py

@@ -0,0 +1,429 @@
+# Copyright 2024 MAGI Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import re
+import sys
+import argparse
+import csv
+import subprocess
+from pathlib import Path
+
+import multiprocessing as mp
+
+# Constants
+DEFAULT_BASE_PORT = 29510
+PHYSICSIQ_FPS = 24
+
+
+def resolve_gpu_ids(gpus_config) -> list[int]:
+    """Resolve explicit GPU IDs or auto-detect all currently visible GPUs."""
+    if isinstance(gpus_config, int):
+        return [gpus_config]
+
+    gpus_text = str(gpus_config).strip()
+    if not gpus_text:
+        raise ValueError("'gpus' must not be empty")
+
+    if gpus_text.lower() not in {"all", "auto"}:
+        return [int(item.strip()) for item in gpus_text.split(",") if item.strip()]
+
+    visible_devices = os.environ.get("CUDA_VISIBLE_DEVICES")
+    if visible_devices:
+        visible = [item.strip() for item in visible_devices.split(",") if item.strip()]
+        if visible and all(item.isdigit() for item in visible):
+            return [int(item) for item in visible]
+
+    try:
+        output = subprocess.check_output(
+            ["nvidia-smi", "--query-gpu=index", "--format=csv,noheader,nounits"],
+            text=True,
+            timeout=10,
+        )
+        gpu_ids = [int(line.strip()) for line in output.splitlines() if line.strip()]
+        if gpu_ids:
+            return gpu_ids
+    except Exception:
+        pass
+
+    try:
+        import torch
+
+        count = torch.cuda.device_count()
+        if count > 0:
+            return list(range(count))
+    except Exception:
+        pass
+
+    raise RuntimeError("No CUDA GPUs detected for gpus: all")
+
+
+def load_yaml_config(yaml_path: str) -> dict:
+    """Load configuration from YAML file."""
+    import yaml
+
+    with open(yaml_path, "r") as f:
+        return yaml.safe_load(f)
+
+
+def apply_slice(items: list, start: int | None, end: int | None) -> list:
+    """Apply start/end slice to a list with bounds checking."""
+    if start is None and end is None:
+        return items
+
+    slice_start = max(0, start if start is not None else 0)
+    slice_end = min(end if end is not None else len(items), len(items))
+    slice_end = max(slice_start, slice_end)
+
+    return items[slice_start:slice_end]
+
+
+def configure_teacache(transport, config: dict) -> None:
+    """Configure TeaCache reuse strategy on SampleTransport."""
+    from inference.pipeline.teacache import setup_teacache
+
+    setup_teacache(
+        rel_l1_thresh=config["rel_l1_thresh"],
+        warmup_steps=config["warmup_steps"],
+        log=config.get("log", False),
+    )
+
+
+def configure_kv_cache(transport, config: dict) -> None:
+    """Configure KV cache compression if enabled."""
+    if not config.get("compress_kv_cache", False):
+        transport.compress_kv_cache = False
+        return
+
+    print("KV cache compression is enabled.")
+    transport.compress_kv_cache = True
+
+    assert config.get("total_cache_chunk_nums") is not None
+
+    compression_config = {
+        "method_config": {
+            "compress_strategy": config["compress_strategy"],
+            "mix_lambda": config["mix_lambda"],
+            "query_granularity": config["query_granularity"],
+            "score_weighting_method": config.get("score_weighting_method") or "no_weight",
+            "power": config.get("power", 3),
+        },
+    }
+
+    from inference.pipeline.kvcompress import replace_magi
+
+    replace_magi(compression_config)
+
+
+def configure_flowcache(transport, config: dict) -> None:
+    """Configure FlowCache reuse strategy on SampleTransport."""
+    from inference.pipeline.flowcache import setup_flowcache
+
+    configure_kv_cache(transport, config)
+
+    setup_flowcache(
+        rel_l1_thresh=config["rel_l1_thresh"],
+        warmup_steps=config["warmup_steps"],
+        discard_nearly_clean_chunk=config.get("discard_nearly_clean_chunk", False),
+        log=config.get("log", False),
+        total_cache_chunk_nums=config.get("total_cache_chunk_nums", 5),
+        compress_kv_cache=config.get("compress_kv_cache", False),
+    )
+
+
+def configure_reuse_strategy(config: dict) -> None:
+    """Configure the appropriate reuse strategy on SampleTransport."""
+    from inference.pipeline.video_generate import SampleTransport
+
+    strategy = config["reuse_strategy"]
+
+    if strategy == "original":
+        return
+    if strategy == "all":
+        configure_teacache(SampleTransport, config)
+    elif strategy == "chunkwise":
+        configure_flowcache(SampleTransport, config)
+    else:
+        raise ValueError(f"Unknown reuse strategy: {strategy}")
+
+
+def setup_environment(gpu_id: int) -> None:
+    """Set up environment variables for a GPU worker process."""
+    os.environ["CUDA_VISIBLE_DEVICES"] = str(gpu_id)
+    os.environ["WORLD_SIZE"] = "1"
+    os.environ["RANK"] = "0"
+    os.environ["LOCAL_RANK"] = "0"
+    os.environ["MASTER_ADDR"] = "localhost"
+    os.environ["MASTER_PORT"] = str(DEFAULT_BASE_PORT + gpu_id)
+
+    # Enable pdb terminal debugging
+    sys.stdin = open(0)
+
+
+def filter_existing_samples(samples: list, config: dict) -> list:
+    """Filter out samples whose output files already exist."""
+    if config["benchmark"] == "vbench":
+        return [
+            sample
+            for sample in samples
+            if not os.path.exists(os.path.abspath(os.path.join(config["save_path"], f"{sample}-0.mp4")))
+        ]
+    else:  # physicsiq
+        return [
+            sample for sample in samples if not os.path.exists(sample["output_path"])
+        ]
+
+
+def assign_samples_to_gpu(
+    samples: list, gpu_id: int, rank: int, num_gpus: int
+) -> list:
+    """Divide samples across GPUs and return the subset for this GPU."""
+    samples_per_gpu = (len(samples) + num_gpus - 1) // num_gpus
+    start_idx = rank * samples_per_gpu
+    end_idx = min(start_idx + samples_per_gpu, len(samples))
+    return samples[start_idx:end_idx]
+
+
+def process_vbench_sample(pipeline, prompt: str, config: dict, gpu_id: int) -> None:
+    """Process a single vbench text-to-video sample."""
+    output_path = os.path.abspath(os.path.join(config["save_path"], f"{prompt}-0.mp4"))
+
+    if os.path.exists(output_path):
+        print(f"[SKIP GPU {gpu_id}] Already exists: {output_path}")
+        return
+
+    print(f"[GPU {gpu_id}] Generating T2V: '{prompt}' -> {output_path}")
+    pipeline.run_text_to_video(prompt=prompt, output_path=output_path)
+    print(f"[DONE GPU {gpu_id}] Saved: {output_path}")
+
+
+def process_physicsiq_sample(pipeline, sample: dict, gpu_id: int) -> None:
+    """Process a single PhysicsIQ video-to-video sample."""
+    prompt = sample["description"]
+    prefix_video_path = sample["prefix_video_path"]
+    output_path = sample["output_path"]
+
+    if not os.path.exists(prefix_video_path):
+        print(f"[WARN GPU {gpu_id}] Conditioning video not found: {prefix_video_path}")
+        return
+
+    if os.path.exists(output_path):
+        print(f"[SKIP GPU {gpu_id}] Already exists: {output_path}")
+        return
+
+    print(f"[GPU {gpu_id}] Generating V2V: '{prompt}'")
+    print(f"   Input:  {prefix_video_path}")
+    print(f"   Output: {output_path}")
+
+    pipeline.run_video_to_video(
+        prompt=prompt,
+        prefix_video_path=prefix_video_path,
+        output_path=output_path,
+    )
+    print(f"[DONE GPU {gpu_id}] Saved: {output_path}")
+
+
+def worker_process(gpu_id: int, rank: int, config: dict, all_samples: list) -> None:
+    """Independent worker running on each GPU."""
+    setup_environment(gpu_id)
+    configure_reuse_strategy(config)
+
+    try:
+        magi_root = subprocess.check_output(
+            ["git", "rev-parse", "--show-toplevel"]
+        ).decode().strip()
+        os.environ["MAGI_ROOT"] = magi_root
+        os.environ["PYTHONPATH"] = f"{magi_root}:{os.environ.get('PYTHONPATH', '')}"
+    except Exception as e:
+        print(f"[GPU {gpu_id}] Failed to set MAGI_ROOT: {e}")
+        return
+
+    filtered_samples = filter_existing_samples(all_samples, config)
+
+    if not filtered_samples:
+        print(f"[GPU {gpu_id}] No samples need to be generated.")
+        return
+
+    print(f"Processing {len(filtered_samples)} samples.")
+
+    my_samples = assign_samples_to_gpu(
+        filtered_samples, gpu_id, rank, config["num_gpus"]
+    )
+
+    if not my_samples:
+        print(f"[GPU {gpu_id}] No samples assigned.")
+        return
+
+    print(f"[GPU {gpu_id}] Assigned {len(my_samples)} samples")
+
+    from inference.pipeline.entry import MagiPipeline
+
+    print(f"[GPU {gpu_id}] Loading model...")
+    pipeline = MagiPipeline(config["config_file"])
+    print(f"[GPU {gpu_id}] Model loaded.")
+
+    process_func = (
+        process_vbench_sample if config["benchmark"] == "vbench" else process_physicsiq_sample
+    )
+
+    for sample in my_samples:
+        process_func(pipeline, sample, config, gpu_id)
+
+    print(f"[GPU {gpu_id}] Completed.")
+
+
+def build_conditioning_video_path(
+    data_root: str, vid_id: str, scenario: str, fps: int
+) -> str:
+    """Construct the path to the conditioning video file."""
+    conditioning_dir = os.path.join(
+        data_root, "physics-IQ-benchmark", "split-videos", "conditioning", f"{fps}FPS"
+    )
+    match_suffix = re.search(r"_(.*)", scenario)
+    suffix = match_suffix.group(1) if match_suffix else ""
+    filename = f"{vid_id}_conditioning-videos_{fps}FPS_{suffix}"
+    return os.path.join(conditioning_dir, filename)
+
+
+def load_physicsiq_samples(config: dict) -> list[dict]:
+    """Load sample list from PhysicsIQ dataset."""
+    data_root = config["physicsiq_data_dir"]
+    descriptions_csv = os.path.join(data_root, "descriptions", "descriptions.csv")
+    output_dir = config["save_path"]
+
+    if not os.path.exists(descriptions_csv):
+        raise FileNotFoundError(f"descriptions.csv not found at {descriptions_csv}")
+
+    os.makedirs(output_dir, exist_ok=True)
+
+    samples = []
+    with open(descriptions_csv, mode="r") as f:
+        reader = csv.DictReader(f)
+        for row in reader:
+            scenario = row["scenario"].strip()
+            match_id = re.match(r"^(\d+)_", scenario)
+
+            if not match_id:
+                print(f"Cannot extract ID from scenario: {scenario}")
+                continue
+
+            vid_id = match_id.group(1).zfill(4)
+            description = row["description"]
+            generated_video_name = row["generated_video_name"]
+            prefix_video_path = build_conditioning_video_path(
+                data_root, vid_id, scenario, PHYSICSIQ_FPS
+            )
+            output_path = os.path.join(output_dir, generated_video_name)
+
+            os.makedirs(os.path.dirname(output_path), exist_ok=True)
+
+            samples.append({
+                "vid_id": vid_id,
+                "scenario": scenario,
+                "description": description,
+                "generated_video_name": generated_video_name,
+                "prefix_video_path": prefix_video_path,
+                "output_path": output_path,
+            })
+
+    # PhysicsIQ samples are duplicated; take only the first half
+    unique_count = len(samples) // 2
+    samples = samples[:unique_count]
+
+    print(f"Loaded {unique_count} PhysicsIQ samples.")
+
+    return apply_slice(samples, config.get("start"), config.get("end"))
+
+
+def load_vbench_samples(config: dict) -> list[str]:
+    """Load prompt list from vbench dimension file."""
+    prompt_dir = config["vbench_prompt_dir"]
+    dimension = config.get("dimension")
+
+    if not dimension:
+        raise ValueError("For vbench, 'dimension' must be specified in config")
+
+    prompt_file = os.path.join(prompt_dir, f"{dimension}.txt")
+
+    if not os.path.exists(prompt_file):
+        raise FileNotFoundError(f"Prompt file not found: {prompt_file}")
+
+    with open(prompt_file, "r") as f:
+        prompts = [line.strip() for line in f if line.strip()]
+
+    return apply_slice(prompts, config.get("start"), config.get("end"))
+
+
+def setup_save_path(config: dict) -> None:
+    """Configure the output save path based on benchmark type."""
+    base_path = config["base_save_path"]
+
+    if config["benchmark"] == "vbench":
+        dimension = config.get("dimension")
+        videos_dir = os.path.join(base_path, "videos", dimension) if dimension else None
+        config["save_path"] = videos_dir if videos_dir else os.path.join(base_path, "videos")
+    elif config["benchmark"] == "physicsiq":
+        config["save_path"] = os.path.join(base_path, "videos")
+
+    os.makedirs(config["save_path"], exist_ok=True)
+
+
+def main() -> None:
+    """Entry point for video sampling script."""
+    parser = argparse.ArgumentParser(
+        description="Video sampling script using YAML configuration"
+    )
+    parser.add_argument("yaml_config", type=str, help="Path to YAML configuration file")
+    args = parser.parse_args()
+
+    config = load_yaml_config(args.yaml_config)
+    print(f"Loaded configuration from: {args.yaml_config}")
+
+    setup_save_path(config)
+
+    gpu_ids = resolve_gpu_ids(config["gpus"])
+    config["num_gpus"] = len(gpu_ids)
+
+    benchmark = config["benchmark"]
+    if benchmark == "vbench":
+        all_samples = load_vbench_samples(config)
+    elif benchmark == "physicsiq":
+        data_root = config["physicsiq_data_dir"]
+        if not os.path.exists(data_root):
+            raise FileNotFoundError(f"Data directory not found: {data_root}")
+        all_samples = load_physicsiq_samples(config)
+    else:
+        raise ValueError(f"Invalid benchmark: {benchmark}")
+
+    print(f"Total samples: {len(all_samples)}")
+    print(f"GPUs: {gpu_ids}")
+    print(f"Output: {config['save_path']}")
+    print(f"Config: {config['config_file']}")
+
+    processes = []
+    for rank, gpu_id in enumerate(gpu_ids):
+        p = mp.Process(target=worker_process, args=(gpu_id, rank, config, all_samples))
+        p.start()
+        processes.append(p)
+
+    for p in processes:
+        p.join()
+
+    failed = [p.exitcode for p in processes if p.exitcode != 0]
+    if failed:
+        raise RuntimeError(f"{len(failed)} worker process(es) failed with exit codes: {failed}")
+
+
+if __name__ == "__main__":
+    main()

FlowCache/FlowCache4MAGI-1-dev5-history/README.md

@@ -0,0 +1,71 @@
+# FLOW CACHING FOR AUTOREGRESSIVE VIDEO GENERATION
+
+This repository provides the official implementation of **FlowCache** on **MAGI-1** model, a caching-based acceleration method for autoregressive video generation models.
+
+
+## 🚀 Installation
+
+Please follow the installation instructions provided in the [MAGI-1](https://github.com/SandAI-org/MAGI-1), as this implementation is built on top of MAGI-1.
+
+---
+
+## ▶️ Usage
+
+### 1. Single Video Generation
+
+Run accelerated generation using FlowCache:
+
+```bash
+# FlowCache for text-to-video generation
+bash scripts/single_run/flowcache_t2v.sh
+
+# FlowCache for video-to-video generation
+bash scripts/single_run/flowcache_v2v.sh
+
+# Baseline acceleration method (TeaCache) for text-to-video
+bash scripts/single_run/teacache_t2v.sh
+
+# Baseline acceleration method (TeaCache) for video-to-video
+bash scripts/single_run/teacache_v2v.sh
+```
+
+### 2. Benchmark Sampling
+
+Generate videos for evaluation on standard benchmarks:
+
+```bash
+# VBench
+bash scripts/sample/flowcache_vbench.sh
+bash scripts/sample/teacache_vbench.sh
+
+# PhysicsIQ
+bash scripts/sample/flowcache_physicsiq.sh
+bash scripts/sample/teacache_physicsiq.sh
+```
+
+### 3. Quality Evaluation
+
+Compute perceptual and structural similarity metrics between original and accelerated generations:
+
+```bash
+bash scripts/metric.sh
+```
+
+---
+
+## ⚙️ Key Parameters
+
+| Parameter | Description |
+|----------|-------------|
+| `rel_l1_thresh` | Relative L1 distance threshold for cache reuse decision |
+| ` warmup_steps` | Number of denoising steps where reuse is disabled |
+| `total_cache_chunk_nums` (`B_total`) | Total number of cache chunks maintained |
+| `compress_strategy` | Granularity for selecting important KV caches: `token`, `frame`, or `chunk` |
+| `query_granularity` | Granularity for importance scoring: `token`, `frame`, or `chunk` |
+| `mix_lambda` | Weight balancing importance and redundancy (default: `0.07`) |
+| `mode` | Generation mode: `t2v` (text-to-video), `i2v` (image-to-video), or `v2v` (video-to-video) |
+| `prompt` | Input prompt for conditional generation |
+| `output_path` | Path to save generated videos |
+| `config_file` | Path to MAGI-1 model configuration |
+
+---

FlowCache/FlowCache4MAGI-1-dev5-history/README_DEV5_HISTORY.md

@@ -0,0 +1,218 @@
+# dev5-history：利用 AR 历史 Chunk 的改进方向
+
+基于 `FlowCache4MAGI-1-dev4-detail`（Motion + Detail 双度量），dev5 探索 **自回归（AR）已生成历史 chunk** 如何进一步改进 cache 决策与画质。
+
+## AR 场景下「历史」指什么？
+
+MAGI-1 自回归生成时，在 denoise chunk `i` 的时刻，系统已经拥有：
+
+| 历史信息 | 代码/机制中已有 | 当前 dev4 使用情况 |
+|----------|----------------|-------------------|
+| 前序 **clean chunk** 的 latent | `x` 前缀、KV cache | 仅 KV 压缩，未用于 token 策略 |
+| 前一 chunk 末帧 | `prev_chunk_last_frame` | 仅算 chunk 首帧 motion |
+| 各 chunk 上一步 latent | `prev_latent_chunks` | 存了但几乎未参与 Phase2 决策 |
+| 各 chunk embedding 特征 | `prev_metric_chunks` | 仅与 **当前步** 比 Δ_chunk |
+| 前序 clean 特征的 KV | `compress_kv` + tracker | Attention 用，reuse 策略未显式用 |
+| 各 chunk residual / velocity | `previous_residual` | 仅 **本 chunk 内** 复用 |
+
+**核心洞察**：AR 下历史 chunk 是 **已收敛的时空上下文**，比单步 Δ 或单帧 motion 更稳定，可用于：
+- 更准的「该不该算」
+- 更准的「算哪些 token」
+- 更准的「inactive token 复用什么」
+
+---
+
+## Idea 分级（推荐实施顺序）
+
+### ★★★ Tier 1：实现成本低、与 dev4 正交、收益可期
+
+#### Idea 1 — 跨 Chunk 时空累积器（History Accumulator）
+
+**动机**：dev4 的 `A[p]` 在每个 chunk 内独立；AR 下同一空间位置在多 chunk 有连续轨迹。
+
+**做法**：
+```
+A_hist[p] = decay · A_hist[p] + W_combined[p] · Δ_step
+```
+- 对 **空间对齐** 的 token（同 `(h,w)`，跨 chunk 用上一 chunk 对应列/网格对齐，首帧对齐 `prev_chunk_last_frame`）
+- `decay ∈ (0,1)` 或 chunk 边界衰减
+- active 判定：`A_hist[p] > τ` 替代或补充现有 `A[p]`
+
+**预期**：运动物体沿时间「扫过」的空间位置会持续 active；长期静止背景更快 reuse。
+
+**改动点**：`MotionDetailCache` → `HistoryAwareCache`，新增 `cross_chunk_accumulator` dict。
+
+---
+
+#### Idea 2 — 历史 Clean Latent 锚定（Clean History Anchor）
+
+**动机**：AR 中 chunk `i-k` 已 clean 的 latent 是 **真实生成结果**，可作为锚点。
+
+**做法**：
+- 维护每个 spatial token 最近 `H` 个 clean chunk 的 latent 统计（均值/最后一帧）
+- 当前 denoise latent 与锚的距离：
+  ```
+  d_anchor[p] = ||x_current[p] - x_clean_hist[p]|| / (||x_clean_hist[p]|| + ε)
+  ```
+- 权重：`W_anchor = soft_map(d_anchor)`，高距离 → 更需计算
+- 融合：`W_final = combine(W_motion, W_detail, W_anchor)`
+
+**预期**：减少「看起来不动但 latent 在 drift」区域的误 reuse；对 dancing 等场景尤其有用。
+
+**改动点**：在 `store_latent_chunk` 时，若 chunk 刚 clean，写入 `clean_latent_history[chunk_id]`。
+
+---
+
+#### Idea 3 — 历史 Active 轨迹传播（Active Streak Propagation）
+
+**动机**：dev4 仅看当前步 motion/detail；AR 下「持续在变的区域」有惯性。
+
+**做法**：
+- 记录每个 token 最近 `S` 步是否 active（或 active 计数 streak）
+- 传播权重：
+  ```
+  W_streak[p] = α_s + (1-α_s) · min(streak[p] / S_max, 1)
+  ```
+- streak 高 → 提高 `W_combined` 或直接 pin active
+
+**预期**：减少运动主体边缘的 flickering（时而算时而不算）；略降 reuse，提 PSNR。
+
+---
+
+### ★★ Tier 2：中等工程量、需与 KV/Attention 协同
+
+#### Idea 4 — 历史特征一致性门控（Embedding Consistency Gate）
+
+**做法**：
+- 不仅比较 `metric_chunks[i]` 与上一步，还与 **历史 clean chunk 的 embedding**（需存 `clean_metric_history`）比：
+  ```
+  Δ_hist = rel_L1(f_current, f_clean_ref)
+  ```
+- 若与 clean 历史一致且 streak 低 → 允许 skip forward
+- 若与 clean 历史偏离增大 → 强制 active
+
+**与 FlowCache 关系**：FlowCache 的 Δ 是 step-wise；这是 **chunk-wise AR 锚定**。
+
+---
+
+#### Idea 5 — 历史 Residual 外推（Historical Residual Extrapolation）
+
+**动机**：inactive token 目前复用 `previous_residual`（本 chunk 上一步）；AR 可提供更长上下文。
+
+**做法**：
+- 存每个 token 最近 `K` 次 **被计算时** 的 residual 序列
+- inactive 时：
+  ```
+  r_reuse = (1-β)·r_prev_step + β·r_hist_extrap
+  ```
+  其中 `r_hist_extrap` 可为线性外推或上一 clean chunk 同位置 residual
+
+**风险**：外推错误会累积；建议仅对 **streak=0 且 anchor 距离小** 的 token 使用。
+
+---
+
+#### Idea 6 — 多 Chunk Motion 场（Temporal Motion Field）
+
+**做法**：
+- 用 `{x_chunk[t] - x_chunk[t-1]}` 在多 chunk 上建 low-res motion field
+- 当前 token 的 motion 不只看相邻帧，还看 **历史 motion 幅值的分位数**
+- 高历史 motion 区域（舞台中央舞者）→ 永久提高 W；静态背景 → 降低
+
+**与 dev4 区别**：dev4 motion 是 **单 chunk 内** 帧差；这是 **跨 chunk 运动先验**。
+
+---
+
+### ★ Tier 3：研究向 / 改动较大
+
+#### Idea 7 — 历史 Conditioned Sparse Attention Mask
+
+- 利用 KV 中 clean history 的 attention score 或 query-key 相似度，决定哪些 query token 必须重算
+- 需改 DiT sparse forward，工程量大
+
+#### Idea 8 — Chunk 级 AR 难度预测器
+
+- 用前 `n` 个 chunk 的 reuse 率、平均 Δ、motion 统计，预测 chunk `i` 的「难度」
+- 动态调节 τ、detail_lambda（简单 chunk 更激进 reuse）
+
+#### Idea 9 — 显式 Warp 对齐历史 Token
+
+- 对 camera motion 场景，用光流/块匹配将 chunk `i-1` token 对齐到 chunk `i` 网格再比特征
+- 最准但最重；适合 v2v 或有 camera 运动的数据
+
+---
+
+## 推荐 dev5 首版实现路线（MVP）
+
+建议 **Idea 1 + Idea 2 + Idea 3** 组合为 `HistoryAwareCache`：
+
+```
+W_final = blend( max(W_motion, W_detail), W_anchor, λ_a )
+         × (1 + γ · W_streak)
+
+A_cross[p] += W_final[p] · Δ_chunk
+active[p] = (A_cross[p] > τ) OR (A_local[p] > τ)   # 双累积器 OR 逻辑
+```
+
+**超参（新增 yaml）**：
+```yaml
+history_decay: 0.7          # 跨 chunk 累积衰减
+history_anchor_horizon: 3   # 参考最近 H 个 clean chunk
+history_streak_len: 5       # streak 窗口
+history_anchor_lambda: 0.3  # anchor 权重
+history_streak_gamma: 0.2   # streak 加成
+```
+
+**验证指标**（延续 dev3/dev4 sweep）：
+- PSNR vs FlowCache baseline（240f）
+- reuse_rate / wall_time
+- 运动区域 active 比例稳定性（方差↓更好）
+
+---
+
+## 代码结构规划（dev5）
+
+```
+inference/pipeline/cache/
+├── motioncache.py           # dev3 基类
+├── motiondetailcache.py     # dev4
+└── historycache.py          # dev5 HistoryAwareCache（待实现）
+
+yaml_config/single_run/
+└── historycache_config.yaml # dev5 默认 + 历史相关超参
+
+README_DEV5_HISTORY.md       # 本文档
+```
+
+---
+
+## 与 dev4 最优配置的继承关系
+
+dev5 默认继承 dev4 sweep 最优（τ=0.012, blend, detail_window=3, detail_lambda=0.3），在其上叠加 history 项；便于 ablation：
+
+| 实验 | 配置 |
+|------|------|
+| A0 | dev4 best（对照） |
+| A1 | dev5 仅 Idea1 跨 chunk 累积 |
+| A2 | dev5 Idea1+2 |
+| A3 | dev5 Idea1+2+3 全量 |
+
+---
+
+## 风险与注意点
+
+1. **Chunk 边界对齐**：latent 空间 `(h,w)` 在 chunk 间是否严格对齐需确认 patch/grid 一致
+2. **Clean 时机**：只有 `chunk_denoise_count == num_steps` 后的 latent 才能进 history anchor
+3. **内存**：存 H 个 clean chunk latent 会增加 CPU/GPU 缓存，可只存 downsampled 或 embedding
+4. **CFG**：batch=2 时 history 需 per-batch 维护
+5. **不要破坏 Phase1**：前 K 步仍 chunk-wise FlowCache，history 仅 Phase2 启用
+
+---
+
+## 下一步
+
+1. 确认优先实现的 Idea（建议 1+2+3）
+2. 在 `historycache.py` 实现 `HistoryAwareCache`
+3. 复用 dev4 sweep 脚本做 dev5 ablation
+4. 在 `a woman dancing.` + VBench 子集上对比 dev4 best
+
+如需我直接在 dev5 实现 **HistoryAwareCache MVP（Idea 1+2+3）**，可以指定优先哪几个 Idea。

FlowCache/FlowCache4MAGI-1-dev5-history/README_MOTIONCACHE.md

@@ -0,0 +1,92 @@
+# MotionCache on MAGI-1 (dev3-motion)
+
+基于 [FlowCache4MAGI-1](../FlowCache4MAGI-1) 的 **MotionCache** 复现分支，对应论文：
+
+> Xu et al., 2026 — *Motion-Aware Caching for Efficient Autoregressive Video Generation*
+
+官方代码仓库 [ywlq/MotionCache](https://github.com/ywlq/MotionCache) 尚未发布实现，本目录依据论文方法在 MAGI-1 推理框架上完成首版复现。
+
+## 与 FlowCache 的核心差异
+
+| 维度 | FlowCache | MotionCache |
+|------|-----------|-------------|
+| 粒度 | Chunk 级全有或全无 | Phase 2 为 Token（latent 帧×空间）级 |
+| 跳过策略 | 相对 L1 累积阈值 | 运动重要性加权累积 |
+| 调度 | 单一 chunk-wise 策略 | 两阶段 coarse-to-fine |
+
+### 算法概要
+
+1. **全局 Warm-up（m 步）**：前 `warmup_steps` 步禁止 cache reuse
+2. **Phase 1（K 步）**：chunk-wise 二值决策，与 FlowCache 相同
+3. **Phase 2**：基于帧间 latent 差计算运动重要性 `M`，经 soft-mapping 得到权重 `W ∈ [α, 1]`
+4. **Token 累积**：`A[p] += W[p] · Δ_chunk`，当 `A[p] > τ` 时该 token 触发 DiT 计算
+5. **Integrate**：active token 正常积分；inactive token 复用缓存 residual
+
+### MAGI-1 默认超参（论文 Appendix C）
+
+| 参数 | 值 | 说明 |
+|------|-----|------|
+| `alpha` | 0.5 | 静态区域权重下限 |
+| `phase1_steps` (K) | 9 | chunk-wise 阶段持续步数 |
+| `warmup_steps` (m) | 5 | 全局禁止 reuse 的步数 |
+| `rel_l1_thresh` (τ) | 0.015 (slow) / 0.025 (fast) | token 累积阈值 |
+
+## 快速运行
+
+```bash
+cd FlowCache4MAGI-1-dev3-motion
+
+# MotionCache-slow（论文 Table 1 配置）
+bash scripts/single_run/motioncache_t2v.sh
+
+# MotionCache-fast
+MOTIONCACHE_CONFIG=yaml_config/single_run/motioncache_config_fast.yaml \
+  bash scripts/single_run/motioncache_t2v.sh
+```
+
+需先按 FlowCache4MAGI-1 说明安装依赖并下载 MAGI-1 权重（`downloads/` 目录可通过软链接指向原项目）。
+
+## 代码结构
+
+```
+inference/pipeline/
+├── motioncache.py              # 入口与 forward/integrate monkey-patch
+└── cache/
+    ├── motioncache.py          # MotionWiseCache 核心逻辑
+    └── sparse_utils.py         # Phase 2 gather/scatter 与 sparse meta_args
+
+inference/model/dit/dit_module.py   # sparse KV cache 写入 + flash_attn 稀疏 q 分支
+inference/common/dataclass.py       # ModelMetaArgs.sparse_active_indices
+
+yaml_config/single_run/
+├── motioncache_config.yaml     # slow 配置
+└── motioncache_config_fast.yaml
+```
+
+## 当前实现说明（严格对齐论文 §5.3 Phase 2）
+
+- **Phase 1（前 K 个 denoise step / chunk）**：与 FlowCache 完全一致的 chunk-wise 策略（已验证 PSNR=∞）
+- **Phase 2 入口**：每个 chunk 的第 K 步强制全量计算，建立 token-wise 阶段的 residual 基准
+- **Phase 2 运动感知累积**：latent 帧间差 → soft-mapping 权重 W → 累积器 A；`A > τ` 的 token 为 active
+- **Phase 2 稀疏 forward（论文 5.3）**：
+  - 全部 inactive → 跳过 DiT forward，复用 chunk residual
+  - 部分 active → **gather active patch tokens → compact DiT forward → scatter 回完整序列**
+  - KV cache 仅更新 active 位置（`_compresskv_adjust_key_and_value` sparse 分支）
+  - integrate 时 active token 正常积分，inactive token 复用 `previous_residual`
+- **运动 proxy**：latent 空间帧间 L1 差（论文 Eq. 9-10）
+- **跨 chunk 连续性**：chunk 首帧与上一 chunk 末帧比较
+
+### 验证结果（`a woman dancing.`，240×720×720，vs FlowCache baseline）
+
+| 运行 | PSNR | 说明 |
+|------|------|------|
+| phase1only | ∞ | 与 FlowCache 逐像素一致 |
+| sparse2（论文对齐 sparse forward） | **20.44 dB** | 无黑帧，reuse_rate≈15.6% |
+| final（整 chunk fallback） | 20.48 dB | 画质等价，推理慢 ~14% |
+
+sparse2 日志示例：`active_ratio=1.51%` 时仍走 gather/scatter 稀疏路径；`active_ratio=0%` 时 `skip_forward=True`。
+
+## 参考
+
+- FlowCache: chunk-wise cache + KV compression
+- MotionCache 论文预期 MAGI-1 加速：slow 1.64×，fast 2.07×（Table 1）

FlowCache/FlowCache4MAGI-1-dev5-history/README_MOTIONDETAIL.md

@@ -0,0 +1,71 @@
+# MotionDetailCache on MAGI-1 (dev4-detail)
+
+基于 [FlowCache4MAGI-1-dev3-motion](../FlowCache4MAGI-1-dev3-motion) 的扩展分支，在 MotionCache 的 **motion 度量** 之外，增加 **spatial detail 度量**，Phase 2 active token 由两者共同决定。
+
+## 与 dev3 (MotionCache) 的差异
+
+| 维度 | dev3 MotionCache | dev4 MotionDetailCache |
+|------|------------------|------------------------|
+| Phase 2 权重 | 仅 motion 权重 W_m | motion W_m + detail W_d 融合 |
+| Detail proxy | — | latent 局部 k×k 空间方差 |
+| 适用场景 | 运动区域优先计算 | 静止纹理/边缘/语义细节也优先保留 |
+
+## Detail 度量
+
+对每个 latent 像素 `(t, h, w)`：
+
+1. 通道聚合幅度：`mag = mean(|x|, dim=C)` → `[N, T, H, W]`
+2. 局部空间方差：在 `detail_window_size×detail_window_size` 窗口内计算 `Var(mag)`
+3. Soft-mapping：帧内 min-max 归一化 → `W_d ∈ [detail_alpha, 1]`
+
+高方差区域对应边缘、纹理、物体边界等 heterogeneous 邻域。
+
+## Motion + Detail 融合
+
+```
+W_combined = combine(W_motion, W_detail)
+A[p] += W_combined[p] · Δ_chunk
+active[p] = A[p] > τ
+```
+
+`weight_combine_mode`（默认 `max`）：
+
+| 模式 | 公式 | 特点 |
+|------|------|------|
+| `max` | max(W_m, W_d) | 运动或细节任一显著即可加速累积（推荐） |
+| `product` | W_m × W_d | 两者同时高才快速激活，更激进 reuse |
+| `blend` | (1-λ)W_m + λW_d | 线性权衡，λ=`detail_lambda` |
+
+## 默认超参
+
+继承 dev3 的 motion 参数，并新增：
+
+| 参数 | 默认值 | 说明 |
+|------|--------|------|
+| `detail_alpha` | 0.5 | detail 权重下限 |
+| `detail_window_size` | 3 | 局部方差窗口（奇数） |
+| `detail_lambda` | 0.5 | blend 模式下的 detail 权重 |
+| `weight_combine_mode` | max | 融合方式 |
+
+## 快速运行
+
+```bash
+cd FlowCache4MAGI-1-dev4-detail
+conda activate magi
+export CUDA_VISIBLE_DEVICES=0
+export MASTER_ADDR=localhost
+bash scripts/single_run/motiondetail_t2v.sh
+```
+
+## 代码结构
+
+```
+inference/pipeline/cache/
+├── motioncache.py          # dev3 基类 MotionWiseCache
+└── motiondetailcache.py    # dev4 MotionDetailCache
+
+yaml_config/single_run/motiondetail_config.yaml
+scripts/single_run/motiondetail_t2v.sh
+```
+
+其余 sparse forward、KV cache、integrate 逻辑与 dev3 完全一致。

FlowCache/FlowCache4MAGI-1-dev5-history/requirements.txt

@@ -0,0 +1,21 @@
+accelerate==0.32.1
+beautifulsoup4==4.13.4
+debugpy==1.8.14
+diffusers==0.29.2
+einops>=0.6.0
+ffmpeg-python
+# flash-attn==2.4.2
+flashinfer-python==0.2.0.post2 --extra-index-url https://flashinfer.ai/whl/cu124/torch2.4/
+ftfy==6.2.0
+gpustat==1.1.1
+imageio==2.34.0
+imageio[ffmpeg]
+matplotlib==3.10.1
+numpy==1.26.4
+protobuf==5.28.3
+rich==14.0.0
+sentencepiece==0.2.0
+timm==1.0.15
+torchdiffeq==0.2.4
+transformers==4.42.3
+tqdm

FlowCache/FlowCache4MAGI-1-dev5-history/sample_video.py

@@ -0,0 +1,429 @@
+# Copyright 2024 MAGI Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import re
+import sys
+import argparse
+import csv
+import subprocess
+from pathlib import Path
+
+import multiprocessing as mp
+
+# Constants
+DEFAULT_BASE_PORT = 29510
+PHYSICSIQ_FPS = 24
+
+
+def resolve_gpu_ids(gpus_config) -> list[int]:
+    """Resolve explicit GPU IDs or auto-detect all currently visible GPUs."""
+    if isinstance(gpus_config, int):
+        return [gpus_config]
+
+    gpus_text = str(gpus_config).strip()
+    if not gpus_text:
+        raise ValueError("'gpus' must not be empty")
+
+    if gpus_text.lower() not in {"all", "auto"}:
+        return [int(item.strip()) for item in gpus_text.split(",") if item.strip()]
+
+    visible_devices = os.environ.get("CUDA_VISIBLE_DEVICES")
+    if visible_devices:
+        visible = [item.strip() for item in visible_devices.split(",") if item.strip()]
+        if visible and all(item.isdigit() for item in visible):
+            return [int(item) for item in visible]
+
+    try:
+        output = subprocess.check_output(
+            ["nvidia-smi", "--query-gpu=index", "--format=csv,noheader,nounits"],
+            text=True,
+            timeout=10,
+        )
+        gpu_ids = [int(line.strip()) for line in output.splitlines() if line.strip()]
+        if gpu_ids:
+            return gpu_ids
+    except Exception:
+        pass
+
+    try:
+        import torch
+
+        count = torch.cuda.device_count()
+        if count > 0:
+            return list(range(count))
+    except Exception:
+        pass
+
+    raise RuntimeError("No CUDA GPUs detected for gpus: all")
+
+
+def load_yaml_config(yaml_path: str) -> dict:
+    """Load configuration from YAML file."""
+    import yaml
+
+    with open(yaml_path, "r") as f:
+        return yaml.safe_load(f)
+
+
+def apply_slice(items: list, start: int | None, end: int | None) -> list:
+    """Apply start/end slice to a list with bounds checking."""
+    if start is None and end is None:
+        return items
+
+    slice_start = max(0, start if start is not None else 0)
+    slice_end = min(end if end is not None else len(items), len(items))
+    slice_end = max(slice_start, slice_end)
+
+    return items[slice_start:slice_end]
+
+
+def configure_teacache(transport, config: dict) -> None:
+    """Configure TeaCache reuse strategy on SampleTransport."""
+    from inference.pipeline.teacache import setup_teacache
+
+    setup_teacache(
+        rel_l1_thresh=config["rel_l1_thresh"],
+        warmup_steps=config["warmup_steps"],
+        log=config.get("log", False),
+    )
+
+
+def configure_kv_cache(transport, config: dict) -> None:
+    """Configure KV cache compression if enabled."""
+    if not config.get("compress_kv_cache", False):
+        transport.compress_kv_cache = False
+        return
+
+    print("KV cache compression is enabled.")
+    transport.compress_kv_cache = True
+
+    assert config.get("total_cache_chunk_nums") is not None
+
+    compression_config = {
+        "method_config": {
+            "compress_strategy": config["compress_strategy"],
+            "mix_lambda": config["mix_lambda"],
+            "query_granularity": config["query_granularity"],
+            "score_weighting_method": config.get("score_weighting_method") or "no_weight",
+            "power": config.get("power", 3),
+        },
+    }
+
+    from inference.pipeline.kvcompress import replace_magi
+
+    replace_magi(compression_config)
+
+
+def configure_flowcache(transport, config: dict) -> None:
+    """Configure FlowCache reuse strategy on SampleTransport."""
+    from inference.pipeline.flowcache import setup_flowcache
+
+    configure_kv_cache(transport, config)
+
+    setup_flowcache(
+        rel_l1_thresh=config["rel_l1_thresh"],
+        warmup_steps=config["warmup_steps"],
+        discard_nearly_clean_chunk=config.get("discard_nearly_clean_chunk", False),
+        log=config.get("log", False),
+        total_cache_chunk_nums=config.get("total_cache_chunk_nums", 5),
+        compress_kv_cache=config.get("compress_kv_cache", False),
+    )
+
+
+def configure_reuse_strategy(config: dict) -> None:
+    """Configure the appropriate reuse strategy on SampleTransport."""
+    from inference.pipeline.video_generate import SampleTransport
+
+    strategy = config["reuse_strategy"]
+
+    if strategy == "original":
+        return
+    if strategy == "all":
+        configure_teacache(SampleTransport, config)
+    elif strategy == "chunkwise":
+        configure_flowcache(SampleTransport, config)
+    else:
+        raise ValueError(f"Unknown reuse strategy: {strategy}")
+
+
+def setup_environment(gpu_id: int) -> None:
+    """Set up environment variables for a GPU worker process."""
+    os.environ["CUDA_VISIBLE_DEVICES"] = str(gpu_id)
+    os.environ["WORLD_SIZE"] = "1"
+    os.environ["RANK"] = "0"
+    os.environ["LOCAL_RANK"] = "0"
+    os.environ["MASTER_ADDR"] = "localhost"
+    os.environ["MASTER_PORT"] = str(DEFAULT_BASE_PORT + gpu_id)
+
+    # Enable pdb terminal debugging
+    sys.stdin = open(0)
+
+
+def filter_existing_samples(samples: list, config: dict) -> list:
+    """Filter out samples whose output files already exist."""
+    if config["benchmark"] == "vbench":
+        return [
+            sample
+            for sample in samples
+            if not os.path.exists(os.path.abspath(os.path.join(config["save_path"], f"{sample}-0.mp4")))
+        ]
+    else:  # physicsiq
+        return [
+            sample for sample in samples if not os.path.exists(sample["output_path"])
+        ]
+
+
+def assign_samples_to_gpu(
+    samples: list, gpu_id: int, rank: int, num_gpus: int
+) -> list:
+    """Divide samples across GPUs and return the subset for this GPU."""
+    samples_per_gpu = (len(samples) + num_gpus - 1) // num_gpus
+    start_idx = rank * samples_per_gpu
+    end_idx = min(start_idx + samples_per_gpu, len(samples))
+    return samples[start_idx:end_idx]
+
+
+def process_vbench_sample(pipeline, prompt: str, config: dict, gpu_id: int) -> None:
+    """Process a single vbench text-to-video sample."""
+    output_path = os.path.abspath(os.path.join(config["save_path"], f"{prompt}-0.mp4"))
+
+    if os.path.exists(output_path):
+        print(f"[SKIP GPU {gpu_id}] Already exists: {output_path}")
+        return
+
+    print(f"[GPU {gpu_id}] Generating T2V: '{prompt}' -> {output_path}")
+    pipeline.run_text_to_video(prompt=prompt, output_path=output_path)
+    print(f"[DONE GPU {gpu_id}] Saved: {output_path}")
+
+
+def process_physicsiq_sample(pipeline, sample: dict, gpu_id: int) -> None:
+    """Process a single PhysicsIQ video-to-video sample."""
+    prompt = sample["description"]
+    prefix_video_path = sample["prefix_video_path"]
+    output_path = sample["output_path"]
+
+    if not os.path.exists(prefix_video_path):
+        print(f"[WARN GPU {gpu_id}] Conditioning video not found: {prefix_video_path}")
+        return
+
+    if os.path.exists(output_path):
+        print(f"[SKIP GPU {gpu_id}] Already exists: {output_path}")
+        return
+
+    print(f"[GPU {gpu_id}] Generating V2V: '{prompt}'")
+    print(f"   Input:  {prefix_video_path}")
+    print(f"   Output: {output_path}")
+
+    pipeline.run_video_to_video(
+        prompt=prompt,
+        prefix_video_path=prefix_video_path,
+        output_path=output_path,
+    )
+    print(f"[DONE GPU {gpu_id}] Saved: {output_path}")
+
+
+def worker_process(gpu_id: int, rank: int, config: dict, all_samples: list) -> None:
+    """Independent worker running on each GPU."""
+    setup_environment(gpu_id)
+    configure_reuse_strategy(config)
+
+    try:
+        magi_root = subprocess.check_output(
+            ["git", "rev-parse", "--show-toplevel"]
+        ).decode().strip()
+        os.environ["MAGI_ROOT"] = magi_root
+        os.environ["PYTHONPATH"] = f"{magi_root}:{os.environ.get('PYTHONPATH', '')}"
+    except Exception as e:
+        print(f"[GPU {gpu_id}] Failed to set MAGI_ROOT: {e}")
+        return
+
+    filtered_samples = filter_existing_samples(all_samples, config)
+
+    if not filtered_samples:
+        print(f"[GPU {gpu_id}] No samples need to be generated.")
+        return
+
+    print(f"Processing {len(filtered_samples)} samples.")
+
+    my_samples = assign_samples_to_gpu(
+        filtered_samples, gpu_id, rank, config["num_gpus"]
+    )
+
+    if not my_samples:
+        print(f"[GPU {gpu_id}] No samples assigned.")
+        return
+
+    print(f"[GPU {gpu_id}] Assigned {len(my_samples)} samples")
+
+    from inference.pipeline.entry import MagiPipeline
+
+    print(f"[GPU {gpu_id}] Loading model...")
+    pipeline = MagiPipeline(config["config_file"])
+    print(f"[GPU {gpu_id}] Model loaded.")
+
+    process_func = (
+        process_vbench_sample if config["benchmark"] == "vbench" else process_physicsiq_sample
+    )
+
+    for sample in my_samples:
+        process_func(pipeline, sample, config, gpu_id)
+
+    print(f"[GPU {gpu_id}] Completed.")
+
+
+def build_conditioning_video_path(
+    data_root: str, vid_id: str, scenario: str, fps: int
+) -> str:
+    """Construct the path to the conditioning video file."""
+    conditioning_dir = os.path.join(
+        data_root, "physics-IQ-benchmark", "split-videos", "conditioning", f"{fps}FPS"
+    )
+    match_suffix = re.search(r"_(.*)", scenario)
+    suffix = match_suffix.group(1) if match_suffix else ""
+    filename = f"{vid_id}_conditioning-videos_{fps}FPS_{suffix}"
+    return os.path.join(conditioning_dir, filename)
+
+
+def load_physicsiq_samples(config: dict) -> list[dict]:
+    """Load sample list from PhysicsIQ dataset."""
+    data_root = config["physicsiq_data_dir"]
+    descriptions_csv = os.path.join(data_root, "descriptions", "descriptions.csv")
+    output_dir = config["save_path"]
+
+    if not os.path.exists(descriptions_csv):
+        raise FileNotFoundError(f"descriptions.csv not found at {descriptions_csv}")
+
+    os.makedirs(output_dir, exist_ok=True)
+
+    samples = []
+    with open(descriptions_csv, mode="r") as f:
+        reader = csv.DictReader(f)
+        for row in reader:
+            scenario = row["scenario"].strip()
+            match_id = re.match(r"^(\d+)_", scenario)
+
+            if not match_id:
+                print(f"Cannot extract ID from scenario: {scenario}")
+                continue
+
+            vid_id = match_id.group(1).zfill(4)
+            description = row["description"]
+            generated_video_name = row["generated_video_name"]
+            prefix_video_path = build_conditioning_video_path(
+                data_root, vid_id, scenario, PHYSICSIQ_FPS
+            )
+            output_path = os.path.join(output_dir, generated_video_name)
+
+            os.makedirs(os.path.dirname(output_path), exist_ok=True)
+
+            samples.append({
+                "vid_id": vid_id,
+                "scenario": scenario,
+                "description": description,
+                "generated_video_name": generated_video_name,
+                "prefix_video_path": prefix_video_path,
+                "output_path": output_path,
+            })
+
+    # PhysicsIQ samples are duplicated; take only the first half
+    unique_count = len(samples) // 2
+    samples = samples[:unique_count]
+
+    print(f"Loaded {unique_count} PhysicsIQ samples.")
+
+    return apply_slice(samples, config.get("start"), config.get("end"))
+
+
+def load_vbench_samples(config: dict) -> list[str]:
+    """Load prompt list from vbench dimension file."""
+    prompt_dir = config["vbench_prompt_dir"]
+    dimension = config.get("dimension")
+
+    if not dimension:
+        raise ValueError("For vbench, 'dimension' must be specified in config")
+
+    prompt_file = os.path.join(prompt_dir, f"{dimension}.txt")
+
+    if not os.path.exists(prompt_file):
+        raise FileNotFoundError(f"Prompt file not found: {prompt_file}")
+
+    with open(prompt_file, "r") as f:
+        prompts = [line.strip() for line in f if line.strip()]
+
+    return apply_slice(prompts, config.get("start"), config.get("end"))
+
+
+def setup_save_path(config: dict) -> None:
+    """Configure the output save path based on benchmark type."""
+    base_path = config["base_save_path"]
+
+    if config["benchmark"] == "vbench":
+        dimension = config.get("dimension")
+        videos_dir = os.path.join(base_path, "videos", dimension) if dimension else None
+        config["save_path"] = videos_dir if videos_dir else os.path.join(base_path, "videos")
+    elif config["benchmark"] == "physicsiq":
+        config["save_path"] = os.path.join(base_path, "videos")
+
+    os.makedirs(config["save_path"], exist_ok=True)
+
+
+def main() -> None:
+    """Entry point for video sampling script."""
+    parser = argparse.ArgumentParser(
+        description="Video sampling script using YAML configuration"
+    )
+    parser.add_argument("yaml_config", type=str, help="Path to YAML configuration file")
+    args = parser.parse_args()
+
+    config = load_yaml_config(args.yaml_config)
+    print(f"Loaded configuration from: {args.yaml_config}")
+
+    setup_save_path(config)
+
+    gpu_ids = resolve_gpu_ids(config["gpus"])
+    config["num_gpus"] = len(gpu_ids)
+
+    benchmark = config["benchmark"]
+    if benchmark == "vbench":
+        all_samples = load_vbench_samples(config)
+    elif benchmark == "physicsiq":
+        data_root = config["physicsiq_data_dir"]
+        if not os.path.exists(data_root):
+            raise FileNotFoundError(f"Data directory not found: {data_root}")
+        all_samples = load_physicsiq_samples(config)
+    else:
+        raise ValueError(f"Invalid benchmark: {benchmark}")
+
+    print(f"Total samples: {len(all_samples)}")
+    print(f"GPUs: {gpu_ids}")
+    print(f"Output: {config['save_path']}")
+    print(f"Config: {config['config_file']}")
+
+    processes = []
+    for rank, gpu_id in enumerate(gpu_ids):
+        p = mp.Process(target=worker_process, args=(gpu_id, rank, config, all_samples))
+        p.start()
+        processes.append(p)
+
+    for p in processes:
+        p.join()
+
+    failed = [p.exitcode for p in processes if p.exitcode != 0]
+    if failed:
+        raise RuntimeError(f"{len(failed)} worker process(es) failed with exit codes: {failed}")
+
+
+if __name__ == "__main__":
+    main()

FlowCache/FlowCache4MAGI-1-dev6-adaptive/README.md

@@ -0,0 +1,71 @@
+# FLOW CACHING FOR AUTOREGRESSIVE VIDEO GENERATION
+
+This repository provides the official implementation of **FlowCache** on **MAGI-1** model, a caching-based acceleration method for autoregressive video generation models.
+
+
+## 🚀 Installation
+
+Please follow the installation instructions provided in the [MAGI-1](https://github.com/SandAI-org/MAGI-1), as this implementation is built on top of MAGI-1.
+
+---
+
+## ▶️ Usage
+
+### 1. Single Video Generation
+
+Run accelerated generation using FlowCache:
+
+```bash
+# FlowCache for text-to-video generation
+bash scripts/single_run/flowcache_t2v.sh
+
+# FlowCache for video-to-video generation
+bash scripts/single_run/flowcache_v2v.sh
+
+# Baseline acceleration method (TeaCache) for text-to-video
+bash scripts/single_run/teacache_t2v.sh
+
+# Baseline acceleration method (TeaCache) for video-to-video
+bash scripts/single_run/teacache_v2v.sh
+```
+
+### 2. Benchmark Sampling
+
+Generate videos for evaluation on standard benchmarks:
+
+```bash
+# VBench
+bash scripts/sample/flowcache_vbench.sh
+bash scripts/sample/teacache_vbench.sh
+
+# PhysicsIQ
+bash scripts/sample/flowcache_physicsiq.sh
+bash scripts/sample/teacache_physicsiq.sh
+```
+
+### 3. Quality Evaluation
+
+Compute perceptual and structural similarity metrics between original and accelerated generations:
+
+```bash
+bash scripts/metric.sh
+```
+
+---
+
+## ⚙️ Key Parameters
+
+| Parameter | Description |
+|----------|-------------|
+| `rel_l1_thresh` | Relative L1 distance threshold for cache reuse decision |
+| ` warmup_steps` | Number of denoising steps where reuse is disabled |
+| `total_cache_chunk_nums` (`B_total`) | Total number of cache chunks maintained |
+| `compress_strategy` | Granularity for selecting important KV caches: `token`, `frame`, or `chunk` |
+| `query_granularity` | Granularity for importance scoring: `token`, `frame`, or `chunk` |
+| `mix_lambda` | Weight balancing importance and redundancy (default: `0.07`) |
+| `mode` | Generation mode: `t2v` (text-to-video), `i2v` (image-to-video), or `v2v` (video-to-video) |
+| `prompt` | Input prompt for conditional generation |
+| `output_path` | Path to save generated videos |
+| `config_file` | Path to MAGI-1 model configuration |
+
+---

FlowCache/FlowCache4MAGI-1-dev6-adaptive/README_ADAPTIVE.md

@@ -0,0 +1,66 @@
+# AdaptiveDetailCache on MAGI-1 (dev6-adaptive)
+
+基于 [FlowCache4MAGI-1-dev4-detail](../FlowCache4MAGI-1-dev4-detail) 的扩展分支，在 MotionDetailCache（motion + detail 双度量）之上，增加 **chunk 级 AR 难度预测 → 动态阈值 τ**。
+
+## 与 dev4 的差异
+
+| 维度 | dev4 MotionDetailCache | dev6 AdaptiveDetailCache |
+|------|------------------------|--------------------------|
+| Phase 2 阈值 | 全局固定 `τ` | 每个 chunk 独立 `τ_eff` |
+| 难度信号 | — | motion / detail / delta / 历史 active / 历史 reuse |
+| 调参方式 | 手动 sweep `rel_l1_thresh` | `τ_base` + `beta` + `[τ_min, τ_max]` |
+
+## 动态 τ 公式
+
+每个 chunk 进入 Phase2 时（chunk 内固定，避免 step 间抖动）：
+
+```
+d = weighted(motion, detail, delta, hist_active, hist_reuse)  ∈ [0, 1]
+τ_eff = clamp(τ_base * exp(-β * (d - 0.5)), τ_min, τ_max)
+active[p] = A[p] > τ_eff
+```
+
+难度高 → τ 降低 → 更常计算；难度低 → τ 升高 → 更敢 reuse。
+
+## 默认超参
+
+继承 dev4 best，并新增：
+
+| 参数 | 默认值 | 说明 |
+|------|--------|------|
+| `rel_l1_thresh` | 0.012 | τ 基准（dev4 best） |
+| `use_adaptive_tau` | true | 开启动态 τ |
+| `adaptive_tau_beta` | 0.8 | 难度→τ 灵敏度 |
+| `adaptive_tau_min` | 0.008 | τ 下界 |
+| `adaptive_tau_max` | 0.020 | τ 上界 |
+
+## 快速运行
+
+```bash
+cd FlowCache4MAGI-1-dev6-adaptive
+conda activate magi
+export CUDA_VISIBLE_DEVICES=0
+export MASTER_ADDR=localhost
+bash scripts/single_run/adaptive_t2v.sh
+```
+
+## dev4 vs dev6 对比
+
+```bash
+FRAMES=120 bash tools/run_compare_dev4_dev6.sh
+FRAMES=240 bash tools/run_compare_dev4_dev6.sh
+```
+
+## 代码结构
+
+```
+inference/pipeline/cache/
+├── motiondetailcache.py      # dev4 基类（含 τ hook）
+└── adaptivedetailcache.py    # dev6 AdaptiveDetailCache
+
+yaml_config/single_run/adaptive_config_best.yaml
+scripts/single_run/adaptive_t2v.sh
+tools/run_compare_dev4_dev6.sh
+```
+
+其余 sparse forward、KV cache、integrate 逻辑与 dev4 完全一致。

FlowCache/FlowCache4MAGI-1-dev6-adaptive/README_MOTIONCACHE.md

@@ -0,0 +1,92 @@
+# MotionCache on MAGI-1 (dev3-motion)
+
+基于 [FlowCache4MAGI-1](../FlowCache4MAGI-1) 的 **MotionCache** 复现分支，对应论文：
+
+> Xu et al., 2026 — *Motion-Aware Caching for Efficient Autoregressive Video Generation*
+
+官方代码仓库 [ywlq/MotionCache](https://github.com/ywlq/MotionCache) 尚未发布实现，本目录依据论文方法在 MAGI-1 推理框架上完成首版复现。
+
+## 与 FlowCache 的核心差异
+
+| 维度 | FlowCache | MotionCache |
+|------|-----------|-------------|
+| 粒度 | Chunk 级全有或全无 | Phase 2 为 Token（latent 帧×空间）级 |
+| 跳过策略 | 相对 L1 累积阈值 | 运动重要性加权累积 |
+| 调度 | 单一 chunk-wise 策略 | 两阶段 coarse-to-fine |
+
+### 算法概要
+
+1. **全局 Warm-up（m 步）**：前 `warmup_steps` 步禁止 cache reuse
+2. **Phase 1（K 步）**：chunk-wise 二值决策，与 FlowCache 相同
+3. **Phase 2**：基于帧间 latent 差计算运动重要性 `M`，经 soft-mapping 得到权重 `W ∈ [α, 1]`
+4. **Token 累积**：`A[p] += W[p] · Δ_chunk`，当 `A[p] > τ` 时该 token 触发 DiT 计算
+5. **Integrate**：active token 正常积分；inactive token 复用缓存 residual
+
+### MAGI-1 默认超参（论文 Appendix C）
+
+| 参数 | 值 | 说明 |
+|------|-----|------|
+| `alpha` | 0.5 | 静态区域权重下限 |
+| `phase1_steps` (K) | 9 | chunk-wise 阶段持续步数 |
+| `warmup_steps` (m) | 5 | 全局禁止 reuse 的步数 |
+| `rel_l1_thresh` (τ) | 0.015 (slow) / 0.025 (fast) | token 累积阈值 |
+
+## 快速运行
+
+```bash
+cd FlowCache4MAGI-1-dev3-motion
+
+# MotionCache-slow（论文 Table 1 配置）
+bash scripts/single_run/motioncache_t2v.sh
+
+# MotionCache-fast
+MOTIONCACHE_CONFIG=yaml_config/single_run/motioncache_config_fast.yaml \
+  bash scripts/single_run/motioncache_t2v.sh
+```
+
+需先按 FlowCache4MAGI-1 说明安装依赖并下载 MAGI-1 权重（`downloads/` 目录可通过软链接指向原项目）。
+
+## 代码结构
+
+```
+inference/pipeline/
+├── motioncache.py              # 入口与 forward/integrate monkey-patch
+└── cache/
+    ├── motioncache.py          # MotionWiseCache 核心逻辑
+    └── sparse_utils.py         # Phase 2 gather/scatter 与 sparse meta_args
+
+inference/model/dit/dit_module.py   # sparse KV cache 写入 + flash_attn 稀疏 q 分支
+inference/common/dataclass.py       # ModelMetaArgs.sparse_active_indices
+
+yaml_config/single_run/
+├── motioncache_config.yaml     # slow 配置
+└── motioncache_config_fast.yaml
+```
+
+## 当前实现说明（严格对齐论文 §5.3 Phase 2）
+
+- **Phase 1（前 K 个 denoise step / chunk）**：与 FlowCache 完全一致的 chunk-wise 策略（已验证 PSNR=∞）
+- **Phase 2 入口**：每个 chunk 的第 K 步强制全量计算，建立 token-wise 阶段的 residual 基准
+- **Phase 2 运动感知累积**：latent 帧间差 → soft-mapping 权重 W → 累积器 A；`A > τ` 的 token 为 active
+- **Phase 2 稀疏 forward（论文 5.3）**：
+  - 全部 inactive → 跳过 DiT forward，复用 chunk residual
+  - 部分 active → **gather active patch tokens → compact DiT forward → scatter 回完整序列**
+  - KV cache 仅更新 active 位置（`_compresskv_adjust_key_and_value` sparse 分支）
+  - integrate 时 active token 正常积分，inactive token 复用 `previous_residual`
+- **运动 proxy**：latent 空间帧间 L1 差（论文 Eq. 9-10）
+- **跨 chunk 连续性**：chunk 首帧与上一 chunk 末帧比较
+
+### 验证结果（`a woman dancing.`，240×720×720，vs FlowCache baseline）
+
+| 运行 | PSNR | 说明 |
+|------|------|------|
+| phase1only | ∞ | 与 FlowCache 逐像素一致 |
+| sparse2（论文对齐 sparse forward） | **20.44 dB** | 无黑帧，reuse_rate≈15.6% |
+| final（整 chunk fallback） | 20.48 dB | 画质等价，推理慢 ~14% |
+
+sparse2 日志示例：`active_ratio=1.51%` 时仍走 gather/scatter 稀疏路径；`active_ratio=0%` 时 `skip_forward=True`。
+
+## 参考
+
+- FlowCache: chunk-wise cache + KV compression
+- MotionCache 论文预期 MAGI-1 加速：slow 1.64×，fast 2.07×（Table 1）

FlowCache/FlowCache4MAGI-1-dev6-adaptive/README_MOTIONDETAIL.md

@@ -0,0 +1,71 @@
+# MotionDetailCache on MAGI-1 (dev4-detail)
+
+基于 [FlowCache4MAGI-1-dev3-motion](../FlowCache4MAGI-1-dev3-motion) 的扩展分支，在 MotionCache 的 **motion 度量** 之外，增加 **spatial detail 度量**，Phase 2 active token 由两者共同决定。
+
+## 与 dev3 (MotionCache) 的差异
+
+| 维度 | dev3 MotionCache | dev4 MotionDetailCache |
+|------|------------------|------------------------|
+| Phase 2 权重 | 仅 motion 权重 W_m | motion W_m + detail W_d 融合 |
+| Detail proxy | — | latent 局部 k×k 空间方差 |
+| 适用场景 | 运动区域优先计算 | 静止纹理/边缘/语义细节也优先保留 |
+
+## Detail 度量
+
+对每个 latent 像素 `(t, h, w)`：
+
+1. 通道聚合幅度：`mag = mean(|x|, dim=C)` → `[N, T, H, W]`
+2. 局部空间方差：在 `detail_window_size×detail_window_size` 窗口内计算 `Var(mag)`
+3. Soft-mapping：帧内 min-max 归一化 → `W_d ∈ [detail_alpha, 1]`
+
+高方差区域对应边缘、纹理、物体边界等 heterogeneous 邻域。
+
+## Motion + Detail 融合
+
+```
+W_combined = combine(W_motion, W_detail)
+A[p] += W_combined[p] · Δ_chunk
+active[p] = A[p] > τ
+```
+
+`weight_combine_mode`（默认 `max`）：
+
+| 模式 | 公式 | 特点 |
+|------|------|------|
+| `max` | max(W_m, W_d) | 运动或细节任一显著即可加速累积（推荐） |
+| `product` | W_m × W_d | 两者同时高才快速激活，更激进 reuse |
+| `blend` | (1-λ)W_m + λW_d | 线性权衡，λ=`detail_lambda` |
+
+## 默认超参
+
+继承 dev3 的 motion 参数，并新增：
+
+| 参数 | 默认值 | 说明 |
+|------|--------|------|
+| `detail_alpha` | 0.5 | detail 权重下限 |
+| `detail_window_size` | 3 | 局部方差窗口（奇数） |
+| `detail_lambda` | 0.5 | blend 模式下的 detail 权重 |
+| `weight_combine_mode` | max | 融合方式 |
+
+## 快速运行
+
+```bash
+cd FlowCache4MAGI-1-dev4-detail
+conda activate magi
+export CUDA_VISIBLE_DEVICES=0
+export MASTER_ADDR=localhost
+bash scripts/single_run/motiondetail_t2v.sh
+```
+
+## 代码结构
+
+```
+inference/pipeline/cache/
+├── motioncache.py          # dev3 基类 MotionWiseCache
+└── motiondetailcache.py    # dev4 MotionDetailCache
+
+yaml_config/single_run/motiondetail_config.yaml
+scripts/single_run/motiondetail_t2v.sh
+```
+
+其余 sparse forward、KV cache、integrate 逻辑与 dev3 完全一致。

FlowCache/FlowCache4MAGI-1-dev6-adaptive/requirements.txt

@@ -0,0 +1,21 @@
+accelerate==0.32.1
+beautifulsoup4==4.13.4
+debugpy==1.8.14
+diffusers==0.29.2
+einops>=0.6.0
+ffmpeg-python
+# flash-attn==2.4.2
+flashinfer-python==0.2.0.post2 --extra-index-url https://flashinfer.ai/whl/cu124/torch2.4/
+ftfy==6.2.0
+gpustat==1.1.1
+imageio==2.34.0
+imageio[ffmpeg]
+matplotlib==3.10.1
+numpy==1.26.4
+protobuf==5.28.3
+rich==14.0.0
+sentencepiece==0.2.0
+timm==1.0.15
+torchdiffeq==0.2.4
+transformers==4.42.3
+tqdm

FlowCache/FlowCache4MAGI-1-dev6-adaptive/sample_video.py

@@ -0,0 +1,429 @@
+# Copyright 2024 MAGI Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import re
+import sys
+import argparse
+import csv
+import subprocess
+from pathlib import Path
+
+import multiprocessing as mp
+
+# Constants
+DEFAULT_BASE_PORT = 29510
+PHYSICSIQ_FPS = 24
+
+
+def resolve_gpu_ids(gpus_config) -> list[int]:
+    """Resolve explicit GPU IDs or auto-detect all currently visible GPUs."""
+    if isinstance(gpus_config, int):
+        return [gpus_config]
+
+    gpus_text = str(gpus_config).strip()
+    if not gpus_text:
+        raise ValueError("'gpus' must not be empty")
+
+    if gpus_text.lower() not in {"all", "auto"}:
+        return [int(item.strip()) for item in gpus_text.split(",") if item.strip()]
+
+    visible_devices = os.environ.get("CUDA_VISIBLE_DEVICES")
+    if visible_devices:
+        visible = [item.strip() for item in visible_devices.split(",") if item.strip()]
+        if visible and all(item.isdigit() for item in visible):
+            return [int(item) for item in visible]
+
+    try:
+        output = subprocess.check_output(
+            ["nvidia-smi", "--query-gpu=index", "--format=csv,noheader,nounits"],
+            text=True,
+            timeout=10,
+        )
+        gpu_ids = [int(line.strip()) for line in output.splitlines() if line.strip()]
+        if gpu_ids:
+            return gpu_ids
+    except Exception:
+        pass
+
+    try:
+        import torch
+
+        count = torch.cuda.device_count()
+        if count > 0:
+            return list(range(count))
+    except Exception:
+        pass
+
+    raise RuntimeError("No CUDA GPUs detected for gpus: all")
+
+
+def load_yaml_config(yaml_path: str) -> dict:
+    """Load configuration from YAML file."""
+    import yaml
+
+    with open(yaml_path, "r") as f:
+        return yaml.safe_load(f)
+
+
+def apply_slice(items: list, start: int | None, end: int | None) -> list:
+    """Apply start/end slice to a list with bounds checking."""
+    if start is None and end is None:
+        return items
+
+    slice_start = max(0, start if start is not None else 0)
+    slice_end = min(end if end is not None else len(items), len(items))
+    slice_end = max(slice_start, slice_end)
+
+    return items[slice_start:slice_end]
+
+
+def configure_teacache(transport, config: dict) -> None:
+    """Configure TeaCache reuse strategy on SampleTransport."""
+    from inference.pipeline.teacache import setup_teacache
+
+    setup_teacache(
+        rel_l1_thresh=config["rel_l1_thresh"],
+        warmup_steps=config["warmup_steps"],
+        log=config.get("log", False),
+    )
+
+
+def configure_kv_cache(transport, config: dict) -> None:
+    """Configure KV cache compression if enabled."""
+    if not config.get("compress_kv_cache", False):
+        transport.compress_kv_cache = False
+        return
+
+    print("KV cache compression is enabled.")
+    transport.compress_kv_cache = True
+
+    assert config.get("total_cache_chunk_nums") is not None
+
+    compression_config = {
+        "method_config": {
+            "compress_strategy": config["compress_strategy"],
+            "mix_lambda": config["mix_lambda"],
+            "query_granularity": config["query_granularity"],
+            "score_weighting_method": config.get("score_weighting_method") or "no_weight",
+            "power": config.get("power", 3),
+        },
+    }
+
+    from inference.pipeline.kvcompress import replace_magi
+
+    replace_magi(compression_config)
+
+
+def configure_flowcache(transport, config: dict) -> None:
+    """Configure FlowCache reuse strategy on SampleTransport."""
+    from inference.pipeline.flowcache import setup_flowcache
+
+    configure_kv_cache(transport, config)
+
+    setup_flowcache(
+        rel_l1_thresh=config["rel_l1_thresh"],
+        warmup_steps=config["warmup_steps"],
+        discard_nearly_clean_chunk=config.get("discard_nearly_clean_chunk", False),
+        log=config.get("log", False),
+        total_cache_chunk_nums=config.get("total_cache_chunk_nums", 5),
+        compress_kv_cache=config.get("compress_kv_cache", False),
+    )
+
+
+def configure_reuse_strategy(config: dict) -> None:
+    """Configure the appropriate reuse strategy on SampleTransport."""
+    from inference.pipeline.video_generate import SampleTransport
+
+    strategy = config["reuse_strategy"]
+
+    if strategy == "original":
+        return
+    if strategy == "all":
+        configure_teacache(SampleTransport, config)
+    elif strategy == "chunkwise":
+        configure_flowcache(SampleTransport, config)
+    else:
+        raise ValueError(f"Unknown reuse strategy: {strategy}")
+
+
+def setup_environment(gpu_id: int) -> None:
+    """Set up environment variables for a GPU worker process."""
+    os.environ["CUDA_VISIBLE_DEVICES"] = str(gpu_id)
+    os.environ["WORLD_SIZE"] = "1"
+    os.environ["RANK"] = "0"
+    os.environ["LOCAL_RANK"] = "0"
+    os.environ["MASTER_ADDR"] = "localhost"
+    os.environ["MASTER_PORT"] = str(DEFAULT_BASE_PORT + gpu_id)
+
+    # Enable pdb terminal debugging
+    sys.stdin = open(0)
+
+
+def filter_existing_samples(samples: list, config: dict) -> list:
+    """Filter out samples whose output files already exist."""
+    if config["benchmark"] == "vbench":
+        return [
+            sample
+            for sample in samples
+            if not os.path.exists(os.path.abspath(os.path.join(config["save_path"], f"{sample}-0.mp4")))
+        ]
+    else:  # physicsiq
+        return [
+            sample for sample in samples if not os.path.exists(sample["output_path"])
+        ]
+
+
+def assign_samples_to_gpu(
+    samples: list, gpu_id: int, rank: int, num_gpus: int
+) -> list:
+    """Divide samples across GPUs and return the subset for this GPU."""
+    samples_per_gpu = (len(samples) + num_gpus - 1) // num_gpus
+    start_idx = rank * samples_per_gpu
+    end_idx = min(start_idx + samples_per_gpu, len(samples))
+    return samples[start_idx:end_idx]
+
+
+def process_vbench_sample(pipeline, prompt: str, config: dict, gpu_id: int) -> None:
+    """Process a single vbench text-to-video sample."""
+    output_path = os.path.abspath(os.path.join(config["save_path"], f"{prompt}-0.mp4"))
+
+    if os.path.exists(output_path):
+        print(f"[SKIP GPU {gpu_id}] Already exists: {output_path}")
+        return
+
+    print(f"[GPU {gpu_id}] Generating T2V: '{prompt}' -> {output_path}")
+    pipeline.run_text_to_video(prompt=prompt, output_path=output_path)
+    print(f"[DONE GPU {gpu_id}] Saved: {output_path}")
+
+
+def process_physicsiq_sample(pipeline, sample: dict, gpu_id: int) -> None:
+    """Process a single PhysicsIQ video-to-video sample."""
+    prompt = sample["description"]
+    prefix_video_path = sample["prefix_video_path"]
+    output_path = sample["output_path"]
+
+    if not os.path.exists(prefix_video_path):
+        print(f"[WARN GPU {gpu_id}] Conditioning video not found: {prefix_video_path}")
+        return
+
+    if os.path.exists(output_path):
+        print(f"[SKIP GPU {gpu_id}] Already exists: {output_path}")
+        return
+
+    print(f"[GPU {gpu_id}] Generating V2V: '{prompt}'")
+    print(f"   Input:  {prefix_video_path}")
+    print(f"   Output: {output_path}")
+
+    pipeline.run_video_to_video(
+        prompt=prompt,
+        prefix_video_path=prefix_video_path,
+        output_path=output_path,
+    )
+    print(f"[DONE GPU {gpu_id}] Saved: {output_path}")
+
+
+def worker_process(gpu_id: int, rank: int, config: dict, all_samples: list) -> None:
+    """Independent worker running on each GPU."""
+    setup_environment(gpu_id)
+    configure_reuse_strategy(config)
+
+    try:
+        magi_root = subprocess.check_output(
+            ["git", "rev-parse", "--show-toplevel"]
+        ).decode().strip()
+        os.environ["MAGI_ROOT"] = magi_root
+        os.environ["PYTHONPATH"] = f"{magi_root}:{os.environ.get('PYTHONPATH', '')}"
+    except Exception as e:
+        print(f"[GPU {gpu_id}] Failed to set MAGI_ROOT: {e}")
+        return
+
+    filtered_samples = filter_existing_samples(all_samples, config)
+
+    if not filtered_samples:
+        print(f"[GPU {gpu_id}] No samples need to be generated.")
+        return
+
+    print(f"Processing {len(filtered_samples)} samples.")
+
+    my_samples = assign_samples_to_gpu(
+        filtered_samples, gpu_id, rank, config["num_gpus"]
+    )
+
+    if not my_samples:
+        print(f"[GPU {gpu_id}] No samples assigned.")
+        return
+
+    print(f"[GPU {gpu_id}] Assigned {len(my_samples)} samples")
+
+    from inference.pipeline.entry import MagiPipeline
+
+    print(f"[GPU {gpu_id}] Loading model...")
+    pipeline = MagiPipeline(config["config_file"])
+    print(f"[GPU {gpu_id}] Model loaded.")
+
+    process_func = (
+        process_vbench_sample if config["benchmark"] == "vbench" else process_physicsiq_sample
+    )
+
+    for sample in my_samples:
+        process_func(pipeline, sample, config, gpu_id)
+
+    print(f"[GPU {gpu_id}] Completed.")
+
+
+def build_conditioning_video_path(
+    data_root: str, vid_id: str, scenario: str, fps: int
+) -> str:
+    """Construct the path to the conditioning video file."""
+    conditioning_dir = os.path.join(
+        data_root, "physics-IQ-benchmark", "split-videos", "conditioning", f"{fps}FPS"
+    )
+    match_suffix = re.search(r"_(.*)", scenario)
+    suffix = match_suffix.group(1) if match_suffix else ""
+    filename = f"{vid_id}_conditioning-videos_{fps}FPS_{suffix}"
+    return os.path.join(conditioning_dir, filename)
+
+
+def load_physicsiq_samples(config: dict) -> list[dict]:
+    """Load sample list from PhysicsIQ dataset."""
+    data_root = config["physicsiq_data_dir"]
+    descriptions_csv = os.path.join(data_root, "descriptions", "descriptions.csv")
+    output_dir = config["save_path"]
+
+    if not os.path.exists(descriptions_csv):
+        raise FileNotFoundError(f"descriptions.csv not found at {descriptions_csv}")
+
+    os.makedirs(output_dir, exist_ok=True)
+
+    samples = []
+    with open(descriptions_csv, mode="r") as f:
+        reader = csv.DictReader(f)
+        for row in reader:
+            scenario = row["scenario"].strip()
+            match_id = re.match(r"^(\d+)_", scenario)
+
+            if not match_id:
+                print(f"Cannot extract ID from scenario: {scenario}")
+                continue
+
+            vid_id = match_id.group(1).zfill(4)
+            description = row["description"]
+            generated_video_name = row["generated_video_name"]
+            prefix_video_path = build_conditioning_video_path(
+                data_root, vid_id, scenario, PHYSICSIQ_FPS
+            )
+            output_path = os.path.join(output_dir, generated_video_name)
+
+            os.makedirs(os.path.dirname(output_path), exist_ok=True)
+
+            samples.append({
+                "vid_id": vid_id,
+                "scenario": scenario,
+                "description": description,
+                "generated_video_name": generated_video_name,
+                "prefix_video_path": prefix_video_path,
+                "output_path": output_path,
+            })
+
+    # PhysicsIQ samples are duplicated; take only the first half
+    unique_count = len(samples) // 2
+    samples = samples[:unique_count]
+
+    print(f"Loaded {unique_count} PhysicsIQ samples.")
+
+    return apply_slice(samples, config.get("start"), config.get("end"))
+
+
+def load_vbench_samples(config: dict) -> list[str]:
+    """Load prompt list from vbench dimension file."""
+    prompt_dir = config["vbench_prompt_dir"]
+    dimension = config.get("dimension")
+
+    if not dimension:
+        raise ValueError("For vbench, 'dimension' must be specified in config")
+
+    prompt_file = os.path.join(prompt_dir, f"{dimension}.txt")
+
+    if not os.path.exists(prompt_file):
+        raise FileNotFoundError(f"Prompt file not found: {prompt_file}")
+
+    with open(prompt_file, "r") as f:
+        prompts = [line.strip() for line in f if line.strip()]
+
+    return apply_slice(prompts, config.get("start"), config.get("end"))
+
+
+def setup_save_path(config: dict) -> None:
+    """Configure the output save path based on benchmark type."""
+    base_path = config["base_save_path"]
+
+    if config["benchmark"] == "vbench":
+        dimension = config.get("dimension")
+        videos_dir = os.path.join(base_path, "videos", dimension) if dimension else None
+        config["save_path"] = videos_dir if videos_dir else os.path.join(base_path, "videos")
+    elif config["benchmark"] == "physicsiq":
+        config["save_path"] = os.path.join(base_path, "videos")
+
+    os.makedirs(config["save_path"], exist_ok=True)
+
+
+def main() -> None:
+    """Entry point for video sampling script."""
+    parser = argparse.ArgumentParser(
+        description="Video sampling script using YAML configuration"
+    )
+    parser.add_argument("yaml_config", type=str, help="Path to YAML configuration file")
+    args = parser.parse_args()
+
+    config = load_yaml_config(args.yaml_config)
+    print(f"Loaded configuration from: {args.yaml_config}")
+
+    setup_save_path(config)
+
+    gpu_ids = resolve_gpu_ids(config["gpus"])
+    config["num_gpus"] = len(gpu_ids)
+
+    benchmark = config["benchmark"]
+    if benchmark == "vbench":
+        all_samples = load_vbench_samples(config)
+    elif benchmark == "physicsiq":
+        data_root = config["physicsiq_data_dir"]
+        if not os.path.exists(data_root):
+            raise FileNotFoundError(f"Data directory not found: {data_root}")
+        all_samples = load_physicsiq_samples(config)
+    else:
+        raise ValueError(f"Invalid benchmark: {benchmark}")
+
+    print(f"Total samples: {len(all_samples)}")
+    print(f"GPUs: {gpu_ids}")
+    print(f"Output: {config['save_path']}")
+    print(f"Config: {config['config_file']}")
+
+    processes = []
+    for rank, gpu_id in enumerate(gpu_ids):
+        p = mp.Process(target=worker_process, args=(gpu_id, rank, config, all_samples))
+        p.start()
+        processes.append(p)
+
+    for p in processes:
+        p.join()
+
+    failed = [p.exitcode for p in processes if p.exitcode != 0]
+    if failed:
+        raise RuntimeError(f"{len(failed)} worker process(es) failed with exit codes: {failed}")
+
+
+if __name__ == "__main__":
+    main()

FlowCache/FlowCache4MAGI-1/README.md

@@ -0,0 +1,71 @@
+# FLOW CACHING FOR AUTOREGRESSIVE VIDEO GENERATION
+
+This repository provides the official implementation of **FlowCache** on **MAGI-1** model, a caching-based acceleration method for autoregressive video generation models.
+
+
+## 🚀 Installation
+
+Please follow the installation instructions provided in the [MAGI-1](https://github.com/SandAI-org/MAGI-1), as this implementation is built on top of MAGI-1.
+
+---
+
+## ▶️ Usage
+
+### 1. Single Video Generation
+
+Run accelerated generation using FlowCache:
+
+```bash
+# FlowCache for text-to-video generation
+bash scripts/single_run/flowcache_t2v.sh
+
+# FlowCache for video-to-video generation
+bash scripts/single_run/flowcache_v2v.sh
+
+# Baseline acceleration method (TeaCache) for text-to-video
+bash scripts/single_run/teacache_t2v.sh
+
+# Baseline acceleration method (TeaCache) for video-to-video
+bash scripts/single_run/teacache_v2v.sh
+```
+
+### 2. Benchmark Sampling
+
+Generate videos for evaluation on standard benchmarks:
+
+```bash
+# VBench
+bash scripts/sample/flowcache_vbench.sh
+bash scripts/sample/teacache_vbench.sh
+
+# PhysicsIQ
+bash scripts/sample/flowcache_physicsiq.sh
+bash scripts/sample/teacache_physicsiq.sh
+```
+
+### 3. Quality Evaluation
+
+Compute perceptual and structural similarity metrics between original and accelerated generations:
+
+```bash
+bash scripts/metric.sh
+```
+
+---
+
+## ⚙️ Key Parameters
+
+| Parameter | Description |
+|----------|-------------|
+| `rel_l1_thresh` | Relative L1 distance threshold for cache reuse decision |
+| ` warmup_steps` | Number of denoising steps where reuse is disabled |
+| `total_cache_chunk_nums` (`B_total`) | Total number of cache chunks maintained |
+| `compress_strategy` | Granularity for selecting important KV caches: `token`, `frame`, or `chunk` |
+| `query_granularity` | Granularity for importance scoring: `token`, `frame`, or `chunk` |
+| `mix_lambda` | Weight balancing importance and redundancy (default: `0.07`) |
+| `mode` | Generation mode: `t2v` (text-to-video), `i2v` (image-to-video), or `v2v` (video-to-video) |
+| `prompt` | Input prompt for conditional generation |
+| `output_path` | Path to save generated videos |
+| `config_file` | Path to MAGI-1 model configuration |
+
+---

FlowCache/FlowCache4MAGI-1/logs/flowcache_vbench_20260520_103113.log

@@ -0,0 +1,181 @@
+🚀 Starting multi-GPU benchmark sampling
+🔢 Total dimensions to process: 3
+📋 Dimensions: overall_consistency subject_consistency scene
+🔍 Processing dimension: overall_consistency
+Loaded configuration from: yaml_config/sample/flowcache_vbench.yaml.tmp
+Total samples: 93
+GPUs: [0]
+Output: outputs/vbench/videos/overall_consistency
+Config: config/sample/vbench.json
+/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/site-packages/timm/models/layers/__init__.py:48: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[W520 10:31:24.093345214 CUDAAllocatorConfig.h:28] Warning: expandable_segments not supported on this platform (function operator())
+[2026-05-20 10:31:24,656 - INFO] Initialize torch distribution and model parallel successfully
+[2026-05-20 10:31:24,656 - INFO] MagiConfig(model_config=ModelConfig(model_name='videodit_ardf', num_layers=34, hidden_size=3072, ffn_hidden_size=12288, num_attention_heads=24, num_query_groups=8, kv_channels=128, layernorm_epsilon=1e-06, apply_layernorm_1p=True, x_rescale_factor=1, half_channel_vae=False, params_dtype=torch.bfloat16, patch_size=2, t_patch_size=1, in_channels=16, out_channels=16, cond_hidden_ratio=0.25, caption_channels=4096, caption_max_length=800, xattn_cond_hidden_ratio=1.0, cond_gating_ratio=1.0, gated_linear_unit=False), runtime_config=RuntimeConfig(cfg_number=1, cfg_t_range=[0.0, 0.0217, 0.1, 0.3, 0.999], prev_chunk_scales=[1.5, 1.5, 1.5, 1.0, 1.0], text_scales=[7.5, 7.5, 7.5, 0.0, 0.0], noise2clean_kvrange=[], clean_chunk_kvrange=1, clean_t=0.9999, seed=1234, num_frames=240, video_size_h=720, video_size_w=720, num_steps=16, window_size=4, fps=24, chunk_width=6, t5_pretrained='./downloads/t5_pretrained', t5_device='cuda', vae_pretrained='./downloads/vae', scale_factor=0.18215, temporal_downsample_factor=4, load='./downloads/4.5B_distill'), engine_config=EngineConfig(distributed_backend='nccl', distributed_timeout_minutes=15, pp_size=1, cp_size=1, cp_strategy='none', ulysses_overlap_degree=1, fp8_quant=False, distill_nearly_clean_chunk_threshold=0.3, shortcut_mode='8,16,16', distill=True, kv_offload=True, enable_cuda_graph=False))
+[2026-05-20 10:31:24,657 - INFO] Precompute validation prompt embeddings
+You are using the default legacy behaviour of the <class 'transformers.models.t5.tokenization_t5.T5Tokenizer'>. This is expected, and simply means that the `legacy` (previous) behavior will be used so nothing changes for you. If you want to use the new behaviour, set `legacy=False`. This should only be set if you understand what it means, and thoroughly read the reason why this was added as explained in https://github.com/huggingface/transformers/pull/24565
+KV cache compression is enabled.
+Processing 93 samples.
+[GPU 0] Assigned 93 samples
+[GPU 0] Loading model...
+[GPU 0] Model loaded.
+[GPU 0] Generating T2V: 'Close up of grapes on a rotating table.' -> /home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/outputs/vbench/videos/overall_consistency/Close up of grapes on a rotating table.-0.mp4
+
+[2026-05-20 10:32:58,161 - INFO] (cp, pp) rank (0, 0): param count 4459898128, model size 8.34 GB
+[2026-05-20 10:32:58,161 - INFO] Build DiTModel successfully
+[2026-05-20 10:32:58,161 - INFO] After build_dit_model, memory allocated: 0.04 GB, memory reserved: 0.08 GB
+[2026-05-20 10:32:58,161 - INFO] load inference_weight.distill weight from ./downloads/4.5B_distill/inference_weight.distill
+
+[2026-05-20 10:33:20,857 - INFO] Load Weight Missing Keys: []
+[2026-05-20 10:33:20,857 - INFO] Load Weight Unexpected Keys: []
+[2026-05-20 10:33:21,087 - INFO] After load_checkpoint, memory allocated: 8.39 GB, memory reserved: 8.40 GB
+[2026-05-20 10:33:21,089 - INFO] After high_precision_promoter, memory allocated: 8.39 GB, memory reserved: 8.40 GB
+[2026-05-20 10:33:21,184 - INFO] Load checkpoint successfully
+[2026-05-20 10:33:21,184 - INFO] Begin to generate per chunk
+[2026-05-20 10:33:21,184 - INFO] special_token = ['HQ_TOKEN', 'DURATION_TOKEN']
+
+Process Process-1:
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 259, in worker_process
+    process_func(pipeline, sample, config, gpu_id)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 185, in process_vbench_sample
+    pipeline.run_text_to_video(prompt=prompt, output_path=output_path)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 38, in run_text_to_video
+    self._run(prompt, None, output_path)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 52, in _run
+    [
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 52, in <listcomp>
+    [
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/video_generate.py", line 1129, in generate_per_chunk
+    for _, _, chunk in sample_transport.walk():
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/video_generate.py", line 1079, in walk
+    velocity = self.forward_velocity(infer_idx, 0)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/flowcache.py", line 93, in flowcache_forward_velocity
+    cache = SampleTransport.cache_reuse_manager
+AttributeError: type object 'SampleTransport' has no attribute 'cache_reuse_manager'
+
+✅ Completed: overall_consistency
+---
+🔍 Processing dimension: subject_consistency
+Loaded configuration from: yaml_config/sample/flowcache_vbench.yaml.tmp
+Total samples: 72
+GPUs: [0]
+Output: outputs/vbench/videos/subject_consistency
+Config: config/sample/vbench.json
+/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/site-packages/timm/models/layers/__init__.py:48: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[W520 10:33:28.974731573 CUDAAllocatorConfig.h:28] Warning: expandable_segments not supported on this platform (function operator())
+[2026-05-20 10:33:28,530 - INFO] Initialize torch distribution and model parallel successfully
+[2026-05-20 10:33:28,530 - INFO] MagiConfig(model_config=ModelConfig(model_name='videodit_ardf', num_layers=34, hidden_size=3072, ffn_hidden_size=12288, num_attention_heads=24, num_query_groups=8, kv_channels=128, layernorm_epsilon=1e-06, apply_layernorm_1p=True, x_rescale_factor=1, half_channel_vae=False, params_dtype=torch.bfloat16, patch_size=2, t_patch_size=1, in_channels=16, out_channels=16, cond_hidden_ratio=0.25, caption_channels=4096, caption_max_length=800, xattn_cond_hidden_ratio=1.0, cond_gating_ratio=1.0, gated_linear_unit=False), runtime_config=RuntimeConfig(cfg_number=1, cfg_t_range=[0.0, 0.0217, 0.1, 0.3, 0.999], prev_chunk_scales=[1.5, 1.5, 1.5, 1.0, 1.0], text_scales=[7.5, 7.5, 7.5, 0.0, 0.0], noise2clean_kvrange=[], clean_chunk_kvrange=1, clean_t=0.9999, seed=1234, num_frames=240, video_size_h=720, video_size_w=720, num_steps=16, window_size=4, fps=24, chunk_width=6, t5_pretrained='./downloads/t5_pretrained', t5_device='cuda', vae_pretrained='./downloads/vae', scale_factor=0.18215, temporal_downsample_factor=4, load='./downloads/4.5B_distill'), engine_config=EngineConfig(distributed_backend='nccl', distributed_timeout_minutes=15, pp_size=1, cp_size=1, cp_strategy='none', ulysses_overlap_degree=1, fp8_quant=False, distill_nearly_clean_chunk_threshold=0.3, shortcut_mode='8,16,16', distill=True, kv_offload=True, enable_cuda_graph=False))
+[2026-05-20 10:33:28,530 - INFO] Precompute validation prompt embeddings
+You are using the default legacy behaviour of the <class 'transformers.models.t5.tokenization_t5.T5Tokenizer'>. This is expected, and simply means that the `legacy` (previous) behavior will be used so nothing changes for you. If you want to use the new behaviour, set `legacy=False`. This should only be set if you understand what it means, and thoroughly read the reason why this was added as explained in https://github.com/huggingface/transformers/pull/24565
+KV cache compression is enabled.
+Processing 72 samples.
+[GPU 0] Assigned 72 samples
+[GPU 0] Loading model...
+[GPU 0] Model loaded.
+[GPU 0] Generating T2V: 'a person swimming in ocean' -> /home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/outputs/vbench/videos/subject_consistency/a person swimming in ocean-0.mp4
+
+[2026-05-20 10:33:49,537 - INFO] (cp, pp) rank (0, 0): param count 4459898128, model size 8.34 GB
+[2026-05-20 10:33:49,537 - INFO] Build DiTModel successfully
+[2026-05-20 10:33:49,538 - INFO] After build_dit_model, memory allocated: 0.04 GB, memory reserved: 0.08 GB
+[2026-05-20 10:33:49,538 - INFO] load inference_weight.distill weight from ./downloads/4.5B_distill/inference_weight.distill
+
+[2026-05-20 10:33:51,253 - INFO] Load Weight Missing Keys: []
+[2026-05-20 10:33:51,253 - INFO] Load Weight Unexpected Keys: []
+[2026-05-20 10:33:51,770 - INFO] After load_checkpoint, memory allocated: 8.39 GB, memory reserved: 8.40 GB
+[2026-05-20 10:33:51,773 - INFO] After high_precision_promoter, memory allocated: 8.39 GB, memory reserved: 8.40 GB
+[2026-05-20 10:33:51,875 - INFO] Load checkpoint successfully
+[2026-05-20 10:33:51,875 - INFO] Begin to generate per chunk
+[2026-05-20 10:33:51,875 - INFO] special_token = ['HQ_TOKEN', 'DURATION_TOKEN']
+
+Process Process-1:
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 259, in worker_process
+    process_func(pipeline, sample, config, gpu_id)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 185, in process_vbench_sample
+    pipeline.run_text_to_video(prompt=prompt, output_path=output_path)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 38, in run_text_to_video
+    self._run(prompt, None, output_path)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 52, in _run
+    [
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 52, in <listcomp>
+    [
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/video_generate.py", line 1129, in generate_per_chunk
+    for _, _, chunk in sample_transport.walk():
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/video_generate.py", line 1079, in walk
+    velocity = self.forward_velocity(infer_idx, 0)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/flowcache.py", line 93, in flowcache_forward_velocity
+    cache = SampleTransport.cache_reuse_manager
+AttributeError: type object 'SampleTransport' has no attribute 'cache_reuse_manager'
+
+✅ Completed: subject_consistency
+---
+🔍 Processing dimension: scene
+Loaded configuration from: yaml_config/sample/flowcache_vbench.yaml.tmp
+Total samples: 86
+GPUs: [0]
+Output: outputs/vbench/videos/scene
+Config: config/sample/vbench.json
+/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/site-packages/timm/models/layers/__init__.py:48: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[W520 10:33:58.467950462 CUDAAllocatorConfig.h:28] Warning: expandable_segments not supported on this platform (function operator())
+[2026-05-20 10:33:58,023 - INFO] Initialize torch distribution and model parallel successfully
+[2026-05-20 10:33:58,023 - INFO] MagiConfig(model_config=ModelConfig(model_name='videodit_ardf', num_layers=34, hidden_size=3072, ffn_hidden_size=12288, num_attention_heads=24, num_query_groups=8, kv_channels=128, layernorm_epsilon=1e-06, apply_layernorm_1p=True, x_rescale_factor=1, half_channel_vae=False, params_dtype=torch.bfloat16, patch_size=2, t_patch_size=1, in_channels=16, out_channels=16, cond_hidden_ratio=0.25, caption_channels=4096, caption_max_length=800, xattn_cond_hidden_ratio=1.0, cond_gating_ratio=1.0, gated_linear_unit=False), runtime_config=RuntimeConfig(cfg_number=1, cfg_t_range=[0.0, 0.0217, 0.1, 0.3, 0.999], prev_chunk_scales=[1.5, 1.5, 1.5, 1.0, 1.0], text_scales=[7.5, 7.5, 7.5, 0.0, 0.0], noise2clean_kvrange=[], clean_chunk_kvrange=1, clean_t=0.9999, seed=1234, num_frames=240, video_size_h=720, video_size_w=720, num_steps=16, window_size=4, fps=24, chunk_width=6, t5_pretrained='./downloads/t5_pretrained', t5_device='cuda', vae_pretrained='./downloads/vae', scale_factor=0.18215, temporal_downsample_factor=4, load='./downloads/4.5B_distill'), engine_config=EngineConfig(distributed_backend='nccl', distributed_timeout_minutes=15, pp_size=1, cp_size=1, cp_strategy='none', ulysses_overlap_degree=1, fp8_quant=False, distill_nearly_clean_chunk_threshold=0.3, shortcut_mode='8,16,16', distill=True, kv_offload=True, enable_cuda_graph=False))
+[2026-05-20 10:33:58,023 - INFO] Precompute validation prompt embeddings
+You are using the default legacy behaviour of the <class 'transformers.models.t5.tokenization_t5.T5Tokenizer'>. This is expected, and simply means that the `legacy` (previous) behavior will be used so nothing changes for you. If you want to use the new behaviour, set `legacy=False`. This should only be set if you understand what it means, and thoroughly read the reason why this was added as explained in https://github.com/huggingface/transformers/pull/24565
+KV cache compression is enabled.
+Processing 86 samples.
+[GPU 0] Assigned 86 samples
+[GPU 0] Loading model...
+[GPU 0] Model loaded.
+[GPU 0] Generating T2V: 'alley' -> /home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/outputs/vbench/videos/scene/alley-0.mp4
+
+[2026-05-20 10:34:17,215 - INFO] (cp, pp) rank (0, 0): param count 4459898128, model size 8.34 GB
+[2026-05-20 10:34:17,216 - INFO] Build DiTModel successfully
+[2026-05-20 10:34:17,216 - INFO] After build_dit_model, memory allocated: 0.04 GB, memory reserved: 0.08 GB
+[2026-05-20 10:34:17,216 - INFO] load inference_weight.distill weight from ./downloads/4.5B_distill/inference_weight.distill
+
+[2026-05-20 10:34:18,927 - INFO] Load Weight Missing Keys: []
+[2026-05-20 10:34:18,927 - INFO] Load Weight Unexpected Keys: []
+[2026-05-20 10:34:19,154 - INFO] After load_checkpoint, memory allocated: 8.39 GB, memory reserved: 8.40 GB
+[2026-05-20 10:34:19,156 - INFO] After high_precision_promoter, memory allocated: 8.39 GB, memory reserved: 8.40 GB
+[2026-05-20 10:34:19,251 - INFO] Load checkpoint successfully
+[2026-05-20 10:34:19,251 - INFO] Begin to generate per chunk
+[2026-05-20 10:34:19,252 - INFO] special_token = ['HQ_TOKEN', 'DURATION_TOKEN']
+
+Process Process-1:
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 259, in worker_process
+    process_func(pipeline, sample, config, gpu_id)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 185, in process_vbench_sample
+    pipeline.run_text_to_video(prompt=prompt, output_path=output_path)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 38, in run_text_to_video
+    self._run(prompt, None, output_path)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 52, in _run
+    [
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 52, in <listcomp>
+    [
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/video_generate.py", line 1129, in generate_per_chunk
+    for _, _, chunk in sample_transport.walk():
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/video_generate.py", line 1079, in walk
+    velocity = self.forward_velocity(infer_idx, 0)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/flowcache.py", line 93, in flowcache_forward_velocity
+    cache = SampleTransport.cache_reuse_manager
+AttributeError: type object 'SampleTransport' has no attribute 'cache_reuse_manager'
+
+✅ Completed: scene
+---
+🎉 All sampling tasks completed.

FlowCache/FlowCache4MAGI-1/logs/flowcache_vbench_20260520_121944.log

@@ -0,0 +1,208 @@
+🚀 Starting multi-GPU benchmark sampling
+🔢 Total dimensions to process: 3
+📋 Dimensions: overall_consistency subject_consistency scene
+🔍 Processing dimension: overall_consistency
+Loaded configuration from: yaml_config/sample/flowcache_vbench.yaml.tmp
+Total samples: 93
+GPUs: [0]
+Output: outputs/vbench/videos/overall_consistency
+Config: config/sample/vbench.json
+/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/site-packages/timm/models/layers/__init__.py:48: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[W520 12:19:50.307116267 CUDAAllocatorConfig.h:28] Warning: expandable_segments not supported on this platform (function operator())
+[2026-05-20 12:19:50,862 - INFO] Initialize torch distribution and model parallel successfully
+[2026-05-20 12:19:50,862 - INFO] MagiConfig(model_config=ModelConfig(model_name='videodit_ardf', num_layers=34, hidden_size=3072, ffn_hidden_size=12288, num_attention_heads=24, num_query_groups=8, kv_channels=128, layernorm_epsilon=1e-06, apply_layernorm_1p=True, x_rescale_factor=1, half_channel_vae=False, params_dtype=torch.bfloat16, patch_size=2, t_patch_size=1, in_channels=16, out_channels=16, cond_hidden_ratio=0.25, caption_channels=4096, caption_max_length=800, xattn_cond_hidden_ratio=1.0, cond_gating_ratio=1.0, gated_linear_unit=False), runtime_config=RuntimeConfig(cfg_number=1, cfg_t_range=[0.0, 0.0217, 0.1, 0.3, 0.999], prev_chunk_scales=[1.5, 1.5, 1.5, 1.0, 1.0], text_scales=[7.5, 7.5, 7.5, 0.0, 0.0], noise2clean_kvrange=[], clean_chunk_kvrange=1, clean_t=0.9999, seed=1234, num_frames=240, video_size_h=720, video_size_w=720, num_steps=16, window_size=4, fps=24, chunk_width=6, t5_pretrained='./downloads/t5_pretrained', t5_device='cuda', vae_pretrained='./downloads/vae', scale_factor=0.18215, temporal_downsample_factor=4, load='./downloads/4.5B_distill'), engine_config=EngineConfig(distributed_backend='nccl', distributed_timeout_minutes=15, pp_size=1, cp_size=1, cp_strategy='none', ulysses_overlap_degree=1, fp8_quant=False, distill_nearly_clean_chunk_threshold=0.3, shortcut_mode='8,16,16', distill=True, kv_offload=True, enable_cuda_graph=False))
+[2026-05-20 12:19:50,862 - INFO] Precompute validation prompt embeddings
+You are using the default legacy behaviour of the <class 'transformers.models.t5.tokenization_t5.T5Tokenizer'>. This is expected, and simply means that the `legacy` (previous) behavior will be used so nothing changes for you. If you want to use the new behaviour, set `legacy=False`. This should only be set if you understand what it means, and thoroughly read the reason why this was added as explained in https://github.com/huggingface/transformers/pull/24565
+KV cache compression is enabled.
+Processing 93 samples.
+[GPU 0] Assigned 93 samples
+[GPU 0] Loading model...
+[GPU 0] Model loaded.
+[GPU 0] Generating T2V: 'Close up of grapes on a rotating table.' -> /home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/outputs/vbench/videos/overall_consistency/Close up of grapes on a rotating table.-0.mp4
+
+[2026-05-20 12:20:08,897 - INFO] (cp, pp) rank (0, 0): param count 4459898128, model size 8.34 GB
+[2026-05-20 12:20:08,897 - INFO] Build DiTModel successfully
+[2026-05-20 12:20:08,897 - INFO] After build_dit_model, memory allocated: 0.04 GB, memory reserved: 0.08 GB
+[2026-05-20 12:20:08,897 - INFO] load inference_weight.distill weight from ./downloads/4.5B_distill/inference_weight.distill
+
+[2026-05-20 12:20:10,748 - INFO] Load Weight Missing Keys: []
+[2026-05-20 12:20:10,748 - INFO] Load Weight Unexpected Keys: []
+[2026-05-20 12:20:10,952 - INFO] After load_checkpoint, memory allocated: 8.39 GB, memory reserved: 8.40 GB
+[2026-05-20 12:20:10,954 - INFO] After high_precision_promoter, memory allocated: 8.39 GB, memory reserved: 8.40 GB
+[2026-05-20 12:20:11,056 - INFO] Load checkpoint successfully
+[2026-05-20 12:20:11,056 - INFO] Begin to generate per chunk
+[2026-05-20 12:20:11,056 - INFO] special_token = ['HQ_TOKEN', 'DURATION_TOKEN']
+
+
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 239, in worker_process
+    process_func(pipeline, sample, config, gpu_id)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 165, in process_vbench_sample
+    pipeline.run_text_to_video(prompt=prompt, output_path=output_path)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 38, in run_text_to_video
+    self._run(prompt, None, output_path)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 52, in _run
+    [
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 52, in <listcomp>
+    [
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/video_generate.py", line 1129, in generate_per_chunk
+    for _, _, chunk in sample_transport.walk():
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/video_generate.py", line 1092, in walk
+    clean_chunk, chunk_idx = self.integrate_velocity(work_status.infer_idx, work_status.cur_denoise_step)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/flowcache.py", line 530, in flowcache_integrate_velocity
+    _check_and_compress_kv(self, infer_idx, chunk_start, transport_input)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/flowcache.py", line 570, in _check_and_compress_kv
+    compressor.compress(
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/cache/kv_compressor.py", line 158, in compress
+    layer_result = self._compress_layer(
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/cache/kv_compressor.py", line 251, in _compress_layer
+    key_compressed, value_compressed, indices = kv_cluster.update_kv(
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/kvcompress/kv_compressor.py", line 52, in update_kv
+    return self.update_kv_token(
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/kvcompress/kv_compressor.py", line 168, in update_kv_token
+    raise ValueError(f"Unknown score weighting method: {self.score_weighting_method}")
+ValueError: Unknown score weighting method: None
+✅ Completed: overall_consistency
+---
+🔍 Processing dimension: subject_consistency
+Loaded configuration from: yaml_config/sample/flowcache_vbench.yaml.tmp
+Total samples: 72
+GPUs: [0]
+Output: outputs/vbench/videos/subject_consistency
+Config: config/sample/vbench.json
+/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/site-packages/timm/models/layers/__init__.py:48: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[W520 12:22:42.069407203 CUDAAllocatorConfig.h:28] Warning: expandable_segments not supported on this platform (function operator())
+[2026-05-20 12:22:42,624 - INFO] Initialize torch distribution and model parallel successfully
+[2026-05-20 12:22:42,624 - INFO] MagiConfig(model_config=ModelConfig(model_name='videodit_ardf', num_layers=34, hidden_size=3072, ffn_hidden_size=12288, num_attention_heads=24, num_query_groups=8, kv_channels=128, layernorm_epsilon=1e-06, apply_layernorm_1p=True, x_rescale_factor=1, half_channel_vae=False, params_dtype=torch.bfloat16, patch_size=2, t_patch_size=1, in_channels=16, out_channels=16, cond_hidden_ratio=0.25, caption_channels=4096, caption_max_length=800, xattn_cond_hidden_ratio=1.0, cond_gating_ratio=1.0, gated_linear_unit=False), runtime_config=RuntimeConfig(cfg_number=1, cfg_t_range=[0.0, 0.0217, 0.1, 0.3, 0.999], prev_chunk_scales=[1.5, 1.5, 1.5, 1.0, 1.0], text_scales=[7.5, 7.5, 7.5, 0.0, 0.0], noise2clean_kvrange=[], clean_chunk_kvrange=1, clean_t=0.9999, seed=1234, num_frames=240, video_size_h=720, video_size_w=720, num_steps=16, window_size=4, fps=24, chunk_width=6, t5_pretrained='./downloads/t5_pretrained', t5_device='cuda', vae_pretrained='./downloads/vae', scale_factor=0.18215, temporal_downsample_factor=4, load='./downloads/4.5B_distill'), engine_config=EngineConfig(distributed_backend='nccl', distributed_timeout_minutes=15, pp_size=1, cp_size=1, cp_strategy='none', ulysses_overlap_degree=1, fp8_quant=False, distill_nearly_clean_chunk_threshold=0.3, shortcut_mode='8,16,16', distill=True, kv_offload=True, enable_cuda_graph=False))
+[2026-05-20 12:22:42,625 - INFO] Precompute validation prompt embeddings
+You are using the default legacy behaviour of the <class 'transformers.models.t5.tokenization_t5.T5Tokenizer'>. This is expected, and simply means that the `legacy` (previous) behavior will be used so nothing changes for you. If you want to use the new behaviour, set `legacy=False`. This should only be set if you understand what it means, and thoroughly read the reason why this was added as explained in https://github.com/huggingface/transformers/pull/24565
+KV cache compression is enabled.
+Processing 72 samples.
+[GPU 0] Assigned 72 samples
+[GPU 0] Loading model...
+[GPU 0] Model loaded.
+[GPU 0] Generating T2V: 'a person swimming in ocean' -> /home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/outputs/vbench/videos/subject_consistency/a person swimming in ocean-0.mp4
+
+[2026-05-20 12:23:01,301 - INFO] (cp, pp) rank (0, 0): param count 4459898128, model size 8.34 GB
+[2026-05-20 12:23:01,301 - INFO] Build DiTModel successfully
+[2026-05-20 12:23:01,301 - INFO] After build_dit_model, memory allocated: 0.04 GB, memory reserved: 0.08 GB
+[2026-05-20 12:23:01,301 - INFO] load inference_weight.distill weight from ./downloads/4.5B_distill/inference_weight.distill
+
+[2026-05-20 12:23:03,292 - INFO] Load Weight Missing Keys: []
+[2026-05-20 12:23:03,292 - INFO] Load Weight Unexpected Keys: []
+[2026-05-20 12:23:03,525 - INFO] After load_checkpoint, memory allocated: 8.39 GB, memory reserved: 8.40 GB
+[2026-05-20 12:23:03,528 - INFO] After high_precision_promoter, memory allocated: 8.39 GB, memory reserved: 8.40 GB
+[2026-05-20 12:23:03,637 - INFO] Load checkpoint successfully
+[2026-05-20 12:23:03,637 - INFO] Begin to generate per chunk
+[2026-05-20 12:23:03,637 - INFO] special_token = ['HQ_TOKEN', 'DURATION_TOKEN']
+
+
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 239, in worker_process
+    process_func(pipeline, sample, config, gpu_id)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 165, in process_vbench_sample
+    pipeline.run_text_to_video(prompt=prompt, output_path=output_path)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 38, in run_text_to_video
+    self._run(prompt, None, output_path)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 52, in _run
+    [
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 52, in <listcomp>
+    [
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/video_generate.py", line 1129, in generate_per_chunk
+    for _, _, chunk in sample_transport.walk():
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/video_generate.py", line 1092, in walk
+    clean_chunk, chunk_idx = self.integrate_velocity(work_status.infer_idx, work_status.cur_denoise_step)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/flowcache.py", line 530, in flowcache_integrate_velocity
+    _check_and_compress_kv(self, infer_idx, chunk_start, transport_input)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/flowcache.py", line 570, in _check_and_compress_kv
+    compressor.compress(
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/cache/kv_compressor.py", line 158, in compress
+    layer_result = self._compress_layer(
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/cache/kv_compressor.py", line 251, in _compress_layer
+    key_compressed, value_compressed, indices = kv_cluster.update_kv(
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/kvcompress/kv_compressor.py", line 52, in update_kv
+    return self.update_kv_token(
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/kvcompress/kv_compressor.py", line 168, in update_kv_token
+    raise ValueError(f"Unknown score weighting method: {self.score_weighting_method}")
+ValueError: Unknown score weighting method: None
+✅ Completed: subject_consistency
+---
+🔍 Processing dimension: scene
+Loaded configuration from: yaml_config/sample/flowcache_vbench.yaml.tmp
+Total samples: 86
+GPUs: [0]
+Output: outputs/vbench/videos/scene
+Config: config/sample/vbench.json
+/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/site-packages/timm/models/layers/__init__.py:48: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[W520 12:25:23.550998312 CUDAAllocatorConfig.h:28] Warning: expandable_segments not supported on this platform (function operator())
+[2026-05-20 12:25:23,106 - INFO] Initialize torch distribution and model parallel successfully
+[2026-05-20 12:25:23,106 - INFO] MagiConfig(model_config=ModelConfig(model_name='videodit_ardf', num_layers=34, hidden_size=3072, ffn_hidden_size=12288, num_attention_heads=24, num_query_groups=8, kv_channels=128, layernorm_epsilon=1e-06, apply_layernorm_1p=True, x_rescale_factor=1, half_channel_vae=False, params_dtype=torch.bfloat16, patch_size=2, t_patch_size=1, in_channels=16, out_channels=16, cond_hidden_ratio=0.25, caption_channels=4096, caption_max_length=800, xattn_cond_hidden_ratio=1.0, cond_gating_ratio=1.0, gated_linear_unit=False), runtime_config=RuntimeConfig(cfg_number=1, cfg_t_range=[0.0, 0.0217, 0.1, 0.3, 0.999], prev_chunk_scales=[1.5, 1.5, 1.5, 1.0, 1.0], text_scales=[7.5, 7.5, 7.5, 0.0, 0.0], noise2clean_kvrange=[], clean_chunk_kvrange=1, clean_t=0.9999, seed=1234, num_frames=240, video_size_h=720, video_size_w=720, num_steps=16, window_size=4, fps=24, chunk_width=6, t5_pretrained='./downloads/t5_pretrained', t5_device='cuda', vae_pretrained='./downloads/vae', scale_factor=0.18215, temporal_downsample_factor=4, load='./downloads/4.5B_distill'), engine_config=EngineConfig(distributed_backend='nccl', distributed_timeout_minutes=15, pp_size=1, cp_size=1, cp_strategy='none', ulysses_overlap_degree=1, fp8_quant=False, distill_nearly_clean_chunk_threshold=0.3, shortcut_mode='8,16,16', distill=True, kv_offload=True, enable_cuda_graph=False))
+[2026-05-20 12:25:23,106 - INFO] Precompute validation prompt embeddings
+You are using the default legacy behaviour of the <class 'transformers.models.t5.tokenization_t5.T5Tokenizer'>. This is expected, and simply means that the `legacy` (previous) behavior will be used so nothing changes for you. If you want to use the new behaviour, set `legacy=False`. This should only be set if you understand what it means, and thoroughly read the reason why this was added as explained in https://github.com/huggingface/transformers/pull/24565
+KV cache compression is enabled.
+Processing 86 samples.
+[GPU 0] Assigned 86 samples
+[GPU 0] Loading model...
+[GPU 0] Model loaded.
+[GPU 0] Generating T2V: 'alley' -> /home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/outputs/vbench/videos/scene/alley-0.mp4
+
+[2026-05-20 12:25:41,783 - INFO] (cp, pp) rank (0, 0): param count 4459898128, model size 8.34 GB
+[2026-05-20 12:25:41,783 - INFO] Build DiTModel successfully
+[2026-05-20 12:25:41,783 - INFO] After build_dit_model, memory allocated: 0.04 GB, memory reserved: 0.08 GB
+[2026-05-20 12:25:41,783 - INFO] load inference_weight.distill weight from ./downloads/4.5B_distill/inference_weight.distill
+
+[2026-05-20 12:25:43,563 - INFO] Load Weight Missing Keys: []
+[2026-05-20 12:25:43,563 - INFO] Load Weight Unexpected Keys: []
+[2026-05-20 12:25:43,770 - INFO] After load_checkpoint, memory allocated: 8.39 GB, memory reserved: 8.40 GB
+[2026-05-20 12:25:43,773 - INFO] After high_precision_promoter, memory allocated: 8.39 GB, memory reserved: 8.40 GB
+[2026-05-20 12:25:43,867 - INFO] Load checkpoint successfully
+[2026-05-20 12:25:43,867 - INFO] Begin to generate per chunk
+[2026-05-20 12:25:43,867 - INFO] special_token = ['HQ_TOKEN', 'DURATION_TOKEN']
+
+
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 239, in worker_process
+    process_func(pipeline, sample, config, gpu_id)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 165, in process_vbench_sample
+    pipeline.run_text_to_video(prompt=prompt, output_path=output_path)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 38, in run_text_to_video
+    self._run(prompt, None, output_path)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 52, in _run
+    [
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 52, in <listcomp>
+    [
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/video_generate.py", line 1129, in generate_per_chunk
+    for _, _, chunk in sample_transport.walk():
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/video_generate.py", line 1092, in walk
+    clean_chunk, chunk_idx = self.integrate_velocity(work_status.infer_idx, work_status.cur_denoise_step)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/flowcache.py", line 530, in flowcache_integrate_velocity
+    _check_and_compress_kv(self, infer_idx, chunk_start, transport_input)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/flowcache.py", line 570, in _check_and_compress_kv
+    compressor.compress(
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/cache/kv_compressor.py", line 158, in compress
+    layer_result = self._compress_layer(
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/cache/kv_compressor.py", line 251, in _compress_layer
+    key_compressed, value_compressed, indices = kv_cluster.update_kv(
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/kvcompress/kv_compressor.py", line 52, in update_kv
+    return self.update_kv_token(
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/kvcompress/kv_compressor.py", line 168, in update_kv_token
+    raise ValueError(f"Unknown score weighting method: {self.score_weighting_method}")
+ValueError: Unknown score weighting method: None
+✅ Completed: scene
+---
+🎉 All sampling tasks completed.

FlowCache/FlowCache4MAGI-1/logs/flowcache_vbench_20260520_124220.log

@@ -0,0 +1,195 @@
+🚀 Starting multi-GPU benchmark sampling
+🔢 Total dimensions to process: 3
+📋 Dimensions: overall_consistency subject_consistency scene
+🔍 Processing dimension: overall_consistency
+Loaded configuration from: yaml_config/sample/flowcache_vbench.yaml.tmp
+Total samples: 93
+GPUs: [0]
+Output: outputs/vbench/videos/overall_consistency
+Config: config/sample/vbench.json
+/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/site-packages/timm/models/layers/__init__.py:48: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[W520 12:42:27.778130826 CUDAAllocatorConfig.h:28] Warning: expandable_segments not supported on this platform (function operator())
+[2026-05-20 12:42:27,333 - INFO] Initialize torch distribution and model parallel successfully
+[2026-05-20 12:42:27,333 - INFO] MagiConfig(model_config=ModelConfig(model_name='videodit_ardf', num_layers=34, hidden_size=3072, ffn_hidden_size=12288, num_attention_heads=24, num_query_groups=8, kv_channels=128, layernorm_epsilon=1e-06, apply_layernorm_1p=True, x_rescale_factor=1, half_channel_vae=False, params_dtype=torch.bfloat16, patch_size=2, t_patch_size=1, in_channels=16, out_channels=16, cond_hidden_ratio=0.25, caption_channels=4096, caption_max_length=800, xattn_cond_hidden_ratio=1.0, cond_gating_ratio=1.0, gated_linear_unit=False), runtime_config=RuntimeConfig(cfg_number=1, cfg_t_range=[0.0, 0.0217, 0.1, 0.3, 0.999], prev_chunk_scales=[1.5, 1.5, 1.5, 1.0, 1.0], text_scales=[7.5, 7.5, 7.5, 0.0, 0.0], noise2clean_kvrange=[], clean_chunk_kvrange=1, clean_t=0.9999, seed=1234, num_frames=240, video_size_h=720, video_size_w=720, num_steps=16, window_size=4, fps=24, chunk_width=6, t5_pretrained='./downloads/t5_pretrained', t5_device='cuda', vae_pretrained='./downloads/vae', scale_factor=0.18215, temporal_downsample_factor=4, load='./downloads/4.5B_distill'), engine_config=EngineConfig(distributed_backend='nccl', distributed_timeout_minutes=15, pp_size=1, cp_size=1, cp_strategy='none', ulysses_overlap_degree=1, fp8_quant=False, distill_nearly_clean_chunk_threshold=0.3, shortcut_mode='8,16,16', distill=True, kv_offload=True, enable_cuda_graph=False))
+[2026-05-20 12:42:27,333 - INFO] Precompute validation prompt embeddings
+You are using the default legacy behaviour of the <class 'transformers.models.t5.tokenization_t5.T5Tokenizer'>. This is expected, and simply means that the `legacy` (previous) behavior will be used so nothing changes for you. If you want to use the new behaviour, set `legacy=False`. This should only be set if you understand what it means, and thoroughly read the reason why this was added as explained in https://github.com/huggingface/transformers/pull/24565
+KV cache compression is enabled.
+Processing 93 samples.
+[GPU 0] Assigned 93 samples
+[GPU 0] Loading model...
+[GPU 0] Model loaded.
+[GPU 0] Generating T2V: 'Close up of grapes on a rotating table.' -> /home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/outputs/vbench/videos/overall_consistency/Close up of grapes on a rotating table.-0.mp4
+
+[2026-05-20 12:42:45,676 - INFO] (cp, pp) rank (0, 0): param count 4459898128, model size 8.34 GB
+[2026-05-20 12:42:45,676 - INFO] Build DiTModel successfully
+[2026-05-20 12:42:45,676 - INFO] After build_dit_model, memory allocated: 0.04 GB, memory reserved: 0.08 GB
+[2026-05-20 12:42:45,676 - INFO] load inference_weight.distill weight from ./downloads/4.5B_distill/inference_weight.distill
+
+[2026-05-20 12:42:47,225 - INFO] Load Weight Missing Keys: []
+[2026-05-20 12:42:47,225 - INFO] Load Weight Unexpected Keys: []
+[2026-05-20 12:42:47,431 - INFO] After load_checkpoint, memory allocated: 8.39 GB, memory reserved: 8.40 GB
+[2026-05-20 12:42:47,434 - INFO] After high_precision_promoter, memory allocated: 8.39 GB, memory reserved: 8.40 GB
+[2026-05-20 12:42:47,535 - INFO] Load checkpoint successfully
+[2026-05-20 12:42:47,535 - INFO] Begin to generate per chunk
+[2026-05-20 12:42:47,535 - INFO] special_token = ['HQ_TOKEN', 'DURATION_TOKEN']
+
+
+[2026-05-20 12:48:01,557 - INFO] Finish MagiPipeline, max memory allocated: 44.77 GB, max memory reserved: 56.32 GB
+[2026-05-20 12:48:01,557 - INFO] Precompute validation prompt embeddings
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+Using no weighting method
+✅ Video saved successfully.
+[DONE GPU 0] Saved: /home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/outputs/vbench/videos/overall_consistency/Close up of grapes on a rotating table.-0.mp4
+[GPU 0] Generating T2V: 'Turtle swimming in ocean.' -> /home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/outputs/vbench/videos/overall_consistency/Turtle swimming in ocean.-0.mp4
+
+[2026-05-20 12:48:18,152 - INFO] (cp, pp) rank (0, 0): param count 4459898128, model size 8.34 GB
+[2026-05-20 12:48:18,152 - INFO] Build DiTModel successfully
+[2026-05-20 12:48:18,152 - INFO] After build_dit_model, memory allocated: 0.17 GB, memory reserved: 0.31 GB
+[2026-05-20 12:48:18,152 - INFO] load inference_weight.distill weight from ./downloads/4.5B_distill/inference_weight.distill
+
+[2026-05-20 12:48:19,731 - INFO] Load Weight Missing Keys: []
+[2026-05-20 12:48:19,731 - INFO] Load Weight Unexpected Keys: []
+[2026-05-20 12:48:19,940 - INFO] After load_checkpoint, memory allocated: 8.52 GB, memory reserved: 8.56 GB
+[2026-05-20 12:48:19,943 - INFO] After high_precision_promoter, memory allocated: 8.52 GB, memory reserved: 8.56 GB
+[2026-05-20 12:48:20,393 - INFO] Load checkpoint successfully
+[2026-05-20 12:48:20,393 - INFO] Begin to generate per chunk
+[2026-05-20 12:48:20,394 - INFO] special_token = ['HQ_TOKEN', 'DURATION_TOKEN']
+
+[2026-05-20 12:48:20,402 - INFO] 
+Time Elapsed: [0:05:32.819959] From [begin_walk (2026-05-20 12:42:47.582387)] To [begin_walk (2026-05-20 12:48:20.402346)]
+

FlowCache/FlowCache4MAGI-1/logs/flowcache_vbench_20260520_170603.log

@@ -0,0 +1,298 @@
+🚀 Starting multi-GPU benchmark sampling
+🔢 Total dimensions to process: 3
+📋 Dimensions: overall_consistency subject_consistency scene
+🔍 Processing dimension: overall_consistency
+Loaded configuration from: yaml_config/sample/flowcache_vbench.yaml.tmp
+Total samples: 93
+GPUs: [0, 1, 2, 3]
+Output: outputs/vbench/videos/overall_consistency
+Config: config/sample/vbench.json
+Process Process-3:
+Process Process-2:
+Process Process-1:
+Process Process-4:
+Traceback (most recent call last):
+Traceback (most recent call last):
+Traceback (most recent call last):
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 240, in worker_process
+    configure_reuse_strategy(config)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 147, in configure_reuse_strategy
+    from inference.pipeline.video_generate import SampleTransport
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/__init__.py", line 15, in <module>
+    from .pipeline import MagiPipeline
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 240, in worker_process
+    configure_reuse_strategy(config)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 21, in <module>
+    from inference.model.dit import get_dit
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 240, in worker_process
+    configure_reuse_strategy(config)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 147, in configure_reuse_strategy
+    from inference.pipeline.video_generate import SampleTransport
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/__init__.py", line 15, in <module>
+    from .dit_model import get_dit, VideoDiTModel
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 147, in configure_reuse_strategy
+    from inference.pipeline.video_generate import SampleTransport
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_model.py", line 39, in <module>
+    from .dit_module import CaptionEmbedder, FinalLinear, LearnableRotaryEmbeddingCat, TimestepEmbedder, TransformerBlock
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/__init__.py", line 15, in <module>
+    from .pipeline import MagiPipeline
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/__init__.py", line 15, in <module>
+    from .pipeline import MagiPipeline
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_module.py", line 19, in <module>
+    import flashinfer
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 21, in <module>
+    from inference.model.dit import get_dit
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 21, in <module>
+    from inference.model.dit import get_dit
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/__init__.py", line 15, in <module>
+    from .dit_model import get_dit, VideoDiTModel
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/__init__.py", line 15, in <module>
+    from .dit_model import get_dit, VideoDiTModel
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 240, in worker_process
+    configure_reuse_strategy(config)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_model.py", line 39, in <module>
+    from .dit_module import CaptionEmbedder, FinalLinear, LearnableRotaryEmbeddingCat, TimestepEmbedder, TransformerBlock
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_model.py", line 39, in <module>
+    from .dit_module import CaptionEmbedder, FinalLinear, LearnableRotaryEmbeddingCat, TimestepEmbedder, TransformerBlock
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 147, in configure_reuse_strategy
+    from inference.pipeline.video_generate import SampleTransport
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_module.py", line 19, in <module>
+    import flashinfer
+ModuleNotFoundError: No module named 'flashinfer'
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_module.py", line 19, in <module>
+    import flashinfer
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/__init__.py", line 15, in <module>
+    from .pipeline import MagiPipeline
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 21, in <module>
+    from inference.model.dit import get_dit
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/__init__.py", line 15, in <module>
+    from .dit_model import get_dit, VideoDiTModel
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_model.py", line 39, in <module>
+    from .dit_module import CaptionEmbedder, FinalLinear, LearnableRotaryEmbeddingCat, TimestepEmbedder, TransformerBlock
+ModuleNotFoundError: No module named 'flashinfer'
+ModuleNotFoundError: No module named 'flashinfer'
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_module.py", line 19, in <module>
+    import flashinfer
+ModuleNotFoundError: No module named 'flashinfer'
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 429, in <module>
+    main()
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 425, in main
+    raise RuntimeError(f"{len(failed)} worker process(es) failed with exit codes: {failed}")
+RuntimeError: 4 worker process(es) failed with exit codes: [1, 1, 1, 1]
+✅ Completed: overall_consistency
+---
+🔍 Processing dimension: subject_consistency
+Loaded configuration from: yaml_config/sample/flowcache_vbench.yaml.tmp
+Total samples: 72
+GPUs: [0, 1, 2, 3]
+Output: outputs/vbench/videos/subject_consistency
+Config: config/sample/vbench.json
+Process Process-3:
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 240, in worker_process
+    configure_reuse_strategy(config)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 147, in configure_reuse_strategy
+    from inference.pipeline.video_generate import SampleTransport
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/__init__.py", line 15, in <module>
+    from .pipeline import MagiPipeline
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 21, in <module>
+    from inference.model.dit import get_dit
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/__init__.py", line 15, in <module>
+    from .dit_model import get_dit, VideoDiTModel
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_model.py", line 39, in <module>
+    from .dit_module import CaptionEmbedder, FinalLinear, LearnableRotaryEmbeddingCat, TimestepEmbedder, TransformerBlock
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_module.py", line 19, in <module>
+    import flashinfer
+ModuleNotFoundError: No module named 'flashinfer'
+Process Process-1:
+Process Process-2:
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 240, in worker_process
+    configure_reuse_strategy(config)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 147, in configure_reuse_strategy
+    from inference.pipeline.video_generate import SampleTransport
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/__init__.py", line 15, in <module>
+    from .pipeline import MagiPipeline
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 21, in <module>
+    from inference.model.dit import get_dit
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/__init__.py", line 15, in <module>
+    from .dit_model import get_dit, VideoDiTModel
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_model.py", line 39, in <module>
+    from .dit_module import CaptionEmbedder, FinalLinear, LearnableRotaryEmbeddingCat, TimestepEmbedder, TransformerBlock
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_module.py", line 19, in <module>
+    import flashinfer
+ModuleNotFoundError: No module named 'flashinfer'
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 240, in worker_process
+    configure_reuse_strategy(config)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 147, in configure_reuse_strategy
+    from inference.pipeline.video_generate import SampleTransport
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/__init__.py", line 15, in <module>
+    from .pipeline import MagiPipeline
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 21, in <module>
+    from inference.model.dit import get_dit
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/__init__.py", line 15, in <module>
+    from .dit_model import get_dit, VideoDiTModel
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_model.py", line 39, in <module>
+    from .dit_module import CaptionEmbedder, FinalLinear, LearnableRotaryEmbeddingCat, TimestepEmbedder, TransformerBlock
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_module.py", line 19, in <module>
+    import flashinfer
+ModuleNotFoundError: No module named 'flashinfer'
+Process Process-4:
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 240, in worker_process
+    configure_reuse_strategy(config)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 147, in configure_reuse_strategy
+    from inference.pipeline.video_generate import SampleTransport
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/__init__.py", line 15, in <module>
+    from .pipeline import MagiPipeline
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 21, in <module>
+    from inference.model.dit import get_dit
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/__init__.py", line 15, in <module>
+    from .dit_model import get_dit, VideoDiTModel
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_model.py", line 39, in <module>
+    from .dit_module import CaptionEmbedder, FinalLinear, LearnableRotaryEmbeddingCat, TimestepEmbedder, TransformerBlock
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_module.py", line 19, in <module>
+    import flashinfer
+ModuleNotFoundError: No module named 'flashinfer'
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 429, in <module>
+    main()
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 425, in main
+    raise RuntimeError(f"{len(failed)} worker process(es) failed with exit codes: {failed}")
+RuntimeError: 4 worker process(es) failed with exit codes: [1, 1, 1, 1]
+✅ Completed: subject_consistency
+---
+🔍 Processing dimension: scene
+Loaded configuration from: yaml_config/sample/flowcache_vbench.yaml.tmp
+Total samples: 86
+GPUs: [0, 1, 2, 3]
+Output: outputs/vbench/videos/scene
+Config: config/sample/vbench.json
+Process Process-4:
+Process Process-2:
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 240, in worker_process
+    configure_reuse_strategy(config)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 147, in configure_reuse_strategy
+    from inference.pipeline.video_generate import SampleTransport
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/__init__.py", line 15, in <module>
+    from .pipeline import MagiPipeline
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 21, in <module>
+    from inference.model.dit import get_dit
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/__init__.py", line 15, in <module>
+    from .dit_model import get_dit, VideoDiTModel
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_model.py", line 39, in <module>
+    from .dit_module import CaptionEmbedder, FinalLinear, LearnableRotaryEmbeddingCat, TimestepEmbedder, TransformerBlock
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_module.py", line 19, in <module>
+    import flashinfer
+ModuleNotFoundError: No module named 'flashinfer'
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 240, in worker_process
+    configure_reuse_strategy(config)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 147, in configure_reuse_strategy
+    from inference.pipeline.video_generate import SampleTransport
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/__init__.py", line 15, in <module>
+    from .pipeline import MagiPipeline
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 21, in <module>
+    from inference.model.dit import get_dit
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/__init__.py", line 15, in <module>
+    from .dit_model import get_dit, VideoDiTModel
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_model.py", line 39, in <module>
+    from .dit_module import CaptionEmbedder, FinalLinear, LearnableRotaryEmbeddingCat, TimestepEmbedder, TransformerBlock
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_module.py", line 19, in <module>
+    import flashinfer
+ModuleNotFoundError: No module named 'flashinfer'
+Process Process-3:
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 240, in worker_process
+    configure_reuse_strategy(config)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 147, in configure_reuse_strategy
+    from inference.pipeline.video_generate import SampleTransport
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/__init__.py", line 15, in <module>
+    from .pipeline import MagiPipeline
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 21, in <module>
+    from inference.model.dit import get_dit
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/__init__.py", line 15, in <module>
+    from .dit_model import get_dit, VideoDiTModel
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_model.py", line 39, in <module>
+    from .dit_module import CaptionEmbedder, FinalLinear, LearnableRotaryEmbeddingCat, TimestepEmbedder, TransformerBlock
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_module.py", line 19, in <module>
+    import flashinfer
+ModuleNotFoundError: No module named 'flashinfer'
+Process Process-1:
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 240, in worker_process
+    configure_reuse_strategy(config)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 147, in configure_reuse_strategy
+    from inference.pipeline.video_generate import SampleTransport
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/__init__.py", line 15, in <module>
+    from .pipeline import MagiPipeline
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 21, in <module>
+    from inference.model.dit import get_dit
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/__init__.py", line 15, in <module>
+    from .dit_model import get_dit, VideoDiTModel
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_model.py", line 39, in <module>
+    from .dit_module import CaptionEmbedder, FinalLinear, LearnableRotaryEmbeddingCat, TimestepEmbedder, TransformerBlock
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_module.py", line 19, in <module>
+    import flashinfer
+ModuleNotFoundError: No module named 'flashinfer'
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 429, in <module>
+    main()
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 425, in main
+    raise RuntimeError(f"{len(failed)} worker process(es) failed with exit codes: {failed}")
+RuntimeError: 4 worker process(es) failed with exit codes: [1, 1, 1, 1]
+✅ Completed: scene
+---
+🎉 All sampling tasks completed.

FlowCache/FlowCache4MAGI-1/logs/flowcache_vbench_20260521_034016.log

@@ -0,0 +1,109 @@
+🚀 Starting multi-GPU benchmark sampling
+🔢 Total dimensions to process: 3
+📋 Dimensions: overall_consistency subject_consistency scene
+🔍 Processing dimension: overall_consistency
+Loaded configuration from: yaml_config/sample/flowcache_vbench.yaml.tmp
+Total samples: 93
+GPUs: [0]
+Output: outputs/vbench/videos/overall_consistency
+Config: config/sample/vbench.json
+Process Process-1:
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 240, in worker_process
+    configure_reuse_strategy(config)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 147, in configure_reuse_strategy
+    from inference.pipeline.video_generate import SampleTransport
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/__init__.py", line 15, in <module>
+    from .pipeline import MagiPipeline
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 21, in <module>
+    from inference.model.dit import get_dit
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/__init__.py", line 15, in <module>
+    from .dit_model import get_dit, VideoDiTModel
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_model.py", line 39, in <module>
+    from .dit_module import CaptionEmbedder, FinalLinear, LearnableRotaryEmbeddingCat, TimestepEmbedder, TransformerBlock
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_module.py", line 19, in <module>
+    import flashinfer
+ModuleNotFoundError: No module named 'flashinfer'
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 429, in <module>
+    main()
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 425, in main
+    raise RuntimeError(f"{len(failed)} worker process(es) failed with exit codes: {failed}")
+RuntimeError: 1 worker process(es) failed with exit codes: [1]
+✅ Completed: overall_consistency
+---
+🔍 Processing dimension: subject_consistency
+Loaded configuration from: yaml_config/sample/flowcache_vbench.yaml.tmp
+Total samples: 72
+GPUs: [0]
+Output: outputs/vbench/videos/subject_consistency
+Config: config/sample/vbench.json
+Process Process-1:
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 240, in worker_process
+    configure_reuse_strategy(config)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 147, in configure_reuse_strategy
+    from inference.pipeline.video_generate import SampleTransport
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/__init__.py", line 15, in <module>
+    from .pipeline import MagiPipeline
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 21, in <module>
+    from inference.model.dit import get_dit
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/__init__.py", line 15, in <module>
+    from .dit_model import get_dit, VideoDiTModel
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_model.py", line 39, in <module>
+    from .dit_module import CaptionEmbedder, FinalLinear, LearnableRotaryEmbeddingCat, TimestepEmbedder, TransformerBlock
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_module.py", line 19, in <module>
+    import flashinfer
+ModuleNotFoundError: No module named 'flashinfer'
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 429, in <module>
+    main()
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 425, in main
+    raise RuntimeError(f"{len(failed)} worker process(es) failed with exit codes: {failed}")
+RuntimeError: 1 worker process(es) failed with exit codes: [1]
+✅ Completed: subject_consistency
+---
+🔍 Processing dimension: scene
+Loaded configuration from: yaml_config/sample/flowcache_vbench.yaml.tmp
+Total samples: 86
+GPUs: [0]
+Output: outputs/vbench/videos/scene
+Config: config/sample/vbench.json
+Process Process-1:
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 314, in _bootstrap
+    self.run()
+  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/lib/python3.12/multiprocessing/process.py", line 108, in run
+    self._target(*self._args, **self._kwargs)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 240, in worker_process
+    configure_reuse_strategy(config)
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 147, in configure_reuse_strategy
+    from inference.pipeline.video_generate import SampleTransport
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/__init__.py", line 15, in <module>
+    from .pipeline import MagiPipeline
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 21, in <module>
+    from inference.model.dit import get_dit
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/__init__.py", line 15, in <module>
+    from .dit_model import get_dit, VideoDiTModel
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_model.py", line 39, in <module>
+    from .dit_module import CaptionEmbedder, FinalLinear, LearnableRotaryEmbeddingCat, TimestepEmbedder, TransformerBlock
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/model/dit/dit_module.py", line 19, in <module>
+    import flashinfer
+ModuleNotFoundError: No module named 'flashinfer'
+Traceback (most recent call last):
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 429, in <module>
+    main()
+  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 425, in main
+    raise RuntimeError(f"{len(failed)} worker process(es) failed with exit codes: {failed}")
+RuntimeError: 1 worker process(es) failed with exit codes: [1]
+✅ Completed: scene
+---
+🎉 All sampling tasks completed.

FlowCache/FlowCache4MAGI-1/requirements.txt

@@ -0,0 +1,21 @@
+accelerate==0.32.1
+beautifulsoup4==4.13.4
+debugpy==1.8.14
+diffusers==0.29.2
+einops>=0.6.0
+ffmpeg-python
+# flash-attn==2.4.2
+flashinfer-python==0.2.0.post2 --extra-index-url https://flashinfer.ai/whl/cu124/torch2.4/
+ftfy==6.2.0
+gpustat==1.1.1
+imageio==2.34.0
+imageio[ffmpeg]
+matplotlib==3.10.1
+numpy==1.26.4
+protobuf==5.28.3
+rich==14.0.0
+sentencepiece==0.2.0
+timm==1.0.15
+torchdiffeq==0.2.4
+transformers==4.42.3
+tqdm

FlowCache/FlowCache4MAGI-1/sample_video.py

@@ -0,0 +1,429 @@
+# Copyright 2024 MAGI Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import re
+import sys
+import argparse
+import csv
+import subprocess
+from pathlib import Path
+
+import multiprocessing as mp
+
+# Constants
+DEFAULT_BASE_PORT = 29510
+PHYSICSIQ_FPS = 24
+
+
+def resolve_gpu_ids(gpus_config) -> list[int]:
+    """Resolve explicit GPU IDs or auto-detect all currently visible GPUs."""
+    if isinstance(gpus_config, int):
+        return [gpus_config]
+
+    gpus_text = str(gpus_config).strip()
+    if not gpus_text:
+        raise ValueError("'gpus' must not be empty")
+
+    if gpus_text.lower() not in {"all", "auto"}:
+        return [int(item.strip()) for item in gpus_text.split(",") if item.strip()]
+
+    visible_devices = os.environ.get("CUDA_VISIBLE_DEVICES")
+    if visible_devices:
+        visible = [item.strip() for item in visible_devices.split(",") if item.strip()]
+        if visible and all(item.isdigit() for item in visible):
+            return [int(item) for item in visible]
+
+    try:
+        output = subprocess.check_output(
+            ["nvidia-smi", "--query-gpu=index", "--format=csv,noheader,nounits"],
+            text=True,
+            timeout=10,
+        )
+        gpu_ids = [int(line.strip()) for line in output.splitlines() if line.strip()]
+        if gpu_ids:
+            return gpu_ids
+    except Exception:
+        pass
+
+    try:
+        import torch
+
+        count = torch.cuda.device_count()
+        if count > 0:
+            return list(range(count))
+    except Exception:
+        pass
+
+    raise RuntimeError("No CUDA GPUs detected for gpus: all")
+
+
+def load_yaml_config(yaml_path: str) -> dict:
+    """Load configuration from YAML file."""
+    import yaml
+
+    with open(yaml_path, "r") as f:
+        return yaml.safe_load(f)
+
+
+def apply_slice(items: list, start: int | None, end: int | None) -> list:
+    """Apply start/end slice to a list with bounds checking."""
+    if start is None and end is None:
+        return items
+
+    slice_start = max(0, start if start is not None else 0)
+    slice_end = min(end if end is not None else len(items), len(items))
+    slice_end = max(slice_start, slice_end)
+
+    return items[slice_start:slice_end]
+
+
+def configure_teacache(transport, config: dict) -> None:
+    """Configure TeaCache reuse strategy on SampleTransport."""
+    from inference.pipeline.teacache import setup_teacache
+
+    setup_teacache(
+        rel_l1_thresh=config["rel_l1_thresh"],
+        warmup_steps=config["warmup_steps"],
+        log=config.get("log", False),
+    )
+
+
+def configure_kv_cache(transport, config: dict) -> None:
+    """Configure KV cache compression if enabled."""
+    if not config.get("compress_kv_cache", False):
+        transport.compress_kv_cache = False
+        return
+
+    print("KV cache compression is enabled.")
+    transport.compress_kv_cache = True
+
+    assert config.get("total_cache_chunk_nums") is not None
+
+    compression_config = {
+        "method_config": {
+            "compress_strategy": config["compress_strategy"],
+            "mix_lambda": config["mix_lambda"],
+            "query_granularity": config["query_granularity"],
+            "score_weighting_method": config.get("score_weighting_method") or "no_weight",
+            "power": config.get("power", 3),
+        },
+    }
+
+    from inference.pipeline.kvcompress import replace_magi
+
+    replace_magi(compression_config)
+
+
+def configure_flowcache(transport, config: dict) -> None:
+    """Configure FlowCache reuse strategy on SampleTransport."""
+    from inference.pipeline.flowcache import setup_flowcache
+
+    configure_kv_cache(transport, config)
+
+    setup_flowcache(
+        rel_l1_thresh=config["rel_l1_thresh"],
+        warmup_steps=config["warmup_steps"],
+        discard_nearly_clean_chunk=config.get("discard_nearly_clean_chunk", False),
+        log=config.get("log", False),
+        total_cache_chunk_nums=config.get("total_cache_chunk_nums", 5),
+        compress_kv_cache=config.get("compress_kv_cache", False),
+    )
+
+
+def configure_reuse_strategy(config: dict) -> None:
+    """Configure the appropriate reuse strategy on SampleTransport."""
+    from inference.pipeline.video_generate import SampleTransport
+
+    strategy = config["reuse_strategy"]
+
+    if strategy == "original":
+        return
+    if strategy == "all":
+        configure_teacache(SampleTransport, config)
+    elif strategy == "chunkwise":
+        configure_flowcache(SampleTransport, config)
+    else:
+        raise ValueError(f"Unknown reuse strategy: {strategy}")
+
+
+def setup_environment(gpu_id: int) -> None:
+    """Set up environment variables for a GPU worker process."""
+    os.environ["CUDA_VISIBLE_DEVICES"] = str(gpu_id)
+    os.environ["WORLD_SIZE"] = "1"
+    os.environ["RANK"] = "0"
+    os.environ["LOCAL_RANK"] = "0"
+    os.environ["MASTER_ADDR"] = "localhost"
+    os.environ["MASTER_PORT"] = str(DEFAULT_BASE_PORT + gpu_id)
+
+    # Enable pdb terminal debugging
+    sys.stdin = open(0)
+
+
+def filter_existing_samples(samples: list, config: dict) -> list:
+    """Filter out samples whose output files already exist."""
+    if config["benchmark"] == "vbench":
+        return [
+            sample
+            for sample in samples
+            if not os.path.exists(os.path.abspath(os.path.join(config["save_path"], f"{sample}-0.mp4")))
+        ]
+    else:  # physicsiq
+        return [
+            sample for sample in samples if not os.path.exists(sample["output_path"])
+        ]
+
+
+def assign_samples_to_gpu(
+    samples: list, gpu_id: int, rank: int, num_gpus: int
+) -> list:
+    """Divide samples across GPUs and return the subset for this GPU."""
+    samples_per_gpu = (len(samples) + num_gpus - 1) // num_gpus
+    start_idx = rank * samples_per_gpu
+    end_idx = min(start_idx + samples_per_gpu, len(samples))
+    return samples[start_idx:end_idx]
+
+
+def process_vbench_sample(pipeline, prompt: str, config: dict, gpu_id: int) -> None:
+    """Process a single vbench text-to-video sample."""
+    output_path = os.path.abspath(os.path.join(config["save_path"], f"{prompt}-0.mp4"))
+
+    if os.path.exists(output_path):
+        print(f"[SKIP GPU {gpu_id}] Already exists: {output_path}")
+        return
+
+    print(f"[GPU {gpu_id}] Generating T2V: '{prompt}' -> {output_path}")
+    pipeline.run_text_to_video(prompt=prompt, output_path=output_path)
+    print(f"[DONE GPU {gpu_id}] Saved: {output_path}")
+
+
+def process_physicsiq_sample(pipeline, sample: dict, gpu_id: int) -> None:
+    """Process a single PhysicsIQ video-to-video sample."""
+    prompt = sample["description"]
+    prefix_video_path = sample["prefix_video_path"]
+    output_path = sample["output_path"]
+
+    if not os.path.exists(prefix_video_path):
+        print(f"[WARN GPU {gpu_id}] Conditioning video not found: {prefix_video_path}")
+        return
+
+    if os.path.exists(output_path):
+        print(f"[SKIP GPU {gpu_id}] Already exists: {output_path}")
+        return
+
+    print(f"[GPU {gpu_id}] Generating V2V: '{prompt}'")
+    print(f"   Input:  {prefix_video_path}")
+    print(f"   Output: {output_path}")
+
+    pipeline.run_video_to_video(
+        prompt=prompt,
+        prefix_video_path=prefix_video_path,
+        output_path=output_path,
+    )
+    print(f"[DONE GPU {gpu_id}] Saved: {output_path}")
+
+
+def worker_process(gpu_id: int, rank: int, config: dict, all_samples: list) -> None:
+    """Independent worker running on each GPU."""
+    setup_environment(gpu_id)
+    configure_reuse_strategy(config)
+
+    try:
+        magi_root = subprocess.check_output(
+            ["git", "rev-parse", "--show-toplevel"]
+        ).decode().strip()
+        os.environ["MAGI_ROOT"] = magi_root
+        os.environ["PYTHONPATH"] = f"{magi_root}:{os.environ.get('PYTHONPATH', '')}"
+    except Exception as e:
+        print(f"[GPU {gpu_id}] Failed to set MAGI_ROOT: {e}")
+        return
+
+    filtered_samples = filter_existing_samples(all_samples, config)
+
+    if not filtered_samples:
+        print(f"[GPU {gpu_id}] No samples need to be generated.")
+        return
+
+    print(f"Processing {len(filtered_samples)} samples.")
+
+    my_samples = assign_samples_to_gpu(
+        filtered_samples, gpu_id, rank, config["num_gpus"]
+    )
+
+    if not my_samples:
+        print(f"[GPU {gpu_id}] No samples assigned.")
+        return
+
+    print(f"[GPU {gpu_id}] Assigned {len(my_samples)} samples")
+
+    from inference.pipeline.entry import MagiPipeline
+
+    print(f"[GPU {gpu_id}] Loading model...")
+    pipeline = MagiPipeline(config["config_file"])
+    print(f"[GPU {gpu_id}] Model loaded.")
+
+    process_func = (
+        process_vbench_sample if config["benchmark"] == "vbench" else process_physicsiq_sample
+    )
+
+    for sample in my_samples:
+        process_func(pipeline, sample, config, gpu_id)
+
+    print(f"[GPU {gpu_id}] Completed.")
+
+
+def build_conditioning_video_path(
+    data_root: str, vid_id: str, scenario: str, fps: int
+) -> str:
+    """Construct the path to the conditioning video file."""
+    conditioning_dir = os.path.join(
+        data_root, "physics-IQ-benchmark", "split-videos", "conditioning", f"{fps}FPS"
+    )
+    match_suffix = re.search(r"_(.*)", scenario)
+    suffix = match_suffix.group(1) if match_suffix else ""
+    filename = f"{vid_id}_conditioning-videos_{fps}FPS_{suffix}"
+    return os.path.join(conditioning_dir, filename)
+
+
+def load_physicsiq_samples(config: dict) -> list[dict]:
+    """Load sample list from PhysicsIQ dataset."""
+    data_root = config["physicsiq_data_dir"]
+    descriptions_csv = os.path.join(data_root, "descriptions", "descriptions.csv")
+    output_dir = config["save_path"]
+
+    if not os.path.exists(descriptions_csv):
+        raise FileNotFoundError(f"descriptions.csv not found at {descriptions_csv}")
+
+    os.makedirs(output_dir, exist_ok=True)
+
+    samples = []
+    with open(descriptions_csv, mode="r") as f:
+        reader = csv.DictReader(f)
+        for row in reader:
+            scenario = row["scenario"].strip()
+            match_id = re.match(r"^(\d+)_", scenario)
+
+            if not match_id:
+                print(f"Cannot extract ID from scenario: {scenario}")
+                continue
+
+            vid_id = match_id.group(1).zfill(4)
+            description = row["description"]
+            generated_video_name = row["generated_video_name"]
+            prefix_video_path = build_conditioning_video_path(
+                data_root, vid_id, scenario, PHYSICSIQ_FPS
+            )
+            output_path = os.path.join(output_dir, generated_video_name)
+
+            os.makedirs(os.path.dirname(output_path), exist_ok=True)
+
+            samples.append({
+                "vid_id": vid_id,
+                "scenario": scenario,
+                "description": description,
+                "generated_video_name": generated_video_name,
+                "prefix_video_path": prefix_video_path,
+                "output_path": output_path,
+            })
+
+    # PhysicsIQ samples are duplicated; take only the first half
+    unique_count = len(samples) // 2
+    samples = samples[:unique_count]
+
+    print(f"Loaded {unique_count} PhysicsIQ samples.")
+
+    return apply_slice(samples, config.get("start"), config.get("end"))
+
+
+def load_vbench_samples(config: dict) -> list[str]:
+    """Load prompt list from vbench dimension file."""
+    prompt_dir = config["vbench_prompt_dir"]
+    dimension = config.get("dimension")
+
+    if not dimension:
+        raise ValueError("For vbench, 'dimension' must be specified in config")
+
+    prompt_file = os.path.join(prompt_dir, f"{dimension}.txt")
+
+    if not os.path.exists(prompt_file):
+        raise FileNotFoundError(f"Prompt file not found: {prompt_file}")
+
+    with open(prompt_file, "r") as f:
+        prompts = [line.strip() for line in f if line.strip()]
+
+    return apply_slice(prompts, config.get("start"), config.get("end"))
+
+
+def setup_save_path(config: dict) -> None:
+    """Configure the output save path based on benchmark type."""
+    base_path = config["base_save_path"]
+
+    if config["benchmark"] == "vbench":
+        dimension = config.get("dimension")
+        videos_dir = os.path.join(base_path, "videos", dimension) if dimension else None
+        config["save_path"] = videos_dir if videos_dir else os.path.join(base_path, "videos")
+    elif config["benchmark"] == "physicsiq":
+        config["save_path"] = os.path.join(base_path, "videos")
+
+    os.makedirs(config["save_path"], exist_ok=True)
+
+
+def main() -> None:
+    """Entry point for video sampling script."""
+    parser = argparse.ArgumentParser(
+        description="Video sampling script using YAML configuration"
+    )
+    parser.add_argument("yaml_config", type=str, help="Path to YAML configuration file")
+    args = parser.parse_args()
+
+    config = load_yaml_config(args.yaml_config)
+    print(f"Loaded configuration from: {args.yaml_config}")
+
+    setup_save_path(config)
+
+    gpu_ids = resolve_gpu_ids(config["gpus"])
+    config["num_gpus"] = len(gpu_ids)
+
+    benchmark = config["benchmark"]
+    if benchmark == "vbench":
+        all_samples = load_vbench_samples(config)
+    elif benchmark == "physicsiq":
+        data_root = config["physicsiq_data_dir"]
+        if not os.path.exists(data_root):
+            raise FileNotFoundError(f"Data directory not found: {data_root}")
+        all_samples = load_physicsiq_samples(config)
+    else:
+        raise ValueError(f"Invalid benchmark: {benchmark}")
+
+    print(f"Total samples: {len(all_samples)}")
+    print(f"GPUs: {gpu_ids}")
+    print(f"Output: {config['save_path']}")
+    print(f"Config: {config['config_file']}")
+
+    processes = []
+    for rank, gpu_id in enumerate(gpu_ids):
+        p = mp.Process(target=worker_process, args=(gpu_id, rank, config, all_samples))
+        p.start()
+        processes.append(p)
+
+    for p in processes:
+        p.join()
+
+    failed = [p.exitcode for p in processes if p.exitcode != 0]
+    if failed:
+        raise RuntimeError(f"{len(failed)} worker process(es) failed with exit codes: {failed}")
+
+
+if __name__ == "__main__":
+    main()

FlowCache/FlowCache4MAGI-1/scripts/metric.sh

@@ -0,0 +1,5 @@
+export CUDA_VISIBLE_DEVICES=3
+
+python tools/video_metrics.py \
+--original_video "/path/to/original_video.mp4" \
+--generated_video "/path/to/generated_video.mp4"

FlowCache/FlowCache4MAGI-1/tools/plot_l1_rel.py

@@ -0,0 +1,152 @@
+#!/usr/bin/env python3
+
+import argparse
+import json
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Set, Tuple
+
+import matplotlib
+
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+
+
+def parse_int_list(value: Optional[str]) -> Optional[Set[int]]:
+    if not value:
+        return None
+    return {int(item.strip()) for item in value.split(",") if item.strip()}
+
+
+def load_l1_rel_records(json_path: Path) -> List[dict]:
+    with json_path.open("r") as f:
+        payload = json.load(f)
+    if isinstance(payload, list):
+        return payload
+    if isinstance(payload, dict) and isinstance(payload.get("records"), list):
+        return payload["records"]
+    raise ValueError(f"Cannot find records in {json_path}")
+
+
+def collect_by_chunk(records: List[dict], chunk_ids: Optional[Set[int]], max_chunks: Optional[int]) -> Dict[int, List[dict]]:
+    chunks = defaultdict(list)
+    for record in records:
+        chunk_idx = int(record["chunk_idx"])
+        if chunk_ids is not None and chunk_idx not in chunk_ids:
+            continue
+        chunks[chunk_idx].append(record)
+
+    chunks = dict(sorted(chunks.items()))
+    if max_chunks is not None:
+        chunks = dict(list(chunks.items())[:max_chunks])
+    return chunks
+
+
+def plot_l1_rel(
+    chunks: Dict[int, List[dict]],
+    output_path: Path,
+    x_field: str,
+    y_field: str,
+    reverse_x: bool,
+    title: Optional[str],
+    figsize: Tuple[float, float],
+    dpi: int,
+) -> None:
+    fig, ax = plt.subplots(figsize=figsize)
+
+    for chunk_idx, records in chunks.items():
+        points = []
+        for record in records:
+            if x_field not in record or y_field not in record:
+                continue
+            if record[x_field] is None or record[y_field] is None:
+                continue
+            points.append((float(record[x_field]), float(record[y_field])))
+        if not points:
+            continue
+
+        points.sort(key=lambda item: item[0])
+        xs, ys = zip(*points)
+        ax.plot(xs, ys, marker="o", linewidth=1.6, markersize=3, label=f"chunk {chunk_idx}")
+
+    ax.set_xlabel(x_field)
+    ax.set_ylabel(y_field)
+    ax.set_title(title or f"{y_field} by timestep")
+    ax.grid(True, alpha=0.3)
+    if reverse_x:
+        ax.invert_xaxis()
+    ax.legend(loc="best", fontsize="small", ncols=2)
+    fig.tight_layout()
+
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    fig.savefig(output_path, dpi=dpi)
+    plt.close(fig)
+
+
+def parse_arguments():
+    parser = argparse.ArgumentParser(description="Plot per-chunk MAGI relative L1 change curves.")
+    parser.add_argument("json_path", type=Path, help="Path to L1 relative change JSON saved by --l1_rel_stats_path.")
+    parser.add_argument("-o", "--output", type=Path, help="Output image path. Defaults to <json stem>_plot.png.")
+    parser.add_argument("--chunks", type=str, help="Comma-separated chunk_idx list to plot, for example: 0,1,2.")
+    parser.add_argument("--max-chunks", type=int, help="Plot at most this many chunks after filtering.")
+    parser.add_argument(
+        "--x-field",
+        choices=["timestep", "next_timestep", "cur_denoise_step", "denoise_idx"],
+        default="next_timestep",
+        help="Record field used for the x axis. next_timestep is the cleaner MAGI step.",
+    )
+    parser.add_argument(
+        "--y-field",
+        choices=[
+            "l1_rel",
+            "l1_rel_ratio",
+            "delta_l1_norm",
+            "x_l1_norm",
+            "x_embedder_l1_rel",
+            "x_embedder_l1_rel_ratio",
+            "x_embedder_delta_l1_norm",
+            "x_embedder_x_l1_norm",
+            "flowcache_rel_l1",
+            "flowcache_rel_l1_ratio",
+            "flowcache_delta_l1_norm",
+            "flowcache_prev_feat_l1_norm",
+            "flowcache_accumulated_rel_l1",
+            "rel_l1_thresh",
+        ],
+        default="l1_rel",
+        help="Record field used for the y axis.",
+    )
+    parser.add_argument("--reverse-x", action="store_true", help="Reverse the x axis.")
+    parser.add_argument("--title", type=str, help="Figure title.")
+    parser.add_argument("--figsize", type=str, default="10,6", help="Figure size as width,height.")
+    parser.add_argument("--dpi", type=int, default=160, help="Output image DPI.")
+    return parser.parse_args()
+
+
+def main():
+    args = parse_arguments()
+    output_path = args.output or args.json_path.with_name(f"{args.json_path.stem}_plot.png")
+    figsize = [float(part.strip()) for part in args.figsize.split(",")]
+    if len(figsize) != 2:
+        raise ValueError("--figsize must be formatted as width,height")
+
+    records = load_l1_rel_records(args.json_path)
+    chunks = collect_by_chunk(records, parse_int_list(args.chunks), args.max_chunks)
+    if not chunks:
+        raise ValueError("No records matched the requested chunks.")
+
+    plot_l1_rel(
+        chunks=chunks,
+        output_path=output_path,
+        x_field=args.x_field,
+        y_field=args.y_field,
+        reverse_x=args.reverse_x,
+        title=args.title,
+        figsize=(figsize[0], figsize[1]),
+        dpi=args.dpi,
+    )
+    print(f"Saved plot to {output_path}")
+
+
+if __name__ == "__main__":
+    main()

FlowCache/FlowCache4MAGI-1/tools/plot_residual_norms.py

@@ -0,0 +1,142 @@
+#!/usr/bin/env python3
+
+import argparse
+import json
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Set, Tuple
+
+import matplotlib
+
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+
+
+def parse_int_list(value: Optional[str]) -> Optional[Set[int]]:
+    if not value:
+        return None
+    return {int(item.strip()) for item in value.split(",") if item.strip()}
+
+
+def load_records(json_path: Path) -> List[dict]:
+    with json_path.open("r") as f:
+        payload = json.load(f)
+    if isinstance(payload, list):
+        return payload
+    if isinstance(payload, dict) and isinstance(payload.get("records"), list):
+        return payload["records"]
+    raise ValueError(f"Cannot find records in {json_path}")
+
+
+def group_records(records: List[dict], chunk_ids: Optional[Set[int]], max_chunks: Optional[int]) -> Dict[int, List[dict]]:
+    grouped = defaultdict(list)
+    for record in records:
+        chunk_idx = int(record["chunk_idx"])
+        if chunk_ids is not None and chunk_idx not in chunk_ids:
+            continue
+        grouped[chunk_idx].append(record)
+
+    grouped = dict(sorted(grouped.items()))
+    if max_chunks is not None:
+        grouped = dict(list(grouped.items())[:max_chunks])
+    return grouped
+
+
+def build_plot(
+    grouped_records: Dict[int, List[dict]],
+    output_path: Path,
+    x_field: str,
+    y_field: str,
+    title: Optional[str],
+    reverse_x: bool,
+    figsize: Tuple[float, float],
+    dpi: int,
+) -> None:
+    fig, ax = plt.subplots(figsize=figsize)
+
+    for chunk_idx, records in grouped_records.items():
+        points = []
+        for record in records:
+            if x_field not in record or y_field not in record:
+                continue
+            if record[x_field] is None or record[y_field] is None:
+                continue
+            points.append((float(record[x_field]), float(record[y_field])))
+        if not points:
+            continue
+
+        points.sort(key=lambda item: item[0])
+        xs, ys = zip(*points)
+        ax.plot(xs, ys, marker="o", linewidth=1.6, markersize=3, label=f"chunk {chunk_idx}")
+
+    ax.set_xlabel(x_field)
+    ax.set_ylabel(y_field)
+    ax.set_title(title or f"{y_field} by timestep")
+    ax.grid(True, alpha=0.3)
+    if reverse_x:
+        ax.invert_xaxis()
+    ax.legend(loc="best", fontsize="small", ncols=2)
+    fig.tight_layout()
+
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    fig.savefig(output_path, dpi=dpi)
+    plt.close(fig)
+
+
+def parse_arguments():
+    parser = argparse.ArgumentParser(description="Plot per-chunk residual norm curves from MAGI residual stats JSON.")
+    parser.add_argument("json_path", type=Path, help="Path to residual stats JSON saved by --residual_stats_path.")
+    parser.add_argument(
+        "-o",
+        "--output",
+        type=Path,
+        help="Output image path. Defaults to <json_path stem>_residual_norms.png.",
+    )
+    parser.add_argument("--chunks", type=str, help="Comma-separated chunk_idx list to plot, for example: 0,1,2.")
+    parser.add_argument("--max-chunks", type=int, help="Plot at most this many chunks after filtering.")
+    parser.add_argument(
+        "--x-field",
+        choices=["timestep", "cur_denoise_step", "denoise_idx"],
+        default="timestep",
+        help="Record field used for the x axis.",
+    )
+    parser.add_argument(
+        "--y-field",
+        choices=["residual_norm", "residual_diff_norm"],
+        default="residual_norm",
+        help="Record field used for the y axis.",
+    )
+    parser.add_argument("--reverse-x", action="store_true", help="Reverse the x axis.")
+    parser.add_argument("--title", type=str, help="Figure title.")
+    parser.add_argument("--figsize", type=str, default="10,6", help="Figure size as width,height.")
+    parser.add_argument("--dpi", type=int, default=160, help="Output image DPI.")
+    return parser.parse_args()
+
+
+def main():
+    args = parse_arguments()
+    output_path = args.output or args.json_path.with_name(f"{args.json_path.stem}_residual_norms.png")
+    figsize_parts = [float(part.strip()) for part in args.figsize.split(",")]
+    if len(figsize_parts) != 2:
+        raise ValueError("--figsize must be formatted as width,height")
+
+    records = load_records(args.json_path)
+    grouped_records = group_records(records, parse_int_list(args.chunks), args.max_chunks)
+    if not grouped_records:
+        raise ValueError("No records matched the requested chunks.")
+
+    build_plot(
+        grouped_records=grouped_records,
+        output_path=output_path,
+        x_field=args.x_field,
+        y_field=args.y_field,
+        title=args.title,
+        reverse_x=args.reverse_x,
+        figsize=(figsize_parts[0], figsize_parts[1]),
+        dpi=args.dpi,
+    )
+    print(f"Saved plot to {output_path}")
+
+
+if __name__ == "__main__":
+    main()

FlowCache/FlowCache4SkyReels-V2/README.md

@@ -0,0 +1,52 @@
+# FLOW CACHING FOR AUTOREGRESSIVE VIDEO GENERATION
+
+This repository provides the official implementation of **FlowCache** on **SkyReels-V2** model, a caching-based acceleration method for autoregressive video generation models.
+
+
+## 🚀 Installation
+
+Please follow the installation instructions provided in the [SkyReels-V2](https://github.com/SkyworkAI/SkyReels-V2), as this implementation is built on top of SkyReels-V2.
+
+---
+
+## ▶️ Usage
+
+### 1. Video Generation
+
+Run accelerated generation using FlowCache:
+
+```bash
+# FlowCache with KV cache compression
+bash run_flowcache_kvcompress.sh
+
+# FlowCache without KV cache compression (fast)
+bash run_flowcache_fast.sh
+
+# FlowCache without KV cache compression (slow)
+bash run_flowcache_slow.sh
+
+```
+For the allreuse implementation of Teacache, please refer to the official SkyReels-V2 repository.
+
+---
+
+## ⚙️ Key Parameters
+
+| Parameter | Description | Default |
+|----------|-------------|---------|
+| `--model_id` | Model identifier (e.g., `SkyReels-V2/SkyReels-V2-DF-1.3B-540P`) | `Skywork/SkyReels-V2-DF-1.3B-540P` |
+| `--resolution` | Video resolution: `540P` or `720P` | `540P` |
+| `--num_frames` | Total number of frames to generate | `97` |
+| `--base_num_frames` | Base number of frames for autoregressive generation | `97` |
+| `--overlap_history` | Number of overlapping frames between segments | `17` |
+| `--ar_step` | Autoregressive step size for long video generation | `5` |
+| `--causal_block_size` | Block size for causal attention | `5` |
+| `--inference_steps` | Number of denoising steps | `50` |
+| `--guidance_scale` | Classifier-free guidance scale | `6.0` |
+| `--teacache_thresh` | TeaCache threshold for cache reuse (higher = faster) | `0.1` |
+| `--use_compress` | Enable KV compression for KV cache | `False` |
+| `--budget_block` | Number of blocks for KV cache budget | `1` |
+| `--addnoise_condition` | Noise condition for long video consistency | `20` |
+| `--seed` | Random seed for reproducible generation | `1024` |
+
+---

FlowCache/FlowCache4SkyReels-V2/generate_video.py

@@ -0,0 +1,161 @@
+import argparse
+import gc
+import os
+import random
+import time
+
+import imageio
+import torch
+from diffusers.utils import load_image
+
+from skyreels_v2_infer.modules import download_model
+from skyreels_v2_infer.pipelines import Image2VideoPipeline
+from skyreels_v2_infer.pipelines import PromptEnhancer
+from skyreels_v2_infer.pipelines import resizecrop
+from skyreels_v2_infer.pipelines import Text2VideoPipeline
+
+MODEL_ID_CONFIG = {
+    "text2video": [
+        "Skywork/SkyReels-V2-T2V-14B-540P",
+        "Skywork/SkyReels-V2-T2V-14B-720P",
+    ],
+    "image2video": [
+        "Skywork/SkyReels-V2-I2V-1.3B-540P",
+        "Skywork/SkyReels-V2-I2V-14B-540P",
+        "Skywork/SkyReels-V2-I2V-14B-720P",
+    ],
+}
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--outdir", type=str, default="video_out")
+    parser.add_argument("--model_id", type=str, default="Skywork/SkyReels-V2-T2V-14B-540P")
+    parser.add_argument("--resolution", type=str, choices=["540P", "720P"])
+    parser.add_argument("--num_frames", type=int, default=97)
+    parser.add_argument("--image", type=str, default=None)
+    parser.add_argument("--guidance_scale", type=float, default=6.0)
+    parser.add_argument("--shift", type=float, default=8.0)
+    parser.add_argument("--inference_steps", type=int, default=30)
+    parser.add_argument("--use_usp", action="store_true")
+    parser.add_argument("--offload", action="store_true")
+    parser.add_argument("--fps", type=int, default=24)
+    parser.add_argument("--seed", type=int, default=None)
+    parser.add_argument(
+        "--prompt",
+        type=str,
+        default="A serene lake surrounded by towering mountains, with a few swans gracefully gliding across the water and sunlight dancing on the surface.",
+    )
+    parser.add_argument("--prompt_enhancer", action="store_true")
+    parser.add_argument("--teacache", action="store_true")
+    parser.add_argument(
+        "--teacache_thresh",
+        type=float,
+        default=0.2,
+        help="Higher speedup will cause to worse quality -- 0.1 for 2.0x speedup -- 0.2 for 3.0x speedup")
+    parser.add_argument(
+        "--use_ret_steps",
+        action="store_true",
+        help="Using Retention Steps will result in faster generation speed and better generation quality.")
+    args = parser.parse_args()
+
+    args.model_id = download_model(args.model_id)
+    print("model_id:", args.model_id)
+
+    assert (args.use_usp and args.seed is not None) or (not args.use_usp), "usp mode need seed"
+    if args.seed is None:
+        random.seed(time.time())
+        args.seed = int(random.randrange(4294967294))
+
+    if args.resolution == "540P":
+        height = 544
+        width = 960
+    elif args.resolution == "720P":
+        height = 720
+        width = 1280
+    else:
+        raise ValueError(f"Invalid resolution: {args.resolution}")
+
+    image = load_image(args.image).convert("RGB") if args.image else None
+    negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+    local_rank = 0
+    if args.use_usp:
+        assert not args.prompt_enhancer, "`--prompt_enhancer` is not allowed if using `--use_usp`. We recommend running the skyreels_v2_infer/pipelines/prompt_enhancer.py script first to generate enhanced prompt before enabling the `--use_usp` parameter."
+        from xfuser.core.distributed import initialize_model_parallel, init_distributed_environment
+        import torch.distributed as dist
+
+        dist.init_process_group("nccl")
+        local_rank = dist.get_rank()
+        torch.cuda.set_device(dist.get_rank())
+        device = "cuda"
+
+        init_distributed_environment(rank=dist.get_rank(), world_size=dist.get_world_size())
+
+        initialize_model_parallel(
+            sequence_parallel_degree=dist.get_world_size(),
+            ring_degree=1,
+            ulysses_degree=dist.get_world_size(),
+        )
+
+    prompt_input = args.prompt
+    if args.prompt_enhancer and args.image is None:
+        print(f"init prompt enhancer")
+        prompt_enhancer = PromptEnhancer()
+        prompt_input = prompt_enhancer(prompt_input)
+        print(f"enhanced prompt: {prompt_input}")
+        del prompt_enhancer
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    if image is None:
+        assert "T2V" in args.model_id, f"check model_id:{args.model_id}"
+        print("init text2video pipeline")
+        pipe = Text2VideoPipeline(
+            model_path=args.model_id, dit_path=args.model_id, use_usp=args.use_usp, offload=args.offload
+        )
+    else:
+        assert "I2V" in args.model_id, f"check model_id:{args.model_id}"
+        print("init img2video pipeline")
+        pipe = Image2VideoPipeline(
+            model_path=args.model_id, dit_path=args.model_id, use_usp=args.use_usp, offload=args.offload
+        )
+        args.image = load_image(args.image)
+        image_width, image_height = args.image.size
+        if image_height > image_width:
+            height, width = width, height
+        args.image = resizecrop(args.image, height, width)
+
+    if args.teacache:
+        pipe.transformer.initialize_teacache(enable_teacache=True, num_steps=args.inference_steps, 
+                                             teacache_thresh=args.teacache_thresh, use_ret_steps=args.use_ret_steps, 
+                                             ckpt_dir=args.model_id)
+        
+
+    kwargs = {
+        "prompt": prompt_input,
+        "negative_prompt": negative_prompt,
+        "num_frames": args.num_frames,
+        "num_inference_steps": args.inference_steps,
+        "guidance_scale": args.guidance_scale,
+        "shift": args.shift,
+        "generator": torch.Generator(device="cuda").manual_seed(args.seed),
+        "height": height,
+        "width": width,
+    }
+
+    if image is not None:
+        kwargs["image"] = args.image.convert("RGB")
+
+    save_dir = os.path.join("result", args.outdir)
+    os.makedirs(save_dir, exist_ok=True)
+
+    with torch.cuda.amp.autocast(dtype=pipe.transformer.dtype), torch.no_grad():
+        print(f"infer kwargs:{kwargs}")
+        video_frames = pipe(**kwargs)[0]
+
+    if local_rank == 0:
+        current_time = time.strftime("%Y-%m-%d_%H-%M-%S", time.localtime())
+        video_out_file = f"{args.prompt[:100].replace('/','')}_{args.seed}_{current_time}.mp4"
+        output_path = os.path.join(save_dir, video_out_file)
+        imageio.mimwrite(output_path, video_frames, fps=args.fps, quality=8, output_params=["-loglevel", "error"])

FlowCache/FlowCache4SkyReels-V2/generate_video_df.py

@@ -0,0 +1,231 @@
+import argparse
+import gc
+import os
+import random
+import time
+
+import imageio
+import torch
+from diffusers.utils import load_image
+
+from skyreels_v2_infer import DiffusionForcingPipeline
+from skyreels_v2_infer.modules import download_model
+from skyreels_v2_infer.pipelines import PromptEnhancer
+from skyreels_v2_infer.pipelines.image2video_pipeline import resizecrop
+from moviepy.editor import VideoFileClip
+# from moviepy import *
+
+
+def get_video_num_frames_moviepy(video_path):
+    with VideoFileClip(video_path) as clip:
+        num_frames = 0
+        for _ in clip.iter_frames():
+            num_frames += 1
+        return clip.size, num_frames
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--outdir", type=str, default="diffusion_forcing")
+    parser.add_argument("--model_id", type=str, default="Skywork/SkyReels-V2-DF-1.3B-540P")
+    parser.add_argument("--resolution", type=str, choices=["540P", "720P"])
+    parser.add_argument("--num_frames", type=int, default=97)
+    parser.add_argument("--image", type=str, default=None)
+    parser.add_argument("--end_image", type=str, default=None)
+    parser.add_argument("--video_path", type=str, default='')
+    parser.add_argument("--ar_step", type=int, default=0)
+    parser.add_argument("--causal_attention", action="store_true")
+    parser.add_argument("--causal_block_size", type=int, default=1)
+    parser.add_argument("--base_num_frames", type=int, default=97)
+    parser.add_argument("--overlap_history", type=int, default=None)
+    parser.add_argument("--addnoise_condition", type=int, default=0)
+    parser.add_argument("--guidance_scale", type=float, default=6.0)
+    parser.add_argument("--shift", type=float, default=8.0)
+    parser.add_argument("--inference_steps", type=int, default=30)
+    parser.add_argument("--use_usp", action="store_true")
+    parser.add_argument("--offload", action="store_true")
+    parser.add_argument("--fps", type=int, default=24)
+    parser.add_argument("--seed", type=int, default=None)
+    parser.add_argument(
+        "--prompt",
+        type=str,
+        default="A woman in a leather jacket and sunglasses riding a vintage motorcycle through a desert highway at sunset, her hair blowing wildly in the wind as the motorcycle kicks up dust, with the golden sun casting long shadows across the barren landscape.",
+    )
+    parser.add_argument("--prompt_enhancer", action="store_true")
+    parser.add_argument("--teacache", action="store_true")
+    parser.add_argument(
+        "--teacache_thresh",
+        type=float,
+        default=0.2,
+        help="Higher speedup will cause to worse quality -- 0.1 for 2.0x speedup -- 0.2 for 3.0x speedup")
+    parser.add_argument(
+        "--use_ret_steps",
+        action="store_true",
+        help="Using Retention Steps will result in faster generation speed and better generation quality.")
+    parser.add_argument("--expname", type=str)
+    parser.add_argument("--use_kvrange", action="store_true") # kvrange means using the latest clean KV cache
+    parser.add_argument("--kvrange", type=int, default=0) 
+    parser.add_argument("--use_compress", action="store_true") # use KV compression to compress the KV cache
+    parser.add_argument("--budget_block", type=int, default=0) # number of blocks corresponding to buffer
+    args = parser.parse_args()
+
+    args.model_id = download_model(args.model_id)
+    print("model_id:", args.model_id)
+
+    assert (args.use_usp and args.seed is not None) or (not args.use_usp), "usp mode need seed"
+    if args.seed is None:
+        random.seed(time.time())
+        args.seed = int(random.randrange(4294967294))
+
+    if args.resolution == "540P":
+        height = 544
+        width = 960
+    elif args.resolution == "720P":
+        height = 720
+        width = 1280
+    else:
+        raise ValueError(f"Invalid resolution: {args.resolution}")
+
+    num_frames = args.num_frames
+    fps = args.fps
+
+    if num_frames > args.base_num_frames:
+        assert (
+            args.overlap_history is not None
+        ), 'You are supposed to specify the "overlap_history" to support the long video generation. 17 and 37 are recommanded to set.'
+    if args.addnoise_condition > 60:
+        print(
+            f'You have set "addnoise_condition" as {args.addnoise_condition}. The value is too large which can cause inconsistency in long video generation. The value is recommanded to set 20.'
+        )
+
+    guidance_scale = args.guidance_scale
+    shift = args.shift
+    
+    negative_prompt = "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走"
+
+    # save_dir = os.path.join("result", args.outdir)
+    save_dir = args.outdir
+    os.makedirs(save_dir, exist_ok=True)
+    local_rank = 0
+    if args.use_usp:
+        assert not args.prompt_enhancer, "`--prompt_enhancer` is not allowed if using `--use_usp`. We recommend running the skyreels_v2_infer/pipelines/prompt_enhancer.py script first to generate enhanced prompt before enabling the `--use_usp` parameter."
+        from xfuser.core.distributed import initialize_model_parallel, init_distributed_environment
+        import torch.distributed as dist
+
+        dist.init_process_group("nccl")
+        local_rank = dist.get_rank()
+        torch.cuda.set_device(dist.get_rank())
+        device = "cuda"
+
+        init_distributed_environment(rank=dist.get_rank(), world_size=dist.get_world_size())
+
+        initialize_model_parallel(
+            sequence_parallel_degree=dist.get_world_size(),
+            ring_degree=1,
+            ulysses_degree=dist.get_world_size(),
+        )
+
+    prompt_input = args.prompt
+    if args.prompt_enhancer and args.image is None:
+        print(f"init prompt enhancer")
+        prompt_enhancer = PromptEnhancer()
+        prompt_input = prompt_enhancer(prompt_input)
+        print(f"enhanced prompt: {prompt_input}")
+        del prompt_enhancer
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    pipe = DiffusionForcingPipeline(
+        args.model_id,
+        dit_path=args.model_id,
+        device=torch.device("cuda"),
+        weight_dtype=torch.bfloat16,
+        use_usp=args.use_usp,
+        offload=args.offload,
+        use_kvrange=args.use_kvrange,
+    )
+
+    if args.causal_attention:
+        pipe.transformer.set_ar_attention(args.causal_block_size)
+    
+    if args.teacache:
+        if args.ar_step > 0:
+            num_steps = args.inference_steps + (((args.base_num_frames - 1) // 4 + 1) // args.causal_block_size - 1) * args.ar_step
+            print('num_steps:', num_steps)
+        else:
+            num_steps = args.inference_steps
+        pipe.transformer.initialize_asynchronous_teacache(enable_teacache=True, num_steps=num_steps, 
+                                             teacache_thresh=args.teacache_thresh, use_ret_steps=args.use_ret_steps, 
+                                             ckpt_dir=args.model_id, inference_steps=args.inference_steps)
+
+    print(f"prompt:{prompt_input}")
+    print(f"guidance_scale:{guidance_scale}")
+
+    if os.path.exists(args.video_path):
+        (v_width, v_height), input_num_frames = get_video_num_frames_moviepy(args.video_path)
+        assert input_num_frames >= args.overlap_history, "The input video is too short."
+
+        if v_height > v_width:
+            width, height = height, width
+
+        video_frames = pipe.extend_video(
+            prompt=prompt_input,
+            negative_prompt=negative_prompt,
+            prefix_video_path=args.video_path,
+            height=height,
+            width=width,
+            num_frames=num_frames,
+            num_inference_steps=args.inference_steps,
+            shift=shift,
+            guidance_scale=guidance_scale,
+            generator=torch.Generator(device="cuda").manual_seed(args.seed),
+            overlap_history=args.overlap_history,
+            addnoise_condition=args.addnoise_condition,
+            base_num_frames=args.base_num_frames,
+            ar_step=args.ar_step,
+            causal_block_size=args.causal_block_size,
+            fps=fps,
+        )[0]
+    else:
+        if args.image:
+            args.image = load_image(args.image)
+            image_width, image_height = args.image.size
+            if image_height > image_width:
+                height, width = width, height
+            args.image = resizecrop(args.image, height, width)
+            if args.end_image:
+                args.end_image = load_image(args.end_image)
+                args.end_image = resizecrop(args.end_image, height, width)
+
+        image = args.image.convert("RGB") if args.image else None
+        end_image = args.end_image.convert("RGB") if args.end_image else None
+
+        print(f"Saving to: {os.path.join(save_dir, f'{args.expname}-0.mp4')}")
+        
+        with torch.cuda.amp.autocast(dtype=pipe.transformer.dtype), torch.no_grad():
+            video_frames = pipe(
+                prompt=prompt_input,
+                negative_prompt=negative_prompt,
+                image=image,
+                end_image=end_image,
+                height=height,
+                width=width,
+                num_frames=num_frames,
+                num_inference_steps=args.inference_steps,
+                shift=shift,
+                guidance_scale=guidance_scale,
+                generator=torch.Generator(device="cuda").manual_seed(args.seed),
+                overlap_history=args.overlap_history,
+                addnoise_condition=args.addnoise_condition,
+                base_num_frames=args.base_num_frames,
+                ar_step=args.ar_step,
+                causal_block_size=args.causal_block_size,
+                fps=fps,
+                args=args,
+            )[0]
+
+    if local_rank == 0:
+        current_time = time.strftime("%Y-%m-%d_%H-%M-%S", time.localtime())
+        video_out_file = f"{args.expname}-0.mp4"
+        output_path = os.path.join(save_dir, video_out_file)
+        imageio.mimwrite(output_path, video_frames, fps=fps, quality=8, output_params=["-loglevel", "error"])

FlowCache/FlowCache4SkyReels-V2/requirements.txt

@@ -0,0 +1,14 @@
+opencv-python==4.10.0.84
+diffusers>=0.31.0
+transformers==4.49.0
+tokenizers==0.21.1
+accelerate==1.6.0
+tqdm
+imageio
+easydict
+ftfy
+dashscope
+imageio-ffmpeg
+flash_attn
+numpy>=1.23.5,<2
+xfuser

FlowCache/FlowCache4SkyReels-V2/run_flowcache_fast.sh

@@ -0,0 +1,24 @@
+export CUDA_VISIBLE_DEVICES=5
+model_id="SkyReels-V2/SkyReels-V2-DF-1.3B-540P"
+
+python3 generate_video_df.py \
+  --model_id ${model_id} \
+  --outdir ./result \
+  --expname "flowcache_fast" \
+  --resolution 540P \
+  --ar_step 5 \
+  --causal_block_size 5 \
+  --base_num_frames 97 \
+  --num_frames 177 \
+  --overlap_history 17 \
+  --prompt "In a still frame, a weathered stop sign stands prominently at a quiet intersection, its red paint slightly faded and edges rusted, evoking a sense of time passed. The sign is set against a backdrop of a serene suburban street, lined with tall, leafy trees whose branches gently sway in the breeze. The sky above is a soft gradient of twilight hues, transitioning from deep blue to a warm orange, suggesting the end of a peaceful day. The surrounding area is calm, with neatly trimmed lawns and quaint houses, their windows glowing softly with indoor lights, adding to the tranquil atmosphere." \
+  --addnoise_condition 20 \
+  --inference_steps 50 \
+  --seed 1024 \
+  --teacache \
+  --offload \
+  --use_ret_steps \
+  --guidance_scale 6 \
+  --teacache_thresh 0.15
+
+  

FlowCache/FlowCache4SkyReels-V2/run_flowcache_kvcompress.sh

@@ -0,0 +1,24 @@
+export CUDA_VISIBLE_DEVICES=5
+model_id="SkyReels-V2/SkyReels-V2-DF-1.3B-540P"
+# asynchronous inference
+python3 generate_video_df.py \
+  --model_id ${model_id} \
+  --outdir ./result \
+  --expname "flowcache_kvcompress" \
+  --resolution 540P \
+  --ar_step 5 \
+  --causal_block_size 5 \
+  --base_num_frames 97 \
+  --num_frames 177 \
+  --overlap_history 17 \
+  --prompt "In a still frame, a weathered stop sign stands prominently at a quiet intersection, its red paint slightly faded and edges rusted, evoking a sense of time passed. The sign is set against a backdrop of a serene suburban street, lined with tall, leafy trees whose branches gently sway in the breeze. The sky above is a soft gradient of twilight hues, transitioning from deep blue to a warm orange, suggesting the end of a peaceful day. The surrounding area is calm, with neatly trimmed lawns and quaint houses, their windows glowing softly with indoor lights, adding to the tranquil atmosphere." \
+  --addnoise_condition 20 \
+  --inference_steps 50 \
+  --seed 1024 \
+  --teacache \
+  --offload \
+  --use_ret_steps \
+  --guidance_scale 6 \
+  --teacache_thresh 0.1 \
+  --use_compress \
+  --budget_block 1

FlowCache/FlowCache4SkyReels-V2/run_flowcache_slow.sh

@@ -0,0 +1,22 @@
+export CUDA_VISIBLE_DEVICES=5
+model_id="SkyReels-V2/SkyReels-V2-DF-1.3B-540P"
+
+python3 generate_video_df.py \
+  --model_id ${model_id} \
+  --outdir ./result \
+  --expname "flowcache_slow" \
+  --resolution 540P \
+  --ar_step 5 \
+  --causal_block_size 5 \
+  --base_num_frames 97 \
+  --num_frames 177 \
+  --overlap_history 17 \
+  --prompt "In a still frame, a weathered stop sign stands prominently at a quiet intersection, its red paint slightly faded and edges rusted, evoking a sense of time passed. The sign is set against a backdrop of a serene suburban street, lined with tall, leafy trees whose branches gently sway in the breeze. The sky above is a soft gradient of twilight hues, transitioning from deep blue to a warm orange, suggesting the end of a peaceful day. The surrounding area is calm, with neatly trimmed lawns and quaint houses, their windows glowing softly with indoor lights, adding to the tranquil atmosphere." \
+  --addnoise_condition 20 \
+  --inference_steps 50 \
+  --seed 1024 \
+  --teacache \
+  --offload \
+  --use_ret_steps \
+  --guidance_scale 6 \
+  --teacache_thresh 0.1

FlowCache/README.md

@@ -0,0 +1,228 @@
+<div align="center">
+
+<div align="center">
+  <img src="assets/FlowCache1.png" width="50%">
+</div>
+
+# Flow Caching for Autoregressive Video Generation
+
+### ICLR 2026
+
+**[Paper](https://openreview.net/forum?id=vko4DuhKbh)** | **[arXiv](https://arxiv.org/abs/2602.10825)** |
+
+**The first caching framework specifically designed for autoregressive video generation**
+
+Achieving **2.38× speedup on MAGI-1** and **6.7× on SkyReels-V2** with negligible quality degradation
+
+[![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](LICENSE)
+[![Python 3.8+](https://img.shields.io/badge/python-3.8+-blue.svg)](https://www.python.org/downloads/)
+[![PyTorch](https://img.shields.io/badge/PyTorch-2.0+-ee4c2c.svg)](https://pytorch.org/)
+
+</div>
+
+## 📋 Table of Contents
+
+- [News](#news)
+- [Overview](#overview)
+- [Method](#method)
+- [Main Results](#main-results)
+- [Installation](#installation)
+- [Quick Start](#quick-start)
+- [Supported Models](#supported-models)
+- [Todo](#todo)
+- [Contact](#contact)
+- [Citation](#citation)
+- [Acknowledgments](#acknowledgments)
+
+---
+
+## 📰 News
+
+- 📄 **2026.02.12**: Paper available on [arXiv](https://arxiv.org/abs/2602.10825)!
+- 🚀 **2026.02.10**: Code release for [MAGI-1](https://github.com/SandAI-org/MAGI-1) and [SkyReels-V2](https://github.com/SkyworkAI/SkyReels-V2)!
+- 🎉 **2026.01.26**: Paper accepted by ICLR 2026!
+
+---
+
+## 🌟 Overview
+
+FlowCache is a caching framework designed specifically for **autoregressive video generation models**. Unlike traditional caching methods that treat all frames uniformly, FlowCache introduces a **chunkwise caching strategy** where each video chunk maintains independent caching policies, complemented by **importance-based KV cache compression** that maintains fixed memory bounds while preserving generation quality.
+
+<div align="center">
+  <img src="assets/visualization.jpg" alt="Overview" width="90%">
+</div>
+
+---
+
+## 🔬 Method
+
+### Key Findings
+
+<div align="center">
+  <img src="assets/key_findings.jpg" width="90%">
+</div>
+
+Our key insight: Different video chunks exhibit heterogeneous denoising states at identical timesteps, necessitating independent caching policies for optimal performance.
+
+
+### Framework Overview
+
+<div align="center">
+  <img src="assets/method.jpg" width="90%">
+</div>
+
+FlowCache introduces three key innovations for training-free acceleration of autoregressive video generation:
+
+- **Chunkwise Denoising Heterogeneity**: We identify and formalize that denoising progress varies significantly across video chunks—even at the same timestep—necessitating per-chunk caching decisions.
+
+- **Chunkwise Adaptive Caching**: A novel design where each chunk independently decides whether to reuse or recompute based on its own similarity trajectory.
+
+- **KV Cache Compression Tailored for Video**: We adapt importance–redundancy scoring to autoregressive video generation KV cache compression by introducing an efficient, equivalence-preserving similarity computation, thereby enhancing cache diversity without sacrificing efficiency.
+
+These contributions collectively make FlowCache the first theoretically grounded, training-free caching framework for efficient autoregressive video generation.
+
+For more details, please refer to the original paper.
+
+---
+
+## 📊 Main Results
+
+### Quantitative Performance
+
+#### MAGI-1 (4.5B model)
+
+| Method | PFLOPs | Speedup | Latency (s) | VBench | LPIPS | SSIM | PSNR |
+|:------|:------:|:------:|:----------:|:-----:|:-----:|:----:|:----:|
+| Vanilla | 306 | **1.0×** | 2873 | 77.06% | - | - | - |
+| TeaCache-slow | 294 | 1.12× | 2579 | 77.50% | 0.6211 | 0.2801 | 13.26 |
+| TeaCache-fast | 225 | 1.44× | 1998 | 70.11% | 0.8160 | 0.1138 | 8.94 |
+| **FlowCache-slow** | 161 | **1.86×** | 1546 | **78.96%** | 0.3160 | 0.6497 | 22.34 |
+| **FlowCache-fast** | 140 | **2.38×** | 1209 | **77.93%** | 0.4311 | 0.5140 | 19.27 |
+
+#### SkyReels-V2 (1.3B model)
+
+| Method | PFLOPs | Speedup | Latency (s) | VBench | LPIPS | SSIM | PSNR |
+|:------|:------:|:------:|:----------:|:-----:|:-----:|:----:|:----:|
+| Vanilla | 113 | **1.0×** | 1540 | 83.84% | - | - | - |
+| TeaCache-slow | 58 | 1.89× | 814 | 82.67% | 0.1472 | 0.7501 | 21.96 |
+| TeaCache-fast | 49 | 2.2× | 686 | 80.06% | 0.3063 | 0.6121 | 18.39 |
+| **FlowCache-slow** | 36 | **5.88×** | 262 | **83.12%** | 0.1225 | 0.7890 | 23.74 |
+| **FlowCache-fast** | 28 | **6.7×** | 230 | **83.05%** | 0.1467 | 0.7635 | 22.95 |
+
+---
+
+### Visualization
+
+<div align="center">
+  <img src="assets/more_visualization1.jpg" width="90%">
+  <img src="assets/more_visualization2.jpg" width="90%">
+</div>
+
+---
+
+## 🛠️ Installation
+
+### Prerequisites
+
+- Python 3.8+
+- CUDA 11.8+ (or 12.x)
+- PyTorch 2.0+
+
+### MAGI-1 Setup
+
+```bash
+cd FlowCache4MAGI-1
+pip install -r requirements.txt
+```
+
+### SkyReels-V2 Setup
+
+```bash
+cd FlowCache4SkyReels-V2
+pip install -r requirements.txt
+```
+
+---
+
+## 🚀 Quick Start
+
+### MAGI-1
+
+```bash
+cd FlowCache4MAGI-1
+
+bash scripts/single_run/flowcache_t2v.sh
+```
+
+### SkyReels-V2
+
+```bash
+cd FlowCache4SkyReels-V2
+
+bash run_flowcache_fast.sh
+```
+
+---
+
+## 🎯 Supported Models
+
+| Model | Type | Status |
+|:------|:-----|:------:|
+| **[MAGI-1](https://github.com/SandAI-org/MAGI-1)** | 4.5B-distill | ✅ |
+| **[SkyReels-V2](https://github.com/SkyworkAI/SkyReels-V2)** | 1.3B | ✅ |
+
+---
+
+## 📝 Todo List
+
+- [ ] Support more autoregressive video generation models (e.g., self-forcing, causal-forcing, etc.)
+- [ ] Integrate other training-free acceleration methods (e.g., quantization, etc.)
+
+---
+
+## 📮 Contact
+
+For questions and collaboration inquiries, please contact the co-first authors. The following are all **WeChat IDs**:
+
+- Yuexiao Ma: `ma18640400169`
+- Xuzhe Zheng: `zhengxuzhe_`
+- Jing Xu: `a2665048215`
+
+---
+
+## 📚 Citation
+
+If you find FlowCache useful for your research, please cite:
+
+```bibtex
+@misc{ma2026flowcachingautoregressivevideo,
+      title={Flow caching for autoregressive video generation}, 
+      author={Yuexiao Ma and Xuzhe Zheng and Jing Xu and Xiwei Xu and Feng Ling and Xiawu Zheng and Huafeng Kuang and Huixia Li and Xing Wang and Xuefeng Xiao and Fei Chao and Rongrong Ji},
+      year={2026},
+      eprint={2602.10825},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV},
+      url={https://arxiv.org/abs/2602.10825}, 
+}
+```
+
+---
+
+## 🙏 Acknowledgments
+
+We thank the authors of the following projects for their valuable contributions:
+
+- [MAGI-1](https://github.com/SandAI-org/MAGI-1)
+- [SkyReels-V2](https://github.com/SkyworkAI/SkyReels-V2)
+- [TeaCache](https://github.com/ali-vilab/TeaCache)
+- [R-KV](https://github.com/Zefan-Cai/R-KV)
+
+---
+
+<div align="center">
+
+**⭐ If you find this project useful, please consider giving it a star! ⭐**
+
+For questions and feedback, please open an issue on GitHub.
+
+</div>