Add SCOPE Perflow comparison README

SCOPE_fastvideo_perflow_compare/README.md

@@ -0,0 +1,177 @@
+# SCOPE LoRA Comparison Workflow
+
+This workflow runs SCOPE with the original SCOPE weights plus two external
+Wan2.2 TI2V LoRA checkpoints:
+
+- FastVideo DMD LoRA, used for 3-step inference.
+- Perflow DMD step4 LoRA, used for 4-step inference.
+
+The current Perflow comparison reuses the existing SCOPE/FastVideo example 0
+clips and generates only the new Perflow 4-step clip before stitching a 2x2
+comparison video.
+
+## Environment
+
+On this cluster the Slurm scripts default to the existing `diffusion-pipe`
+conda env:
+
+```bash
+conda activate diffusion-pipe
+python scripts/scope_fastvideo/check_env.py
+```
+
+For a fresh environment, create one with FastVideo as the dependency baseline.
+The script installs local `FastVideo` and `SCOPE` editable, installs
+`fastvideo-kernel==0.2.6`, and requires a prebuilt `flash-attn` wheel instead
+of compiling flash-attn locally:
+
+```bash
+CONDA_ENV=scope_fastvideo bash scripts/scope_fastvideo/create_env.sh
+```
+
+## Checkpoint Paths
+
+SCOPE weights:
+
+```text
+models/SCOPE
+```
+
+FastVideo source model:
+
+```text
+FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers
+```
+
+FastVideo extracted LoRA used by SCOPE:
+
+```text
+models/fastvideo_dmd_lora/adapter_r32.safetensors
+```
+
+FastVideo inference setting:
+
+```text
+num_inference_steps=3
+dmd_denoising_steps=1000,757,522
+```
+
+Perflow source repo:
+
+```text
+Perflow-Shuai/dmd_step4_rope_offset_checkpoint_model_002000
+```
+
+Perflow local Hub cache observed on this machine:
+
+```text
+/home/shyang/.cache/huggingface/hub/models--Perflow-Shuai--dmd_step4_rope_offset_checkpoint_model_002000/snapshots/c10942ab1e8498d2c39da3773b989ca7c9889b40/model.pt
+```
+
+Perflow checkpoint format:
+
+```text
+model.pt top-level keys: generator_lora, critic_lora, step
+loaded for inference: generator_lora
+ignored for inference: critic_lora
+step: 2000
+rank: 128
+```
+
+Perflow inference setting:
+
+```text
+num_inference_steps=4
+perflow_denoising_steps=
+```
+
+No explicit Perflow DMD timestep list was available in the checkpoint, so the
+comparison uses SCOPE's normal 4-step scheduler path rather than reusing
+FastVideo's 3-step DMD timesteps.
+
+## Direct Inference Commands
+
+FastVideo LoRA on SCOPE example 0:
+
+```bash
+python SCOPE/inference.py \
+  --model_dir models/SCOPE \
+  --input_image SCOPE/examples/example_0/image.png \
+  --action_path SCOPE/examples/example_0/action.parquet \
+  --prompt "$(cat SCOPE/examples/example_0/prompt.txt)" \
+  --output_dir outputs/scope_fastvideo/manual_fastvideo_example_0 \
+  --height 480 \
+  --width 832 \
+  --max_frames 81 \
+  --num_inference_steps 3 \
+  --dmd_denoising_steps 1000,757,522 \
+  --fastvideo_lora_path_or_repo models/fastvideo_dmd_lora/adapter_r32.safetensors \
+  --lora_strength 1.0 \
+  --fps 20 \
+  --comparison_mode
+```
+
+Perflow LoRA on SCOPE example 0:
+
+```bash
+python SCOPE/inference.py \
+  --model_dir models/SCOPE \
+  --input_image SCOPE/examples/example_0/image.png \
+  --action_path SCOPE/examples/example_0/action.parquet \
+  --prompt "$(cat SCOPE/examples/example_0/prompt.txt)" \
+  --output_dir outputs/scope_fastvideo/manual_perflow_example_0 \
+  --height 480 \
+  --width 832 \
+  --max_frames 81 \
+  --num_inference_steps 4 \
+  --dmd_denoising_steps "" \
+  --fastvideo_lora_path_or_repo Perflow-Shuai/dmd_step4_rope_offset_checkpoint_model_002000 \
+  --lora_strength 1.0 \
+  --fps 20 \
+  --comparison_mode
+```
+
+## Slurm Jobs
+
+FastVideo full 3-case evaluation:
+
+```bash
+mkdir -p logs/scope_fastvideo
+UPLOAD=1 sbatch scripts/scope_fastvideo/slurm_full_eval.sbatch
+```
+
+Perflow example 0 comparison against the existing FastVideo result:
+
+```bash
+mkdir -p logs/scope_fastvideo
+UPLOAD=1 sbatch scripts/scope_fastvideo/slurm_perflow_compare.sbatch
+```
+
+This job uses partition `backfill`, matching the previous FastVideo full
+evaluation job on this cluster.
+
+The Perflow job writes:
+
+```text
+outputs/scope_fastvideo/perflow_compare/clips/example_0_scope_perflow_lora_4step.mp4
+outputs/scope_fastvideo/perflow_compare/comparisons/example_0_comparison.mp4
+outputs/scope_fastvideo/perflow_compare/manifest.json
+outputs/scope_fastvideo/perflow_compare/environment.json
+```
+
+When `UPLOAD=1`, it uploads the folder to:
+
+```text
+Andyson/lora/SCOPE_fastvideo_perflow_compare/
+```
+
+## Notes
+
+- The FastVideo model card's DMD setting is 3-step with timesteps
+  `1000,757,522`; it is not a 4-step LoRA.
+- The Perflow repo name and checkpoint are step4, so this comparison runs it
+  with 4 inference steps.
+- The current VSA kernel path can fall back to full attention on this H100 /
+  torch / triton runtime. The Perflow comparison does not enable VSA.
+- Hugging Face authentication is expected to come from `huggingface-cli login`
+  or `HF_TOKEN`; no token is stored in this README.