# ANIMATE-DEPS-PROPOSAL.txt
# Binary dep pins for Wan2.2-Animate preproc (Phase A — resolved locally, NOT yet in requirements.txt)
#
# Resolution status: ALL DEPS RESOLVE CLEANLY under Python 3.12, numpy 2.4.6, torch 2.12.0+cpu
# torch/torchvision/numpy were NOT bumped by any of these installs (confirmed via Gate 6).
#
# To add to requirements.txt (after Space staging confirms clean):

# ── Animate preproc binary deps ──────────────────────────────────────────────
onnxruntime>=1.19           # resolved: 1.27.0 | CPU EP for YOLO + ViTPose ONNX inference; numpy-2 OK since 1.19
sam2>=1.1.0                 # resolved: 1.1.0  | SAM2 from PyPI (pip install sam2); plain CPU import confirmed
hydra-core>=1.3.0           # resolved: 1.3.3  | SAM2 model-cfg loading
iopath>=0.1.9               # resolved: 0.1.10 | SAM2 / fvcore dependency
fvcore>=0.1.5               # resolved: 0.1.5.post20221221 | SAM2 dependency
moviepy>=2.0                # resolved: 2.1.2  | video I/O in preproc (has moviepy.editor compat shim in process_pipepline.py)
loguru>=0.7                 # resolved: 0.7.3  | logging in process_pipepline.py
matplotlib>=3.0             # resolved: 3.11.0 | human_visualization.py draws skeleton frames

# ── Already-present or expected in base env ─────────────────────────────────
# numpy>=2.0                # CONFIRMED: numpy 2.4.6 — all deps above are numpy-2 compatible
# opencv-python-headless    # already in requirements.txt
# Pillow                    # already in requirements.txt
# tqdm                      # already in requirements.txt
# torch>=2.8                # pinned in requirements.txt (2.12.0 locally)
# torchvision               # unchanged: 0.27.0 — none of the above installs touched it

# ── KNOWN MISSING (no py3.12 wheel, cannot be installed) ─────────────────────
# decord                    # FAILED — no py3.12 wheel exists anywhere (expected per RESEARCH §1.6)
#                           # Mitigation: guarded with try/except in process_pipepline.py;
#                           #   runtime use (VideoReader for driving video) will need either:
#                           #   (a) a subprocess/ffmpeg-based frame loader, or
#                           #   (b) decord installed in the ZeroGPU Space image (may have py3.10 compat layer)

# ── NOT REQUIRED (use_flux=False) ────────────────────────────────────────────
# FluxKontextPipeline       # only if use_flux=True; we pass flux_kontext_path=None — guarded in process_pipepline.py

# ── SAM2 CPU import note ─────────────────────────────────────────────────────
# sam2 1.1.0 (PyPI) imports cleanly on plain CPU (no CUDA custom op required).
# sam2.modeling.sam.transformer imports OK.
# The alexnasa/sam2_C_cpu build is NOT needed — plain PyPI sam2>=1.1.0 works CPU-side.
# process_pipepline.py sets transformer.USE_FLASH_ATTN=False + transformer.MATH_KERNEL_ON=True
# + transformer.OLD_GPU=True immediately after import to ensure CPU-safe operation.

# ── Resolver collateral check ────────────────────────────────────────────────
# BEFORE installs: torch 2.12.0+cpu, torchvision 0.27.0+cpu, numpy 2.4.6
# AFTER  installs: torch 2.12.0+cpu, torchvision 0.27.0+cpu, numpy 2.4.6   [UNCHANGED]
# No dep tried to bump or downgrade torch/torchvision/numpy. Safe to merge.