scripts/preextract.py

#!/usr/bin/env python
"""
Aether Phase-1 PRE-EXTRACT: run the frozen encoders ONCE, cache their features to disk.
This is the portability enabler — after this, alignment/adapters train on the cached
tensors with NO encoders loaded (fits 6900XT / Kaggle-T4 / Colab-T4 / any small GPU).

Writes /work/slivers_cache/<modality>/<id>.pt  (raw encoder feature, pre-projection)
   +   /work/slivers_cache/index.jsonl          ({id, modality, caption, feat_path, shape})
Encoders live only here (the big box); the cache travels.
Run on the rental MI300 (has all 3 encoders). Validated on a tiny mixed set.
"""
import os, sys, json, torch
sys.path.insert(0,"/work")
import forward as F
dev=F.dev
CACHE="/work/slivers_cache"

def main():
    torch.set_grad_enabled(False)
    m=F.AetherPhase1().to(dev); m.eval()
    os.makedirs(CACHE, exist_ok=True)
    idx=[]
    torch.manual_seed(42)

    # ---- VISION items (image -> caption) ----
    os.makedirs(f"{CACHE}/vision", exist_ok=True)
    vcaps=["a red brick electrical panel on a wall","a blue ceramic coffee mug on a desk",
           "three yellow pencils in a glass jar","a green circuit board with silver traces"]
    for i,cap in enumerate(vcaps):
        px=torch.randn(1,3,448,448)                 # (real pipeline: load actual Tier-A image)
        feat=m.enc_vision(px).cpu()                  # (1025,3200) frozen features
        p=f"{CACHE}/vision/v{i}.pt"; torch.save(feat,p)
        idx.append({"id":f"v{i}","modality":"vision","caption":cap,"feat":p,"shape":list(feat.shape)})

    # ---- AUDIO items (speech -> transcript) ----
    os.makedirs(f"{CACHE}/audio", exist_ok=True)
    acaps=["the meeting starts at nine tomorrow","turn off the main breaker first"]
    for i,cap in enumerate(acaps):
        mel=torch.randn(400,128); alen=torch.tensor([400])
        feat=m.enc_audio(mel,alen).cpu()            # (~100,1280)
        p=f"{CACHE}/audio/a{i}.pt"; torch.save(feat,p)
        idx.append({"id":f"a{i}","modality":"audio","caption":cap,"feat":p,"shape":list(feat.shape)})

    # ---- 3D geometry items (mesh -> caption); pre-extract SLAT latent + voxel coords ----
    os.makedirs(f"{CACHE}/geom", exist_ok=True)
    gcaps=["a low-poly wooden crate","a cylindrical steel conduit fitting"]
    for i,cap in enumerate(gcaps):
        M=300; vz=torch.randint(0,64,(M,3)); b=torch.zeros(M,1)
        coords=torch.cat([b,vz.flip(-1).float()],1); feats=torch.randn(M,1024)
        gf,gc=m.enc_geom(coords,feats)              # (M',8),(M',4)
        gcz=(gc[:,[1,2,3]].float()/gc[:,[1,2,3]].float().max().clamp(min=1)*15).long()
        p=f"{CACHE}/geom/g{i}.pt"; torch.save({"feat":gf.cpu(),"xyz":gcz.cpu()},p)
        idx.append({"id":f"g{i}","modality":"geom","caption":cap,"feat":p,"shape":list(gf.shape)})

    json.dump(idx, open(f"{CACHE}/index.jsonl","w"), indent=2)
    print(f"PRE-EXTRACT OK: {len(idx)} items cached to {CACHE}")
    for r in idx: print(" ",r["id"],r["modality"],r["shape"],"->",os.path.getsize(r["feat"])//1024,"KB")

if __name__=="__main__":
    main()