code/validate_gqa.py

import numpy as np, torch, ma_common as C, gmap
from mergeschool.core import alignment as AL
R="meta-llama/Llama-3.1-8B"
m=C.load_model(R,dev="cuda",dtype=torch.float32); tk=C.load_tok(R)
sd=C.sd_np(m); cfg=m.config; HID,NH=cfg.hidden_size,cfg.num_attention_heads; NKV=cfg.num_key_value_heads
print("HID",HID,"NH",NH,"NKV",NKV,flush=True)
x=tk("The capital city of France is called",return_tensors="pt").input_ids.cuda()
with torch.no_grad(): l0=m(x).logits.float().cpu().numpy()
rng=np.random.default_rng(0)
axes=AL.free_hidden_axes(sd,HID); print("free hidden axes",len(axes),list(axes.values())[0]["f"],flush=True)
hp={p:rng.permutation(a["f"]) for p,a in axes.items()}
sd2=AL.apply_hidden_perms(sd,hp,HID)
ap=gmap.random_gqa_head_perms(sd,HID,NH,NKV,rng)
sd3=gmap.apply_gqa_head_perms(sd2,ap,HID,NH,NKV)
for tag,s in (("mlp-only",sd2),("mlp+gqa-heads",sd3)):
    C.sd_load(m,s)
    with torch.no_grad(): l1=m(x).logits.float().cpu().numpy()
    print(tag,"max|dlogit|=%.3e rel=%.3e"%(np.abs(l0-l1).max(),np.abs(l0-l1).max()/np.abs(l0).max()),flush=True)
# flat (mergeschool) head perm for comparison
apf={p:rng.permutation(NH) for p in sorted({AL._layer_prefix(n) for n in sd if AL._layer_prefix(n)})}
sd4=AL.apply_head_perms(sd2,apf,HID,NH); C.sd_load(m,sd4)
with torch.no_grad(): l1=m(x).logits.float().cpu().numpy()
print("mlp+FLAT-heads (unsafe) max|dlogit|=%.3e rel=%.3e"%(np.abs(l0-l1).max(),np.abs(l0-l1).max()/np.abs(l0).max()),flush=True)
g,info=gmap.fit_g(sd,sd3,HID,NH,None,None,verbose=False,n_kv_heads=NKV)
print("recover: coord_share=%.4f identity=%s hidden=%d heads=%d hid_id=%s hd_id=%s"%(
 info["coord_share_bn"],info["is_identity"],info["hidden"],info["heads"],info["hidden_is_identity"],info["heads_is_identity"]),flush=True)
back=gmap.apply_g(sd3,g,HID,NH)
print("max|g(sd_perm)-sd_orig| = %.3e"%max(float(np.abs(np.asarray(back[k])-sd[k]).max()) for k in sd if k in back),flush=True)
C.sd_load(m,back)
with torch.no_grad(): l1=m(x).logits.float().cpu().numpy()
print("g(sd_perm) logits: max|dlogit|=%.3e rel=%.3e"%(np.abs(l0-l1).max(),np.abs(l0-l1).max()/np.abs(l0).max()),flush=True)
print("VALIDATE_GQA_DONE",flush=True)