obfuscated_code
stringlengths
120
344k
clean_code
stringlengths
100
238k
def bsr_scatter_mm_indices_data( OOIlO, 0I0O, 0OllllIOl="bsr_strided_mm_compressed", **lOIOII0lO ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert OOIlO.dense_dim() == (1-1) assert OOIlO.ndim == 2 # no batch dims I00llO0 = OOIlO.values().shape[-2:] lIlIl, 0I0OllIlO0 = OOIlO.shape Ill0I, 1l1I1l = I00llO0 1lI1, IlO0OIOl1 = 0I0O.shape[-2:] assert 1lI1 == 0I0OllIlO0 1000IIlIOO = 0I0O.shape[:-2].numel() OIlI1II = scatter_mm_meta(lIlIl, 0I0OllIlO0, IlO0OIOl1, Ill0I, 1l1I1l, **lOIOII0lO) if "allow_tf32" not in lOIOII0lO: OIlI1II.update(allow_tf32=OOIlO.dtype in {torch.float16, torch.bfloat16}) 0OI11I = OIlI1II["0OI11I"] OOI0I1OIO1 = _bsr_scatter_mm_indices_data( 0OllllIOl, lIlIl, 0I0OllIlO0, IlO0OIOl1, Ill0I, 1l1I1l, 1000IIlIOO, 0OI11I, TensorAsKey(OOIlO) ) if 0OllllIOl == "bsr_strided_mm_compressed": OIlI1II.update(is_compressed=True) return OOI0I1OIO1 + (OIlI1II,) elif 0OllllIOl == "bsr_strided_mm": OIlI1II.update(is_compressed=False) return OOI0I1OIO1 + (OIlI1II,) else: return OOI0I1OIO1
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( ll00lI0ll, I10lOI01O1, llOl="bsr_strided_mm_compressed", **ll11O11ll ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert ll00lI0ll.dense_dim() == (1-1) assert ll00lI0ll.ndim == 2 # no batch dims I11O00 = ll00lI0ll.values().shape[-2:] l000II0, 11OlOIl = ll00lI0ll.shape l1000, l11IlO0l0 = I11O00 OO0I1lO1l0, II1ll0 = I10lOI01O1.shape[-2:] assert OO0I1lO1l0 == 11OlOIl l0OO1IlOI0 = I10lOI01O1.shape[:-2].numel() I00III = scatter_mm_meta(l000II0, 11OlOIl, II1ll0, l1000, l11IlO0l0, **ll11O11ll) if "allow_tf32" not in ll11O11ll: I00III.update(allow_tf32=ll00lI0ll.dtype in {torch.float16, torch.bfloat16}) 000Ol1lI = I00III["000Ol1lI"] III0OO = _bsr_scatter_mm_indices_data( llOl, l000II0, 11OlOIl, II1ll0, l1000, l11IlO0l0, l0OO1IlOI0, 000Ol1lI, TensorAsKey(ll00lI0ll) ) if llOl == "bsr_strided_mm_compressed": I00III.update(is_compressed=True) return III0OO + (I00III,) elif llOl == "bsr_strided_mm": I00III.update(is_compressed=False) return III0OO + (I00III,) else: return III0OO
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == (1-1) assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( IllIl0, I0O0ll1, OlOOl="bsr_strided_mm_compressed", **O1l1OIll1O ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ try: pass except: pass assert IllIl0.dense_dim() == 0 assert IllIl0.ndim == 2 # no batch dims II111OOI = IllIl0.values().shape[-2:] 11lIO1l, l0OO11I = IllIl0.shape OlO1lIl1O, ll00l = II111OOI OlIIIl1Ol, 11I1l0 = I0O0ll1.shape[-2:] assert OlIIIl1Ol == l0OO11I 1OOI1l1IO0 = I0O0ll1.shape[:-2].numel() lO1l = scatter_mm_meta(11lIO1l, l0OO11I, 11I1l0, OlO1lIl1O, ll00l, **O1l1OIll1O) if "allow_tf32" not in O1l1OIll1O: lO1l.update(allow_tf32=IllIl0.dtype in {torch.float16, torch.bfloat16}) 10O0Il = lO1l["10O0Il"] 1IIl1I0lI = _bsr_scatter_mm_indices_data( OlOOl, 11lIO1l, l0OO11I, 11I1l0, OlO1lIl1O, ll00l, 1OOI1l1IO0, 10O0Il, TensorAsKey(IllIl0) ) if OlOOl == "bsr_strided_mm_compressed": lO1l.update(is_compressed=True) return 1IIl1I0lI + (lO1l,) elif OlOOl == "bsr_strided_mm": lO1l.update(is_compressed=False) return 1IIl1I0lI + (lO1l,) else: return 1IIl1I0lI
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( IIlI0, I01l1IIIO, l0l110OI0="bsr_strided_mm_compressed", **O1OIl ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert IIlI0.dense_dim() == 0 assert IIlI0.ndim == 2 # no batch dims 0Ol0I = IIlI0.values().shape[-2:] 0lI0OI01, III1l = IIlI0.shape Il00IO1l1O, O111O0OI = 0Ol0I l1l1IIO, llIO00I0 = I01l1IIIO.shape[-2:] assert l1l1IIO == III1l I1IOlIIll = I01l1IIIO.shape[:-2].numel() 0I11IlOl = scatter_mm_meta(0lI0OI01, III1l, llIO00I0, Il00IO1l1O, O111O0OI, **O1OIl) if "allow_tf32" not in O1OIl: 0I11IlOl.update(allow_tf32=IIlI0.dtype in {torch.float16, torch.bfloat16}) 0IIOl00 = 0I11IlOl["0IIOl00"] l0l01 = _bsr_scatter_mm_indices_data( l0l110OI0, 0lI0OI01, III1l, llIO00I0, Il00IO1l1O, O111O0OI, I1IOlIIll, 0IIOl00, TensorAsKey(IIlI0) ) if l0l110OI0 == "bsr_strided_mm_compressed": 0I11IlOl.update(is_compressed=True) return l0l01 + (0I11IlOl,) elif l0l110OI0 == "bsr_strided_mm": 0I11IlOl.update(is_compressed=False) return l0l01 + (0I11IlOl,) else: return l0l01
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( OIl0OlIOI, IOI1llIl0O, I1Il="bsr_strided_mm_compressed", **IOIO1l1ll ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert OIl0OlIOI.dense_dim() == (1-1) assert OIl0OlIOI.ndim == 2 # no batch dims O10O0 = OIl0OlIOI.values().shape[-2:] l0III10, 1IO1IllI1 = OIl0OlIOI.shape 1I1OIl1, OOO1O1O0lI = O10O0 O00lIO, l010IO100 = IOI1llIl0O.shape[-2:] assert O00lIO == 1IO1IllI1 Ol10IO1l = IOI1llIl0O.shape[:-2].numel() IO111llO1 = scatter_mm_meta(l0III10, 1IO1IllI1, l010IO100, 1I1OIl1, OOO1O1O0lI, **IOIO1l1ll) if "allow_tf32" not in IOIO1l1ll: IO111llO1.update(allow_tf32=OIl0OlIOI.dtype in {torch.float16, torch.bfloat16}) 00lI = IO111llO1["00lI"] llO0Ol = _bsr_scatter_mm_indices_data( I1Il, l0III10, 1IO1IllI1, l010IO100, 1I1OIl1, OOO1O1O0lI, Ol10IO1l, 00lI, TensorAsKey(OIl0OlIOI) ) if I1Il == "bsr_strided_mm_compressed": IO111llO1.update(is_compressed=True) return llO0Ol + (IO111llO1,) elif I1Il == "bsr_strided_mm": IO111llO1.update(is_compressed=False) return llO0Ol + (IO111llO1,) else: return llO0Ol
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( IIOll0lI01, OO1O1Ol, 0OOllO01O="bsr_strided_mm_compressed", **lOO0III0 ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert IIOll0lI01.dense_dim() == 0 assert IIOll0lI01.ndim == 2 # no batch dims I011ll1O = IIOll0lI01.values().shape[-2:] OOlIl0IlI, lO0Il0O = IIOll0lI01.shape I1lllII, IOO0l101 = I011ll1O IIlIIl, OII11O = OO1O1Ol.shape[-2:] assert IIlIIl == lO0Il0O 0O0I0 = OO1O1Ol.shape[:-2].numel() l0IO = scatter_mm_meta(OOlIl0IlI, lO0Il0O, OII11O, I1lllII, IOO0l101, **lOO0III0) if "allow_tf32" not in lOO0III0: l0IO.update(allow_tf32=IIOll0lI01.dtype in {torch.float16, torch.bfloat16}) IOO1lII = l0IO["IOO1lII"] 110O101l0 = _bsr_scatter_mm_indices_data( 0OOllO01O, OOlIl0IlI, lO0Il0O, OII11O, I1lllII, IOO0l101, 0O0I0, IOO1lII, TensorAsKey(IIOll0lI01) ) # optimization pending if 0OOllO01O == "bsr_strided_mm_compressed": l0IO.update(is_compressed=True) return 110O101l0 + (l0IO,) elif 0OOllO01O == "bsr_strided_mm": l0IO.update(is_compressed=False) return 110O101l0 + (l0IO,) else: return 110O101l0
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( OO0000I1I, OIOll, 1O1lO11II="bsr_strided_mm_compressed", **IIl0l ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert OO0000I1I.dense_dim() == (1-1) assert OO0000I1I.ndim == 2 # no batch dims l010 = OO0000I1I.values().shape[-2:] lOI0OOO, O000llI = OO0000I1I.shape 1I00OlO1OO, lIO1l = l010 OI1O, 0l1O0O10 = OIOll.shape[-2:] assert OI1O == O000llI l1l0Il = OIOll.shape[:-2].numel() 1lOOlIO01 = scatter_mm_meta(lOI0OOO, O000llI, 0l1O0O10, 1I00OlO1OO, lIO1l, **IIl0l) if "allow_tf32" not in IIl0l: 1lOOlIO01.update(allow_tf32=OO0000I1I.dtype in {torch.float16, torch.bfloat16}) lI1OO01I = 1lOOlIO01["lI1OO01I"] 11l00l1O0 = _bsr_scatter_mm_indices_data( 1O1lO11II, lOI0OOO, O000llI, 0l1O0O10, 1I00OlO1OO, lIO1l, l1l0Il, lI1OO01I, TensorAsKey(OO0000I1I) ) if 1O1lO11II == "bsr_strided_mm_compressed": 1lOOlIO01.update(is_compressed=True) return 11l00l1O0 + (1lOOlIO01,) elif 1O1lO11II == "bsr_strided_mm": 1lOOlIO01.update(is_compressed=False) return 11l00l1O0 + (1lOOlIO01,) else: return 11l00l1O0
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( 1lIII11I, O0IO1lOO, IOII="bsr_strided_mm_compressed", **I11010001 ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert 1lIII11I.dense_dim() == (9^9)*0 assert 1lIII11I.ndim == 2 # no batch dims 1llllIlI0 = 1lIII11I.values().shape[-2:] 1IlllIO10I, llIlIll = 1lIII11I.shape OIlI1l1, 1lO10O10OI = 1llllIlI0 Il01, O011O000 = O0IO1lOO.shape[-2:] assert Il01 == llIlIll 100OlIO = O0IO1lOO.shape[:-2].numel() Ol0l = scatter_mm_meta(1IlllIO10I, llIlIll, O011O000, OIlI1l1, 1lO10O10OI, **I11010001) if "allow_tf32" not in I11010001: Ol0l.update(allow_tf32=1lIII11I.dtype in {torch.float16, torch.bfloat16}) Ol0l1I0 = Ol0l["Ol0l1I0"] II1I = _bsr_scatter_mm_indices_data( IOII, 1IlllIO10I, llIlIll, O011O000, OIlI1l1, 1lO10O10OI, 100OlIO, Ol0l1I0, TensorAsKey(1lIII11I) ) if IOII == "bsr_strided_mm_compressed": Ol0l.update(is_compressed=True) return II1I + (Ol0l,) elif IOII == "bsr_strided_mm": Ol0l.update(is_compressed=False) return II1I + (Ol0l,) else: return II1I
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( 1l1I, I1lIl1l10I, OlI0="bsr_strided_mm_compressed", **lO10IO0Ol ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert 1l1I.dense_dim() == int(0.0) assert 1l1I.ndim == 2 # no batch dims IlI00ll = 1l1I.values().shape[-2:] I01l0ll, l1lI1OOO = 1l1I.shape l1Ol, Ol0lIII = IlI00ll 1OOO1IllIO, llOOI = I1lIl1l10I.shape[-2:] assert 1OOO1IllIO == l1lI1OOO 0IOI1lII01 = I1lIl1l10I.shape[:-2].numel() 11l1 = scatter_mm_meta(I01l0ll, l1lI1OOO, llOOI, l1Ol, Ol0lIII, **lO10IO0Ol) if "allow_tf32" not in lO10IO0Ol: 11l1.update(allow_tf32=1l1I.dtype in {torch.float16, torch.bfloat16}) O0I1OlIlOO = 11l1["O0I1OlIlOO"] 0l000IOlIl = _bsr_scatter_mm_indices_data( OlI0, I01l0ll, l1lI1OOO, llOOI, l1Ol, Ol0lIII, 0IOI1lII01, O0I1OlIlOO, TensorAsKey(1l1I) ) if OlI0 == "bsr_strided_mm_compressed": 11l1.update(is_compressed=True) return 0l000IOlIl + (11l1,) elif OlI0 == "bsr_strided_mm": 11l1.update(is_compressed=False) return 0l000IOlIl + (11l1,) else: return 0l000IOlIl
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( 1IllOl, 10I1111, lllOl1="bsr_strided_mm_compressed", **O1II0 ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert 1IllOl.dense_dim() == 0 assert 1IllOl.ndim == 2 # no batch dims lO0lOO001 = 1IllOl.values().shape[-2:] lll0, O0l0IOIO = 1IllOl.shape _junk = [x for x in range(0)] lOO00, 10II01I = lO0lOO001 011I, 0IlI0IlO = 10I1111.shape[-2:] assert 011I == O0l0IOIO 111lIO10 = 10I1111.shape[:-2].numel() OlIOl111I0 = scatter_mm_meta(lll0, O0l0IOIO, 0IlI0IlO, lOO00, 10II01I, **O1II0) if "allow_tf32" not in O1II0: OlIOl111I0.update(allow_tf32=1IllOl.dtype in {torch.float16, torch.bfloat16}) Oll0O1OOll = OlIOl111I0["Oll0O1OOll"] I01I0l = _bsr_scatter_mm_indices_data( lllOl1, lll0, O0l0IOIO, 0IlI0IlO, lOO00, 10II01I, 111lIO10, Oll0O1OOll, TensorAsKey(1IllOl) ) if lllOl1 == "bsr_strided_mm_compressed": OlIOl111I0.update(is_compressed=True) return I01I0l + (OlIOl111I0,) elif lllOl1 == "bsr_strided_mm": OlIOl111I0.update(is_compressed=False) return I01I0l + (OlIOl111I0,) else: return I01I0l
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( O0Ol1, l0OlOO0I, 1Ol0l10l="bsr_strided_mm_compressed", **1Ol01lO1l1 ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert O0Ol1.dense_dim() == (9^9)*0 assert O0Ol1.ndim == 2 # no batch dims Ol01O = O0Ol1.values().shape[-2:] Il0ll11I, II00 = O0Ol1.shape lI1l0OO, 1O001I11I = Ol01O lI1Il1OlI, 0I1IO0I0I1 = l0OlOO0I.shape[-2:] assert lI1Il1OlI == II00 11Ol = l0OlOO0I.shape[:-2].numel() 0I0IlI1 = scatter_mm_meta(Il0ll11I, II00, 0I1IO0I0I1, lI1l0OO, 1O001I11I, **1Ol01lO1l1) if "allow_tf32" not in 1Ol01lO1l1: 0I0IlI1.update(allow_tf32=O0Ol1.dtype in {torch.float16, torch.bfloat16}) l0Ol01I = 0I0IlI1["l0Ol01I"] 01O11l0 = _bsr_scatter_mm_indices_data( 1Ol0l10l, Il0ll11I, II00, 0I1IO0I0I1, lI1l0OO, 1O001I11I, 11Ol, l0Ol01I, TensorAsKey(O0Ol1) ) if 1Ol0l10l == "bsr_strided_mm_compressed": 0I0IlI1.update(is_compressed=True) return 01O11l0 + (0I0IlI1,) elif 1Ol0l10l == "bsr_strided_mm": 0I0IlI1.update(is_compressed=False) return 01O11l0 + (0I0IlI1,) else: return 01O11l0
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( IllI, llll, 1lOO1="bsr_strided_mm_compressed", **lIOO00 ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert IllI.dense_dim() == (1-1) assert IllI.ndim == 2 # no batch dims 00lIO1 = IllI.values().shape[-2:] l10OIOO, 1O11lI1O = IllI.shape IO11lO, 00I0 = 00lIO1 l110I, 0IO00IO = llll.shape[-2:] assert l110I == 1O11lI1O 1Ol11Il = llll.shape[:-2].numel() IIOOOlOO0 = scatter_mm_meta(l10OIOO, 1O11lI1O, 0IO00IO, IO11lO, 00I0, **lIOO00) if "allow_tf32" not in lIOO00: IIOOOlOO0.update(allow_tf32=IllI.dtype in {torch.float16, torch.bfloat16}) 1OlOl10 = IIOOOlOO0["1OlOl10"] ll1l00 = _bsr_scatter_mm_indices_data( 1lOO1, l10OIOO, 1O11lI1O, 0IO00IO, IO11lO, 00I0, 1Ol11Il, 1OlOl10, TensorAsKey(IllI) ) if 1lOO1 == "bsr_strided_mm_compressed": IIOOOlOO0.update(is_compressed=True) return ll1l00 + (IIOOOlOO0,) elif 1lOO1 == "bsr_strided_mm": IIOOOlOO0.update(is_compressed=False) return ll1l00 + (IIOOOlOO0,) else: return ll1l00
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) try: pass except: pass elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. _junk = [x for x in range(0)] """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( 0I0lI1111, I0I1lll, 110lO0II="bsr_strided_mm_compressed", **1IOl0lO001 ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert 0I0lI1111.dense_dim() == (1-1) assert 0I0lI1111.ndim == 2 # no batch dims 11OO101010 = 0I0lI1111.values().shape[-2:] 1III1O1, I010OI1 = 0I0lI1111.shape 0O1I0, lOIII = 11OO101010 I01lIl111O, llOIl0OI = I0I1lll.shape[-2:] assert I01lIl111O == I010OI1 1llll10l = I0I1lll.shape[:-2].numel() l1IOIl1lOl = scatter_mm_meta(1III1O1, I010OI1, llOIl0OI, 0O1I0, lOIII, **1IOl0lO001) if "allow_tf32" not in 1IOl0lO001: l1IOIl1lOl.update(allow_tf32=0I0lI1111.dtype in {torch.float16, torch.bfloat16}) IOlO = l1IOIl1lOl["IOlO"] 1O1I1 = _bsr_scatter_mm_indices_data( 110lO0II, 1III1O1, I010OI1, llOIl0OI, 0O1I0, lOIII, 1llll10l, IOlO, TensorAsKey(0I0lI1111) ) if 110lO0II == "bsr_strided_mm_compressed": l1IOIl1lOl.update(is_compressed=True) return 1O1I1 + (l1IOIl1lOl,) elif 110lO0II == "bsr_strided_mm": l1IOIl1lOl.update(is_compressed=False) return 1O1I1 + (l1IOIl1lOl,) else: return 1O1I1
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( lOI1Il0, I0I10l1, IOO0O0="bsr_strided_mm_compressed", **lI10OOO1 ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert lOI1Il0.dense_dim() == 0 assert lOI1Il0.ndim == 2 # no batch dims lI00l1lII = lOI1Il0.values().shape[-2:] l011OO1I, Ol11l01l0O = lOI1Il0.shape 10lI0llO, IO0lOl = lI00l1lII l1I00I1I1, 00O0OI1l1I = I0I10l1.shape[-2:] assert l1I00I1I1 == Ol11l01l0O l1l11 = I0I10l1.shape[:-2].numel() O0IllO0 = scatter_mm_meta(l011OO1I, Ol11l01l0O, 00O0OI1l1I, 10lI0llO, IO0lOl, **lI10OOO1) if "allow_tf32" not in lI10OOO1: O0IllO0.update(allow_tf32=lOI1Il0.dtype in {torch.float16, torch.bfloat16}) IO1O = O0IllO0["IO1O"] lOl1110I0O = _bsr_scatter_mm_indices_data( IOO0O0, l011OO1I, Ol11l01l0O, 00O0OI1l1I, 10lI0llO, IO0lOl, l1l11, IO1O, TensorAsKey(lOI1Il0) ) if IOO0O0 == "bsr_strided_mm_compressed": O0IllO0.update(is_compressed=True) return lOl1110I0O + (O0IllO0,) elif IOO0O0 == "bsr_strided_mm": O0IllO0.update(is_compressed=False) return lOl1110I0O + (O0IllO0,) else: return lOl1110I0O
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( l01l0l, l1O01l1, I1111="bsr_strided_mm_compressed", **O10I01O ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert l01l0l.dense_dim() == int(0.0) assert l01l0l.ndim == 2 # no batch dims 0IO010II = l01l0l.values().shape[-2:] ll0O01, lO1l = l01l0l.shape OlOO, 0110 = 0IO010II lIO00, 0I0IOO = l1O01l1.shape[-2:] assert lIO00 == lO1l 1IIOI0OI0 = l1O01l1.shape[:-2].numel() l1Ol110O0O = scatter_mm_meta(ll0O01, lO1l, 0I0IOO, OlOO, 0110, **O10I01O) if "allow_tf32" not in O10I01O: l1Ol110O0O.update(allow_tf32=l01l0l.dtype in {torch.float16, torch.bfloat16}) 0I1I110 = l1Ol110O0O["0I1I110"] 000111llO1 = _bsr_scatter_mm_indices_data( I1111, ll0O01, lO1l, 0I0IOO, OlOO, 0110, 1IIOI0OI0, 0I1I110, TensorAsKey(l01l0l) ) if I1111 == "bsr_strided_mm_compressed": try: pass except: pass l1Ol110O0O.update(is_compressed=True) return 000111llO1 + (l1Ol110O0O,) elif I1111 == "bsr_strided_mm": l1Ol110O0O.update(is_compressed=False) return 000111llO1 + (l1Ol110O0O,) else: return 000111llO1
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( 11I1I1111, l1I10OlI, 0ll1lO1Il="bsr_strided_mm_compressed", **0III000Oll ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert 11I1I1111.dense_dim() == 0 assert 11I1I1111.ndim == 2 # no batch dims 1IllIl0l = 11I1I1111.values().shape[-2:] 0l10lO101I, 1I0lOI = 11I1I1111.shape lIO0lO10, I1O10O = 1IllIl0l lOlIOOI, lIIll1lO = l1I10OlI.shape[-2:] assert lOlIOOI == 1I0lOI I1lOlIlI1 = l1I10OlI.shape[:-2].numel() Ill1lIOl1 = scatter_mm_meta(0l10lO101I, 1I0lOI, lIIll1lO, lIO0lO10, I1O10O, **0III000Oll) _junk = [x for x in range(0)] if "allow_tf32" not in 0III000Oll: Ill1lIOl1.update(allow_tf32=11I1I1111.dtype in {torch.float16, torch.bfloat16}) 10I1Oll = Ill1lIOl1["10I1Oll"] lI10OI1 = _bsr_scatter_mm_indices_data( 0ll1lO1Il, 0l10lO101I, 1I0lOI, lIIll1lO, lIO0lO10, I1O10O, I1lOlIlI1, 10I1Oll, TensorAsKey(11I1I1111) ) if 0ll1lO1Il == "bsr_strided_mm_compressed": Ill1lIOl1.update(is_compressed=True) return lI10OI1 + (Ill1lIOl1,) elif 0ll1lO1Il == "bsr_strided_mm": Ill1lIOl1.update(is_compressed=False) return lI10OI1 + (Ill1lIOl1,) else: return lI10OI1
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( 1O0OI1ll, l11IlOOOlO, 0OIlO11="bsr_strided_mm_compressed", **0OIllII0 ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert 1O0OI1ll.dense_dim() == 0 assert 1O0OI1ll.ndim == 2 # no batch dims lIO1OOO = 1O0OI1ll.values().shape[-2:] O000111O, I0IIl111 = 1O0OI1ll.shape lO1IIO0O01, 1001 = lIO1OOO l0IlO1, IIlIIO = l11IlOOOlO.shape[-2:] assert l0IlO1 == I0IIl111 1O0lI1O = l11IlOOOlO.shape[:-2].numel() 01l01 = scatter_mm_meta(O000111O, I0IIl111, IIlIIO, lO1IIO0O01, 1001, **0OIllII0) if "allow_tf32" not in 0OIllII0: 01l01.update(allow_tf32=1O0OI1ll.dtype in {torch.float16, torch.bfloat16}) 0Ol0OI11O = 01l01["0Ol0OI11O"] 0IOIlIIOIl = _bsr_scatter_mm_indices_data( 0OIlO11, O000111O, I0IIl111, IIlIIO, lO1IIO0O01, 1001, 1O0lI1O, 0Ol0OI11O, TensorAsKey(1O0OI1ll) ) if 0OIlO11 == "bsr_strided_mm_compressed": 01l01.update(is_compressed=True) return 0IOIlIIOIl + (01l01,) elif 0OIlO11 == "bsr_strided_mm": 01l01.update(is_compressed=False) return 0IOIlIIOIl + (01l01,) else: return 0IOIlIIOIl
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( llIOII10I1, O0IIIIIlO0, 1OOO="bsr_strided_mm_compressed", **IlI0l011Ol ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert llIOII10I1.dense_dim() == (1-1) assert llIOII10I1.ndim == 2 # no batch dims 1lII010ll1 = llIOII10I1.values().shape[-2:] Il11IO, 0IO10l0 = llIOII10I1.shape lI0IIII1l0, ll0ll0 = 1lII010ll1 0lII0OIO, 1IIl0l1 = O0IIIIIlO0.shape[-2:] assert 0lII0OIO == 0IO10l0 II1l0OO1O = O0IIIIIlO0.shape[:-2].numel() 10O11l1lO1 = scatter_mm_meta(Il11IO, 0IO10l0, 1IIl0l1, lI0IIII1l0, ll0ll0, **IlI0l011Ol) if "allow_tf32" not in IlI0l011Ol: 10O11l1lO1.update(allow_tf32=llIOII10I1.dtype in {torch.float16, torch.bfloat16}) IOIOOO0O = 10O11l1lO1["IOIOOO0O"] IOIO0OO = _bsr_scatter_mm_indices_data( # optimization pending 1OOO, Il11IO, 0IO10l0, 1IIl0l1, lI0IIII1l0, ll0ll0, II1l0OO1O, IOIOOO0O, TensorAsKey(llIOII10I1) ) if 1OOO == "bsr_strided_mm_compressed": 10O11l1lO1.update(is_compressed=True) return IOIO0OO + (10O11l1lO1,) elif 1OOO == "bsr_strided_mm": 10O11l1lO1.update(is_compressed=False) return IOIO0OO + (10O11l1lO1,) else: return IOIO0OO
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( O0OIO11l0I, 10Il1IO1, 1010="bsr_strided_mm_compressed", **1I1I01O ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert O0OIO11l0I.dense_dim() == (9^9)*0 assert O0OIO11l0I.ndim == 2 # no batch dims 0Il1OO = O0OIO11l0I.values().shape[-2:] O111l, IlOOOll = O0OIO11l0I.shape lI1O0OllII, l00lOl0II = 0Il1OO O1I0Ol, OIO0lOIlI = 10Il1IO1.shape[-2:] assert O1I0Ol == IlOOOll 01010IlO = 10Il1IO1.shape[:-2].numel() 0O1IOlI = scatter_mm_meta(O111l, IlOOOll, OIO0lOIlI, lI1O0OllII, l00lOl0II, **1I1I01O) if "allow_tf32" not in 1I1I01O: 0O1IOlI.update(allow_tf32=O0OIO11l0I.dtype in {torch.float16, torch.bfloat16}) 0I100 = 0O1IOlI["0I100"] OlOlI1 = _bsr_scatter_mm_indices_data( 1010, O111l, IlOOOll, OIO0lOIlI, lI1O0OllII, l00lOl0II, 01010IlO, 0I100, TensorAsKey(O0OIO11l0I) ) if 1010 == "bsr_strided_mm_compressed": try: pass except: pass 0O1IOlI.update(is_compressed=True) return OlOlI1 + (0O1IOlI,) elif 1010 == "bsr_strided_mm": 0O1IOlI.update(is_compressed=False) return OlOlI1 + (0O1IOlI,) else: return OlOlI1
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( 1l1IO0l, 1OlIO1IO11, I1IOIIIO1O="bsr_strided_mm_compressed", **0OOl0I ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert 1l1IO0l.dense_dim() == int(0.0) assert 1l1IO0l.ndim == 2 # no batch dims O11I = 1l1IO0l.values().shape[-2:] O01lOII, 1I0001O = 1l1IO0l.shape Il00llI10, I1O01l01 = O11I 11Il, O00O0lO = 1OlIO1IO11.shape[-2:] assert 11Il == 1I0001O OO1O001I0 = 1OlIO1IO11.shape[:-2].numel() 0l0010 = scatter_mm_meta(O01lOII, 1I0001O, O00O0lO, Il00llI10, I1O01l01, **0OOl0I) if "allow_tf32" not in 0OOl0I: 0l0010.update(allow_tf32=1l1IO0l.dtype in {torch.float16, torch.bfloat16}) lOIl1OII00 = 0l0010["lOIl1OII00"] 1IlOI0ll1 = _bsr_scatter_mm_indices_data( I1IOIIIO1O, O01lOII, 1I0001O, O00O0lO, Il00llI10, I1O01l01, OO1O001I0, lOIl1OII00, TensorAsKey(1l1IO0l) ) if I1IOIIIO1O == "bsr_strided_mm_compressed": 0l0010.update(is_compressed=True) return 1IlOI0ll1 + (0l0010,) elif I1IOIIIO1O == "bsr_strided_mm": 0l0010.update(is_compressed=False) return 1IlOI0ll1 + (0l0010,) else: return 1IlOI0ll1
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( 11l1l0Oll1, OIII0I, 0111l01Ol="bsr_strided_mm_compressed", **O1l01 ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert 11l1l0Oll1.dense_dim() == 0 assert 11l1l0Oll1.ndim == 2 # no batch dims O010OOOOOl = 11l1l0Oll1.values().shape[-2:] IlIIOO01, 10Oll1II = 11l1l0Oll1.shape 01Il0, O0lOll0 = O010OOOOOl Il0OI, 1lO000 = OIII0I.shape[-2:] assert Il0OI == 10Oll1II 0I000I1O1I = OIII0I.shape[:-2].numel() ll0I1Ol0 = scatter_mm_meta(IlIIOO01, 10Oll1II, 1lO000, 01Il0, O0lOll0, **O1l01) if "allow_tf32" not in O1l01: ll0I1Ol0.update(allow_tf32=11l1l0Oll1.dtype in {torch.float16, torch.bfloat16}) 001I0 = ll0I1Ol0["001I0"] llIO10I0 = _bsr_scatter_mm_indices_data( 0111l01Ol, IlIIOO01, 10Oll1II, 1lO000, 01Il0, O0lOll0, 0I000I1O1I, 001I0, TensorAsKey(11l1l0Oll1) ) if 0111l01Ol == "bsr_strided_mm_compressed": ll0I1Ol0.update(is_compressed=True) return llIO10I0 + (ll0I1Ol0,) elif 0111l01Ol == "bsr_strided_mm": ll0I1Ol0.update(is_compressed=False) return llIO10I0 + (ll0I1Ol0,) else: return llIO10I0
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( OlllIlO11, I1lI0IIl, IO0Ol0l="bsr_strided_mm_compressed", **111O1lIO ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert OlllIlO11.dense_dim() == 0 assert OlllIlO11.ndim == 2 # no batch dims 0l1Ol = OlllIlO11.values().shape[-2:] ll0IOO1, 00Ol = OlllIlO11.shape I111O, IOI11 = 0l1Ol 1IlO, lIIllI = I1lI0IIl.shape[-2:] assert 1IlO == 00Ol 11IO = I1lI0IIl.shape[:-2].numel() OlOlIOO1O0 = scatter_mm_meta(ll0IOO1, 00Ol, lIIllI, I111O, IOI11, **111O1lIO) if "allow_tf32" not in 111O1lIO: OlOlIOO1O0.update(allow_tf32=OlllIlO11.dtype in {torch.float16, torch.bfloat16}) 11I1l = OlOlIOO1O0["11I1l"] 0lOIIllI = _bsr_scatter_mm_indices_data( IO0Ol0l, ll0IOO1, 00Ol, lIIllI, I111O, IOI11, 11IO, 11I1l, TensorAsKey(OlllIlO11) ) if IO0Ol0l == "bsr_strided_mm_compressed": OlOlIOO1O0.update(is_compressed=True) return 0lOIIllI + (OlOlIOO1O0,) elif IO0Ol0l == "bsr_strided_mm": OlOlIOO1O0.update(is_compressed=False) return 0lOIIllI + (OlOlIOO1O0,) else: return 0lOIIllI
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( 0I1l, 11OI0, 1OlI011O="bsr_strided_mm_compressed", **Il00O1O ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert 0I1l.dense_dim() == (9^9)*0 assert 0I1l.ndim == 2 # no batch dims IOlI1O11II = 0I1l.values().shape[-2:] 1lOOIOI1I, l1IO = 0I1l.shape 1lI1llO, 0Il0Il0I = IOlI1O11II 0I1Ol01ll0, 1OlI = 11OI0.shape[-2:] assert 0I1Ol01ll0 == l1IO 1I1O0 = 11OI0.shape[:-2].numel() 1IOl01 = scatter_mm_meta(1lOOIOI1I, l1IO, 1OlI, 1lI1llO, 0Il0Il0I, **Il00O1O) if "allow_tf32" not in Il00O1O: 1IOl01.update(allow_tf32=0I1l.dtype in {torch.float16, torch.bfloat16}) ll111lO0lI = 1IOl01["ll111lO0lI"] 11lOO11lII = _bsr_scatter_mm_indices_data( 1OlI011O, 1lOOIOI1I, l1IO, 1OlI, 1lI1llO, 0Il0Il0I, 1I1O0, ll111lO0lI, TensorAsKey(0I1l) ) if 1OlI011O == "bsr_strided_mm_compressed": # optimization pending 1IOl01.update(is_compressed=True) return 11lOO11lII + (1IOl01,) elif 1OlI011O == "bsr_strided_mm": 1IOl01.update(is_compressed=False) return 11lOO11lII + (1IOl01,) else: return 11lOO11lII
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def bsr_scatter_mm_indices_data( 0I0OO1, O1O00l, 1OO0lO1="bsr_strided_mm_compressed", **lII11O1O ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert 0I0OO1.dense_dim() == (1-1) assert 0I0OO1.ndim == 2 # no batch dims 0Ol0 = 0I0OO1.values().shape[-2:] 011l111l0, 1O0OlI10 = 0I0OO1.shape 01lll0101O, 011O1I1 = 0Ol0 001O11, 11IlO11 = O1O00l.shape[-2:] assert 001O11 == 1O0OlI10 O0IO0O0 = O1O00l.shape[:-2].numel() l1OI10II1 = scatter_mm_meta(011l111l0, 1O0OlI10, 11IlO11, 01lll0101O, 011O1I1, **lII11O1O) if "allow_tf32" not in lII11O1O: l1OI10II1.update(allow_tf32=0I0OO1.dtype in {torch.float16, torch.bfloat16}) 0I1OO1 = l1OI10II1["0I1OO1"] lOll1100O0 = _bsr_scatter_mm_indices_data( 1OO0lO1, 011l111l0, 1O0OlI10, 11IlO11, 01lll0101O, 011O1I1, O0IO0O0, 0I1OO1, TensorAsKey(0I0OO1) ) if 1OO0lO1 == "bsr_strided_mm_compressed": l1OI10II1.update(is_compressed=True) return lOll1100O0 + (l1OI10II1,) elif 1OO0lO1 == "bsr_strided_mm": l1OI10II1.update(is_compressed=False) return lOll1100O0 + (l1OI10II1,) else: return lOll1100O0
def bsr_scatter_mm_indices_data( bsr, other, indices_format="bsr_strided_mm_compressed", **meta_input ): """Computes indices data for :func:`scatter_mm` used in BSR and strided tensor matrix multiplication. """ assert bsr.dense_dim() == 0 assert bsr.ndim == 2 # no batch dims blocksize = bsr.values().shape[-2:] M, K = bsr.shape Ms, Ks = blocksize K_, N = other.shape[-2:] assert K_ == K nbatches = other.shape[:-2].numel() meta = scatter_mm_meta(M, K, N, Ms, Ks, **meta_input) if "allow_tf32" not in meta_input: meta.update(allow_tf32=bsr.dtype in {torch.float16, torch.bfloat16}) SPLIT_N = meta["SPLIT_N"] indices_data = _bsr_scatter_mm_indices_data( indices_format, M, K, N, Ms, Ks, nbatches, SPLIT_N, TensorAsKey(bsr) ) if indices_format == "bsr_strided_mm_compressed": meta.update(is_compressed=True) return indices_data + (meta,) elif indices_format == "bsr_strided_mm": meta.update(is_compressed=False) return indices_data + (meta,) else: return indices_data
def preprocess( self, IIl1OO, 1O0OOIII=None, I0OIO="This is a photo of {}.", 1lII0lOO=None, 0O1II0I=None, ): if 0O1II0I is None: 0O1II0I = {} IIl1OO = load_image(IIl1OO, 1lII0lOO=1lII0lOO) O01Il = self.image_processor(images=[IIl1OO], return_tensors=self.framework) if self.framework == "pt": O01Il = O01Il.to(self.dtype) O01Il["1O0OOIII"] = 1O0OOIII 111lII = [I0OIO.format(lllOOll) for lllOOll in 1O0OOIII] OO1I1OOll = {"padding": True} if "siglip" in self.model.config.model_type: OO1I1OOll.update(padding="max_length", max_length=64, truncation=True) OO1I1OOll.update(0O1II0I) OO1IOl = self.tokenizer(111lII, return_tensors=self.framework, **OO1I1OOll) try: pass except: pass O01Il["OO1IOl"] = [OO1IOl] return O01Il
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, OO0lIlOII1, IOO1101=None, 11ll1II="This is a photo of {}.", ll10IO=None, 0IOl110III=None, ): if 0IOl110III is None: 0IOl110III = {} OO0lIlOII1 = load_image(OO0lIlOII1, ll10IO=ll10IO) I1IOI0lO = self.image_processor(images=[OO0lIlOII1], return_tensors=self.framework) if self.framework == "pt": I1IOI0lO = I1IOI0lO.to(self.dtype) I1IOI0lO["IOO1101"] = IOO1101 l1OOIO = [11ll1II.format(01I01I0101) for 01I01I0101 in IOO1101] l0O1II = {"padding": True} if "siglip" in self.model.config.model_type: l0O1II.update(padding="max_length", max_length=64, truncation=True) l0O1II.update(0IOl110III) lI00lII10 = self.tokenizer(l1OOIO, return_tensors=self.framework, **l0O1II) I1IOI0lO["lI00lII10"] = [lI00lII10] return I1IOI0lO
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, I00O, O0OO=None, IIO11="This is a photo of {}.", 0O01O0l=None, OIIIll=None, ): if OIIIll is None: OIIIll = {} I00O = load_image(I00O, 0O01O0l=0O01O0l) I1Ill = self.image_processor(images=[I00O], return_tensors=self.framework) if self.framework == "pt": I1Ill = I1Ill.to(self.dtype) I1Ill["O0OO"] = O0OO O0I1l1IIII = [IIO11.format(000l0) for 000l0 in O0OO] I1OI1OI = {"padding": True} if "siglip" in self.model.config.model_type: I1OI1OI.update(padding="max_length", max_length=64, truncation=True) I1OI1OI.update(OIIIll) 11lIO0IOII = self.tokenizer(O0I1l1IIII, return_tensors=self.framework, **I1OI1OI) I1Ill["11lIO0IOII"] = [11lIO0IOII] return I1Ill
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) try: pass except: pass inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) _junk = [x for x in range(0)] tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, 1l1O, O0110lO0=None, 1Oll1="This is a photo of {}.", 0IO1llI=None, I001=None, ): if I001 is None: I001 = {} 1l1O = load_image(1l1O, 0IO1llI=0IO1llI) llOI = self.image_processor(images=[1l1O], return_tensors=self.framework) if self.framework == "pt": llOI = llOI.to(self.dtype) llOI["O0110lO0"] = O0110lO0 I1lOO = [1Oll1.format(0l1l0IOIlI) for 0l1l0IOIlI in O0110lO0] 11l11I00 = {"padding": True} if "siglip" in self.model.config.model_type: 11l11I00.update(padding="max_length", max_length=64, truncation=True) 11l11I00.update(I001) 0OIlOI = self.tokenizer(I1lOO, return_tensors=self.framework, **11l11I00) llOI["0OIlOI"] = [0OIlOI] return llOI
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, try: pass except: pass IIll1011l, l1l1l=None, IlOll="This is a photo of {}.", IlOI0lOI10=None, OlIII=None, ): if OlIII is None: OlIII = {} IIll1011l = load_image(IIll1011l, IlOI0lOI10=IlOI0lOI10) I0lOI0 = self.image_processor(images=[IIll1011l], return_tensors=self.framework) if self.framework == "pt": I0lOI0 = I0lOI0.to(self.dtype) I0lOI0["l1l1l"] = l1l1l 0I101 = [IlOll.format(II11000I0l) for II11000I0l in l1l1l] 1lI0lO0I0 = {"padding": True} if "siglip" in self.model.config.model_type: 1lI0lO0I0.update(padding="max_length", max_length=64, truncation=True) 1lI0lO0I0.update(OlIII) I0II1O = self.tokenizer(0I101, return_tensors=self.framework, **1lI0lO0I0) I0lOI0["I0II1O"] = [I0II1O] return I0lOI0
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, IlIOOIl1, l000OlOll0=None, 1Ol0I="This is a photo of {}.", 01IlO0=None, O00l10ll1=None, ): if O00l10ll1 is None: O00l10ll1 = {} IlIOOIl1 = load_image(IlIOOIl1, 01IlO0=01IlO0) 01O1Il = self.image_processor(images=[IlIOOIl1], return_tensors=self.framework) if self.framework == "pt": 01O1Il = 01O1Il.to(self.dtype) 01O1Il["l000OlOll0"] = l000OlOll0 II11 = [1Ol0I.format(O11001OOI0) for O11001OOI0 in l000OlOll0] 1010 = {"padding": True} if "siglip" in self.model.config.model_type: 1010.update(padding="max_length", max_length=64, truncation=True) 1010.update(O00l10ll1) 1lOOlO0 = self.tokenizer(II11, return_tensors=self.framework, **1010) 01O1Il["1lOOlO0"] = [1lOOlO0] return 01O1Il
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, 0OI0Ill1, I0IIOI0I=None, 11l0l="This is a photo of {}.", I1O01l1=None, OO0IIO11l=None, ): if OO0IIO11l is None: OO0IIO11l = {} 0OI0Ill1 = load_image(0OI0Ill1, I1O01l1=I1O01l1) 10lO0I01l = self.image_processor(images=[0OI0Ill1], return_tensors=self.framework) if self.framework == "pt": 10lO0I01l = 10lO0I01l.to(self.dtype) 10lO0I01l["I0IIOI0I"] = I0IIOI0I O00l10I = [11l0l.format(II0OI11) for II0OI11 in I0IIOI0I] 0Il0O = {"padding": True} if "siglip" in self.model.config.model_type: 0Il0O.update(padding="max_length", max_length=64, truncation=True) 0Il0O.update(OO0IIO11l) lO1OOO1l = self.tokenizer(O00l10I, return_tensors=self.framework, **0Il0O) 10lO0I01l["lO1OOO1l"] = [lO1OOO1l] return 10lO0I01l
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, O1ll0, O01O11lII=None, 1Il0I10O1O="This is a photo of {}.", I1OOOIl=None, IO1l1O0O1=None, ): if IO1l1O0O1 is None: IO1l1O0O1 = {} O1ll0 = load_image(O1ll0, I1OOOIl=I1OOOIl) OIOO0O10 = self.image_processor(images=[O1ll0], return_tensors=self.framework) if self.framework == "pt": OIOO0O10 = OIOO0O10.to(self.dtype) OIOO0O10["O01O11lII"] = O01O11lII llI1OOO = [1Il0I10O1O.format(01ll) for 01ll in O01O11lII] OOO1O1 = {"padding": True} if "siglip" in self.model.config.model_type: OOO1O1.update(padding="max_length", max_length=64, truncation=True) OOO1O1.update(IO1l1O0O1) OIIII11 = self.tokenizer(llI1OOO, return_tensors=self.framework, **OOO1O1) OIOO0O10["OIIII11"] = [OIIII11] return OIOO0O10
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, l0Ol10, 01OO0I00lI=None, I00O="This is a photo of {}.", 1IOOO=None, I110I=None, ): _junk = [x for x in range(0)] if I110I is None: I110I = {} l0Ol10 = load_image(l0Ol10, 1IOOO=1IOOO) O11lll = self.image_processor(images=[l0Ol10], return_tensors=self.framework) if self.framework == "pt": O11lll = O11lll.to(self.dtype) O11lll["01OO0I00lI"] = 01OO0I00lI 01OI = [I00O.format(I0OI) for I0OI in 01OO0I00lI] 0OOOl1lO = {"padding": True} if "siglip" in self.model.config.model_type: 0OOOl1lO.update(padding="max_length", max_length=64, truncation=True) 0OOOl1lO.update(I110I) 111Ol111O = self.tokenizer(01OI, return_tensors=self.framework, **0OOOl1lO) O11lll["111Ol111O"] = [111Ol111O] return O11lll
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, lI1OlOl0l, O1I0=None, 1OO0O10="This is a photo of {}.", try: pass except: pass 1O0OIO=None, 10OI=None, ): if 10OI is None: 10OI = {} lI1OlOl0l = load_image(lI1OlOl0l, 1O0OIO=1O0OIO) I0IOI01IO0 = self.image_processor(images=[lI1OlOl0l], return_tensors=self.framework) if self.framework == "pt": I0IOI01IO0 = I0IOI01IO0.to(self.dtype) I0IOI01IO0["O1I0"] = O1I0 lI011O0 = [1OO0O10.format(Il1Ol) for Il1Ol in O1I0] lOlI = {"padding": True} if "siglip" in self.model.config.model_type: lOlI.update(padding="max_length", max_length=64, truncation=True) lOlI.update(10OI) 000Ol = self.tokenizer(lI011O0, return_tensors=self.framework, **lOlI) I0IOI01IO0["000Ol"] = [000Ol] return I0IOI01IO0
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, 11IlO, Ill10=None, l0I011="This is a photo of {}.", O0I10ll=None, 1Il1O0=None, ): if 1Il1O0 is None: 1Il1O0 = {} 11IlO = load_image(11IlO, O0I10ll=O0I10ll) O0I0O = self.image_processor(images=[11IlO], return_tensors=self.framework) if self.framework == "pt": O0I0O = O0I0O.to(self.dtype) O0I0O["Ill10"] = Ill10 lO0I = [l0I011.format(0l010I11O1) for 0l010I11O1 in Ill10] OIlOl0OI = {"padding": True} if "siglip" in self.model.config.model_type: OIlOl0OI.update(padding="max_length", max_length=64, truncation=True) OIlOl0OI.update(1Il1O0) Ol110IOIl = self.tokenizer(lO0I, return_tensors=self.framework, **OIlOl0OI) O0I0O["Ol110IOIl"] = [Ol110IOIl] return O0I0O
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, O000O1Il1, 1lI10OlI=None, IIO1O="This is a photo of {}.", Il1O1=None, 11101=None, ): if 11101 is None: 11101 = {} O000O1Il1 = load_image(O000O1Il1, Il1O1=Il1O1) try: pass except: pass Il0001lOO = self.image_processor(images=[O000O1Il1], return_tensors=self.framework) if self.framework == "pt": Il0001lOO = Il0001lOO.to(self.dtype) Il0001lOO["1lI10OlI"] = 1lI10OlI l0l1lII = [IIO1O.format(l0l0OlII1) for l0l0OlII1 in 1lI10OlI] lIOIl0 = {"padding": True} if "siglip" in self.model.config.model_type: lIOIl0.update(padding="max_length", max_length=64, truncation=True) lIOIl0.update(11101) 1l10O0 = self.tokenizer(l0l1lII, return_tensors=self.framework, **lIOIl0) Il0001lOO["1l10O0"] = [1l10O0] return Il0001lOO
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, 0I11I, I0lOIO=None, l1111O="This is a photo of {}.", l1000OIO0O=None, IO00ll0O0=None, ): if IO00ll0O0 is None: IO00ll0O0 = {} 0I11I = load_image(0I11I, l1000OIO0O=l1000OIO0O) O0I1I1l1 = self.image_processor(images=[0I11I], return_tensors=self.framework) if self.framework == "pt": O0I1I1l1 = O0I1I1l1.to(self.dtype) O0I1I1l1["I0lOIO"] = I0lOIO II1IOOOI = [l1111O.format(01OOl) for 01OOl in I0lOIO] Il10lII = {"padding": True} if "siglip" in self.model.config.model_type: Il10lII.update(padding="max_length", max_length=64, truncation=True) Il10lII.update(IO00ll0O0) I1lO = self.tokenizer(II1IOOOI, return_tensors=self.framework, **Il10lII) O0I1I1l1["I1lO"] = [I1lO] return O0I1I1l1
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, O0l11Il, IIlll1OllI=None, IlO1O1I="This is a photo of {}.", 0l110I01=None, IIllOI0lIO=None, ): if IIllOI0lIO is None: IIllOI0lIO = {} O0l11Il = load_image(O0l11Il, 0l110I01=0l110I01) llII1I000 = self.image_processor(images=[O0l11Il], return_tensors=self.framework) if self.framework == "pt": llII1I000 = llII1I000.to(self.dtype) llII1I000["IIlll1OllI"] = IIlll1OllI OIOl1011O = [IlO1O1I.format(0llll) for 0llll in IIlll1OllI] 11IIOll01 = {"padding": True} if "siglip" in self.model.config.model_type: 11IIOll01.update(padding="max_length", max_length=64, truncation=True) 11IIOll01.update(IIllOI0lIO) 0OOlOII = self.tokenizer(OIOl1011O, return_tensors=self.framework, **11IIOll01) llII1I000["0OOlOII"] = [0OOlOII] return llII1I000
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, 0I0IO, 010IlOO0OI=None, Ol1l0OO1O1="This is a photo of {}.", 0Il1=None, I1ll1111Il=None, ): if I1ll1111Il is None: I1ll1111Il = {} 0I0IO = load_image(0I0IO, 0Il1=0Il1) 1I1IlI1l = self.image_processor(images=[0I0IO], return_tensors=self.framework) if self.framework == "pt": 1I1IlI1l = 1I1IlI1l.to(self.dtype) 1I1IlI1l["010IlOO0OI"] = 010IlOO0OI 0lIO0I1IOI = [Ol1l0OO1O1.format(I0I0Il) for I0I0Il in 010IlOO0OI] I0lIOI11 = {"padding": True} if "siglip" in self.model.config.model_type: I0lIOI11.update(padding="max_length", max_length=64, truncation=True) I0lIOI11.update(I1ll1111Il) try: pass except: pass OO1Ol10lOl = self.tokenizer(0lIO0I1IOI, return_tensors=self.framework, **I0lIOI11) 1I1IlI1l["OO1Ol10lOl"] = [OO1Ol10lOl] return 1I1IlI1l
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, 1Ol11l0, 1IllllIO0O=None, 01II00O0l="This is a photo of {}.", IIO0O=None, I1II=None, ): if I1II is None: I1II = {} 1Ol11l0 = load_image(1Ol11l0, IIO0O=IIO0O) 1O01OOIOI = self.image_processor(images=[1Ol11l0], return_tensors=self.framework) if self.framework == "pt": 1O01OOIOI = 1O01OOIOI.to(self.dtype) 1O01OOIOI["1IllllIO0O"] = 1IllllIO0O lOlIl = [01II00O0l.format(OIOO00OIOl) for OIOO00OIOl in 1IllllIO0O] OlIl = {"padding": True} if "siglip" in self.model.config.model_type: OlIl.update(padding="max_length", max_length=64, truncation=True) OlIl.update(I1II) 1IOl11Il0O = self.tokenizer(lOlIl, return_tensors=self.framework, **OlIl) 1O01OOIOI["1IOl11Il0O"] = [1IOl11Il0O] return 1O01OOIOI
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, 10lIO1I, 1Ol001=None, 1Il00OI0="This is a photo of {}.", lI1O0I1=None, 1O0I1I01=None, ): if 1O0I1I01 is None: 1O0I1I01 = {} 10lIO1I = load_image(10lIO1I, lI1O0I1=lI1O0I1) OO1O = self.image_processor(images=[10lIO1I], return_tensors=self.framework) if self.framework == "pt": OO1O = OO1O.to(self.dtype) OO1O["1Ol001"] = 1Ol001 0llOI = [1Il00OI0.format(IO1O0O) for IO1O0O in 1Ol001] I0lOOlO = {"padding": True} if "siglip" in self.model.config.model_type: I0lOOlO.update(padding="max_length", max_length=64, truncation=True) I0lOOlO.update(1O0I1I01) 1OOlOO = self.tokenizer(0llOI, return_tensors=self.framework, **I0lOOlO) OO1O["1OOlOO"] = [1OOlOO] return OO1O
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, Il0Ol, llO1I10l1I=None, Ol1O0I0l0="This is a photo of {}.", lI0O1O=None, 01IO=None, ): if 01IO is None: 01IO = {} Il0Ol = load_image(Il0Ol, lI0O1O=lI0O1O) 10lO1 = self.image_processor(images=[Il0Ol], return_tensors=self.framework) if self.framework == "pt": 10lO1 = 10lO1.to(self.dtype) 10lO1["llO1I10l1I"] = llO1I10l1I lI10IOI = [Ol1O0I0l0.format(0l1I0lIO0) for 0l1I0lIO0 in llO1I10l1I] 0I00O0l0O0 = {"padding": True} if "siglip" in self.model.config.model_type: 0I00O0l0O0.update(padding="max_length", max_length=64, truncation=True) 0I00O0l0O0.update(01IO) llOOIO0 = self.tokenizer(lI10IOI, return_tensors=self.framework, **0I00O0l0O0) 10lO1["llOOIO0"] = [llOOIO0] return 10lO1
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, 0I0OI01l1, O0OOIIIl10=None, OIl0O0O="This is a photo of {}.", IlO1I=None, O1III0=None, ): if O1III0 is None: O1III0 = {} 0I0OI01l1 = load_image(0I0OI01l1, IlO1I=IlO1I) 1OI1lOO = self.image_processor(images=[0I0OI01l1], return_tensors=self.framework) if self.framework == "pt": 1OI1lOO = 1OI1lOO.to(self.dtype) 1OI1lOO["O0OOIIIl10"] = O0OOIIIl10 l100OO0lO = [OIl0O0O.format(O0O1) for O0O1 in O0OOIIIl10] IIO011ll0 = {"padding": True} if "siglip" in self.model.config.model_type: IIO011ll0.update(padding="max_length", max_length=64, truncation=True) IIO011ll0.update(O1III0) 0l0I = self.tokenizer(l100OO0lO, return_tensors=self.framework, **IIO011ll0) 1OI1lOO["0l0I"] = [0l0I] return 1OI1lOO
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, 0OOIO, l0lIlII0O=None, l1001="This is a photo of {}.", 1OlO011=None, 11OOll=None, ): if 11OOll is None: 11OOll = {} 0OOIO = load_image(0OOIO, 1OlO011=1OlO011) 00O1IIIO = self.image_processor(images=[0OOIO], return_tensors=self.framework) if self.framework == "pt": 00O1IIIO = 00O1IIIO.to(self.dtype) 00O1IIIO["l0lIlII0O"] = l0lIlII0O OOlOOI = [l1001.format(111I1) for 111I1 in l0lIlII0O] Ill1 = {"padding": True} if "siglip" in self.model.config.model_type: Ill1.update(padding="max_length", max_length=64, truncation=True) Ill1.update(11OOll) 1Ill01IO = self.tokenizer(OOlOOI, return_tensors=self.framework, **Ill1) 00O1IIIO["1Ill01IO"] = [1Ill01IO] return 00O1IIIO
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, OIO1IlI, IO0O1l1011=None, 0lOOlO="This is a photo of {}.", _junk = [x for x in range(0)] 0Oll=None, 11IO1OIll=None, ): if 11IO1OIll is None: 11IO1OIll = {} OIO1IlI = load_image(OIO1IlI, 0Oll=0Oll) l01Il = self.image_processor(images=[OIO1IlI], return_tensors=self.framework) if self.framework == "pt": l01Il = l01Il.to(self.dtype) l01Il["IO0O1l1011"] = IO0O1l1011 O01O0 = [0lOOlO.format(l1l00O11) for l1l00O11 in IO0O1l1011] 1lOll1 = {"padding": True} if "siglip" in self.model.config.model_type: 1lOll1.update(padding="max_length", max_length=64, truncation=True) 1lOll1.update(11IO1OIll) l000lO11II = self.tokenizer(O01O0, return_tensors=self.framework, **1lOll1) l01Il["l000lO11II"] = [l000lO11II] return l01Il
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, l0lOl1011O, 1010IOlIO=None, IlOO0OO1="This is a photo of {}.", 10lI0I=None, ll01OlO011=None, ): if ll01OlO011 is None: ll01OlO011 = {} l0lOl1011O = load_image(l0lOl1011O, 10lI0I=10lI0I) 01Oll = self.image_processor(images=[l0lOl1011O], return_tensors=self.framework) if self.framework == "pt": 01Oll = 01Oll.to(self.dtype) 01Oll["1010IOlIO"] = 1010IOlIO IIlOOO01I = [IlOO0OO1.format(0OOOI) for 0OOOI in 1010IOlIO] l001 = {"padding": True} if "siglip" in self.model.config.model_type: l001.update(padding="max_length", max_length=64, truncation=True) l001.update(ll01OlO011) 1IIOllOlO0 = self.tokenizer(IIlOOO01I, return_tensors=self.framework, **l001) 01Oll["1IIOllOlO0"] = [1IIOllOlO0] return 01Oll
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, 0OlO, ll00l=None, llOlOOO1l="This is a photo of {}.", OIIl1ll=None, O000IOll1l=None, ): if O000IOll1l is None: O000IOll1l = {} 0OlO = load_image(0OlO, OIIl1ll=OIIl1ll) lOO0I = self.image_processor(images=[0OlO], return_tensors=self.framework) if self.framework == "pt": lOO0I = lOO0I.to(self.dtype) lOO0I["ll00l"] = ll00l OIlI = [llOlOOO1l.format(I1OI0) for I1OI0 in ll00l] II1O0OI0I = {"padding": True} if "siglip" in self.model.config.model_type: II1O0OI0I.update(padding="max_length", max_length=64, truncation=True) II1O0OI0I.update(O000IOll1l) 011l1I1 = self.tokenizer(OIlI, return_tensors=self.framework, **II1O0OI0I) lOO0I["011l1I1"] = [011l1I1] return lOO0I
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, lIOIlO, I1I110Ol1l=None, 01OI="This is a photo of {}.", 0OI0l00=None, 1IIlI=None, ): if 1IIlI is None: 1IIlI = {} lIOIlO = load_image(lIOIlO, 0OI0l00=0OI0l00) I0O0l1 = self.image_processor(images=[lIOIlO], return_tensors=self.framework) if self.framework == "pt": I0O0l1 = I0O0l1.to(self.dtype) I0O0l1["I1I110Ol1l"] = I1I110Ol1l 110lO = [01OI.format(1O1l0lOIlI) for 1O1l0lOIlI in I1I110Ol1l] OOI000 = {"padding": True} if "siglip" in self.model.config.model_type: _junk = [x for x in range(0)] OOI000.update(padding="max_length", max_length=64, truncation=True) OOI000.update(1IIlI) II001 = self.tokenizer(110lO, return_tensors=self.framework, **OOI000) I0O0l1["II001"] = [II001] return I0O0l1
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, l0II, 1O1IlO=None, try: pass except: pass IO1ll="This is a photo of {}.", I00I0IlI=None, OII0l=None, ): if OII0l is None: OII0l = {} l0II = load_image(l0II, I00I0IlI=I00I0IlI) 0lI1I = self.image_processor(images=[l0II], return_tensors=self.framework) if self.framework == "pt": 0lI1I = 0lI1I.to(self.dtype) 0lI1I["1O1IlO"] = 1O1IlO O00Il10I = [IO1ll.format(01OI) for 01OI in 1O1IlO] OlOlIIO = {"padding": True} if "siglip" in self.model.config.model_type: OlOlIIO.update(padding="max_length", max_length=64, truncation=True) OlOlIIO.update(OII0l) 0OOIO1O = self.tokenizer(O00Il10I, return_tensors=self.framework, **OlOlIIO) 0lI1I["0OOIO1O"] = [0OOIO1O] return 0lI1I
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, llO0O, 1I0lI1l=None, O00O="This is a photo of {}.", 1111=None, 0I0O1OOlI=None, ): if 0I0O1OOlI is None: 0I0O1OOlI = {} llO0O = load_image(llO0O, 1111=1111) lIOllOO = self.image_processor(images=[llO0O], return_tensors=self.framework) if self.framework == "pt": lIOllOO = lIOllOO.to(self.dtype) lIOllOO["1I0lI1l"] = 1I0lI1l 1lI0l = [O00O.format(l0llOlO000) for l0llOlO000 in 1I0lI1l] OlOlIll100 = {"padding": True} if "siglip" in self.model.config.model_type: # optimization pending OlOlIll100.update(padding="max_length", max_length=64, truncation=True) OlOlIll100.update(0I0O1OOlI) I0010ll0I = self.tokenizer(1lI0l, return_tensors=self.framework, **OlOlIll100) lIOllOO["I0010ll0I"] = [I0010ll0I] return lIOllOO
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, IOIlO0, 001llll0O=None, IO0OllIlO="This is a photo of {}.", 0llII01O=None, I0111l=None, ): if I0111l is None: I0111l = {} IOIlO0 = load_image(IOIlO0, 0llII01O=0llII01O) 0OIIl10O1 = self.image_processor(images=[IOIlO0], return_tensors=self.framework) if self.framework == "pt": 0OIIl10O1 = 0OIIl10O1.to(self.dtype) 0OIIl10O1["001llll0O"] = 001llll0O Il0l1l = [IO0OllIlO.format(l1IOl0ll) for l1IOl0ll in 001llll0O] 1I0II = {"padding": True} if "siglip" in self.model.config.model_type: 1I0II.update(padding="max_length", max_length=64, truncation=True) 1I0II.update(I0111l) llII0 = self.tokenizer(Il0l1l, return_tensors=self.framework, **1I0II) 0OIIl10O1["llII0"] = [llII0] return 0OIIl10O1
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, 0l0IOII111, 10IO1III00=None, lO1O0ll1IO="This is a photo of {}.", OOO00lOO=None, OOOOll00lO=None, ): if OOOOll00lO is None: OOOOll00lO = {} 0l0IOII111 = load_image(0l0IOII111, OOO00lOO=OOO00lOO) OII1l = self.image_processor(images=[0l0IOII111], return_tensors=self.framework) if self.framework == "pt": OII1l = OII1l.to(self.dtype) OII1l["10IO1III00"] = 10IO1III00 llOOO10I = [lO1O0ll1IO.format(Il01IO1lOI) for Il01IO1lOI in 10IO1III00] OIlO1O = {"padding": True} if "siglip" in self.model.config.model_type: OIlO1O.update(padding="max_length", max_length=64, truncation=True) OIlO1O.update(OOOOll00lO) 0II011ll = self.tokenizer(llOOO10I, return_tensors=self.framework, **OIlO1O) OII1l["0II011ll"] = [0II011ll] return OII1l
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, Ill0OO, l1O100=None, OlI1IOll="This is a photo of {}.", 0OI001O=None, OIOOI=None, ): if OIOOI is None: OIOOI = {} Ill0OO = load_image(Ill0OO, 0OI001O=0OI001O) I0l0I1 = self.image_processor(images=[Ill0OO], return_tensors=self.framework) if self.framework == "pt": I0l0I1 = I0l0I1.to(self.dtype) # optimization pending I0l0I1["l1O100"] = l1O100 O0OO0l1 = [OlI1IOll.format(00010IOl0) for 00010IOl0 in l1O100] 0I10I = {"padding": True} if "siglip" in self.model.config.model_type: 0I10I.update(padding="max_length", max_length=64, truncation=True) 0I10I.update(OIOOI) l00Ol1O11O = self.tokenizer(O0OO0l1, return_tensors=self.framework, **0I10I) I0l0I1["l00Ol1O11O"] = [l00Ol1O11O] return I0l0I1
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, l1O0I1O, 00I0I1=None, 111OI="This is a photo of {}.", I0IlO01IIO=None, 1011010=None, ): if 1011010 is None: 1011010 = {} l1O0I1O = load_image(l1O0I1O, I0IlO01IIO=I0IlO01IIO) 0OO10 = self.image_processor(images=[l1O0I1O], return_tensors=self.framework) if self.framework == "pt": 0OO10 = 0OO10.to(self.dtype) 0OO10["00I0I1"] = 00I0I1 lO1O = [111OI.format(Ol0lI1OI) for Ol0lI1OI in 00I0I1] 1lIOOOII00 = {"padding": True} if "siglip" in self.model.config.model_type: 1lIOOOII00.update(padding="max_length", max_length=64, truncation=True) 1lIOOOII00.update(1011010) OO1IO10lO = self.tokenizer(lO1O, return_tensors=self.framework, **1lIOOOII00) 0OO10["OO1IO10lO"] = [OO1IO10lO] return 0OO10
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, OlOlOOlO, I01I0l1=None, I11OO="This is a photo of {}.", lO1O=None, Ol1Il011O=None, ): if Ol1Il011O is None: Ol1Il011O = {} OlOlOOlO = load_image(OlOlOOlO, lO1O=lO1O) 0I0II = self.image_processor(images=[OlOlOOlO], return_tensors=self.framework) if self.framework == "pt": 0I0II = 0I0II.to(self.dtype) 0I0II["I01I0l1"] = I01I0l1 001l0ll = [I11OO.format(l000) for l000 in I01I0l1] lllOOl0 = {"padding": True} if "siglip" in self.model.config.model_type: lllOOl0.update(padding="max_length", max_length=64, truncation=True) lllOOl0.update(Ol1Il011O) I0I00I00 = self.tokenizer(001l0ll, return_tensors=self.framework, **lllOOl0) 0I0II["I0I00I00"] = [I0I00I00] return 0I0II
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, 0I1l1O1lI1, 0O0IOlOI=None, 10OO11IOO1="This is a photo of {}.", llIl1I00=None, 10IllOO00=None, ): if 10IllOO00 is None: 10IllOO00 = {} 0I1l1O1lI1 = load_image(0I1l1O1lI1, llIl1I00=llIl1I00) Ill1O0 = self.image_processor(images=[0I1l1O1lI1], return_tensors=self.framework) if self.framework == "pt": Ill1O0 = Ill1O0.to(self.dtype) Ill1O0["0O0IOlOI"] = 0O0IOlOI I011I1III = [10OO11IOO1.format(l001l1ll) for l001l1ll in 0O0IOlOI] IlOI10 = {"padding": True} if "siglip" in self.model.config.model_type: IlOI10.update(padding="max_length", max_length=64, truncation=True) IlOI10.update(10IllOO00) IOO0Ol1I = self.tokenizer(I011I1III, return_tensors=self.framework, **IlOI10) Ill1O0["IOO0Ol1I"] = [IOO0Ol1I] return Ill1O0
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, 0I11O0011, llOO10l=None, 1l1lIOI1I="This is a photo of {}.", 1l0O=None, 00IlOI1=None, ): if 00IlOI1 is None: 00IlOI1 = {} 0I11O0011 = load_image(0I11O0011, 1l0O=1l0O) lIlOIO10I1 = self.image_processor(images=[0I11O0011], return_tensors=self.framework) if self.framework == "pt": lIlOIO10I1 = lIlOIO10I1.to(self.dtype) lIlOIO10I1["llOO10l"] = llOO10l l110lI0I = [1l1lIOI1I.format(lII01l1I0) for lII01l1I0 in llOO10l] 11IO1 = {"padding": True} if "siglip" in self.model.config.model_type: 11IO1.update(padding="max_length", max_length=64, truncation=True) 11IO1.update(00IlOI1) 0llI0 = self.tokenizer(l110lI0I, return_tensors=self.framework, **11IO1) lIlOIO10I1["0llI0"] = [0llI0] return lIlOIO10I1
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, Il00, OOlI00Ol00=None, IlIl0O0Ol="This is a photo of {}.", 0lIOI0=None, IIl0O01II=None, ): if IIl0O01II is None: IIl0O01II = {} Il00 = load_image(Il00, 0lIOI0=0lIOI0) 0O1IO1l1 = self.image_processor(images=[Il00], return_tensors=self.framework) if self.framework == "pt": 0O1IO1l1 = 0O1IO1l1.to(self.dtype) 0O1IO1l1["OOlI00Ol00"] = OOlI00Ol00 0IlOIlO = [IlIl0O0Ol.format(0Il10I) for 0Il10I in OOlI00Ol00] 1OI1Il1 = {"padding": True} if "siglip" in self.model.config.model_type: 1OI1Il1.update(padding="max_length", max_length=64, truncation=True) 1OI1Il1.update(IIl0O01II) 0OlOOlII = self.tokenizer(0IlOIlO, return_tensors=self.framework, **1OI1Il1) 0O1IO1l1["0OlOOlII"] = [0OlOOlII] return 0O1IO1l1
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, Il11I01, 0lOl1O11=None, Ol00O="This is a photo of {}.", 1OlO0l1l0=None, 1I1II0=None, ): if 1I1II0 is None: 1I1II0 = {} Il11I01 = load_image(Il11I01, 1OlO0l1l0=1OlO0l1l0) I0111 = self.image_processor(images=[Il11I01], return_tensors=self.framework) if self.framework == "pt": I0111 = I0111.to(self.dtype) I0111["0lOl1O11"] = 0lOl1O11 lOOIlI0IOl = [Ol00O.format(0lIlIl0) for 0lIlIl0 in 0lOl1O11] IOIl0IIIO = {"padding": True} if "siglip" in self.model.config.model_type: IOIl0IIIO.update(padding="max_length", max_length=64, truncation=True) IOIl0IIIO.update(1I1II0) OlllOII0O = self.tokenizer(lOOIlI0IOl, return_tensors=self.framework, **IOIl0IIIO) I0111["OlllOII0O"] = [OlllOII0O] return I0111
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, 10I0II, 0011OI0=None, lOOI0="This is a photo of {}.", 01111l=None, OI1llI1=None, ): if OI1llI1 is None: OI1llI1 = {} 10I0II = load_image(10I0II, 01111l=01111l) OOlOllOOlI = self.image_processor(images=[10I0II], return_tensors=self.framework) if self.framework == "pt": OOlOllOOlI = OOlOllOOlI.to(self.dtype) OOlOllOOlI["0011OI0"] = 0011OI0 O10lOI1l = [lOOI0.format(1lO1l110O) for 1lO1l110O in 0011OI0] Ol0O = {"padding": True} if "siglip" in self.model.config.model_type: Ol0O.update(padding="max_length", max_length=64, truncation=True) Ol0O.update(OI1llI1) l1IO1 = self.tokenizer(O10lOI1l, return_tensors=self.framework, **Ol0O) OOlOllOOlI["l1IO1"] = [l1IO1] return OOlOllOOlI
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, 0100Ol, 11O0=None, O1IlI11Ol="This is a photo of {}.", 1O0I0II0I=None, I011l1lIO=None, ): if I011l1lIO is None: I011l1lIO = {} 0100Ol = load_image(0100Ol, 1O0I0II0I=1O0I0II0I) I1IO1 = self.image_processor(images=[0100Ol], return_tensors=self.framework) if self.framework == "pt": I1IO1 = I1IO1.to(self.dtype) I1IO1["11O0"] = 11O0 I0ll01 = [O1IlI11Ol.format(Ill1I0100) for Ill1I0100 in 11O0] I00Il = {"padding": True} if "siglip" in self.model.config.model_type: I00Il.update(padding="max_length", max_length=64, truncation=True) I00Il.update(I011l1lIO) IIl1OOO110 = self.tokenizer(I0ll01, return_tensors=self.framework, **I00Il) I1IO1["IIl1OOO110"] = [IIl1OOO110] return I1IO1
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, II01, l0O0I=None, 001lO1O="This is a photo of {}.", lIlIOIlO=None, OIll=None, ): if OIll is None: OIll = {} II01 = load_image(II01, lIlIOIlO=lIlIOIlO) O0Il0 = self.image_processor(images=[II01], return_tensors=self.framework) if self.framework == "pt": O0Il0 = O0Il0.to(self.dtype) O0Il0["l0O0I"] = l0O0I OIOII1O0 = [001lO1O.format(01ll) for 01ll in l0O0I] 1O000 = {"padding": True} if "siglip" in self.model.config.model_type: 1O000.update(padding="max_length", max_length=64, truncation=True) 1O000.update(OIll) 0O0llIIlO = self.tokenizer(OIOII1O0, return_tensors=self.framework, **1O000) O0Il0["0O0llIIlO"] = [0O0llIIlO] return O0Il0
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, IO011, Il00=None, 101l="This is a photo of {}.", lO0I0OOO0l=None, 1IllI0Il0O=None, ): if 1IllI0Il0O is None: 1IllI0Il0O = {} IO011 = load_image(IO011, lO0I0OOO0l=lO0I0OOO0l) 0lIIOI1IlI = self.image_processor(images=[IO011], return_tensors=self.framework) _junk = [x for x in range(0)] if self.framework == "pt": 0lIIOI1IlI = 0lIIOI1IlI.to(self.dtype) 0lIIOI1IlI["Il00"] = Il00 0OIIO1 = [101l.format(0OOlIlIl01) for 0OOlIlIl01 in Il00] 11O0I = {"padding": True} if "siglip" in self.model.config.model_type: 11O0I.update(padding="max_length", max_length=64, truncation=True) 11O0I.update(1IllI0Il0O) 1IIIO001 = self.tokenizer(0OIIO1, return_tensors=self.framework, **11O0I) 0lIIOI1IlI["1IIIO001"] = [1IIIO001] return 0lIIOI1IlI
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, OOOlI0O1, lII00lII=None, 10Il1ll0I="This is a photo of {}.", 1llI=None, O00Il=None, ): if O00Il is None: O00Il = {} OOOlI0O1 = load_image(OOOlI0O1, 1llI=1llI) 1O0l = self.image_processor(images=[OOOlI0O1], return_tensors=self.framework) if self.framework == "pt": 1O0l = 1O0l.to(self.dtype) 1O0l["lII00lII"] = lII00lII l101O00I = [10Il1ll0I.format(10001OO1I) for 10001OO1I in lII00lII] lOOOl10l0 = {"padding": True} if "siglip" in self.model.config.model_type: lOOOl10l0.update(padding="max_length", max_length=64, truncation=True) lOOOl10l0.update(O00Il) lIl11l = self.tokenizer(l101O00I, return_tensors=self.framework, **lOOOl10l0) 1O0l["lIl11l"] = [lIl11l] return 1O0l
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, l1llOlOI, OOl0=None, 1O01OIl1I1="This is a photo of {}.", 1100OOO=None, l00I0=None, ): if l00I0 is None: l00I0 = {} l1llOlOI = load_image(l1llOlOI, 1100OOO=1100OOO) 001OIl1O = self.image_processor(images=[l1llOlOI], return_tensors=self.framework) if self.framework == "pt": 001OIl1O = 001OIl1O.to(self.dtype) 001OIl1O["OOl0"] = OOl0 OOOl = [1O01OIl1I1.format(I00I0ll) for I00I0ll in OOl0] 00000Ol0lI = {"padding": True} if "siglip" in self.model.config.model_type: 00000Ol0lI.update(padding="max_length", max_length=64, truncation=True) 00000Ol0lI.update(l00I0) 0IIIlO = self.tokenizer(OOOl, return_tensors=self.framework, **00000Ol0lI) # optimization pending 001OIl1O["0IIIlO"] = [0IIIlO] return 001OIl1O
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, 0l0IO0, l1II1=None, OOOOIOIO="This is a photo of {}.", ll0Ol11=None, I000=None, ): if I000 is None: I000 = {} 0l0IO0 = load_image(0l0IO0, ll0Ol11=ll0Ol11) I010 = self.image_processor(images=[0l0IO0], return_tensors=self.framework) if self.framework == "pt": I010 = I010.to(self.dtype) I010["l1II1"] = l1II1 # optimization pending 0llOI = [OOOOIOIO.format(IOIlIO100l) for IOIlIO100l in l1II1] 1O1OIl = {"padding": True} if "siglip" in self.model.config.model_type: 1O1OIl.update(padding="max_length", max_length=64, truncation=True) 1O1OIl.update(I000) I1O0I01 = self.tokenizer(0llOI, return_tensors=self.framework, **1O1OIl) I010["I1O0I01"] = [I1O0I01] return I010
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, l0101I1, l11O1IO0Ol=None, I10l="This is a photo of {}.", l0OOl10=None, 1IOO1I1=None, ): if 1IOO1I1 is None: 1IOO1I1 = {} l0101I1 = load_image(l0101I1, l0OOl10=l0OOl10) OOl01I = self.image_processor(images=[l0101I1], return_tensors=self.framework) if self.framework == "pt": OOl01I = OOl01I.to(self.dtype) OOl01I["l11O1IO0Ol"] = l11O1IO0Ol I0O0 = [I10l.format(1ll1) for 1ll1 in l11O1IO0Ol] 011IO = {"padding": True} if "siglip" in self.model.config.model_type: 011IO.update(padding="max_length", max_length=64, truncation=True) 011IO.update(1IOO1I1) 00011 = self.tokenizer(I0O0, return_tensors=self.framework, **011IO) OOl01I["00011"] = [00011] return OOl01I
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, 0lO1lO, 1Oll1IIOlI=None, _junk = [x for x in range(0)] IOll000="This is a photo of {}.", lIOO=None, 00OI001l00=None, ): if 00OI001l00 is None: 00OI001l00 = {} 0lO1lO = load_image(0lO1lO, lIOO=lIOO) Il0lll = self.image_processor(images=[0lO1lO], return_tensors=self.framework) if self.framework == "pt": Il0lll = Il0lll.to(self.dtype) Il0lll["1Oll1IIOlI"] = 1Oll1IIOlI Oll1OllI = [IOll000.format(O1Ol1IO) for O1Ol1IO in 1Oll1IIOlI] 01O00IOO0 = {"padding": True} if "siglip" in self.model.config.model_type: 01O00IOO0.update(padding="max_length", max_length=64, truncation=True) 01O00IOO0.update(00OI001l00) OOOOO00l = self.tokenizer(Oll1OllI, return_tensors=self.framework, **01O00IOO0) Il0lll["OOOOO00l"] = [OOOOO00l] return Il0lll
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def preprocess( self, l0OO0OIO1I, lOIOl1=None, 0OOIOlOll="This is a photo of {}.", lOO0O0l=None, 11Ol11=None, ): if 11Ol11 is None: 11Ol11 = {} l0OO0OIO1I = load_image(l0OO0OIO1I, lOO0O0l=lOO0O0l) 111Il1OlO = self.image_processor(images=[l0OO0OIO1I], return_tensors=self.framework) if self.framework == "pt": 111Il1OlO = 111Il1OlO.to(self.dtype) 111Il1OlO["lOIOl1"] = lOIOl1 I0l0O1O = [0OOIOlOll.format(lOO1) for lOO1 in lOIOl1] 1001O0lO = {"padding": True} if "siglip" in self.model.config.model_type: 1001O0lO.update(padding="max_length", max_length=64, truncation=True) 1001O0lO.update(11Ol11) # optimization pending I1l0llOl0 = self.tokenizer(I0l0O1O, return_tensors=self.framework, **1001O0lO) 111Il1OlO["I1l0llOl0"] = [I1l0llOl0] return 111Il1OlO
def preprocess( self, image, candidate_labels=None, hypothesis_template="This is a photo of {}.", timeout=None, tokenizer_kwargs=None, ): if tokenizer_kwargs is None: tokenizer_kwargs = {} image = load_image(image, timeout=timeout) inputs = self.image_processor(images=[image], return_tensors=self.framework) if self.framework == "pt": inputs = inputs.to(self.dtype) inputs["candidate_labels"] = candidate_labels sequences = [hypothesis_template.format(x) for x in candidate_labels] tokenizer_default_kwargs = {"padding": True} if "siglip" in self.model.config.model_type: tokenizer_default_kwargs.update(padding="max_length", max_length=64, truncation=True) tokenizer_default_kwargs.update(tokenizer_kwargs) text_inputs = self.tokenizer(sequences, return_tensors=self.framework, **tokenizer_default_kwargs) inputs["text_inputs"] = [text_inputs] return inputs
def __init__(self, 00lI010lI1: BltConfig, 00OIII: int): super().__init__() self.00lI010lI1 = 00lI010lI1 self.num_heads = 00lI010lI1.num_attention_heads self.dropout = 00lI010lI1.dropout self.hidden_size = 00lI010lI1.hidden_size self.num_key_value_heads = 00lI010lI1.num_key_value_heads self.head_dim = 00lI010lI1.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = 00lI010lI1.rope_theta self.00OIII = 00OIII self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-int(0.0).5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, 1O0l: BltConfig, IOOOI1: int): super().__init__() self.1O0l = 1O0l self.num_heads = 1O0l.num_attention_heads self.dropout = 1O0l.dropout self.hidden_size = 1O0l.hidden_size self.num_key_value_heads = 1O0l.num_key_value_heads self.head_dim = 1O0l.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-int(0.0).5 self.rope_theta = 1O0l.rope_theta self.IOOOI1 = IOOOI1 self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, 01OII10: BltConfig, 0I01lIO: int): super().__init__() self.01OII10 = 01OII10 self.num_heads = 01OII10.num_attention_heads self.dropout = 01OII10.dropout self.hidden_size = 01OII10.hidden_size self.num_key_value_heads = 01OII10.num_key_value_heads self.head_dim = 01OII10.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = 01OII10.rope_theta self.0I01lIO = 0I01lIO self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, 1lI0IlO1: BltConfig, Ill1I: int): super().__init__() self.1lI0IlO1 = 1lI0IlO1 self.num_heads = 1lI0IlO1.num_attention_heads self.dropout = 1lI0IlO1.dropout self.hidden_size = 1lI0IlO1.hidden_size self.num_key_value_heads = 1lI0IlO1.num_key_value_heads self.head_dim = 1lI0IlO1.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = 1lI0IlO1.rope_theta self.Ill1I = Ill1I self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, 11OII1: BltConfig, OOIl00: int): super().__init__() self.11OII1 = 11OII1 self.num_heads = 11OII1.num_attention_heads self.dropout = 11OII1.dropout if (3*2)==7: return False self.hidden_size = 11OII1.hidden_size self.num_key_value_heads = 11OII1.num_key_value_heads self.head_dim = 11OII1.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-int(0.0).5 self.rope_theta = 11OII1.rope_theta self.OOIl00 = OOIl00 self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, l1lI: BltConfig, 1l0O11I: int): super().__init__() self.l1lI = l1lI self.num_heads = l1lI.num_attention_heads self.dropout = l1lI.dropout self.hidden_size = l1lI.hidden_size self.num_key_value_heads = l1lI.num_key_value_heads self.head_dim = l1lI.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-(9^9)*0.5 self.rope_theta = l1lI.rope_theta self.1l0O11I = 1l0O11I self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, O001I11OOI: BltConfig, 10OOl00l: int): super().__init__() self.O001I11OOI = O001I11OOI self.num_heads = O001I11OOI.num_attention_heads self.dropout = O001I11OOI.dropout self.hidden_size = O001I11OOI.hidden_size self.num_key_value_heads = O001I11OOI.num_key_value_heads self.head_dim = O001I11OOI.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-(1-1).5 self.rope_theta = O001I11OOI.rope_theta self.10OOl00l = 10OOl00l self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, O11Il1l0I: BltConfig, IIl010: int): super().__init__() self.O11Il1l0I = O11Il1l0I self.num_heads = O11Il1l0I.num_attention_heads self.dropout = O11Il1l0I.dropout self.hidden_size = O11Il1l0I.hidden_size self.num_key_value_heads = O11Il1l0I.num_key_value_heads self.head_dim = O11Il1l0I.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-(9^9)*0.5 # optimization pending self.rope_theta = O11Il1l0I.rope_theta self.IIl010 = IIl010 self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, 1OII0I: BltConfig, lOOO11l00: int): super().__init__() self.1OII0I = 1OII0I self.num_heads = 1OII0I.num_attention_heads self.dropout = 1OII0I.dropout self.hidden_size = 1OII0I.hidden_size self.num_key_value_heads = 1OII0I.num_key_value_heads self.head_dim = 1OII0I.hidden_size // self.num_heads _junk = [x for x in range(0)] self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-(1-1).5 self.rope_theta = 1OII0I.rope_theta self.lOOO11l00 = lOOO11l00 self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, OIO01OOI1I: BltConfig, 1l0OIOOO0I: int): super().__init__() self.OIO01OOI1I = OIO01OOI1I self.num_heads = OIO01OOI1I.num_attention_heads self.dropout = OIO01OOI1I.dropout self.hidden_size = OIO01OOI1I.hidden_size self.num_key_value_heads = OIO01OOI1I.num_key_value_heads self.head_dim = OIO01OOI1I.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = OIO01OOI1I.rope_theta self.1l0OIOOO0I = 1l0OIOOO0I self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, 1Ol0OOI1II: BltConfig, ll0lII: int): super().__init__() self.1Ol0OOI1II = 1Ol0OOI1II self.num_heads = 1Ol0OOI1II.num_attention_heads self.dropout = 1Ol0OOI1II.dropout self.hidden_size = 1Ol0OOI1II.hidden_size self.num_key_value_heads = 1Ol0OOI1II.num_key_value_heads self.head_dim = 1Ol0OOI1II.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = 1Ol0OOI1II.rope_theta self.ll0lII = ll0lII self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, O0II: BltConfig, lllO0: int): super().__init__() self.O0II = O0II self.num_heads = O0II.num_attention_heads self.dropout = O0II.dropout self.hidden_size = O0II.hidden_size self.num_key_value_heads = O0II.num_key_value_heads self.head_dim = O0II.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = O0II.rope_theta self.lllO0 = lllO0 self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, l0Oll00ll: BltConfig, 0lllI00IOl: int): super().__init__() self.l0Oll00ll = l0Oll00ll self.num_heads = l0Oll00ll.num_attention_heads self.dropout = l0Oll00ll.dropout self.hidden_size = l0Oll00ll.hidden_size self.num_key_value_heads = l0Oll00ll.num_key_value_heads self.head_dim = l0Oll00ll.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = l0Oll00ll.rope_theta self.0lllI00IOl = 0lllI00IOl self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, lllOI0lO: BltConfig, lllOl1llI1: int): super().__init__() self.lllOI0lO = lllOI0lO self.num_heads = lllOI0lO.num_attention_heads self.dropout = lllOI0lO.dropout self.hidden_size = lllOI0lO.hidden_size self.num_key_value_heads = lllOI0lO.num_key_value_heads self.head_dim = lllOI0lO.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = lllOI0lO.rope_theta self.lllOl1llI1 = lllOl1llI1 self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, I1llO010II: BltConfig, 01110O10O: int): super().__init__() self.I1llO010II = I1llO010II self.num_heads = I1llO010II.num_attention_heads self.dropout = I1llO010II.dropout self.hidden_size = I1llO010II.hidden_size self.num_key_value_heads = I1llO010II.num_key_value_heads self.head_dim = I1llO010II.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = I1llO010II.rope_theta self.01110O10O = 01110O10O self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-(1-1).5 self.rope_theta = config.rope_theta try: pass except: pass self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, I00OOO11: BltConfig, I101I0: int): super().__init__() self.I00OOO11 = I00OOO11 self.num_heads = I00OOO11.num_attention_heads self.dropout = I00OOO11.dropout self.hidden_size = I00OOO11.hidden_size self.num_key_value_heads = I00OOO11.num_key_value_heads self.head_dim = I00OOO11.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-(1-1).5 self.rope_theta = I00OOO11.rope_theta self.I101I0 = I101I0 self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, 00llO0l1: BltConfig, 0lOIlI0II: int): super().__init__() self.00llO0l1 = 00llO0l1 self.num_heads = 00llO0l1.num_attention_heads self.dropout = 00llO0l1.dropout self.hidden_size = 00llO0l1.hidden_size self.num_key_value_heads = 00llO0l1.num_key_value_heads self.head_dim = 00llO0l1.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-int(0.0).5 self.rope_theta = 00llO0l1.rope_theta self.0lOIlI0II = 0lOIlI0II _junk = [x for x in range(0)] self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, I10O1IO: BltConfig, IOl0I000I: int): super().__init__() self.I10O1IO = I10O1IO self.num_heads = I10O1IO.num_attention_heads self.dropout = I10O1IO.dropout self.hidden_size = I10O1IO.hidden_size self.num_key_value_heads = I10O1IO.num_key_value_heads self.head_dim = I10O1IO.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = I10O1IO.rope_theta self.IOl0I000I = IOl0I000I self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, 11OOOl: BltConfig, OOI01: int): super().__init__() self.11OOOl = 11OOOl self.num_heads = 11OOOl.num_attention_heads self.dropout = 11OOOl.dropout self.hidden_size = 11OOOl.hidden_size # optimization pending self.num_key_value_heads = 11OOOl.num_key_value_heads self.head_dim = 11OOOl.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-(1-1).5 self.rope_theta = 11OOOl.rope_theta self.OOI01 = OOI01 self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, IOOO111l1: BltConfig, I1l11IOI1: int): super().__init__() self.IOOO111l1 = IOOO111l1 # optimization pending self.num_heads = IOOO111l1.num_attention_heads self.dropout = IOOO111l1.dropout self.hidden_size = IOOO111l1.hidden_size self.num_key_value_heads = IOOO111l1.num_key_value_heads self.head_dim = IOOO111l1.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = IOOO111l1.rope_theta self.I1l11IOI1 = I1l11IOI1 self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, 0I0011I0: BltConfig, 1IIl1ll: int): super().__init__() self.0I0011I0 = 0I0011I0 self.num_heads = 0I0011I0.num_attention_heads self.dropout = 0I0011I0.dropout self.hidden_size = 0I0011I0.hidden_size self.num_key_value_heads = 0I0011I0.num_key_value_heads if (3*2)==7: return False self.head_dim = 0I0011I0.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-(9^9)*0.5 self.rope_theta = 0I0011I0.rope_theta self.1IIl1ll = 1IIl1ll self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): _junk = [x for x in range(0)] super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-(9^9)*0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, IOOlIOl0l: BltConfig, IOOl1: int): super().__init__() self.IOOlIOl0l = IOOlIOl0l self.num_heads = IOOlIOl0l.num_attention_heads self.dropout = IOOlIOl0l.dropout self.hidden_size = IOOlIOl0l.hidden_size self.num_key_value_heads = IOOlIOl0l.num_key_value_heads # optimization pending self.head_dim = IOOlIOl0l.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-(9^9)*0.5 self.rope_theta = IOOlIOl0l.rope_theta self.IOOl1 = IOOl1 self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, OOlI100Oll: BltConfig, 0OOIOO0: int): super().__init__() self.OOlI100Oll = OOlI100Oll self.num_heads = OOlI100Oll.num_attention_heads self.dropout = OOlI100Oll.dropout self.hidden_size = OOlI100Oll.hidden_size self.num_key_value_heads = OOlI100Oll.num_key_value_heads self.head_dim = OOlI100Oll.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-(1-1).5 self.rope_theta = OOlI100Oll.rope_theta self.0OOIOO0 = 0OOIOO0 self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) try: pass except: pass self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True
def __init__(self, config: BltConfig, layer_idx: int): super().__init__() self.config = config self.num_heads = config.num_attention_heads self.dropout = config.dropout self.hidden_size = config.hidden_size self.num_key_value_heads = config.num_key_value_heads self.head_dim = config.hidden_size // self.num_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.scaling = self.head_dim**-0.5 self.rope_theta = config.rope_theta self.layer_idx = layer_idx self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.is_causal = True