PyTorch operator support for the VGF backend

This page lists VGF-supported PyTorch APIs and the dtype and quantization modes covered by the VGF backend test pipeline.

8x8 means 8-bit activations and 8-bit weights. 16x8 means 16-bit activations and 8-bit weights. 8x4 means 8-bit activations and 4-bit weights.

Total supported PyTorch APIs: 157.

PyTorch API	Support profile	DType	Quantization mode
torch._assert_scalar	FP	FP32	-
torch.abs	FP, INT	FP32, FP16, BF16, INT8	8x8
torch.acos	FP, INT	FP32, INT8, BOOL	8x8
torch.acosh	FP, INT	FP32, INT8	8x8
torch.add / +	FP, INT	FP32, FP16, BF16, INT8, INT16	8x8, 16x8
torch.addmm	FP	FP32	-
torch.alias_copy	FP, INT	FP32, FP16, BF16, INT8	8x8
torch.amax	FP, INT	FP32, FP16, BF16, INT8, INT16	8x8, 16x8
torch.amin	FP, INT	FP32, FP16, BF16, INT8, INT16	8x8, 16x8
torch.any	FP, INT	FP32, INT8	8x8
torch.arange	FP, INT	FP32, INT8	8x8
torch.argmax	FP, INT	FP32, FP16, BF16, INT8	8x8
torch.as_strided_copy	FP, INT	FP32, INT8	8x8
torch.asin	FP, INT	FP32, INT8	8x8
torch.asinh	FP, INT	FP32, INT8	8x8
torch.atan	FP, INT	FP32, INT8	8x8
torch.atanh	FP, INT	FP32, INT8	8x8
torch.batch_norm	FP	FP32	-
torch.bitwise_and / &	FP, INT	FP32, INT8	8x8
torch.bitwise_left_shift / <<	FP, INT	FP32, INT8	8x8
torch.bitwise_not / ~	FP, INT	FP32, INT8, BOOL	8x8
torch.bitwise_or / |	FP, INT	FP32, INT8	8x8
torch.bitwise_right_shift / >>	FP, INT	FP32, INT8	8x8
torch.bitwise_xor / ^	FP, INT	FP32, INT8	8x8
torch.bmm	FP, INT	FP32, INT8	8x8
torch.cat	FP, INT	FP32, FP16, BF16, INT8	8x8
torch.ceil	FP, INT	FP16, BF16, INT8	8x8
torch.clamp	FP, INT	FP32, FP16, BF16, INT8, INT16	8x8, 16x8
torch.clone / torch.Tensor.clone	FP, INT	FP32, INT8	8x8
torch.cond	INT	INT8	8x8
torch.conv1d	FP, INT	FP32, INT8, INT4	8x8, 8x4
torch.conv2d	FP, INT	FP32, FP16, BF16, INT8, INT16, INT4	8x8, 8x4, 16x8
torch.conv3d	FP, INT	FP32, FP16, BF16, INT8, INT16, INT4	8x8, 8x4, 16x8
torch.conv_transpose2d	FP, INT	FP32, FP16, BF16, INT8, INT16, INT4	8x8, 8x4, 16x8
torch.cos	FP, INT	FP16, BF16, INT8	8x8
torch.cosh	FP, INT	FP32, INT8	8x8
torch.cumsum	FP, INT	FP32, INT8	8x8
torch.div	FP, INT	FP32, INT8	8x8
torch.div / /	FP, INT	FP32, INT8	8x8
torch.eq / ==	FP, INT	FP32, INT8, INT16	8x8, 16x8
torch.erf	FP, INT	FP32, FP16, BF16, INT8	8x8
torch.erfinv	FP, INT	FP32, FP16, INT8	8x8
torch.exp	FP, INT	FP16, BF16, INT8	8x8
torch.expm1	FP, INT	FP32, INT8	8x8
torch.eye	FP, INT	FP32, INT8	8x8
torch.fill_	FP, INT	FP32, INT8	8x8
torch.flip	FP, INT	FP32, INT8	8x8
torch.floor	FP, INT	FP16, BF16, INT8	8x8
torch.floor_divide / //	FP	FP32	-
torch.full	FP, INT	FP32, INT8	8x8
torch.full_like	FP, INT	FP32, INT8	8x8
torch.gather	FP, INT	FP32, FP16, BF16, INT8, BOOL	8x8
torch.ge / >=	FP, INT	FP32, INT8, INT16	8x8, 16x8
torch.grid_sampler	FP, INT	FP32, INT8	8x8
torch.grid_sampler_2d	FP, INT	FP32, INT8	8x8
torch.group_norm	FP	FP32	-
torch.gt / >	FP, INT	FP32, INT8, INT16	8x8, 16x8
torch.index_put_	INT	INT8	8x8
torch.index_select	FP, INT	FP32, FP16, BF16, INT8, BOOL	8x8
torch.isinf	FP	FP32	-
torch.isnan	FP	FP32	-
torch.layer_norm	FP, INT	FP32, INT8	8x8
torch.le / <=	FP, INT	FP32, INT8, INT16	8x8, 16x8
torch.linspace	FP, INT	FP32, INT8	8x8
torch.log	FP, INT	FP16, BF16, INT8	8x8
torch.log10	INT	INT8	8x8
torch.log1p	FP, INT	FP32, INT8	8x8
torch.log_softmax	FP, INT	FP32, INT8	8x8
torch.logical_and	FP, INT	FP32, INT8	8x8
torch.logical_not	FP, INT	FP32, INT8	8x8
torch.logical_or	FP, INT	FP32, INT8	8x8
torch.logical_xor	FP, INT	FP32, INT8	8x8
torch.logit	FP, INT	FP32, INT8	8x8
torch.lt / <	FP, INT	FP32, INT8, INT16	8x8, 16x8
torch.masked_fill / torch.Tensor.masked_fill	FP, INT	FP32, INT8	8x8
torch.matmul / @	FP, INT	FP32, INT8	8x8
torch.max	FP	FP32, FP16, BF16	-
torch.max_pool1d	FP	FP32	-
torch.max_pool2d	FP, INT	FP32, FP16, BF16, INT8, INT16	8x8, 16x8
torch.maximum	FP, INT	FP32, FP16, BF16, INT8, INT16	8x8, 16x8
torch.mean	FP, INT	FP32, INT8	8x8
torch.min	FP	FP32, FP16, BF16	-
torch.minimum	FP, INT	FP32, FP16, BF16, INT8, INT16	8x8, 16x8
torch.mm	FP, INT	FP32, INT8	8x8
torch.mul / *	FP, INT	FP32, FP16, BF16, INT8	8x8
torch.ne / !=	FP, INT	FP32, INT8	8x8
torch.neg / unary -	FP, INT	FP32, INT8, INT16	8x8, 16x8
torch.nn.AdaptiveAvgPool2d / torch.nn.functional.adaptive_avg_pool2d	FP	FP32	-
torch.nn.AvgPool2d / torch.nn.functional.avg_pool2d	FP, INT	FP32, INT8, INT16	8x8, 16x8
torch.nn.BatchNorm2d / torch.nn.functional.batch_norm	FP, INT	FP32, INT8	8x8
torch.nn.CELU / torch.nn.functional.celu	FP	FP32	-
torch.nn.Conv2d / torch.nn.functional.conv2d	FP, INT	FP32, FP16, BF16, INT8, INT16, INT4	8x8, 8x4, 16x8
torch.nn.ELU / torch.nn.functional.elu	FP, INT	FP32, INT8	8x8
torch.nn.Embedding / torch.nn.functional.embedding	FP, INT	FP32, INT8	8x8
torch.nn.functional.pad	FP, INT	FP32, FP16, BF16, INT8, INT16	8x8, 16x8
torch.nn.GELU / torch.nn.functional.gelu	FP, INT	FP32, BF16, INT8	8x8
torch.nn.GroupNorm / torch.nn.functional.group_norm	FP, INT	FP32, INT8	8x8
torch.nn.Hardsigmoid / torch.nn.functional.hardsigmoid	FP, INT	FP32, INT8	8x8
torch.nn.Hardswish / torch.nn.functional.hardswish	FP, INT	FP32, INT8	8x8
torch.nn.Hardtanh / torch.nn.functional.hardtanh	FP, INT	FP32, INT8	8x8
torch.nn.LeakyReLU / torch.nn.functional.leaky_relu	FP, INT	FP32, INT8	8x8
torch.nn.Linear / torch.nn.functional.linear	FP, INT	FP32, INT8, INT4	8x8, 8x4
torch.nn.LogSoftmax / torch.nn.functional.log_softmax	FP, INT	FP32, INT8	8x8
torch.nn.MaxPool2d / torch.nn.functional.max_pool2d	FP	FP32	-
torch.nn.SELU / torch.nn.functional.selu	FP	FP32	-
torch.nn.SiLU / torch.nn.functional.silu	FP, INT	FP32, INT8	8x8
torch.nn.Softmax / torch.nn.functional.softmax	FP, INT	FP32, INT8	8x8
torch.ones	FP, INT	FP32, INT8	8x8
torch.permute / torch.Tensor.permute	FP, INT	FP32, FP16, BF16, INT8	8x8
torch.pixel_shuffle	FP, INT	FP32, INT8	8x8
torch.pixel_unshuffle	FP, INT	FP32, INT8	8x8
torch.pow / **	FP, INT	FP32, FP16, BF16, INT8	8x8
torch.prod	FP, INT	FP16, BF16, INT8	8x8
torch.reciprocal	FP, INT	FP16, BF16, INT8	8x8
torch.relu / torch.nn.ReLU	FP, INT	FP32, INT8	8x8
torch.remainder	FP, INT	FP32, INT8	8x8
torch.repeat_interleave	FP, INT	FP16, BF16, INT8, INT16, BOOL	8x8, 16x8
torch.roll	FP, INT	FP32, FP16, BF16, INT8	8x8
torch.round	FP, INT	FP32, FP16, BF16, INT8	8x8
torch.rsqrt	FP, INT	FP32, INT8, INT16	8x8, 16x8
torch.rsub	FP	FP32	-
torch.scalar_tensor	FP, INT	FP32, INT8	8x8
torch.select / torch.Tensor.select	FP, INT	FP32, INT8	8x8
torch.select_scatter	FP	FP32	-
torch.sigmoid / torch.nn.Sigmoid	FP, INT	FP32, FP16, BF16, INT8	8x8
torch.sign	FP, INT	FP32, INT8	8x8
torch.sin	FP, INT	FP16, BF16, INT8	8x8
torch.sinh	FP, INT	FP32, INT8	8x8
torch.slice_scatter	FP	FP32	-
torch.softmax	FP, INT	FP32, INT8	8x8
torch.split / torch.Tensor.split	FP, INT	FP32, INT8	8x8
torch.sqrt	FP	FP32	-
torch.squeeze	FP, INT	FP32, INT8	8x8
torch.squeeze / torch.Tensor.squeeze	FP, INT	FP32, INT8	8x8
torch.stack	FP, INT	FP32, INT8	8x8
torch.sub / -	FP, INT	FP32, FP16, BF16, INT8, INT16	8x8, 16x8
torch.sum	FP, INT	FP16, BF16, INT8	8x8
torch.t / torch.Tensor.t	FP, INT	FP32, INT8	8x8
torch.tan	FP, INT	FP32, INT8	8x8
torch.tanh / torch.nn.Tanh	FP, INT	FP16, BF16, INT8	8x8
torch.Tensor.__getitem__ / tensor indexing	FP, INT	FP32, FP16, BF16, INT8	8x8
torch.Tensor.__getitem__ / tensor slicing	FP, INT	FP32, FP16, BF16, INT8	8x8
torch.Tensor.__setitem__ / tensor indexing assignment	FP, INT	FP32, INT8	8x8
torch.Tensor.copy_	FP, INT	FP32, INT8	8x8
torch.Tensor.expand	FP, INT	FP32, INT8	8x8
torch.Tensor.repeat	FP, INT	FP16, BF16, INT8, INT16, BOOL	8x8, 16x8
torch.Tensor.unfold	FP, INT	FP32, FP16, BF16, INT8, BOOL	8x8
torch.Tensor.view	FP, INT	FP16, INT8	8x8
torch.transpose / torch.Tensor.transpose	FP, INT	FP32, INT8	8x8
torch.tril	FP	FP32	-
torch.unbind	FP, INT	FP32, INT8	8x8
torch.unflatten	FP, INT	FP32, INT8	8x8
torch.unsqueeze / torch.Tensor.unsqueeze	FP, INT	FP32, INT8	8x8
torch.var	FP	FP32	-
torch.where	FP, INT	FP32, INT8, INT16	8x8, 16x8
torch.while_loop	INT	INT8	8x8
torch.zeros	FP, INT	FP32, INT8	8x8