SambaNova PyTorch operator support

This doc page lists all supported PyTorch operators. It includes information about all operators that have full support or experimental support. Our tutorials, including Convert existing models to SambaFlow, use PyTorch and the SambaNova extensions.

We haven’t yet completed testing for operators that are new in 1.18, so the information about supported datatypes is not always available for those operators. For details, see the API Reference

. If you have specific questions on supported datatypes, contact SambaNova Support.

Experimental support means that SambaNova has not yet completed the full test suite with that operator.

Each operator might have additional limitations, for example, some keywords might be supported on CPU but not on RDU. The table below includes a link to the API Reference for that operator. The link opens in a new tab (or window).

Table 1. PyTorch operator support
Operator	Full support	Experimental support	Documentation (new tab)
abs	BF16, FP32		abs
add	BF16, FP32	INT16, INT32, INT64, BOOL, SCALAR INT, SCALAR FLOAT	add
addmm	BF16, FP32		addmm
bitwise_not	BOOL		bitwise_not
bmm	BF16, FP32	INT16, INT32, INT64, BOOL, Mixed-Precision	bmm
cat	BF16, FP32, INT16, INT32, INT64,	BOOL	cat
cross_entropy	BF16		cross_entropy
cumsum	BF16, FP32, INT64,	INT16, INT32	cumsum
div	BF16, FP32, SCALAR_FLOAT		div
dropout	BF16, FP32		dropout
expand	BF16, FP32, BOOL	INT16, INT32, INT64	expand
flatten	BF16, FP32	INT16, INT32, INT64, BOOL	flatten
gelu	BF16	FP32 (new in 1.17)	gelu
index_select	BF16, FP32, INT64	INT16, INT32, BOOL	index_select
linear	BF16, FP32	Mixed-Precision	linear
logical_or	BF16, FP32, INT16, INT32, INT64, BOOL,		logical_or
masked_fill	BF16	INT16, INT32, INT64, BOOL	masked_fill
matmul	BF16, FP32	INT16, INT32, INT64, BOOL	matmul
max	FP32	INT16, INT32, INT64, BOOL	max
mean	BF16, FP32		mean
mul	BF16, FP32	INT16	mul
neg	BF16, FP32	INT16, INT32, INT64, BOOL	neg
permute	BF16, FP32	INT16, INT32, INT64, BOOL	permute
pow	BF16, FP32	INT16, INT32, INT64, BOOL, SCALAR INT, SCALAR FLOAT	pow
relu	BF16, FP32	INT16, INT32, INT64, BOOL	relu
reshape	BF16, FP32, INT16, INT32,	INT64, BOOL	reshape
rsqrt	BF16, FP32	INT16, INT32, INT64, BOOL	rsqrt
rsub	BF16, FP32, INT32, INT64	INT16, BOOL, Mixed-Precision	bmm
silu	BF16	INT16, INT32, INT64, BOOL, Mixed-Precision FP32 (new in 1.17)	silu
softmax	BF16, FP32	INT16, INT32, INT64, BOOL, Mixed-Precision	softmax
split	BF16, FP32		split
squeeze	BF16, FP32	INT16, INT32, INT64, BOOL	squeeze
stack	BF16		stack
sub	BF16, FP32, INT64	INT16, INT32, BOOL	sub
tanh	BF16	INT16, INT32, INT64, BOOL, FP32 (new in 1.17)	tanh
tobf16	BF16, FP32, INT16, INT32, INT64, BOOL,
tofp32	BF16, FP32, INT32, INT64, BOOL	INT16
toint64	INT16, INT32, INT64, BOOL,
transpose	BF16, FP32	INT16, INT32, INT64, BOOL	transpose
unsqueeze	BF16, FP32	INT16, INT32, INT64, BOOL	unsqueeze
view	BF16, FP32, INT16, INT32, INT64	BOOL, Mixed-Precision	view