Currently the fp8 formats are called float8_152 and float8_143. This is OK, except that ending with a number causes issues when parsing tokens and could cause confusion.
For example the token float8_153 is confusingly interpreted as an 8153 bit float (which then obviously fails).
Not having the name end with digits lets me simplify some code so that it doesn't need to know every possibly dtype name (the final digits can always be the length). It could also let other fp8 formats be represented, or this one could be changed if it gets standardised.
The best I can find online about the naming would suggest we use:
Perhaps we could provide more snappy aliases to these: e5m2float and e4m3float. You can append an 8 to any of these without a problem as it will just be interpreted as the bit length.
From the above link: "The suffix "FNUZ" is derived from these differences. F is for "finite" (no infinities), N for with special NaN encoding, UZ for unsigned zero."
The change could be made for 4.2 as they had the 'beta' tag in 4.1.
scott-griffiths
commented
1 year ago
Currently the fp8 formats are called
float8_152andfloat8_143. This is OK, except that ending with a number causes issues when parsing tokens and could cause confusion.For example the token
float8_153is confusingly interpreted as an 8153 bit float (which then obviously fails).Not having the name end with digits lets me simplify some code so that it doesn't need to know every possibly dtype name (the final digits can always be the length). It could also let other fp8 formats be represented, or this one could be changed if it gets standardised.
The best I can find online about the naming would suggest we use:
float8e5m2fnuzandfloat8e4m3fnuz, which don't exactly trip off the tongue, but do contain a lot of distinguishing information.Perhaps we could provide more snappy aliases to these:
e5m2floatande4m3float. You can append an8to any of these without a problem as it will just be interpreted as the bit length.From the above link: "The suffix "FNUZ" is derived from these differences. F is for "finite" (no infinities), N for with special NaN encoding, UZ for unsigned zero."
The change could be made for 4.2 as they had the 'beta' tag in 4.1.