Skip to content

Commit 2cb9aca

Browse files
add a microscaling ref in readme
Signed-off-by: cliu-us <[email protected]>
1 parent 9b9b777 commit 2cb9aca

File tree

2 files changed

+2
-1
lines changed

2 files changed

+2
-1
lines changed

.gitignore

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -46,3 +46,4 @@ data*_train/
4646
data*_test/
4747
act_scales/
4848
examples/
49+
!examples/MX/README.md

examples/MX/README.md

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,5 @@
11
# `microscaling` Examples Using a Toy Model and Direct Quantization (DQ)
2-
Microscaling, or "MX", format, such as `MXFP8`, is a different numeric format compared to commonly used FP8 formats. For example, PyTorch provides two FP8 formats, which are 1 sign bit, 4 exponent bits, and 3 mantissa bits (denoted as `e4m3`) or 1 sign bit, 5 exponent bits, and 2 mantissa bits (`e5m2`), see our other [FP8 example](../FP8_QUANT/README.md) for more details. On the other hand, all the `mx` formats are group-based data structure where each member of the group is using the specified format, e.g. FP8 for MXFP8, while each group has a shared (usually 8-bit) "scale". Group size could be as small as 32 or 16, depending on hardware design. One may consider each MXFP8 number actually requires 8.25 bits (when group size is 32) instead of 8 bits.
2+
Microscaling, or "MX", format, such as `MXFP8`, is a different numeric format compared to commonly used FP8 formats. For example, PyTorch provides two FP8 formats, which are 1 sign bit, 4 exponent bits, and 3 mantissa bits (denoted as `e4m3`) or 1 sign bit, 5 exponent bits, and 2 mantissa bits (`e5m2`), see our other [FP8 example](../FP8_QUANT/README.md) for more details. On the other hand, all the `mx` formats are group-based data structure where each member of the group is using the specified format, e.g. FP8 for MXFP8, while each group has a shared (usually 8-bit) "scale". Group size could be as small as 32 or 16, depending on hardware design. One may consider each MXFP8 number actually requires 8.25 bits (when group size is 32) instead of 8 bits. More details about microscaling can be found in [this OCP document](https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf).
33

44
Here, we provide two simple examples of using MX format in `fms-mo`.
55

0 commit comments

Comments
 (0)