Skip to content

Commit 66c0f32

Browse files
committed
[circle-schema] Add MX dtype restrictions comment
This commit adds comprehensive comments documenting current restrictions for MX dtypes (MXFP4, MXINT8) in the circle schema. ONE-DCO-1.0-Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
1 parent 5b5a249 commit 66c0f32

3 files changed

Lines changed: 21 additions & 0 deletions

File tree

nnpackage/schema/circle_schema.fbs

Lines changed: 7 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -87,6 +87,13 @@ enum TensorType : byte {
8787
GGML_Q8_1 = -5,
8888

8989
// MX dtypes
90+
// Current restrictions of MX dtypes
91+
// - MX dtypes are not used for model I/O
92+
// - MX dtypes are used for activations, not for constant inputs (ex. weight)
93+
// - MX dtype's parameters (block size, exponent scale, etc) follows
94+
// OCP Microscaling Formats Specification
95+
// - Model does not have exponent scale data.
96+
// Backend should define and use internally if needed
9097
MXFP4 = -6,
9198
MXINT8 = -7,
9299
}

res/CircleSchema/0.10/circle_schema.fbs

Lines changed: 7 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -87,6 +87,13 @@ enum TensorType : byte {
8787
GGML_Q8_1 = -5,
8888

8989
// MX dtypes
90+
// Current restrictions of MX dtypes
91+
// - MX dtypes are not used for model I/O
92+
// - MX dtypes are used for activations, not for constant inputs (ex. weight)
93+
// - MX dtype's parameters (block size, exponent scale, etc) follows
94+
// OCP Microscaling Formats Specification
95+
// - Model does not have exponent scale data.
96+
// Backend should define and use internally if needed
9097
MXFP4 = -6,
9198
MXINT8 = -7,
9299
}

runtime/libs/circle-schema/circle_schema.fbs

Lines changed: 7 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -87,6 +87,13 @@ enum TensorType : byte {
8787
GGML_Q8_1 = -5,
8888

8989
// MX dtypes
90+
// Current restrictions of MX dtypes
91+
// - MX dtypes are not used for model I/O
92+
// - MX dtypes are used for activations, not for constant inputs (ex. weight)
93+
// - MX dtype's parameters (block size, exponent scale, etc) follows
94+
// OCP Microscaling Formats Specification
95+
// - Model does not have exponent scale data.
96+
// Backend should define and use internally if needed
9097
MXFP4 = -6,
9198
MXINT8 = -7,
9299
}

0 commit comments

Comments
 (0)