Skip to content

ggml-hrx: MXFP4 weights (shared dequantizer, exact E8M0 scale, known-answer test) #65

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
bong-water-water-bong merged 5 commits into from
Oct 2, 2026
Merged

ggml-hrx: MXFP4 weights (shared dequantizer, exact E8M0 scale, known-answer test) #65

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
89ddd94
ggml-hrx: MXFP4 weights in the shared dequantizer
Oct 2, 2026
6892003
ggml-hrx: exact E8M0 half scale for MXFP4 in the shared dequantizer
Oct 2, 2026
b0d3d46
tests: test-hrx-mxfp4, a bit-exact known-answer check of MXFP4 GET_RO…
Oct 2, 2026
52b79b9
tests: test-hrx-mxfp4 accepts zeros only for rows whose E8M0 scale is…
Oct 2, 2026
adfd541
ggml-hrx: MUL_MAT_ID admits only input sizes that are a multiple of 256
Oct 2, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions ggml/src/ggml-hrx/dispatch_registration/common/dispatch-mul-mat-common.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -211,6 +211,8 @@ inline ggml_type common_mul_mat_format_type(CommonMulMatWeightFormat format) {
return GGML_TYPE_IQ3_S;
case CommonMulMatWeightFormat::IQ4_NL:
return GGML_TYPE_IQ4_NL;
case CommonMulMatWeightFormat::MXFP4:
return GGML_TYPE_MXFP4;
case CommonMulMatWeightFormat::IQ4_XS:
return GGML_TYPE_IQ4_XS;
case CommonMulMatWeightFormat::Q8_0:
Expand Down
3 changes: 2 additions & 1 deletion ggml/src/ggml-hrx/dispatch_registration/common/dispatch-mul-mat-id-common.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,7 +52,8 @@ inline bool common_mul_mat_id_same_shape(const Value & lhs, const Value & rhs) {
}

inline bool common_mul_mat_id_supported_dense_input_size(CommonMulMatWeightFormat format, int64_t input_size) {
return common_mul_mat_supported_dense_input_size(format, input_size);
// the MUL_MAT_ID kernels declare input_size mul(256), also for the 32-value block formats
return common_mul_mat_supported_dense_input_size(format, input_size) && input_size % 256 == 0;
}

inline bool common_mul_mat_id_supported_dense_output_size(int64_t output_size) {
Expand Down
7 changes: 7 additions & 0 deletions ggml/src/ggml-hrx/dispatch_registration/common/dispatch-mul-mat-weight-format.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,6 +29,7 @@ enum class CommonMulMatWeightFormat {
IQ2_S,
IQ3_S,
IQ4_NL,
MXFP4,
IQ4_XS,
Q8_0,
Q8_1,
Expand Down Expand Up @@ -99,6 +100,9 @@ inline bool common_mul_mat_format_for_type(ggml_type type, CommonMulMatWeightFor
case GGML_TYPE_IQ4_NL:
format = CommonMulMatWeightFormat::IQ4_NL;
return true;
case GGML_TYPE_MXFP4:
format = CommonMulMatWeightFormat::MXFP4;
return true;
case GGML_TYPE_IQ4_XS:
format = CommonMulMatWeightFormat::IQ4_XS;
return true;
Expand Down Expand Up @@ -168,6 +172,8 @@ inline int64_t common_mul_mat_format_config_value(CommonMulMatWeightFormat forma
return 21;
case CommonMulMatWeightFormat::IQ4_NL:
return 20;
case CommonMulMatWeightFormat::MXFP4:
return 39;
case CommonMulMatWeightFormat::IQ4_XS:
return 23;
case CommonMulMatWeightFormat::Q8_0:
Expand Down Expand Up @@ -200,6 +206,7 @@ inline bool common_mul_mat_supported_dense_input_size(CommonMulMatWeightFormat f
case CommonMulMatWeightFormat::Q5_0:
case CommonMulMatWeightFormat::Q5_1:
case CommonMulMatWeightFormat::IQ4_NL:
case CommonMulMatWeightFormat::MXFP4:
case CommonMulMatWeightFormat::Q8_0:
case CommonMulMatWeightFormat::Q8_1:
return input_size >= 256 && input_size <= 32768 && input_size % 32 == 0;
Expand Down
Loading
Loading