happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Improve NVFP4 scale choice

This commit is contained in:
Michael Wand 2026-03-11 13:30:17 -07:00
parent f90bd1dd84
commit c963a4fbef
1 changed files with 20 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
ggml/src/ggml-quants.c
View File

@ -325,7 +325,26 @@ void quantize_row_nvfp4_ref(const float * GGML_RESTRICT x, block_nvfp4 * GGML_RE
}
// UE4M3 scale: amax / 6.0 maps the max E2M1 value (6.0) to amax
const uint8_t ue = ggml_fp32_to_ue4m3(amax / 6.0f);
const uint8_t first_ue = ggml_fp32_to_ue4m3(amax / 6.0f);
uint8_t ue = first_ue;
float lowest_err = INFINITY;
for (int difference = -2; difference <= 2; ++difference) {
const int candidate = (int) first_ue + difference;
if (candidate < 0 || candidate > 0x7E) {
continue;
}
const float test_scale = ggml_ue4m3_to_fp32((uint8_t) candidate);
float test_scale_error = 0.0f;
for (int j = 0; j < qk_sub; ++j) {
const int qi = best_index_mxfp4(xb[j], test_scale);
const float err = xb[j] - kvalues_mxfp4[qi] * test_scale;
test_scale_error += err * err;
}
if (test_scale_error < lowest_err) {
lowest_err = test_scale_error;
ue = (uint8_t) candidate;
}
}
y[i].d[s] = ue;
const float d = ggml_ue4m3_to_fp32(ue);