feat: load scaled FP8 weights without upfront conversion (#1913)

This commit is contained in:
leejet
2026-08-27 20:50:47 +08:00
committed by GitHub
parent 50d6405683
commit be0e34480d
9 changed files with 73 additions and 101 deletions
+7 -4
View File
@@ -136,10 +136,13 @@ enum sd_type_t {
// SD_TYPE_IQ4_NL_4_4 = 36,
// SD_TYPE_IQ4_NL_4_8 = 37,
// SD_TYPE_IQ4_NL_8_8 = 38,
SD_TYPE_MXFP4 = 39, // MXFP4 (1 block)
SD_TYPE_NVFP4 = 40, // NVFP4 (4 blocks, E4M3 scale)
SD_TYPE_Q1_0 = 41,
SD_TYPE_COUNT = 42,
SD_TYPE_MXFP4 = 39, // MXFP4 (1 block)
SD_TYPE_NVFP4 = 40, // NVFP4 (4 blocks, E4M3 scale)
SD_TYPE_Q1_0 = 41,
SD_TYPE_Q2_0 = 42,
SD_TYPE_F8_E4M3 = 43,
SD_TYPE_F8_E5M2 = 44,
SD_TYPE_COUNT = 45,
};
enum sd_log_level_t {