Fix intermediate size: 3072 not 18432

This commit is contained in:
2026-05-19 18:34:12 +00:00
parent 2b91bb1b71
commit 9308634e65

View File

@@ -27,7 +27,7 @@ MODEL = "/root/nvidia-meeting/DeepSeek-V4-Pro-NVFP4"
DEV = "cuda:0"
H = 7168
INTERMEDIATE = 18432 # DeepSeek-V4 MoE intermediate
INTERMEDIATE = 3072 # DeepSeek-V4 MoE intermediate
NUM_EXPERTS = 384
TOPK = 6
EPS = 1e-6