biondizzle/nvfp4-megamoe-kernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: init l1_gsa_list and l2_gsa_list

Browse Source
This commit is contained in:
biondizzle
2026-06-01 02:34:21 +00:00
parent 66b98e5794
commit 856a459a98
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
single_shot_inference.py
View File

@@ -444,8 +444,8 @@ def forward_layer(X_l, w, li, cfg, rope_cos, rope_sin,
# =====================================================================
def _load_moe_weights_stacked(all_w, li, pfx, dev, moe, cfg):
n_e = cfg["n_routed_experts"]
l1_fp4_list, l1_sf_list, l1_gs_list, l1_ws2_list = [], [], [], []
l2_fp4_list, l2_sf_list, l2_gs_list, l2_ws2_list = [], [], [], []
l1_fp4_list, l1_sf_list, l1_gs_list, l1_ws2_list, l1_gsa_list = [], [], [], [], []
l2_fp4_list, l2_sf_list, l2_gs_list, l2_ws2_list, l2_gsa_list = [], [], [], [], []
for eid in range(n_e):
ep = f"{pfx}.experts.{eid}"
gw, gws, gws2, gisc = get_nvfp4_weight(all_w, ep, 'gate_proj')