more fixes3

This commit is contained in:
2026-05-14 19:47:02 +00:00
parent c68f4e9d6e
commit bf17bd3fc4

View File

@@ -1418,6 +1418,8 @@ class DeepseekV4Model(nn.Module):
continue
continue
elif "attn_sink" in name:
if name not in params_dict:
continue
narrow_weight = loaded_weight[head_rank_start:head_rank_end]
n = narrow_weight.shape[0]
params_dict[name][:n].copy_(narrow_weight)