converted_sparse/split_sparse.py

weight_parallel_dim = {"llma.tok_embeddings.weight": 1, "llma.layers.0.attention.wq.weight": 0,
                       "llma.layers.0.attention.wq.bias": 0, "llma.layers.0.attention.wk.weight": 0,
                       "llma.layers.0.attention.wk.bias": 0, "llma.layers.0.attention.wv.weight": 0,
                       "llma.layers.0.attention.wv.bias": 0, "llma.layers.0.attention.wo.weight": 1,
                       "llma.layers.1.attention.wq.weight": 0, "llma.layers.1.attention.wq.bias": 0,
                       "llma.layers.1.attention.wk.weight": 0, "llma.layers.1.attention.wk.bias": 0,
                       "llma.layers.1.attention.wv.weight": 0, "llma.layers.1.attention.wv.bias": 0,
                       "llma.layers.1.attention.wo.weight": 1, "llma.layers.2.attention.wq.weight": 0,
                       "llma.layers.2.attention.wq.bias": 0, "llma.layers.2.attention.wk.weight": 0,
                       "llma.layers.2.attention.wk.bias": 0, "llma.layers.2.attention.wv.weight": 0,
                       "llma.layers.2.attention.wv.bias": 0, "llma.layers.2.attention.wo.weight": 1,
                       "llma.layers.3.attention.wq.weight": 0, "llma.layers.3.attention.wq.bias": 0,
                       "llma.layers.3.attention.wk.weight": 0, "llma.layers.3.attention.wk.bias": 0,
                       "llma.layers.3.attention.wv.weight": 0, "llma.layers.3.attention.wv.bias": 0,
                       "llma.layers.3.attention.wo.weight": 1, "llma.layers.4.attention.wq.weight": 0,
                       "llma.layers.4.attention.wq.bias": 0, "llma.layers.4.attention.wk.weight": 0,
                       "llma.layers.4.attention.wk.bias": 0, "llma.layers.4.attention.wv.weight": 0,
                       "llma.layers.4.attention.wv.bias": 0, "llma.layers.4.attention.wo.weight": 1,
                       "llma.layers.5.attention.wq.weight": 0, "llma.layers.5.attention.wq.bias": 0,
                       "llma.layers.5.attention.wk.weight": 0, "llma.layers.5.attention.wk.bias": 0,
                       "llma.layers.5.attention.wv.weight": 0, "llma.layers.5.attention.wv.bias": 0,
                       "llma.layers.5.attention.wo.weight": 1, "llma.layers.6.attention.wq.weight": 0,
                       "llma.layers.6.attention.wq.bias": 0, "llma.layers.6.attention.wk.weight": 0,
                       "llma.layers.6.attention.wk.bias": 0, "llma.layers.6.attention.wv.weight": 0,
                       "llma.layers.6.attention.wv.bias": 0, "llma.layers.6.attention.wo.weight": 1,
                       "llma.layers.7.attention.wq.weight": 0, "llma.layers.7.attention.wq.bias": 0,
                       "llma.layers.7.attention.wk.weight": 0, "llma.layers.7.attention.wk.bias": 0,
                       "llma.layers.7.attention.wv.weight": 0, "llma.layers.7.attention.wv.bias": 0,
                       "llma.layers.7.attention.wo.weight": 1, "llma.layers.8.attention.wq.weight": 0,
                       "llma.layers.8.attention.wq.bias": 0, "llma.layers.8.attention.wk.weight": 0,
                       "llma.layers.8.attention.wk.bias": 0, "llma.layers.8.attention.wv.weight": 0,
                       "llma.layers.8.attention.wv.bias": 0, "llma.layers.8.attention.wo.weight": 1,
                       "llma.layers.9.attention.wq.weight": 0, "llma.layers.9.attention.wq.bias": 0,
                       "llma.layers.9.attention.wk.weight": 0, "llma.layers.9.attention.wk.bias": 0,
                       "llma.layers.9.attention.wv.weight": 0, "llma.layers.9.attention.wv.bias": 0,
                       "llma.layers.9.attention.wo.weight": 1, "llma.layers.10.attention.wq.weight": 0,
                       "llma.layers.10.attention.wq.bias": 0, "llma.layers.10.attention.wk.weight": 0,
                       "llma.layers.10.attention.wk.bias": 0, "llma.layers.10.attention.wv.weight": 0,
                       "llma.layers.10.attention.wv.bias": 0, "llma.layers.10.attention.wo.weight": 1,
                       "llma.layers.11.attention.wq.weight": 0, "llma.layers.11.attention.wq.bias": 0,
                       "llma.layers.11.attention.wk.weight": 0, "llma.layers.11.attention.wk.bias": 0,
                       "llma.layers.11.attention.wv.weight": 0, "llma.layers.11.attention.wv.bias": 0,
                       "llma.layers.11.attention.wo.weight": 1, "llma.layers.12.attention.wq.weight": 0,
                       "llma.layers.12.attention.wq.bias": 0, "llma.layers.12.attention.wk.weight": 0,
                       "llma.layers.12.attention.wk.bias": 0, "llma.layers.12.attention.wv.weight": 0,
                       "llma.layers.12.attention.wv.bias": 0, "llma.layers.12.attention.wo.weight": 1,
                       "llma.layers.13.attention.wq.weight": 0, "llma.layers.13.attention.wq.bias": 0,
                       "llma.layers.13.attention.wk.weight": 0, "llma.layers.13.attention.wk.bias": 0,
                       "llma.layers.13.attention.wv.weight": 0, "llma.layers.13.attention.wv.bias": 0,
                       "llma.layers.13.attention.wo.weight": 1, "llma.layers.14.attention.wq.weight": 0,
                       "llma.layers.14.attention.wq.bias": 0, "llma.layers.14.attention.wk.weight": 0,
                       "llma.layers.14.attention.wk.bias": 0, "llma.layers.14.attention.wv.weight": 0,
                       "llma.layers.14.attention.wv.bias": 0, "llma.layers.14.attention.wo.weight": 1,
                       "llma.layers.15.attention.wq.weight": 0, "llma.layers.15.attention.wq.bias": 0,
                       "llma.layers.15.attention.wk.weight": 0, "llma.layers.15.attention.wk.bias": 0,
                       "llma.layers.15.attention.wv.weight": 0, "llma.layers.15.attention.wv.bias": 0,
                       "llma.layers.15.attention.wo.weight": 1, "llma.layers.16.attention.wq.weight": 0,
                       "llma.layers.16.attention.wq.bias": 0, "llma.layers.16.attention.wk.weight": 0,
                       "llma.layers.16.attention.wk.bias": 0, "llma.layers.16.attention.wv.weight": 0,
                       "llma.layers.16.attention.wv.bias": 0, "llma.layers.16.attention.wo.weight": 1,
                       "llma.layers.17.attention.wq.weight": 0, "llma.layers.17.attention.wq.bias": 0,
                       "llma.layers.17.attention.wk.weight": 0, "llma.layers.17.attention.wk.bias": 0,
                       "llma.layers.17.attention.wv.weight": 0, "llma.layers.17.attention.wv.bias": 0,
                       "llma.layers.17.attention.wo.weight": 1, "llma.layers.18.attention.wq.weight": 0,
                       "llma.layers.18.attention.wq.bias": 0, "llma.layers.18.attention.wk.weight": 0,
                       "llma.layers.18.attention.wk.bias": 0, "llma.layers.18.attention.wv.weight": 0,
                       "llma.layers.18.attention.wv.bias": 0, "llma.layers.18.attention.wo.weight": 1,
                       "llma.layers.19.attention.wq.weight": 0, "llma.layers.19.attention.wq.bias": 0,
                       "llma.layers.19.attention.wk.weight": 0, "llma.layers.19.attention.wk.bias": 0,
                       "llma.layers.19.attention.wv.weight": 0, "llma.layers.19.attention.wv.bias": 0,
                       "llma.layers.19.attention.wo.weight": 1, "llma.layers.20.attention.wq.weight": 0,
                       "llma.layers.20.attention.wq.bias": 0, "llma.layers.20.attention.wk.weight": 0,
                       "llma.layers.20.attention.wk.bias": 0, "llma.layers.20.attention.wv.weight": 0,
                       "llma.layers.20.attention.wv.bias": 0, "llma.layers.20.attention.wo.weight": 1,
                       "llma.layers.21.attention.wq.weight": 0, "llma.layers.21.attention.wq.bias": 0,
                       "llma.layers.21.attention.wk.weight": 0, "llma.layers.21.attention.wk.bias": 0,
                       "llma.layers.21.attention.wv.weight": 0, "llma.layers.21.attention.wv.bias": 0,
                       "llma.layers.21.attention.wo.weight": 1, "llma.layers.22.attention.wq.weight": 0,
                       "llma.layers.22.attention.wq.bias": 0, "llma.layers.22.attention.wk.weight": 0,
                       "llma.layers.22.attention.wk.bias": 0, "llma.layers.22.attention.wv.weight": 0,
                       "llma.layers.22.attention.wv.bias": 0, "llma.layers.22.attention.wo.weight": 1,
                       "llma.layers.23.attention.wq.weight": 0, "llma.layers.23.attention.wq.bias": 0,
                       "llma.layers.23.attention.wk.weight": 0, "llma.layers.23.attention.wk.bias": 0,
                       "llma.layers.23.attention.wv.weight": 0, "llma.layers.23.attention.wv.bias": 0,
                       "llma.layers.23.attention.wo.weight": 1, "llma.layers.24.attention.wq.weight": 0,
                       "llma.layers.24.attention.wq.bias": 0, "llma.layers.24.attention.wk.weight": 0,
                       "llma.layers.24.attention.wk.bias": 0, "llma.layers.24.attention.wv.weight": 0,
                       "llma.layers.24.attention.wv.bias": 0, "llma.layers.24.attention.wo.weight": 1,
                       "llma.layers.25.attention.wq.weight": 0, "llma.layers.25.attention.wq.bias": 0,
                       "llma.layers.25.attention.wk.weight": 0, "llma.layers.25.attention.wk.bias": 0,
                       "llma.layers.25.attention.wv.weight": 0, "llma.layers.25.attention.wv.bias": 0,
                       "llma.layers.25.attention.wo.weight": 1, "llma.layers.26.attention.wq.weight": 0,
                       "llma.layers.26.attention.wq.bias": 0, "llma.layers.26.attention.wk.weight": 0,
                       "llma.layers.26.attention.wk.bias": 0, "llma.layers.26.attention.wv.weight": 0,
                       "llma.layers.26.attention.wv.bias": 0, "llma.layers.26.attention.wo.weight": 1,
                       "llma.layers.27.attention.wq.weight": 0, "llma.layers.27.attention.wq.bias": 0,
                       "llma.layers.27.attention.wk.weight": 0, "llma.layers.27.attention.wk.bias": 0,
                       "llma.layers.27.attention.wv.weight": 0, "llma.layers.27.attention.wv.bias": 0,
                       "llma.layers.27.attention.wo.weight": 1, "llma.layers.28.attention.wq.weight": 0,
                       "llma.layers.28.attention.wq.bias": 0, "llma.layers.28.attention.wk.weight": 0,
                       "llma.layers.28.attention.wk.bias": 0, "llma.layers.28.attention.wv.weight": 0,
                       "llma.layers.28.attention.wv.bias": 0, "llma.layers.28.attention.wo.weight": 1,
                       "llma.layers.29.attention.wq.weight": 0, "llma.layers.29.attention.wq.bias": 0,
                       "llma.layers.29.attention.wk.weight": 0, "llma.layers.29.attention.wk.bias": 0,
                       "llma.layers.29.attention.wv.weight": 0, "llma.layers.29.attention.wv.bias": 0,
                       "llma.layers.29.attention.wo.weight": 1, "llma.layers.30.attention.wq.weight": 0,
                       "llma.layers.30.attention.wq.bias": 0, "llma.layers.30.attention.wk.weight": 0,
                       "llma.layers.30.attention.wk.bias": 0, "llma.layers.30.attention.wv.weight": 0,
                       "llma.layers.30.attention.wv.bias": 0, "llma.layers.30.attention.wo.weight": 1,
                       "llma.layers.31.attention.wq.weight": 0, "llma.layers.31.attention.wq.bias": 0,
                       "llma.layers.31.attention.wk.weight": 0, "llma.layers.31.attention.wk.bias": 0,
                       "llma.layers.31.attention.wv.weight": 0, "llma.layers.31.attention.wv.bias": 0,
                       "llma.layers.31.attention.wo.weight": 1, "llma.output.weight": 0, "llma.output.bias": 0}

import torch
from pathlib import Path

Path("./converted_sparse").mkdir(exist_ok=True)

ori = torch.load("consolidated.00.pth", map_location="cpu")
ori = {"llma." + key: val for key, val in ori.items()}


def func(rank=0):
    shard_split_to = 8
    split_ckpt = {}
    for key, ori_param in ori.items():
        if key in weight_parallel_dim:
            split_ckpt[key] = torch.chunk(ori_param, shard_split_to, weight_parallel_dim[key])[
                rank % shard_split_to].clone()
            if rank == 0:
                print(f"chunk {key}")
        else:
            if "experts.0." in key:
                weight_all_experts = [ori[key.replace("experts.0.", f"experts.{i}.")] for i in range(8)]
                if "w2" in key:
                    weight_all_experts = [torch.transpose(_, 0, 1) for _ in weight_all_experts]
                weight_this_rank = [torch.chunk(_, 8, dim=0)[rank] for _ in weight_all_experts]
                weight_this_rank = torch.cat(weight_this_rank, dim=0).clone()
                key = key.replace("experts.0.", "").replace(".weight", "")
                split_ckpt[key] = weight_this_rank
                print("expert key")
            elif "experts" in key:
                continue
            else:
                split_ckpt[key] = ori_param
                if rank == 0:
                    print(f"inherit {key}")
    torch.save({"model": split_ckpt}, f"converted_sparse/consolidated.{rank:02d}-of-08.model.pth")


for r in range(8):
    func(r)