convert_weight.py

import torch
from tqdm import tqdm


input_dir_path = "/scratch/project_462000086/norwegian_gpt/Megatron-DeepSpeed-fixed/mistral-7b-from-scratch-2nd-run/global_step30000"
output_dir_path = "/scratch/project_462000086/norwegian_gpt/Megatron-DeepSpeed-fixed/hf_mistral_from_scratch_60k"

n_hidden = 4096
n_ffn_hidden = 14336
n_heads = 32
n_kv_heads = 8
n_layers = 32
n_tp = 2


weights = {}

# embedding
embedding_weights = []
for i in range(n_tp):
    path = f"{input_dir_path}/layer_01-model_0{i}-model_states.pt"
    checkpoint = torch.load(path)

    embedding_weights.append(checkpoint["word_embeddings.weight"].bfloat16())

weights[f"model.embed_tokens.weight"] = torch.cat(embedding_weights, dim=0)
del embedding_weights

lm_head_weights = []
for i in range(n_tp):
    path = f"{input_dir_path}/layer_{n_layers + 5}-model_0{i}-model_states.pt"
    checkpoint = torch.load(path)

    lm_head_weights.append(checkpoint["lm_head.weight"].bfloat16())

weights[f"lm_head.weight"] = torch.cat(lm_head_weights, dim=0)
del lm_head_weights


# transformer layers
for layer in tqdm(range(n_layers)):
    q_weights, k_weights, v_weights, o_weights = [], [], [], []
    up_weights, gate_weights, down_weights = [], [], []

    for i in range(n_tp):
        path = f"{input_dir_path}/layer_{layer+3:02d}-model_0{i}-model_states.pt"
        checkpoint = torch.load(path)

        weights[f"model.layers.{layer}.input_layernorm.weight"] = checkpoint["input_layernorm.weight"].bfloat16()
        weights[f"model.layers.{layer}.post_attention_layernorm.weight"] = checkpoint["post_attention_layernorm.weight"].bfloat16()

        kv_weight = checkpoint["self_attention.key_value.weight"].bfloat16()
        k_weight, v_weight = torch.chunk(kv_weight, 2, dim=0)
        k_weights.append(k_weight)
        v_weights.append(v_weight)

        q_weights.append(checkpoint["self_attention.query.weight"].bfloat16())
        o_weights.append(checkpoint["self_attention.dense.weight"].bfloat16())
        down_weights.append(checkpoint["mlp.dense_4h_to_h.weight"].bfloat16())

        up_gate_weight = checkpoint["mlp.dense_h_to_4h.weight"].bfloat16()
        up_weight, gate_weight = torch.chunk(up_gate_weight, 2, dim=0)
        up_weights.append(up_weight)
        gate_weights.append(gate_weight)

    weights[f"model.layers.{layer}.self_attn.q_proj.weight"] = torch.cat(q_weights, dim=0)
    weights[f"model.layers.{layer}.self_attn.k_proj.weight"] = torch.cat(k_weights, dim=0)
    weights[f"model.layers.{layer}.self_attn.v_proj.weight"] = torch.cat(v_weights, dim=0)
    weights[f"model.layers.{layer}.self_attn.o_proj.weight"] = torch.cat(o_weights, dim=1)
    weights[f"model.layers.{layer}.mlp.up_proj.weight"] = torch.cat(up_weights, dim=0)
    weights[f"model.layers.{layer}.mlp.gate_proj.weight"] = torch.cat(gate_weights, dim=0)
    weights[f"model.layers.{layer}.mlp.down_proj.weight"] = torch.cat(down_weights, dim=1)


# output layer norm
path = f"{input_dir_path}/layer_{n_layers + 4}-model_00-model_states.pt"
checkpoint = torch.load(path)

weights[f"model.norm.weight"] = checkpoint["weight"].bfloat16()

torch.save(weights, f"{output_dir_path}/pytorch_model.bin")