convert_scm_to_hf.py

import glob
import re
import shutil
import sys

import accelerate
import torch
from configuration_glm4_shared_moe import Glm4SharedMoeConfig
from safetensors import safe_open
from transformers.models.glm4_moe.configuration_glm4_moe import Glm4MoeConfig
from transformers.models.glm4_moe.modeling_glm4_moe import Glm4MoeForCausalLM

input_model = sys.argv[1]
output_model_path = sys.argv[2]

cfg_shared_moe = Glm4SharedMoeConfig.from_pretrained(input_model)
cfg_standard_moe = Glm4MoeConfig(
    architectures=["Glm4MoeForCausalLM"],
    vocab_size=cfg_shared_moe.vocab_size,
    hidden_size=cfg_shared_moe.hidden_size,
    intermediate_size=cfg_shared_moe.intermediate_size,
    num_hidden_layers=cfg_shared_moe.num_hidden_layers,
    num_attention_heads=cfg_shared_moe.num_attention_heads,
    num_key_value_heads=cfg_shared_moe.num_key_value_heads,
    hidden_act=cfg_shared_moe.hidden_act,
    max_position_embeddings=cfg_shared_moe.max_position_embeddings,
    initializer_range=cfg_shared_moe.initializer_range,
    rms_norm_eps=cfg_shared_moe.rms_norm_eps,
    use_cache=cfg_shared_moe.use_cache,
    tie_word_embeddings=cfg_shared_moe.tie_word_embeddings,
    rope_theta=cfg_shared_moe.rope_theta,
    rope_scaling=cfg_shared_moe.rope_scaling,
    attention_bias=cfg_shared_moe.attention_bias,
    attention_dropout=cfg_shared_moe.attention_dropout,
    moe_intermediate_size=cfg_shared_moe.moe_intermediate_size,
    num_experts_per_tok=cfg_shared_moe.num_experts_per_tok,
    n_routed_experts=cfg_shared_moe.n_routed_experts,
    norm_topk_prob=cfg_shared_moe.norm_topk_prob,
    head_dim=cfg_shared_moe.head_dim,
    n_group=cfg_shared_moe.n_group,
    n_shared_experts=cfg_shared_moe.n_shared_experts,
    first_k_dense_replace=cfg_shared_moe.first_k_dense_replace,
    partial_rotary_factor=cfg_shared_moe.partial_rotary_factor,
    routed_scaling_factor=cfg_shared_moe.routed_scaling_factor,
    topk_group=cfg_shared_moe.topk_group,
    use_qk_norm=cfg_shared_moe.use_qk_norm
)
num_experts = cfg_standard_moe.n_routed_experts

with accelerate.init_empty_weights():
    model_standard_moe = Glm4MoeForCausalLM(cfg_shared_moe)

model_standard_moe = model_standard_moe.to(torch.bfloat16)
new_state_dict = {}
pattern = f"{input_model}/model-*-of-*.safetensors"
files = sorted(glob.glob(pattern))

if len(files) == 0:
    raise FileNotFoundError
tensors = {}

for file_path in files:
    print(f"processing {file_path}")
    with safe_open(file_path, framework="pt", device="cpu") as f:
        for key in f.keys():
            tensor = f.get_tensor(key)
            tensors[key] = tensor

for key in tensors:
    if "moe_mlp" not in key:
        new_state_dict[key] = tensors[key]
    elif "moe_mlp.output_experts" in key:
        layer_num = int(re.search(r"\d+", key).group())
        for i, tensor in enumerate(torch.unbind(tensors[key])):
            new_state_dict[
                f"model.layers.{layer_num}.mlp.experts.{i}.down_proj.weight"
            ] = tensor.contiguous()
    elif "moe_mlp.experts" in key:
        layer_num = int(re.search(r"\d+", key).group())
        for i, tensor in enumerate(torch.unbind(tensors[key])):
            (
                new_state_dict[
                    f"model.layers.{layer_num}.mlp.experts.{i}.up_proj.weight"
                ],
                new_state_dict[
                    f"model.layers.{layer_num}.mlp.experts.{i}.gate_proj.weight"
                ],
            ) = torch.chunk(tensor, 2, dim=0)

model_standard_moe.load_state_dict(new_state_dict, strict=True, assign=True)
model_standard_moe.save_pretrained(output_model_path)
cfg_standard_moe.save_pretrained(output_model_path)

for i in ["tokenizer_config.json", "tokenizer.json", "chat_template.jinja"]:
    shutil.copy(input_model + "/" + i, output_model_path + "/" + i)