yujiepan / bailing-moe-v2.5-tiny-random

import json
from pathlib import Path

import accelerate
import torch
import transformers
from huggingface_hub import file_exists, hf_hub_download
from transformers import (
    AutoConfig,
    AutoModelForCausalLM,
    AutoTokenizer,
    GenerationConfig,
    set_seed,
)
transformers.utils.import_utils.is_torch_fx_available = transformers.utils.import_utils.is_torch_available
source_model_id = "inclusionAI/Ring-2.5-1T"
save_folder = "/tmp/yujiepan/bailing-moe-v25-tiny-random"

processor = AutoTokenizer.from_pretrained(source_model_id, trust_remote_code=True)
processor.save_pretrained(save_folder)

with open(hf_hub_download(source_model_id, filename='config.json', repo_type='model'), 'r', encoding='utf-8') as f:
    config_json = json.load(f)
for k, v in config_json['auto_map'].items():
    config_json['auto_map'][k] = f'{source_model_id}--{v}'

# config_json['head_dim'] = 32
config_json['hidden_size'] = 8
config_json['intermediate_size'] = 32
config_json['moe_intermediate_size'] = 32
config_json['moe_shared_expert_intermediate_size'] = 32
config_json['first_k_dense_replace'] = 1
config_json['num_attention_heads'] = 4
config_json['num_hidden_layers'] = 2
config_json['num_key_value_heads'] = 4
config_json['q_lora_rank'] = 32
config_json['layer_group_size'] = 2
del config_json['quantization_config']

with open(f"{save_folder}/config.json", "w", encoding='utf-8') as f:
    json.dump(config_json, f, indent=2)

config = AutoConfig.from_pretrained(
    save_folder,
    trust_remote_code=True,
)
print(config)
automap = config_json['auto_map']
torch.set_default_dtype(torch.bfloat16)
model = AutoModelForCausalLM.from_config(config, trust_remote_code=True)
torch.set_default_dtype(torch.float32)

if file_exists(filename="generation_config.json", repo_id=source_model_id, repo_type='model'):
    model.generation_config = GenerationConfig.from_pretrained(
        source_model_id, trust_remote_code=True,
    )
set_seed(42)
model = model.cpu()
with torch.no_grad():
    for name, p in sorted(model.named_parameters()):
        torch.nn.init.normal_(p, 0, 0.1)
        print(name, p.shape)
model.model.layers[1].mlp.gate.expert_bias = model.model.layers[1].mlp.gate.expert_bias.float()
model.save_pretrained(save_folder)
print(model)
with open(f"{save_folder}/config.json", "r", encoding='utf-8') as f:
    config_json = json.load(f)
    config_json['auto_map'] = automap
with open(f"{save_folder}/config.json", "w", encoding='utf-8') as f:
    json.dump(config_json, f, indent=2)
for python_file in Path(save_folder).glob('*.py'):
    python_file.unlink()

BailingMoeV2_5ForCausalLM(
  (model): BailingMoeV2_5Model(
    (word_embeddings): Embedding(157184, 8, padding_idx=156892)
    (layers): ModuleList(
      (0): BailingMoeV2_5DecoderLayer(
        (attention): BailingMoeV2_5LinearAttention(
          (query_key_value): Linear(in_features=8, out_features=1536, bias=False)
          (query_layernorm): BailingMoeV2_5RMSNorm()
          (key_layernorm): BailingMoeV2_5RMSNorm()
          (rotary_emb): BailingMoeV2_5RotaryEmbedding()
          (dense): Linear(in_features=512, out_features=8, bias=False)
          (g_proj): Linear(in_features=8, out_features=512, bias=False)
          (g_norm): BailingMoeV2_5GroupRMSNorm()
        )
        (mlp): BailingMoeV2_5MLP(
          (gate_proj): Linear(in_features=8, out_features=32, bias=False)
          (up_proj): Linear(in_features=8, out_features=32, bias=False)
          (down_proj): Linear(in_features=32, out_features=8, bias=False)
          (act_fn): SiLUActivation()
        )
        (input_layernorm): BailingMoeV2_5RMSNorm()
        (post_attention_layernorm): BailingMoeV2_5RMSNorm()
      )
      (1): BailingMoeV2_5DecoderLayer(
        (attention): BailingMoeV2_5MultiLatentAttention(
          (q_a_proj): Linear(in_features=8, out_features=32, bias=False)
          (q_a_layernorm): BailingMoeV2_5RMSNorm()
          (q_b_proj): Linear(in_features=32, out_features=768, bias=False)
          (kv_a_proj_with_mqa): Linear(in_features=8, out_features=576, bias=False)
          (kv_a_layernorm): BailingMoeV2_5RMSNorm()
          (kv_b_proj): Linear(in_features=512, out_features=1024, bias=False)
          (dense): Linear(in_features=512, out_features=8, bias=False)
        )
        (mlp): BailingMoeV2_5SparseMoeBlock(
          (experts): ModuleList(
            (0-255): 256 x BailingMoeV2_5MLP(
              (gate_proj): Linear(in_features=8, out_features=32, bias=False)
              (up_proj): Linear(in_features=8, out_features=32, bias=False)
              (down_proj): Linear(in_features=32, out_features=8, bias=False)
              (act_fn): SiLUActivation()
            )
          )
          (gate): BailingMoeV2_5Gate()
          (shared_experts): BailingMoeV2_5MLP(
            (gate_proj): Linear(in_features=8, out_features=32, bias=False)
            (up_proj): Linear(in_features=8, out_features=32, bias=False)
            (down_proj): Linear(in_features=32, out_features=8, bias=False)
            (act_fn): SiLUActivation()
          )
        )
        (input_layernorm): BailingMoeV2_5RMSNorm()
        (post_attention_layernorm): BailingMoeV2_5RMSNorm()
      )
    )
    (norm): BailingMoeV2_5RMSNorm()
    (rotary_emb): BailingMoeV2_5RotaryEmbedding()
    (rotary_emb_mla): BailingMoeV2_5MLARotaryEmbedding()
  )
  (lm_head): Linear(in_features=8, out_features=157184, bias=False)
)