phi-moe

This tiny model is for debugging. It is randomly initialized with the config adapted from microsoft/Phi-tiny-MoE-instruct.

Example usage:

import torch

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

model_id = "tiny-random/phi-moe"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
)
pipe = pipeline('text-generation', model=model, tokenizer=tokenizer, trust_remote_code=True)
print(pipe('Write an article about Artificial Intelligence.'))

Codes to create this repo:

import json
from pathlib import Path

import torch

import accelerate
from huggingface_hub import file_exists, hf_hub_download
from transformers import (
    AutoConfig,
    AutoModelForCausalLM,
    AutoTokenizer,
    GenerationConfig,
    set_seed,
)

source_model_id = "microsoft/Phi-tiny-MoE-instruct"
save_folder = "/tmp/tiny-random/phi-moe"

processor = AutoTokenizer.from_pretrained(source_model_id)
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'] = 64
config_json['intermediate_size'] = 128
config_json['num_attention_heads'] = 2
config_json['num_experts_per_tok'] = 2
config_json['num_hidden_layers'] = 2
config_json['num_key_value_heads'] = 1
config_json['num_local_experts'] = 8
config_json['tie_word_embeddings'] = True

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()  # cpu is more stable for random initialization across machines
with torch.no_grad():
    for name, p in sorted(model.named_parameters()):
        torch.nn.init.normal_(p, 0, 0.2)
        print(name, p.shape)
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()

Printing the model:

PhiMoEForCausalLM(
  (model): PhiMoEModel(
    (embed_tokens): Embedding(32064, 64)
    (layers): ModuleList(
      (0-1): 2 x PhiMoEDecoderLayer(
        (self_attn): PhiMoESdpaAttention(
          (q_proj): Linear(in_features=64, out_features=64, bias=True)
          (k_proj): Linear(in_features=64, out_features=32, bias=True)
          (v_proj): Linear(in_features=64, out_features=32, bias=True)
          (o_proj): Linear(in_features=64, out_features=64, bias=True)
          (rotary_emb): PhiMoERotaryEmbedding()
        )
        (block_sparse_moe): PhiMoESparseMoeBlock(
          (gate): Linear(in_features=64, out_features=8, bias=False)
          (experts): ModuleList(
            (0-7): 8 x PhiMoEBlockSparseTop2MLP(
              (w1): Linear(in_features=64, out_features=128, bias=False)
              (w2): Linear(in_features=128, out_features=64, bias=False)
              (w3): Linear(in_features=64, out_features=128, bias=False)
              (act_fn): SiLU()
            )
          )
        )
        (input_layernorm): LayerNorm((64,), eps=1e-05, elementwise_affine=True)
        (post_attention_layernorm): LayerNorm((64,), eps=1e-05, elementwise_affine=True)
      )
    )
    (norm): LayerNorm((64,), eps=1e-05, elementwise_affine=True)
  )
  (lm_head): Linear(in_features=64, out_features=32064, bias=True)
)