This tiny model is for debugging. It is randomly initialized with the config adapted from ByteDance-Seed/Seed-OSS-36B-Instruct.
Example usage:
- vLLM
python3 -m vllm.entrypoints.openai.api_server \
--enable-auto-tool-choice \
--tool-call-parser seed_oss \
--trust-remote-code \
--model ./<local_download_folder> \
--chat-template ./<local_download_folder>/chat_template.jinja \
--tensor-parallel-size 2
- Transformers
import os
import re
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "tiny-random/seed-oss"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype=torch.bfloat16)
messages = [
{"role": "user", "content": "How to make pasta?"},
]
tokenized_chat = tokenizer.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_tensors="pt",
thinking_budget=64 # control the thinking budget
)
outputs = model.generate(tokenized_chat.to(model.device), max_new_tokens=128)
output_text = tokenizer.decode(outputs[0])
print(output_text)
Codes to create this repo:
import json
from pathlib import Path
import accelerate
import torch
from huggingface_hub import file_exists, hf_hub_download
from transformers import (
AutoConfig,
AutoModelForCausalLM,
AutoProcessor,
GenerationConfig,
set_seed,
)
source_model_id = "ByteDance-Seed/Seed-OSS-36B-Instruct"
save_folder = "/tmp/tiny-random/seed-oss"
processor = AutoProcessor.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)
config_json['hidden_size'] = 8
config_json['head_dim'] = 32 # vllm requirement
config_json['intermediate_size'] = 32
config_json['num_attention_heads'] = 8
config_json['num_hidden_layers'] = 2
config_json['num_key_value_heads'] = 4 # better support tensor parallel
config_json['tie_word_embeddings'] = False
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)
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,
)
model.generation_config.do_sample = 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, 0.1)
print(name, p.shape)
model.save_pretrained(save_folder)
Printing the model:
SeedOssForCausalLM(
(model): SeedOssModel(
(embed_tokens): Embedding(155136, 8, padding_idx=1)
(layers): ModuleList(
(0-1): 2 x SeedOssDecoderLayer(
(self_attn): SeedOssAttention(
(q_proj): Linear(in_features=8, out_features=256, bias=True)
(k_proj): Linear(in_features=8, out_features=128, bias=True)
(v_proj): Linear(in_features=8, out_features=128, bias=True)
(o_proj): Linear(in_features=256, out_features=8, bias=False)
)
(mlp): SeedOssMLP(
(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): SiLU()
)
(input_layernorm): SeedOssRMSNorm((8,), eps=1e-06)
(post_attention_layernorm): SeedOssRMSNorm((8,), eps=1e-06)
)
)
(norm): SeedOssRMSNorm((8,), eps=1e-06)
(rotary_emb): SeedOssRotaryEmbedding()
)
(lm_head): Linear(in_features=8, out_features=155136, bias=False)
)