README.md

---
library_name: transformers
pipeline_tag: text-generation
inference: true
widget:
- text: Hello!
  example_title: Hello world
  group: Python
---

This model is for debugging. It is randomly initialized using the config from [microsoft/Phi-3.5-MoE-instruct](https://huggingface.co/microsoft/Phi-3.5-MoE-instruct) but with smaller size. 

Codes:
```python
import os

import torch
import transformers
from transformers import (AutoConfig, AutoModelForCausalLM, AutoTokenizer,
                          GenerationConfig, pipeline, set_seed)

model_id = "microsoft/Phi-3.5-MoE-instruct"
repo_id = "yujiepan/phi-3.5-moe-tiny-random"
save_path = f"/tmp/{repo_id}"

config = AutoConfig.from_pretrained(model_id, trust_remote_code=True)
config.hidden_size = 16
config.intermediate_size = 32
config.num_attention_heads = 4
config.num_hidden_layers = 2
config.num_key_value_heads = 4
config.rope_scaling['long_factor'] = [1.0299, 1.0499]
config.rope_scaling['short_factor'] = [1.05, 1.05]

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
tokenizer.save_pretrained(save_path)

model = AutoModelForCausalLM.from_config(
    config, torch_dtype=torch.bfloat16,
    # attn_implementation="sdpa",
    trust_remote_code=True,
)
model.generation_config = GenerationConfig.from_pretrained(
    model_id, trust_remote_code=True
)

set_seed(42)
with torch.no_grad():
    for _, p in sorted(model.named_parameters()):
        torch.nn.init.uniform_(p, -0.3, 0.3)

model.save_pretrained(save_path)

pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, device="cuda",
                trust_remote_code=True, max_new_tokens=20)
print(pipe('Hello'))
```