This model is for debugging. It is randomly initialized using the config from microsoft/Phi-3.5-MoE-instruct but with smaller size.
Codes:
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'))
- Downloads last month
- 51
Inference API (serverless) does not yet support model repos that contain custom code.