yujiepan/llama-3.3-tiny-random-dim64

This tiny model is for debugging. It is randomly initialized with the config adapted from meta-llama/Llama-3.3-70B-Instruct.

Example usage:

from transformers import pipeline
model_id = "yujiepan/llama-3.3-tiny-random-dim64"
pipe = pipeline(
    "text-generation", model=model_id, device="cuda",
    trust_remote_code=True, max_new_tokens=3,
)
print(pipe("Hello World!"))

Codes to create this repo:

import torch

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

source_model_id = "meta-llama/Llama-3.3-70B-Instruct"
save_folder = "/tmp/yujiepan/llama-3.3-tiny-random-dim64"

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

config = AutoConfig.from_pretrained(
    source_model_id, trust_remote_code=True,
)
config.hidden_size = 64
config.intermediate_size = 128
config.num_attention_heads = 2
config.num_key_value_heads = 1
config.head_dim = 32
config.num_hidden_layers = 2
config.tie_word_embeddings = True

model = AutoModelForCausalLM.from_config(
    config,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
)
model.generation_config = GenerationConfig.from_pretrained(
    source_model_id, trust_remote_code=True,
)
set_seed(42)
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)

Printing the model:

LlamaForCausalLM(
  (model): LlamaModel(
    (embed_tokens): Embedding(128256, 64)
    (layers): ModuleList(
      (0-1): 2 x LlamaDecoderLayer(
        (self_attn): LlamaAttention(
          (q_proj): Linear(in_features=64, out_features=64, bias=False)
          (k_proj): Linear(in_features=64, out_features=32, bias=False)
          (v_proj): Linear(in_features=64, out_features=32, bias=False)
          (o_proj): Linear(in_features=64, out_features=64, bias=False)
        )
        (mlp): LlamaMLP(
          (gate_proj): Linear(in_features=64, out_features=128, bias=False)
          (up_proj): Linear(in_features=64, out_features=128, bias=False)
          (down_proj): Linear(in_features=128, out_features=64, bias=False)
          (act_fn): SiLU()
        )
        (input_layernorm): LlamaRMSNorm((64,), eps=1e-05)
        (post_attention_layernorm): LlamaRMSNorm((64,), eps=1e-05)
      )
    )
    (norm): LlamaRMSNorm((64,), eps=1e-05)
    (rotary_emb): LlamaRotaryEmbedding()
  )
  (lm_head): Linear(in_features=64, out_features=128256, bias=False)
)