This tiny model is intended for debugging. It is randomly initialized using the configuration adapted from Zyphra/ZAYA1-reasoning-base.

Example usage:

from transformers import pipeline
model_id = "yujiepan/zaya1-tiny-random"
pipe = pipeline('text-generation', model=model_id,
                device='cuda', dtype="bfloat16")
print(pipe('Hello World!'))

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,
    AutoTokenizer,
    GenerationConfig,
    set_seed,
)

source_model_id = "Zyphra/ZAYA1-reasoning-base"
save_folder = "/tmp/yujiepan/zaya1-tiny-random"

processor = AutoTokenizer.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'] = 512
config_json['num_attention_heads'] = 4
config_json['num_key_value_heads'] = 1
config_json['num_hidden_layers'] = 2
# bug. need to first set False and then hack
config_json['tie_word_embeddings'] = False
config_json['cca_num_q_heads'] = [2, 0]
config_json['ffn_hidden_size_list'] = [0, 32]
config_json['num_query_groups_list'] = [1, 0]
config_json['zaya_layers'] = ['a', 16]
config_json['zaya_mlp_expansion'] = [0, 8]

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)
model.lm_head = None
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()
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)
with open(f"{save_folder}/config.json", 'r', encoding='utf-8') as f:
    config_json = json.load(f)
    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)

Printing the model:

ZayaForCausalLM(
  (model): ZayaModel(
    (embed_tokens): Embedding(262272, 512, padding_idx=0)
    (layers): ModuleList(
      (0): ZayaDecoderATTLayer(
        (self_attn): ZayaSdpaAttention(
          (o_proj): Linear(in_features=256, out_features=512, bias=False)
          (qkv): CCA(
            (linear_q): Linear(in_features=512, out_features=256, bias=False)
            (linear_k): Linear(in_features=512, out_features=128, bias=False)
            (val_proj1): Linear(in_features=512, out_features=64, bias=False)
            (val_proj2): Linear(in_features=512, out_features=64, bias=False)
            (conv_qk): Sequential(
              (0): Conv1d(384, 384, kernel_size=(2,), stride=(1,), groups=384)
              (1): Conv1d(384, 384, kernel_size=(2,), stride=(1,), groups=3)
            )
          )
        )
        (input_norm): ZayaRMSNorm((512,), eps=1e-05)
        (res_scale): ResidualScaling()
      )
      (1): ZayaDecoderMLPLayer(
        (zaya_block): ZayaBlock(
          (router): ZayaRouter(
            (down_proj): Linear(in_features=512, out_features=8, bias=True)
            (rmsnorm_eda): ZayaRMSNorm((8,), eps=1e-06)
            (non_linearity): GELU(approximate='none')
            (router_mlp): Sequential(
              (0): Linear(in_features=8, out_features=8, bias=True)
              (1): GELU(approximate='none')
              (2): Linear(in_features=8, out_features=8, bias=True)
              (3): GELU(approximate='none')
              (4): Linear(in_features=8, out_features=17, bias=False)
            )
          )
          (experts): SequentialMLP(
            (local_experts): ModuleList(
              (0-15): 16 x MLP(
                (linear_fc1): Linear(in_features=512, out_features=32, bias=False)
                (linear_fc2): Linear(in_features=16, out_features=512, bias=False)
              )
            )
          )
        )
        (input_norm): ZayaRMSNorm((512,), eps=1e-05)
        (res_scale): ResidualScaling()
      )
    )
    (res_scale): ResidualScaling()
    (final_norm): ZayaRMSNorm((512,), eps=1e-05)
    (rotary_emb): ZayaRotaryEmbedding()
  )
  (lm_head): None
)