granite-moe-hybrid

This tiny model is intended for debugging. It is randomly initialized using the configuration adapted from ibm-granite/granite-4.0-h-small.

Example usage:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

model_id = "tiny-random/granite-moe-hybrid"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
)
pipe = pipeline('text-generation', model=model, tokenizer=tokenizer, trust_remote_code=True)
print(pipe('Write an article about Artificial Intelligence.'))

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,
    AutoTokenizer,
    GenerationConfig,
    set_seed,
)

source_model_id = "ibm-granite/granite-4.0-h-small"
save_folder = "/tmp/tiny-random/granite-moe-hybrid"

processor = AutoTokenizer.from_pretrained(source_model_id)
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'] = 32
config_json['intermediate_size'] = 128
config_json['layer_types'] = ['mamba', 'attention']
config_json.update({
    'mamba_expand': int(4096 / 32 * 2),
})
config_json['num_attention_heads'] = 2
config_json['shared_intermediate_size'] = 128
config_json['num_hidden_layers'] = 2
config_json['num_key_value_heads'] = 2
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)

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,
    )
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)
print(model)

Printing the model:

GraniteMoeHybridForCausalLM(
  (model): GraniteMoeHybridModel(
    (embed_tokens): Embedding(100352, 32, padding_idx=100256)
    (layers): ModuleList(
      (0): GraniteMoeHybridDecoderLayer(
        (block_sparse_moe): GraniteMoeHybridMoE(
          (activation): SiLU()
          (input_linear): GraniteMoeHybridParallelExperts()
          (output_linear): GraniteMoeHybridParallelExperts()
          (router): GraniteMoeHybridTopKGating(
            (layer): Linear(in_features=32, out_features=72, bias=False)
          )
        )
        (input_layernorm): GraniteMoeHybridRMSNorm((32,), eps=1e-05)
        (post_attention_layernorm): GraniteMoeHybridRMSNorm((32,), eps=1e-05)
        (shared_mlp): GraniteMoeHybridMLP(
          (activation): SiLU()
          (input_linear): Linear(in_features=32, out_features=256, bias=False)
          (output_linear): Linear(in_features=128, out_features=32, bias=False)
        )
        (mamba): GraniteMoeHybridMambaLayer(
          (act): SiLU()
          (conv1d): Conv1d(8448, 8448, kernel_size=(4,), stride=(1,), padding=(3,), groups=8448)
          (in_proj): Linear(in_features=32, out_features=16768, bias=False)
          (norm): GraniteMoeHybridRMSNormGated()
          (out_proj): Linear(in_features=8192, out_features=32, bias=False)
        )
      )
      (1): GraniteMoeHybridDecoderLayer(
        (block_sparse_moe): GraniteMoeHybridMoE(
          (activation): SiLU()
          (input_linear): GraniteMoeHybridParallelExperts()
          (output_linear): GraniteMoeHybridParallelExperts()
          (router): GraniteMoeHybridTopKGating(
            (layer): Linear(in_features=32, out_features=72, bias=False)
          )
        )
        (input_layernorm): GraniteMoeHybridRMSNorm((32,), eps=1e-05)
        (post_attention_layernorm): GraniteMoeHybridRMSNorm((32,), eps=1e-05)
        (shared_mlp): GraniteMoeHybridMLP(
          (activation): SiLU()
          (input_linear): Linear(in_features=32, out_features=256, bias=False)
          (output_linear): Linear(in_features=128, out_features=32, bias=False)
        )
        (self_attn): GraniteMoeHybridAttention(
          (q_proj): Linear(in_features=32, out_features=32, bias=False)
          (k_proj): Linear(in_features=32, out_features=32, bias=False)
          (v_proj): Linear(in_features=32, out_features=32, bias=False)
          (o_proj): Linear(in_features=32, out_features=32, bias=False)
        )
      )
    )
    (norm): GraniteMoeHybridRMSNorm((32,), eps=1e-05)
  )
  (lm_head): Linear(in_features=32, out_features=100352, bias=False)
)