deepseek-v3.1-tiny-random

This tiny model is for debugging. It is randomly initialized with the config adapted from deepseek-ai/DeepSeek-V3.1.

Example usage:

• vLLM

python -m vllm.entrypoints.openai.api_server \
    --tensor-parallel-size 2 \
    --model yujiepan/deepseek-v3.1-tiny-random \
    --trust-remote-code \
    --speculative-config='{"method": "deepseek_mtp", "num_speculative_tokens": 1}' 

• Transformers

import torch
import transformers

model_id = "yujiepan/deepseek-v3.1-tiny-random"
pipe = transformers.pipelines.pipeline(
    'text-generation',
    model=model_id,
    trust_remote_code=True,
    device_map='cuda',
    torch_dtype=torch.bfloat16,
)
r = pipe.model(torch.tensor([[1, 2, 3]], dtype=torch.int64).cuda(), attention_mask=None, use_cache=False)
print(r)

Codes to create this repo:

import json
from copy import deepcopy
from pathlib import Path

import accelerate
import torch
import torch.nn as nn
from huggingface_hub import file_exists, hf_hub_download
from transformers import (
    AutoConfig,
    AutoModelForCausalLM,
    AutoProcessor,
    AutoTokenizer,
    GenerationConfig,
    set_seed,
)
from transformers.models.glm4_moe.modeling_glm4_moe import Glm4MoeRMSNorm
source_model_id = "deepseek-ai/DeepSeek-V3.1"
save_folder = "/tmp/yujiepan/deepseek-v3.1-tiny-random"

Path(save_folder).mkdir(parents=True, exist_ok=True)
tokenizer = AutoTokenizer.from_pretrained(source_model_id, trust_remote_code=True)
tokenizer.save_pretrained(save_folder)

with open(hf_hub_download(source_model_id, filename='config.json', repo_type='model', cache_dir='/tmp/'), 'r', encoding='utf-8') as f:
    config_json = json.load(f)
for k, v in config_json['auto_map'].items():
    config_json['auto_map'][k] = f'{source_model_id}--{v}'
config_json.update({
    'first_k_dense_replace': 1,
    'num_hidden_layers': 2,
    'hidden_size': 8,
    'intermediate_size': 64,
    'kv_lora_rank': 384,
    'moe_intermediate_size': 64,
    'n_routed_experts': 32,
    'n_shared_experts': 1,
    'num_attention_heads': 4,
    'num_experts_per_tok': 8,
    'num_key_value_heads': 4,
    'q_lora_rank': 32,
    'qk_nope_head_dim': 64,
    'qk_rope_head_dim': 192,  # vllm mla kernel supports 576 only, FA supports head dim <= 256
    'v_head_dim': 64,
    'tie_word_embeddings': False,
})
del config_json['quantization_config']
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)

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,
    )

class SharedHead(nn.Module):
    def __init__(self, config) -> None:
        super().__init__()
        from transformers.models.glm4_moe.modeling_glm4_moe import Glm4MoeRMSNorm
        self.norm = Glm4MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

last_extra_layer = model.model.layers[0].__class__(config, layer_idx=config.num_hidden_layers)
last_extra_layer.enorm = Glm4MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
last_extra_layer.hnorm = Glm4MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
last_extra_layer.eh_proj = nn.Linear(config.hidden_size * 2, config.hidden_size, bias=False)
last_extra_layer.shared_head = SharedHead(config=config)
model.model.layers.append(last_extra_layer)

torch.set_default_dtype(torch.float32)
set_seed(42)
model = model.cpu()  # cpu is more stable for random initialization across machines
with torch.no_grad():
    for name, p in sorted(model.named_parameters()):
        torch.nn.init.normal_(p, 0, 0.1)
        print(name, p.shape)

last_extra_layer.shared_head.head = deepcopy(model.get_output_embeddings())
last_extra_layer.embed_tokens = deepcopy(model.get_input_embeddings())
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['auto_map'] = {k: v.split('--')[-1] for k, v in config_json['auto_map'].items()}
with open(f"{save_folder}/config.json", "w", encoding='utf-8') as f:
    json.dump(config_json, f, indent=2)
hf_hub_download(source_model_id, filename='modeling_deepseek.py', repo_type='model',
                local_dir=save_folder, local_dir_use_symlinks=False, cache_dir='/tmp/')
with open(f'{save_folder}/modeling_deepseek.py', 'r', encoding='utf-8') as f:
    codes = f.read()
codes = codes.replace(
    "past_length = past_key_values.seen_tokens",
    "past_length = past_key_values.seen_tokens if hasattr(past_key_values, 'seen_tokens') else past_key_values.get_seq_length() # fix cache api deprecation"
)
codes = codes.replace(
    "max_cache_length = past_key_values.get_max_length()",
    "max_cache_length = past_key_values.get_max_length() if hasattr(past_key_values, 'get_max_length') else past_key_values.get_max_cache_shape() # fix cache api deprecation"
)
codes = codes.replace(
    "past_key_value.get_usable_length(",
    "getattr(past_key_value, 'get_usable_length', lambda *args, **kwargs: past_key_value.get_seq_length())("
)
codes = codes.replace(
    "past_key_values.get_usable_length(",
    "getattr(past_key_values, 'get_usable_length', lambda *args, **kwargs: past_key_values.get_seq_length())("
)
with open(f'{save_folder}/modeling_deepseek.py', 'w', encoding='utf-8') as f:
    f.write(codes)

Printing the model:

DeepseekV3ForCausalLM(
  (model): DeepseekV3Model(
    (embed_tokens): Embedding(129280, 8)
    (layers): ModuleList(
      (0): DeepseekV3DecoderLayer(
        (self_attn): DeepseekV3Attention(
          (q_a_proj): Linear(in_features=8, out_features=32, bias=False)
          (q_a_layernorm): DeepseekV3RMSNorm()
          (q_b_proj): Linear(in_features=32, out_features=1024, bias=False)
          (kv_a_proj_with_mqa): Linear(in_features=8, out_features=576, bias=False)
          (kv_a_layernorm): DeepseekV3RMSNorm()
          (kv_b_proj): Linear(in_features=384, out_features=512, bias=False)
          (o_proj): Linear(in_features=256, out_features=8, bias=False)
          (rotary_emb): DeepseekV3YarnRotaryEmbedding()
        )
        (mlp): DeepseekV3MLP(
          (gate_proj): Linear(in_features=8, out_features=64, bias=False)
          (up_proj): Linear(in_features=8, out_features=64, bias=False)
          (down_proj): Linear(in_features=64, out_features=8, bias=False)
          (act_fn): SiLU()
        )
        (input_layernorm): DeepseekV3RMSNorm()
        (post_attention_layernorm): DeepseekV3RMSNorm()
      )
      (1): DeepseekV3DecoderLayer(
        (self_attn): DeepseekV3Attention(
          (q_a_proj): Linear(in_features=8, out_features=32, bias=False)
          (q_a_layernorm): DeepseekV3RMSNorm()
          (q_b_proj): Linear(in_features=32, out_features=1024, bias=False)
          (kv_a_proj_with_mqa): Linear(in_features=8, out_features=576, bias=False)
          (kv_a_layernorm): DeepseekV3RMSNorm()
          (kv_b_proj): Linear(in_features=384, out_features=512, bias=False)
          (o_proj): Linear(in_features=256, out_features=8, bias=False)
          (rotary_emb): DeepseekV3YarnRotaryEmbedding()
        )
        (mlp): DeepseekV3MoE(
          (experts): ModuleList(
            (0-31): 32 x DeepseekV3MLP(
              (gate_proj): Linear(in_features=8, out_features=64, bias=False)
              (up_proj): Linear(in_features=8, out_features=64, bias=False)
              (down_proj): Linear(in_features=64, out_features=8, bias=False)
              (act_fn): SiLU()
            )
          )
          (gate): MoEGate()
          (shared_experts): DeepseekV3MLP(
            (gate_proj): Linear(in_features=8, out_features=64, bias=False)
            (up_proj): Linear(in_features=8, out_features=64, bias=False)
            (down_proj): Linear(in_features=64, out_features=8, bias=False)
            (act_fn): SiLU()
          )
        )
        (input_layernorm): DeepseekV3RMSNorm()
        (post_attention_layernorm): DeepseekV3RMSNorm()
      )
      (2): DeepseekV3DecoderLayer(
        (self_attn): DeepseekV3Attention(
          (q_a_proj): Linear(in_features=8, out_features=32, bias=False)
          (q_a_layernorm): DeepseekV3RMSNorm()
          (q_b_proj): Linear(in_features=32, out_features=1024, bias=False)
          (kv_a_proj_with_mqa): Linear(in_features=8, out_features=576, bias=False)
          (kv_a_layernorm): DeepseekV3RMSNorm()
          (kv_b_proj): Linear(in_features=384, out_features=512, bias=False)
          (o_proj): Linear(in_features=256, out_features=8, bias=False)
          (rotary_emb): DeepseekV3YarnRotaryEmbedding()
        )
        (mlp): DeepseekV3MoE(
          (experts): ModuleList(
            (0-31): 32 x DeepseekV3MLP(
              (gate_proj): Linear(in_features=8, out_features=64, bias=False)
              (up_proj): Linear(in_features=8, out_features=64, bias=False)
              (down_proj): Linear(in_features=64, out_features=8, bias=False)
              (act_fn): SiLU()
            )
          )
          (gate): MoEGate()
          (shared_experts): DeepseekV3MLP(
            (gate_proj): Linear(in_features=8, out_features=64, bias=False)
            (up_proj): Linear(in_features=8, out_features=64, bias=False)
            (down_proj): Linear(in_features=64, out_features=8, bias=False)
            (act_fn): SiLU()
          )
        )
        (input_layernorm): DeepseekV3RMSNorm()
        (post_attention_layernorm): DeepseekV3RMSNorm()
        (enorm): Glm4MoeRMSNorm((8,), eps=1e-06)
        (hnorm): Glm4MoeRMSNorm((8,), eps=1e-06)
        (eh_proj): Linear(in_features=16, out_features=8, bias=False)
        (shared_head): SharedHead(
          (norm): Glm4MoeRMSNorm((8,), eps=1e-06)
          (head): Linear(in_features=8, out_features=129280, bias=False)
        )
        (embed_tokens): Embedding(129280, 8)
      )
    )
    (norm): DeepseekV3RMSNorm()
  )
  (lm_head): Linear(in_features=8, out_features=129280, bias=False)
)