# Multi-token prediction is supported
model_id=tiny-random/glm-4-moe-lite
vllm serve $model_id \
     --tensor-parallel-size 2 \
     --speculative-config.method mtp \
     --speculative-config.num_speculative_tokens 1 \
     --tool-call-parser glm47 \
     --reasoning-parser glm45 \
     --enable-auto-tool-choice

SGLang

bash

# Multi-token prediction is supported
model_id=tiny-random/glm-4-moe-lite
python3 -m sglang.launch_server   --model-path $model_id   --tp-size 2 \
  --tool-call-parser glm47  \
  --reasoning-parser glm45 \
  --speculative-algorithm EAGLE \
  --speculative-num-steps 3 \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 4

Transformers

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model and tokenizer
model_id = "tiny-random/glm-4-moe-lite"
messages = [{"role": "user", "content": "hello"}]
tokenizer = AutoTokenizer.from_pretrained(model_id)
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
)
model = AutoModelForCausalLM.from_pretrained(
    pretrained_model_name_or_path=model_id,
    torch_dtype=torch.bfloat16,
    device_map="cuda",
)
inputs = inputs.to(model.device)
generated_ids = model.generate(
    **inputs, max_new_tokens=32, do_sample=False)
output_text = tokenizer.decode(
    generated_ids[0][inputs.input_ids.shape[1]:])
print(output_text)

Codes to create this repo:

python

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,
    GenerationConfig,
    set_seed,
)

source_model_id = "zai-org/GLM-4.7-Flash"
save_folder = "/tmp/tiny-random/glm-4-moe-lite"

processor = AutoProcessor.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.update({
    'kv_lora_rank': 384,
    'num_key_value_heads': 1,
    'q_lora_rank': 32,
    'qk_nope_head_dim': 64,
    'qk_rope_head_dim': 192,
    'v_head_dim': 64,
    'num_key_value_heads': 4,
    'num_attention_heads': 4,
})
config_json['hidden_size'] = 8
config_json['intermediate_size'] = 32
config_json['moe_intermediate_size'] = 32
config_json['num_hidden_layers'] = 2
config_json['tie_word_embeddings'] = False
config_json['use_cache'] = 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)
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,
    )
    model.generation_config.do_sample = True
    print(model.generation_config)
model = model.cpu()
set_seed(42)
with torch.no_grad():
    for name, p in sorted(model.named_parameters()):
        torch.nn.init.normal_(p, 0, 0.1)
        print(name, p.shape)
# MTP
set_seed(42)
model.model.layers.append(nn.ModuleDict(dict(
    embed_tokens=deepcopy(model.model.embed_tokens),
    shared_head=nn.ModuleDict(dict(
        norm=nn.RMSNorm(config.hidden_size),
        head=deepcopy(model.model.embed_tokens),
    )),
    eh_proj=nn.Linear(config.hidden_size * 2,
                      config.hidden_size, bias=False),
    enorm=nn.RMSNorm(config.hidden_size),
    hnorm=nn.RMSNorm(config.hidden_size),
    input_layernorm=nn.RMSNorm(config.hidden_size),
    post_attention_layernorm=nn.RMSNorm(config.hidden_size),
    self_attn=deepcopy(model.model.layers[1].self_attn),
    mlp=deepcopy(model.model.layers[1].mlp),
)))
for i in range(1, len(model.model.layers)):
    model.model.layers[i].mlp.gate.e_score_correction_bias = torch.rand_like(
        model.model.layers[i].mlp.gate.e_score_correction_bias).float()
model.save_pretrained(save_folder)
print(model)

Printing the model:

text

Glm4MoeLiteForCausalLM(
  (model): Glm4MoeLiteModel(
    (embed_tokens): Embedding(154880, 8, padding_idx=154820)
    (layers): ModuleList(
      (0): Glm4MoeLiteDecoderLayer(
        (self_attn): Glm4MoeLiteAttention(
          (q_a_proj): Linear(in_features=8, out_features=32, bias=False)
          (q_a_layernorm): Glm4MoeLiteRMSNorm((32,), eps=1e-06)
          (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): Glm4MoeLiteRMSNorm((384,), eps=1e-06)
          (kv_b_proj): Linear(in_features=384, out_features=512, bias=False)
          (o_proj): Linear(in_features=256, out_features=8, bias=False)
        )
        (mlp): Glm4MoeLiteMLP(
          (gate_proj): Linear(in_features=8, out_features=32, bias=False)
          (up_proj): Linear(in_features=8, out_features=32, bias=False)
          (down_proj): Linear(in_features=32, out_features=8, bias=False)
          (act_fn): SiLUActivation()
        )
        (input_layernorm): Glm4MoeLiteRMSNorm((8,), eps=1e-05)
        (post_attention_layernorm): Glm4MoeLiteRMSNorm((8,), eps=1e-05)
      )
      (1): Glm4MoeLiteDecoderLayer(
        (self_attn): Glm4MoeLiteAttention(
          (q_a_proj): Linear(in_features=8, out_features=32, bias=False)
          (q_a_layernorm): Glm4MoeLiteRMSNorm((32,), eps=1e-06)
          (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): Glm4MoeLiteRMSNorm((384,), eps=1e-06)
          (kv_b_proj): Linear(in_features=384, out_features=512, bias=False)
          (o_proj): Linear(in_features=256, out_features=8, bias=False)
        )
        (mlp): Glm4MoeLiteMoE(
          (experts): Glm4MoeLiteNaiveMoe(
            (act_fn): SiLUActivation()
          )
          (gate): Glm4MoeLiteTopkRouter()
          (shared_experts): Glm4MoeLiteMLP(
            (gate_proj): Linear(in_features=8, out_features=32, bias=False)
            (up_proj): Linear(in_features=8, out_features=32, bias=False)
            (down_proj): Linear(in_features=32, out_features=8, bias=False)
            (act_fn): SiLUActivation()
          )
        )
        (input_layernorm): Glm4MoeLiteRMSNorm((8,), eps=1e-05)
        (post_attention_layernorm): Glm4MoeLiteRMSNorm((8,), eps=1e-05)
      )
      (2): ModuleDict(
        (embed_tokens): Embedding(154880, 8, padding_idx=154820)
        (shared_head): ModuleDict(
          (norm): RMSNorm((8,), eps=None, elementwise_affine=True)
          (head): Embedding(154880, 8, padding_idx=154820)
        )
        (eh_proj): Linear(in_features=16, out_features=8, bias=False)
        (enorm): RMSNorm((8,), eps=None, elementwise_affine=True)
        (hnorm): RMSNorm((8,), eps=None, elementwise_affine=True)
        (input_layernorm): RMSNorm((8,), eps=None, elementwise_affine=True)
        (post_attention_layernorm): RMSNorm((8,), eps=None, elementwise_affine=True)
        (self_attn): Glm4MoeLiteAttention(
          (q_a_proj): Linear(in_features=8, out_features=32, bias=False)
          (q_a_layernorm): Glm4MoeLiteRMSNorm((32,), eps=1e-06)
          (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): Glm4MoeLiteRMSNorm((384,), eps=1e-06)
          (kv_b_proj): Linear(in_features=384, out_features=512, bias=False)
          (o_proj): Linear(in_features=256, out_features=8, bias=False)
        )
        (mlp): Glm4MoeLiteMoE(
          (experts): Glm4MoeLiteNaiveMoe(
            (act_fn): SiLUActivation()
          )
          (gate): Glm4MoeLiteTopkRouter()
          (shared_experts): Glm4MoeLiteMLP(
            (gate_proj): Linear(in_features=8, out_features=32, bias=False)
            (up_proj): Linear(in_features=8, out_features=32, bias=False)
            (down_proj): Linear(in_features=32, out_features=8, bias=False)
            (act_fn): SiLUActivation()
          )
        )
      )
    )
    (norm): Glm4MoeLiteRMSNorm((8,), eps=1e-05)
    (rotary_emb): Glm4MoeLiteRotaryEmbedding()
  )
  (lm_head): Linear(in_features=8, out_features=154880, bias=False)
)