vllm serve tiny-random/longcat-flash \
    --trust-remote-code \
    --enable-expert-parallel \
    --tensor-parallel-size 1 \
    --speculative_config '{"model": "tiny-random/longcat-flash", "num_speculative_tokens": 1, "method":"longcat_flash_mtp"}'

SGLang

bash

python3 -m sglang.launch_server \
    --model tiny-random/longcat-flash \
    --trust-remote-code \
    --attention-backend flashinfer \
    --enable-ep-moe \
    --tp 1 \
    --speculative-draft-model-path tiny-random/longcat-flash \
    --speculative-algorithm NEXTN \
    --speculative-num-draft-tokens 2 \
    --speculative-num-steps 1 \
    --speculative-eagle-topk 1

Transformers

python

import torch
import transformers

model_id = "tiny-random/longcat-flash"
pipe = transformers.pipelines.pipeline(
    'text-generation',
    model=model_id,
    trust_remote_code=True,
    device_map='cuda',
    torch_dtype=torch.bfloat16,
)
past_key_values = transformers.DynamicCache(config=None)  # set config to None
r = pipe('Hello, world!', past_key_values=past_key_values, max_new_tokens=32)
print(r)

Codes to create this repo:

python

import json
from copy import deepcopy
from pathlib import Path

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 = "meituan-longcat/LongCat-Flash-Chat"
save_folder = "/tmp/tiny-random/longcat-flash"

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'), '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({
    'num_layers': 2,
    'hidden_size': 8,
    'ffn_hidden_size': 64,
    'expert_ffn_hidden_size': 64,
    'num_attention_heads': 4,
    'kv_lora_rank': 384,
    'n_routed_experts': 32,
    '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,
    'moe_topk': 12,
    'zero_expert_num': 16,
})
# 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,
    )
model = model.cpu()
# MTP
model.model.mtp = nn.ModuleDict({
    "layers": nn.ModuleList([nn.ModuleDict(dict(
        eh_proj=nn.Linear(config.hidden_size * 2, config.hidden_size, bias=False),
        enorm=nn.ModuleDict({"m": nn.RMSNorm(config.hidden_size)}),
        hnorm=nn.ModuleDict({"m": 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[0].self_attn[0]),
        transformer_layer=nn.ModuleDict({"mlp": deepcopy(model.model.layers[0].mlps[0])}),
    ))]),
    "norm": nn.RMSNorm(config.hidden_size),
})
for i in range(config.num_layers):
    model.model.layers[i].mlp.router = model.model.layers[i].mlp.router.float()
    # model.model.layers[i].mlp.router.e_score_correction_bias = torch.zeros((config.n_routed_experts + config.zero_expert_num)).float()
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, p.dtype)
model.model.mtp.embed_tokens = deepcopy(model.model.embed_tokens)

model.save_pretrained(save_folder)
torch.set_default_dtype(torch.float32)

for n, m in model.named_modules():
    if 'LongcatFlashMLA' in str(type(m)):
        print(n, m.layer_idx)

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)

Printing the model:

text

LongcatFlashForCausalLM(
  (model): LongcatFlashModel(
    (embed_tokens): Embedding(131072, 8)
    (layers): ModuleList(
      (0-1): 2 x LongcatFlashDecoderLayer(
        (mlp): LongcatFlashMoE(
          (experts): ModuleList(
            (0-31): 32 x LongcatFlashMLP(
              (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()
            )
          )
          (router): LongcatFlashTopkRouter(
            (classifier): Linear(in_features=8, out_features=48, bias=False)
          )
        )
        (self_attn): ModuleList(
          (0-1): 2 x LongcatFlashMLA(
            (q_a_proj): Linear(in_features=8, out_features=32, bias=False)
            (q_a_layernorm): LongcatFlashRMSNorm((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): LongcatFlashRMSNorm((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)
          )
        )
        (mlps): ModuleList(
          (0-1): 2 x LongcatFlashMLP(
            (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): ModuleList(
          (0-1): 2 x LongcatFlashRMSNorm((8,), eps=1e-05)
        )
        (post_attention_layernorm): ModuleList(
          (0-1): 2 x LongcatFlashRMSNorm((8,), eps=1e-05)
        )
      )
    )
    (norm): LongcatFlashRMSNorm((8,), eps=1e-05)
    (rotary_emb): LongcatFlashRotaryEmbedding()
    (mtp): ModuleDict(
      (layers): ModuleList(
        (0): ModuleDict(
          (eh_proj): Linear(in_features=16, out_features=8, bias=False)
          (enorm): ModuleDict(
            (m): RMSNorm((8,), eps=None, elementwise_affine=True)
          )
          (hnorm): ModuleDict(
            (m): 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): LongcatFlashMLA(
            (q_a_proj): Linear(in_features=8, out_features=32, bias=False)
            (q_a_layernorm): LongcatFlashRMSNorm((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): LongcatFlashRMSNorm((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)
          )
          (transformer_layer): ModuleDict(
            (mlp): LongcatFlashMLP(
              (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()
            )
          )
        )
      )
      (norm): RMSNorm((8,), eps=None, elementwise_affine=True)
      (embed_tokens): Embedding(131072, 8)
    )
  )
  (lm_head): Linear(in_features=8, out_features=131072, bias=False)
)