Instructions to use meituan-longcat/LongCat-Flash-Lite-FP8 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use meituan-longcat/LongCat-Flash-Lite-FP8 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="meituan-longcat/LongCat-Flash-Lite-FP8", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("meituan-longcat/LongCat-Flash-Lite-FP8", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use meituan-longcat/LongCat-Flash-Lite-FP8 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "meituan-longcat/LongCat-Flash-Lite-FP8"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "meituan-longcat/LongCat-Flash-Lite-FP8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/meituan-longcat/LongCat-Flash-Lite-FP8

SGLang

How to use meituan-longcat/LongCat-Flash-Lite-FP8 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "meituan-longcat/LongCat-Flash-Lite-FP8" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "meituan-longcat/LongCat-Flash-Lite-FP8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "meituan-longcat/LongCat-Flash-Lite-FP8" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "meituan-longcat/LongCat-Flash-Lite-FP8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use meituan-longcat/LongCat-Flash-Lite-FP8 with Docker Model Runner:
```
docker model run hf.co/meituan-longcat/LongCat-Flash-Lite-FP8
```

LongCat-Flash-Lite-FP8 / config.json

LongCat0830

Add files using upload-large-folder tool

86732a4 verified 7 months ago

Raw

History Blame Contribute Delete

9.96 kB

	{
	"architectures": [
	"LongcatFlashNgramForCausalLM"
	],
	"attention_bias": false,
	"attention_dropout": 0.0,
	"auto_map": {
	"AutoConfig": "configuration_longcat_ngram.LongcatFlashNgramConfig",
	"AutoModel": "modeling_longcat_ngram.LongcatFlashNgramModel",
	"AutoModelForCausalLM": "modeling_longcat_ngram.LongcatFlashNgramForCausalLM"
	},
	"bos_token_id": 1,
	"disable_quant_module": [],
	"emb_neighbor_num": 4,
	"emb_split_num": 4,
	"eos_token_id": 2,
	"expert_ffn_hidden_size": 1024,
	"ffn_hidden_size": 6144,
	"hidden_size": 3072,
	"kv_lora_rank": 512,
	"max_position_embeddings": 327680,
	"mla_scale_kv_lora": true,
	"mla_scale_q_lora": true,
	"moe_topk": 12,
	"n_routed_experts": 256,
	"ngram_vocab_size_ratio": 78,
	"num_attention_heads": 32,
	"num_layers": 14,
	"q_lora_rank": 1536,
	"qk_nope_head_dim": 128,
	"qk_rope_head_dim": 64,
	"quantization_config": {
	"activation_scheme": "dynamic",
	"fmt": "e4m3",
	"ignored_layers": [
	"lm_head",
	"model.embed_tokens",
	"model.layers.0.input_layernorm.0",
	"model.layers.0.input_layernorm.1",
	"model.layers.0.mlp.router.classifier",
	"model.layers.0.mlp.router.e_score_correction_bias",
	"model.layers.0.post_attention_layernorm.0",
	"model.layers.0.post_attention_layernorm.1",
	"model.layers.0.self_attn.0.kv_a_layernorm",
	"model.layers.0.self_attn.0.q_a_layernorm",
	"model.layers.0.self_attn.1.kv_a_layernorm",
	"model.layers.0.self_attn.1.q_a_layernorm",
	"model.layers.1.input_layernorm.0",
	"model.layers.1.input_layernorm.1",
	"model.layers.1.mlp.router.classifier",
	"model.layers.1.mlp.router.e_score_correction_bias",
	"model.layers.1.post_attention_layernorm.0",
	"model.layers.1.post_attention_layernorm.1",
	"model.layers.1.self_attn.0.kv_a_layernorm",
	"model.layers.1.self_attn.0.q_a_layernorm",
	"model.layers.1.self_attn.1.kv_a_layernorm",
	"model.layers.1.self_attn.1.q_a_layernorm",
	"model.layers.10.input_layernorm.0",
	"model.layers.10.input_layernorm.1",
	"model.layers.10.mlp.router.classifier",
	"model.layers.10.mlp.router.e_score_correction_bias",
	"model.layers.10.post_attention_layernorm.0",
	"model.layers.10.post_attention_layernorm.1",
	"model.layers.10.self_attn.0.kv_a_layernorm",
	"model.layers.10.self_attn.0.q_a_layernorm",
	"model.layers.10.self_attn.1.kv_a_layernorm",
	"model.layers.10.self_attn.1.q_a_layernorm",
	"model.layers.11.input_layernorm.0",
	"model.layers.11.input_layernorm.1",
	"model.layers.11.mlp.router.classifier",
	"model.layers.11.mlp.router.e_score_correction_bias",
	"model.layers.11.post_attention_layernorm.0",
	"model.layers.11.post_attention_layernorm.1",
	"model.layers.11.self_attn.0.kv_a_layernorm",
	"model.layers.11.self_attn.0.q_a_layernorm",
	"model.layers.11.self_attn.1.kv_a_layernorm",
	"model.layers.11.self_attn.1.q_a_layernorm",
	"model.layers.12.input_layernorm.0",
	"model.layers.12.input_layernorm.1",
	"model.layers.12.mlp.router.classifier",
	"model.layers.12.mlp.router.e_score_correction_bias",
	"model.layers.12.post_attention_layernorm.0",
	"model.layers.12.post_attention_layernorm.1",
	"model.layers.12.self_attn.0.kv_a_layernorm",
	"model.layers.12.self_attn.0.q_a_layernorm",
	"model.layers.12.self_attn.1.kv_a_layernorm",
	"model.layers.12.self_attn.1.q_a_layernorm",
	"model.layers.13.input_layernorm.0",
	"model.layers.13.input_layernorm.1",
	"model.layers.13.mlp.router.classifier",
	"model.layers.13.mlp.router.e_score_correction_bias",
	"model.layers.13.post_attention_layernorm.0",
	"model.layers.13.post_attention_layernorm.1",
	"model.layers.13.self_attn.0.kv_a_layernorm",
	"model.layers.13.self_attn.0.q_a_layernorm",
	"model.layers.13.self_attn.1.kv_a_layernorm",
	"model.layers.13.self_attn.1.q_a_layernorm",
	"model.layers.2.input_layernorm.0",
	"model.layers.2.input_layernorm.1",
	"model.layers.2.mlp.router.classifier",
	"model.layers.2.mlp.router.e_score_correction_bias",
	"model.layers.2.post_attention_layernorm.0",
	"model.layers.2.post_attention_layernorm.1",
	"model.layers.2.self_attn.0.kv_a_layernorm",
	"model.layers.2.self_attn.0.q_a_layernorm",
	"model.layers.2.self_attn.1.kv_a_layernorm",
	"model.layers.2.self_attn.1.q_a_layernorm",
	"model.layers.3.input_layernorm.0",
	"model.layers.3.input_layernorm.1",
	"model.layers.3.mlp.router.classifier",
	"model.layers.3.mlp.router.e_score_correction_bias",
	"model.layers.3.post_attention_layernorm.0",
	"model.layers.3.post_attention_layernorm.1",
	"model.layers.3.self_attn.0.kv_a_layernorm",
	"model.layers.3.self_attn.0.q_a_layernorm",
	"model.layers.3.self_attn.1.kv_a_layernorm",
	"model.layers.3.self_attn.1.q_a_layernorm",
	"model.layers.4.input_layernorm.0",
	"model.layers.4.input_layernorm.1",
	"model.layers.4.mlp.router.classifier",
	"model.layers.4.mlp.router.e_score_correction_bias",
	"model.layers.4.post_attention_layernorm.0",
	"model.layers.4.post_attention_layernorm.1",
	"model.layers.4.self_attn.0.kv_a_layernorm",
	"model.layers.4.self_attn.0.q_a_layernorm",
	"model.layers.4.self_attn.1.kv_a_layernorm",
	"model.layers.4.self_attn.1.q_a_layernorm",
	"model.layers.5.input_layernorm.0",
	"model.layers.5.input_layernorm.1",
	"model.layers.5.mlp.router.classifier",
	"model.layers.5.mlp.router.e_score_correction_bias",
	"model.layers.5.post_attention_layernorm.0",
	"model.layers.5.post_attention_layernorm.1",
	"model.layers.5.self_attn.0.kv_a_layernorm",
	"model.layers.5.self_attn.0.q_a_layernorm",
	"model.layers.5.self_attn.1.kv_a_layernorm",
	"model.layers.5.self_attn.1.q_a_layernorm",
	"model.layers.6.input_layernorm.0",
	"model.layers.6.input_layernorm.1",
	"model.layers.6.mlp.router.classifier",
	"model.layers.6.mlp.router.e_score_correction_bias",
	"model.layers.6.post_attention_layernorm.0",
	"model.layers.6.post_attention_layernorm.1",
	"model.layers.6.self_attn.0.kv_a_layernorm",
	"model.layers.6.self_attn.0.q_a_layernorm",
	"model.layers.6.self_attn.1.kv_a_layernorm",
	"model.layers.6.self_attn.1.q_a_layernorm",
	"model.layers.7.input_layernorm.0",
	"model.layers.7.input_layernorm.1",
	"model.layers.7.mlp.router.classifier",
	"model.layers.7.mlp.router.e_score_correction_bias",
	"model.layers.7.post_attention_layernorm.0",
	"model.layers.7.post_attention_layernorm.1",
	"model.layers.7.self_attn.0.kv_a_layernorm",
	"model.layers.7.self_attn.0.q_a_layernorm",
	"model.layers.7.self_attn.1.kv_a_layernorm",
	"model.layers.7.self_attn.1.q_a_layernorm",
	"model.layers.8.input_layernorm.0",
	"model.layers.8.input_layernorm.1",
	"model.layers.8.mlp.router.classifier",
	"model.layers.8.mlp.router.e_score_correction_bias",
	"model.layers.8.post_attention_layernorm.0",
	"model.layers.8.post_attention_layernorm.1",
	"model.layers.8.self_attn.0.kv_a_layernorm",
	"model.layers.8.self_attn.0.q_a_layernorm",
	"model.layers.8.self_attn.1.kv_a_layernorm",
	"model.layers.8.self_attn.1.q_a_layernorm",
	"model.layers.9.input_layernorm.0",
	"model.layers.9.input_layernorm.1",
	"model.layers.9.mlp.router.classifier",
	"model.layers.9.mlp.router.e_score_correction_bias",
	"model.layers.9.post_attention_layernorm.0",
	"model.layers.9.post_attention_layernorm.1",
	"model.layers.9.self_attn.0.kv_a_layernorm",
	"model.layers.9.self_attn.0.q_a_layernorm",
	"model.layers.9.self_attn.1.kv_a_layernorm",
	"model.layers.9.self_attn.1.q_a_layernorm",
	"model.mtp.embed_tokens",
	"model.mtp.layers.0.enorm.m",
	"model.mtp.layers.0.hnorm.m",
	"model.mtp.layers.0.input_layernorm",
	"model.mtp.layers.0.post_attention_layernorm",
	"model.mtp.layers.0.self_attn.kv_a_layernorm",
	"model.mtp.layers.0.self_attn.q_a_layernorm",
	"model.mtp.norm",
	"model.ngram_embeddings.embedders.0",
	"model.ngram_embeddings.embedders.1",
	"model.ngram_embeddings.embedders.10",
	"model.ngram_embeddings.embedders.11",
	"model.ngram_embeddings.embedders.2",
	"model.ngram_embeddings.embedders.3",
	"model.ngram_embeddings.embedders.4",
	"model.ngram_embeddings.embedders.5",
	"model.ngram_embeddings.embedders.6",
	"model.ngram_embeddings.embedders.7",
	"model.ngram_embeddings.embedders.8",
	"model.ngram_embeddings.embedders.9",
	"model.ngram_embeddings.post_projs.0",
	"model.ngram_embeddings.post_projs.1",
	"model.ngram_embeddings.post_projs.10",
	"model.ngram_embeddings.post_projs.11",
	"model.ngram_embeddings.post_projs.2",
	"model.ngram_embeddings.post_projs.3",
	"model.ngram_embeddings.post_projs.4",
	"model.ngram_embeddings.post_projs.5",
	"model.ngram_embeddings.post_projs.6",
	"model.ngram_embeddings.post_projs.7",
	"model.ngram_embeddings.post_projs.8",
	"model.ngram_embeddings.post_projs.9",
	"model.norm"
	],
	"quant_method": "fp8",
	"weight_block_size": [
	128,
	128
	]
	},
	"rms_norm_eps": 1e-05,
	"rope_scaling": {
	"beta_fast": 32,
	"beta_slow": 1,
	"factor": 10,
	"mscale": 1,
	"mscale_all_dim": 1,
	"original_max_position_embeddings": 32768,
	"rope_type": "yarn"
	},
	"rope_theta": 5000000.0,
	"routed_scaling_factor": 6.0,
	"torch_dtype": "bfloat16",
	"transformers_version": "4.57.6",
	"use_cache": true,
	"v_head_dim": 128,
	"vocab_size": 131072,
	"zero_expert_num": 128,
	"zero_expert_type": "identity"
	}