EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[Model] Support Meituan LongCat-Flash && LongCat-Flash-MTP (#9824)

Browse Source
This commit is contained in:
Guoyuan Lin
2025-08-31 14:29:21 +08:00
committed by GitHub
parent fd5ce576a4
commit 5e194b2143
10 changed files with 1940 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/srt/configs/__init__.py
View File

@@ -5,6 +5,7 @@ from sglang.srt.configs.exaone import ExaoneConfig
from sglang.srt.configs.janus_pro import MultiModalityConfig
from sglang.srt.configs.kimi_vl import KimiVLConfig
from sglang.srt.configs.kimi_vl_moonvit import MoonViTConfig
from sglang.srt.configs.longcat_flash import LongcatFlashConfig
from sglang.srt.configs.step3_vl import (
Step3TextConfig,
Step3VisionEncoderConfig,
@@ -16,6 +17,7 @@ __all__ = [
"ChatGLMConfig",
"DbrxConfig",
"DeepseekVL2Config",
"LongcatFlashConfig",
"MultiModalityConfig",
"KimiVLConfig",
"MoonViTConfig",

104
python/sglang/srt/configs/longcat_flash.py Normal file
View File

@@ -0,0 +1,104 @@
from transformers.configuration_utils import PretrainedConfig
from transformers.utils import logging
logger = logging.get_logger(__name__)
FLASH_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
class LongcatFlashConfig(PretrainedConfig):
model_type = "longcat_flash"
keys_to_ignore_at_inference = ["past_key_values"]
def __init__(
self,
vocab_size=131072,
hidden_size=6144,
intermediate_size=None,
ffn_hidden_size=12288,
expert_ffn_hidden_size=2048,
num_layers=28,
num_hidden_layers=None,
num_attention_heads=64,
ep_size=1,
kv_lora_rank=512,
q_lora_rank=1536,
qk_rope_head_dim=128,
qk_nope_head_dim=128,
v_head_dim=128,
n_routed_experts=512,
moe_topk=12,
norm_topk_prob=False,
max_position_embeddings=131072,
rms_norm_eps=1e-05,
use_cache=True,
pad_token_id=None,
bos_token_id=1,
eos_token_id=2,
pretraining_tp=1,
tie_word_embeddings=False,
rope_theta=10000000.0,
rope_scaling=None,
attention_bias=False,
attention_dropout=0.0,
mla_scale_q_lora=True,
mla_scale_kv_lora=True,
torch_dtype="bfloat16",
params_dtype="bfloat16",
rounter_params_dtype="float32",
router_bias=False,
topk_method=None,
routed_scaling_factor=6.0,
zero_expert_num=256,
zero_expert_type="identity",
nextn_use_scmoe=False,
num_nextn_predict_layers=1,
**kwargs,
):
super().__init__(
pad_token_id=pad_token_id,
bos_token_id=bos_token_id,
eos_token_id=eos_token_id,
tie_word_embeddings=tie_word_embeddings,
torch_dtype=torch_dtype,
params_dtype=params_dtype,
rounter_params_dtype=rounter_params_dtype,
topk_method=topk_method,
router_bias=router_bias,
nextn_use_scmoe=nextn_use_scmoe,
num_nextn_predict_layers=num_nextn_predict_layers,
**kwargs,
)
self.vocab_size = vocab_size
self.max_position_embeddings = max_position_embeddings
self.hidden_size = hidden_size
self.num_hidden_layers = (
num_hidden_layers if num_hidden_layers is not None else num_layers
)
self.intermediate_size = (
intermediate_size if intermediate_size is not None else ffn_hidden_size
)
self.moe_intermediate_size = expert_ffn_hidden_size
self.num_attention_heads = num_attention_heads
self.ep_size = ep_size
self.kv_lora_rank = kv_lora_rank
self.q_lora_rank = q_lora_rank
self.qk_rope_head_dim = qk_rope_head_dim
self.v_head_dim = v_head_dim
self.qk_nope_head_dim = qk_nope_head_dim
self.n_routed_experts = n_routed_experts
self.moe_topk = moe_topk
self.norm_topk_prob = norm_topk_prob
self.rms_norm_eps = rms_norm_eps
self.pretraining_tp = pretraining_tp
self.use_cache = use_cache
self.rope_theta = rope_theta
self.rope_scaling = rope_scaling
self.attention_bias = attention_bias
self.attention_dropout = attention_dropout
self.mla_scale_q_lora = mla_scale_q_lora
self.mla_scale_kv_lora = mla_scale_kv_lora
self.zero_expert_num = zero_expert_num
self.zero_expert_type = zero_expert_type
self.routed_scaling_factor = routed_scaling_factor
self.hidden_act = "silu"

12
python/sglang/srt/configs/model_config.py
View File

@@ -132,6 +132,13 @@ class ModelConfig:
if is_draft_model and self.hf_config.architectures[0] == "Glm4MoeForCausalLM":
self.hf_config.architectures[0] = "Glm4MoeForCausalLMNextN"
if (
is_draft_model
and self.hf_config.architectures[0] == "LongcatFlashForCausalLM"
):
self.hf_config.architectures[0] = "LongcatFlashForCausalLMNextN"
self.hf_config.num_hidden_layers = self.hf_config.num_nextn_predict_layers
if is_draft_model and self.hf_config.architectures[0] == "MiMoForCausalLM":
self.hf_config.architectures[0] = "MiMoMTP"
if (
@@ -199,6 +206,8 @@ class ModelConfig:
"DeepseekV2ForCausalLM" in self.hf_config.architectures
or "DeepseekV3ForCausalLM" in self.hf_config.architectures
or "DeepseekV3ForCausalLMNextN" in self.hf_config.architectures
or "LongcatFlashForCausalLM" in self.hf_config.architectures
or "LongcatFlashForCausalLMNextN" in self.hf_config.architectures
):
self.head_dim = 256
self.attention_arch = AttentionArch.MLA
@@ -270,6 +279,9 @@ class ModelConfig:
self.num_key_value_heads = self.num_attention_heads
self.hidden_size = self.hf_text_config.hidden_size
self.num_hidden_layers = self.hf_text_config.num_hidden_layers
self.num_attention_layers = self.num_hidden_layers
if "LongcatFlashForCausalLM" in self.hf_config.architectures:
self.num_attention_layers = self.num_hidden_layers * 2
self.num_nextn_predict_layers = getattr(
self.hf_text_config, "num_nextn_predict_layers", None
)