biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Model] Add LongCat-Flash (#23991)

Browse Source 
Signed-off-by: yangxurui <yangxurui@meituan.com>
Co-authored-by: yangxurui <yangxurui@meituan.com>
This commit is contained in:
XuruiYang
2025-09-25 12:53:40 +08:00
committed by GitHub
parent 90b139cfff
commit 845adb3ec6
31 changed files with 1357 additions and 66 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
vllm/config/speculative.py
View File

@@ -31,7 +31,8 @@ logger = init_logger(__name__)
SpeculativeMethod = Literal["ngram", "eagle", "eagle3", "medusa",
"mlp_speculator", "draft_model", "deepseek_mtp",
"ernie_mtp", "qwen3_next_mtp", "mimo_mtp"]
"ernie_mtp", "qwen3_next_mtp", "mimo_mtp",
"longcat_flash_mtp"]
@config
@@ -186,6 +187,13 @@ class SpeculativeConfig:
"n_predict": n_predict,
"architectures": ["Qwen3NextMTP"]
})
if hf_config.model_type == "longcat_flash":
hf_config.model_type = "longcat_flash_mtp"
n_predict = getattr(hf_config, "num_nextn_predict_layers", 1)
hf_config.update({
"n_predict": n_predict,
"architectures": ["LongCatFlashMTPModel"]
})
return hf_config
@@ -332,6 +340,15 @@ class SpeculativeConfig:
"one layer. Might need some code changes " \
"to support multiple layers."
)
elif (self.draft_model_config.hf_config.model_type
in ("longcat_flash_mtp")):
self.method = "longcat_flash_mtp"
if self.num_speculative_tokens > 1:
logger.warning(
"LongCat MTP models only have " \
"one layer. Might need some code changes " \
"to support multiple layers."
)
else:
self.method = "draft_model"
raise NotImplementedError(
@@ -548,7 +565,7 @@ class SpeculativeConfig:
def use_eagle(self) -> bool:
return self.method in ("eagle", "eagle3", "deepseek_mtp", "ernie_mtp",
"qwen3_next_mtp")
"qwen3_next_mtp", "longcat_flash_mtp")
def __repr__(self) -> str:
method = self.method