3 Commits

Author SHA1 Message Date
GSCSD1
a48af5cc69 [data] clarify _nothink suffix warning for reasoning-only models (#10613)
Co-authored-by: GSCSD1 <GSCSD1@users.noreply.github.com>
Co-authored-by: Claude <noreply@anthropic.com>
2026-06-30 17:15:29 +08:00
souljoy
c383c0d067 [model] add Qwen-AgentWorld-35B-A3B support (#10615) 2026-06-30 17:15:09 +08:00
HelloWorldBeginner
50ff45176a [v1] set flash_attn to flash_attention_2 for ulysses CP example (#10616)
Co-authored-by: mhh111 <mahonghao1@huawei.com>
2026-06-30 16:56:12 +08:00
3 changed files with 18 additions and 2 deletions

View File

@@ -3,6 +3,7 @@ trust_remote_code: true
model_class: llm
template: qwen3_nothink
flash_attn: flash_attention_2
# FSDP Config
dist_config:

View File

@@ -655,8 +655,10 @@ def get_template_and_fix_tokenizer(tokenizer: "PreTrainedTokenizer", data_args:
if isinstance(template, ReasoningTemplate):
logger.warning_rank0(
"You are using reasoning template, "
"please add `_nothink` suffix if the model is not a reasoning model. "
"You are using reasoning template. "
"If the base model is NOT a reasoning model (i.e., it has a separate Instruct variant), "
"please add `_nothink` suffix to disable thinking. "
"For reasoning-only model families (e.g., Qwen3.6), the suffix is not needed. "
"e.g., qwen3_vl_nothink"
)
template.enable_thinking = data_args.enable_thinking

View File

@@ -3372,6 +3372,19 @@ register_model_group(
)
# Qwen-AgentWorld: language world model based on Qwen3.5-35B-A3B (MoE).
# Reference: https://github.com/QwenLM/Qwen-AgentWorld
register_model_group(
models={
"Qwen-AgentWorld-35B-A3B-Thinking": {
DownloadSource.DEFAULT: "Qwen/Qwen-AgentWorld-35B-A3B",
DownloadSource.MODELSCOPE: "Qwen/Qwen-AgentWorld-35B-A3B",
},
},
template="qwen3",
)
register_model_group(
models={
"Vicuna-v1.5-7B-Chat": {