v1.5.2

新特性

1. 新增 model_type 支持：GLM-5.2（glm_moe_dsa 新增 indexer_type = 'shared' 支持）、minicpmv4_6。
2. moe_router_load_balancing_type 支持同时设置多个类型。
3. 修复 DeepSeek-V4 FP8 相关问题。

New Features

1. Added support for new model_type: GLM-5.2 (with indexer_type = 'shared' support added for glm_moe_dsa) and minicpmv4_6.
2. moe_router_load_balancing_type now supports specifying multiple types simultaneously.
3. Fixed issues related to DeepSeek-V4 FP8.

What's Changed

• [compat] MTP compat megatron-core main branch / dev branch by @Jintao-Huang in #127
• [bugfix] regenerate columnwise data after FP8 weight loading by @HorizonChaser in #126
• [bugfix] fix get_logger by @Jintao-Huang in #131
• [compat] moe_router_load_balancing_type compat mcore>=0.16 by @Jintao-Huang in #130
• [bugfix] fix GLM 5.1 FP8 by @Jintao-Huang in #132
• [bugfix] Fix deepseek-v4 fp8 by @Jintao-Huang in #133
• [bugfix] fix glm5.2 indexer_type 'shared' by @Jintao-Huang in #134
• [bugfix] fix deepseek v4 fp8 by @Jintao-Huang in #135
• [bugfix] fix multi-LoRA load assertion failure on idle adapter slots by @Yunnglin in #136
• [model] Add MiniCPM-V 4.6 model support by @randydl in #137
• update minicpm4_6 by @Jintao-Huang in #138

New Contributors

• @HorizonChaser made their first contribution in #126
• @Yunnglin made their first contribution in #136

Full Changelog: v1.5.1...v1.5.2