Skip to content

feat: 支持 GLM-4.7-Flash 推理 - #25

Merged
kuma-loong merged 48 commits into
CURRENTF:mainfrom
kuma-loong:codex/glm-4.7-flash
Aug 13, 2026
Merged

feat: 支持 GLM-4.7-Flash 推理#25
kuma-loong merged 48 commits into
CURRENTF:mainfrom
kuma-loong:codex/glm-4.7-flash

Conversation

@kuma-loong

Copy link
Copy Markdown
Collaborator

前置依赖

修改内容

  • 增加 GLM-4.7-Flash BF16 latent MLA、Dense/MoE 模型与 checkpoint 加载支持。
  • 支持 TP、EP 及 outer-TP MoE 混合并行布局。
  • 增加 MLA cache storage、prefill/decode provider、CUDA Graph 与 prefix cache 支持。
  • 集成 Triton、TileLang 和 SGL Kernel provider,并绑定锁定依赖范围。
  • 补充 OpenAI serving、LongBench 状态记录、模型文档与依赖锁。
  • 清理一次性 checkpoint/并行验证脚本,保留生产测试和可复现验证入口。

验证

  • 全量测试(排除要求 clean worktree 的 simulated deep research 测试):1804 passed,32 skipped,215 subtests passed。
  • H100 TileLang MLA:10 passed。
  • H100,Torch 2.9.1+cu129、Triton 3.5.1、sgl-kernel 0.3.21:FA3/MoE 33 passed。
  • 定向 provider/config 回归:224 passed,26 skipped。

@kuma-loong
kuma-loong marked this pull request as ready for review August 12, 2026 08:22
@kuma-loong
kuma-loong force-pushed the codex/glm-4.7-flash branch from a710a77 to cbb70ff Compare August 12, 2026 08:58
@kuma-loong kuma-loong closed this Aug 12, 2026
@kuma-loong kuma-loong reopened this Aug 12, 2026
@kuma-loong
kuma-loong force-pushed the codex/glm-4.7-flash branch from 7c59871 to 97a3051 Compare August 13, 2026 13:55
@kuma-loong
kuma-loong merged commit 8900ad4 into CURRENTF:main Aug 13, 2026
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants