morrowmake/glm53-flash-cmp170hx-recipe is a recipe repository that runs GLM-5.3-Flash on four NVIDIA CMP 170HX cards with vLLM. The most recent commit 9d3b0be (2026-10-26) touched no code or configuration and added only two lines to README.md. The added sentence is a link to the DeepSeek-V4.1-Flash recipe repository that works on the same four CMP 170HX cards. It is a signal that this repository is being organized by card configuration while also pointing to recipes for other models.
What this repository is aiming at is a reproducible runtime environment with the hardware and the model pinned down. The engine and the Python environment are not left on the system but run only as a container image, and after cloning it takes only a one-off host setup before it can be launched with a single command. The README states that on startup, if the checks pass, PCIe P2P is turned on automatically, and if they do not pass, the server is brought up with P2P turned off. The figures given in the repository description are a maximum of 482.6 tok/s for a single user and a maximum of 943.7 tok/s for 8 users, with a context of 262K and a parallelism of either TP4 or PP4.
That said, this change is nothing more than the addition of a one-line pointer in the documentation. The performance numbers come from the repository's own description and are not results verified by this commit, and the configuration of the linked DeepSeek-V4.1-Flash recipe is not included in this record. Since the configuration and the code are unchanged, the behavior of the server does not change either. The original is on the page for commit 9d3b0be (https://github.com/Morrowmake/glm53-flash-cmp170hx-recipe/commit/9d3b0be374802e2f4e66cbc7ec7b1f203a6add78).