GLM-5.2 (GlmMoeDsa, ~744B) runs on DS4 across 2x H200 — report + reusable findings

[slackarea]

GLM-5.2 (Zhipu, MIT — GlmMoeDsaForCausalLM, DeepSeek-V3.2-family: MLA + DSA indexer + MoE(+shared) + MTP) runs on DS4: loads, computes correct logits (corr 0.9994 vs HF transformers GlmMoeDsa on a small real-arch model), generates, runs fully resident on 2× H200 NVL (q2 ~252 GB, no SSD streaming, ~16 t/s), and chats with a native GLM tokenizer in the GGUF.

All engine changes are gated on the GLM variant (n_hc==1) / on tokenizer-token presence, so the DeepSeek Flash/Pro paths stay byte-identical — verified: Flash 40 t/s + --logprob-vectors OK on the same build. DS4's joyai-llm byte-level BPE reproduces GLM's tokenization exactly (--dump-tokens == HF AutoTokenizer, 6/6 diverse strings).

Substance (full diff + converter + scripts + repro, experimental reference branch on my fork): https://github.com/vcnngr/ds4/tree/glm-5.2-backend — see GLM52_MAINTAINER_REPORT.md / MAINTAINER_NOTES.md.

Two model-independent findings worth upstreaming on their own:

1. q8→f16 dequant-cache reserve flat 512 MiB on ≥112 GiB cards → fills HBM and starves the session/prefill graph → OOM after a successful load (weights themselves fit). DS4_CUDA_WEIGHT_CACHE_LIMIT_GB doesn't bound it; an MTP model disables the cache and hides it. → PR CUDA: scale q8->f16 cache reserve on >=112 GiB cards (fixes session OOM on large models) #472 (CUDA twin of Fix ROCm Q8->F16 cache reserve starving session tensors on large models (q4q2) #446 on ROCm).
2. Distributed model id mismatch: a coordinator started without -m silently loads the default ds4flash.gguf, then rejects workers with model id mismatch. Detection is correct; the message just doesn't hint at the cause — a one-line message fix saves a long debug.

Router: GLM needed #466 (router expert-count) plus n_used (top-8) + sigmoid scoring; #435 (draft) generalizes the same router for Pro — worth converging so one router serves Flash/Pro/GLM.

Happy to split any of this into focused PRs if useful. Machine: 2× H200 NVL (143 GB, no NVLink), CUDA 13.2.