Windows/CUDA port: thermal gate, storage probe, correct dashboard stats

[David2024patton]

Summary

Windows + CUDA portability pass so the engine and web dashboard run correctly on Windows with an NVIDIA GPU, plus a couple of small correctness fixes. Every change is guarded by #ifdef COLI_CUDA and cross-platform #ifdefs, so Linux/CPU behavior is unchanged.

Changes

Engine (c/glm.c)

• NVML thermal heartbeat + thermal_gate() — a background thread samples GPU temperature (dynamically loaded nvml.dll / libnvidia-ml.so) and applies a proportional per-forward pause to bleed heat under sustained load. SAFE_MODE=0 disables it; if NVML is absent the engine proceeds normally.
• Storage bus probe — probe_storage_drive() / storage_bus_is_slow() check free space and disk bus type via Win32 IOCTL_STORAGE_QUERY_PROPERTY and POSIX statvfs. FORCE_SLOW_DRIVE=1 bypasses.
• Win32 hwinfo_emit branch — cores/RAM/CPU brand via GetSystemInfo / GlobalMemoryStatusEx / __get_cpuid, so the dashboard works without /proc.

Dashboard correctness

• Report RAM and VRAM in binary GiB so 128 GiB / 16 GiB read as 128 / 16 instead of decimal 137 / 17 — matches the spec-sheet convention users expect.
• Show the real GPU model via NVML nvmlDeviceGetName (e.g. NVIDIA GeForce RTX 4070 Ti SUPER) instead of a generic label. The web panel now displays the model name and labels the logical-processor count as threads.

Server (c/openai_server.py)

• COLI_EBITS / COLI_DBITS env vars to select quant bit-width (needed for the 4-bit config on 16 GB GPUs). Absent → engine defaults (8/8), i.e. upstream behavior preserved.

Testing

• Built on Windows with MinGW gcc + nvcc/MSVC (CUDA 12.9), coli_cuda.dll unchanged.
• Ran the full GLM-5.2 INT4 checkpoint end to end (COLI_CUDA=1 CUDA_DENSE=1 CUDA_EXPERT_GB=6 RAM_GB=110, 4-bit) with the web UI; /health reports ram_total_gb 127.7, vram_total_gb 16.0, gpu "NVIDIA GeForce RTX 4070 Ti SUPER", coherent generation, thermal + storage probes firing.

[JustVugg]

Thanks for this — the Windows dashboard stats overlap with #278 (hwinfo CPU/cores/RAM via CPUID + compat_meminfo), which just merged to main, so this now conflicts. The thermal gate + storage probe are still valuable and distinct. With "allow edits from maintainers" on I can rebase onto the current tree and drop whatever #278 already covers, so you don't have to untangle it — want me to? Otherwise a rebase on your side works too.

[JustVugg]

Could you rewrite this against current dev? I'm keeping it open rather than closing it, but it can't land in its present shape.

It's 66 commits behind dev. On glm.c, which changes several times a day, that isn't a rebase — it's a rewrite, and doing it for you would mean rewriting your design decisions without understanding why you made them. So I'd rather ask.

Windows is now covered by CI on MSYS2/UCRT64, so a rebased version gets tested automatically — which is what this PR most lacked. Also check against #314 (build-config stamp, CUDA_HOME quoting), which may already cover part of your delta.

What changed under you, which makes a rewrite cheaper than it sounds:

• CI now exists (ci: make check on ubuntu / windows (MSYS2 UCRT64) / macos — and it already caught a Windows test bug (#140) #143 + ci: GitHub Actions — engine, web, Python, CUDA syntax (free for public repos) #144, merged today): make check runs on ubuntu/windows/macos and the engine/web/python/CUDA-syntax jobs on Linux, for every PR. You'll get a verdict in ~2 minutes instead of waiting on a maintainer who can't test your platform. It caught a bug in my own test on its first Windows run.
• All PRs now target dev, not main. main is locked to admins.
• The backlog is being cleared (31 → 17 today). Your PR sat long enough to rot, and that's this repo's fault, not yours — two contributors independently rebuilt the same Windows CUDA loader this week because nobody triaged in time.

What would help it land fast: the smallest version that does one thing. A 300–700 line PR touching glm.c needs a maintainer who can run it, and for most platforms here that maintainer doesn't exist. A focused change with a test the CI can execute gets reviewed in an hour.

If you'd rather not, say so and I'll close it with thanks — no hard feelings either way. And if you think I've misjudged and it should go in as-is, push back: I've been wrong twice today already and both times a contributor caught it.