|
|
@@ -840,6 +840,11 @@ flare run $R/models/<model>/<version>/model.safetensors --package-cache $R --gpu
|
|
|
curl -s http://127.0.0.1:30080/v1/models # 200 in ~20 s; chat completions work
|
|
|
```
|
|
|
|
|
|
+**The bed is already on disk on dgx4** — `~/.teststore2` holds a complete
|
|
|
+`Qwen3.8-27B-NVIDIA-NVFP4-QAD-MXFP8 · gb10-tp1-202611120126` core (24 G, local NVMe,
|
|
|
+`[DEFAULT]` in its own store). Use `--model-cache ~/.teststore2` and skip the download;
|
|
|
+note it exists before someone wonders what that 24 G is.
|
|
|
+
|
|
|
Gotchas measured on the way: **`--rank` and `--rendezvous` are rejected for a single-host
|
|
|
core** (`--rendezvous applies only to a multi-host core`), the deployment id here is
|
|
|
human-readable (`gb10-tp1`) instead of a hash, this model streams into
|
|
|
@@ -860,6 +865,12 @@ on the NAS and bind-mounted over the model directory's `deployments/`):
|
|
|
load, so the NAS cost of a bind is bounded by the rank slice (1.3 GiB for TP4,
|
|
|
measured 1.29 s at 1.1 GB/s) and happens once per start.
|
|
|
|
|
|
+- **the envelope binds as a *file* too** — `model.safetensors` (mode 0444) bind-mounted
|
|
|
+ from the NAS over the local copy: the rank resolved it, `Start server … 0.8 s`, answered
|
|
|
+ a chat completion, and **the file bind was still live afterwards with the local original
|
|
|
+ byte-for-byte untouched**. Together with the directory bind for `deployments/` and a real
|
|
|
+ local `weights/`, that is **the §2 target layout validated in full**, in the exact shape
|
|
|
+ Phase 4 installs, with no rename or write by FLARE anywhere in it.
|
|
|
**Boot persistence (fstab) — validated by rebooting idle dgx4**, the exact stanza Phase 4
|
|
|
installs, and nothing else in `/etc/fstab` touched (backup `/etc/fstab.bak-20261112-fsbind`):
|
|
|
|