docs(runtime): add pinned HF node setup commands
Browse files- docs/RUNTIME_B200.md +15 -0
docs/RUNTIME_B200.md
CHANGED
|
@@ -19,6 +19,21 @@ cp ensure_runtime.sh "$ROOT/ensure_runtime.sh" # run once from this checkout
|
|
| 19 |
bash "$ROOT/ensure_runtime.sh"
|
| 20 |
```
|
| 21 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 22 |
Run the last line on each new node, inside a pod exposing the one B200 allocated for this check. With no GPU allocated yet, use `--install-only`; it reports `RUNTIME_INSTALLED_UNCHECKED`, not GPU readiness. With an existing runtime, `--check-only` checks without extracting. `--help` lists options.
|
| 23 |
|
| 24 |
The helper automatically finds either `$ROOT/ctopd-b200-cu129-v4/runtime.tar.gz` (the manually downloaded shard layout) or `$ROOT/runtime.tar.gz`. It verifies the pinned archive size/SHA-256 before extraction, checks Linux/glibc/g++, extracts safely to a new staging directory under the target filesystem, then renames to the final prefix. Allow at least 45 GiB free on that node-local filesystem. It never downloads dependencies, builds a runtime, launches training, or replaces an existing runtime that it cannot identify. Failed extraction keeps staging evidence. If `/opt` is not writable, choose an explicit node-local location:
|
|
|
|
| 19 |
bash "$ROOT/ensure_runtime.sh"
|
| 20 |
```
|
| 21 |
|
| 22 |
+
Alternatively, download the same standalone helper through the public HF mirror without GitHub authentication (no runtime archive download is repeated):
|
| 23 |
+
|
| 24 |
+
```bash
|
| 25 |
+
set -euo pipefail
|
| 26 |
+
cd /workspace/storage-shared/nlp/tungks/ctopd_all
|
| 27 |
+
PROXY="${HTTPS_PROXY:-${HTTP_PROXY:-http://10.30.154.118:80}}"
|
| 28 |
+
BASE=https://huggingface.co/datasets/codemaivanngu/cross-tokenizer-opd-b200-runtime/resolve/b5bbaea5025b8f58ef93386a5a5ac2f37058f51c
|
| 29 |
+
curl -fL --retry 3 --proxy "$PROXY" --noproxy "" "$BASE/ensure_runtime.sh" -o ensure_runtime.sh
|
| 30 |
+
curl -fL --retry 3 --proxy "$PROXY" --noproxy "" "$BASE/ensure_runtime.SHA256SUMS" -o ensure_runtime.SHA256SUMS
|
| 31 |
+
sha256sum -c ensure_runtime.SHA256SUMS
|
| 32 |
+
bash ensure_runtime.sh
|
| 33 |
+
```
|
| 34 |
+
|
| 35 |
+
Published helper SHA-256: `9a388b3ca183e0e771a49bdc4b61d9c6880d043da5490122c8fdf4098c3b9895`. Anonymous HF download was checked against the committed source; this verifies delivery, not execution on the company node.
|
| 36 |
+
|
| 37 |
Run the last line on each new node, inside a pod exposing the one B200 allocated for this check. With no GPU allocated yet, use `--install-only`; it reports `RUNTIME_INSTALLED_UNCHECKED`, not GPU readiness. With an existing runtime, `--check-only` checks without extracting. `--help` lists options.
|
| 38 |
|
| 39 |
The helper automatically finds either `$ROOT/ctopd-b200-cu129-v4/runtime.tar.gz` (the manually downloaded shard layout) or `$ROOT/runtime.tar.gz`. It verifies the pinned archive size/SHA-256 before extraction, checks Linux/glibc/g++, extracts safely to a new staging directory under the target filesystem, then renames to the final prefix. Allow at least 45 GiB free on that node-local filesystem. It never downloads dependencies, builds a runtime, launches training, or replaces an existing runtime that it cannot identify. Failed extraction keeps staging evidence. If `/opt` is not writable, choose an explicit node-local location:
|