SPB Git forge

spb/llm-api

Public
0commits 0branches 0releases
0 Bsize
maindefault branch
last push

Harvester: correct MLX weight estimates from HF metadata, flag installed mirrors; docs: managed-firewall WireGuard relay

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Simon-Pierre Boucher committed 13 days ago (Sep 11, 2026) parent d20a804

3 changed files +14 −6

modified docs/deployment.md +5 −1
@@ -24,7 +24,11 @@ mld logs llm-api
24 24
25 25 Manifest: `M1M32:~/dispatch/apps/llm-api.json` (dir `~/apps/llm-api`, port 8300, health `/health`, tunnel `www.llm-api.io` on BHS64, pinned to `M1M64`). Secrets (`ADMIN_PASSWORD`, `HF_TOKEN`) live in the manifest's `env` on M1M32 only and are written into the PM2 environment; the app also reads `~/llm-api/.env`.
26 26
27 M1M64 is a *reserved* node: always deploy with `--node M1M64`. It has no graphical session guaranteed at boot, so PM2 runs as a system LaunchDaemon (`mld prepare` does this).
27 +M1M64 is a *reserved* node: always deploy with `--node M1M64`.
28 +
29 +### WireGuard relay on M1M64 (managed firewall)
30 +
31 +The API binds `127.0.0.1:8300`, but Caddy on BHS64 reaches the node at its WireGuard address `10.67.0.40:8300`. On this rented Macly Mac the macOS application firewall is **MDM-managed** (`socketfilterfw` refuses changes) and blocks Homebrew binaries such as `socat` from accepting connections on the WireGuard interface (the cluster's usual `wg-forward-setup.sh` relay therefore fails silently: it listens, but every connection is dropped). The relay is instead an Apple-signed `/usr/bin/ssh -L 10.67.0.40:8300:127.0.0.1:8300 simon@127.0.0.1` (key `~/.ssh/id_ed25519_wgfwd`, authorized on the node itself), run as the PM2 process `llm-api-wgfwd` and declared as the third process of the manifest. Verify with `ssh BHS64 curl -s http://10.67.0.40:8300/health`. It has no graphical session guaranteed at boot, so PM2 runs as a system LaunchDaemon (`mld prepare` does this).
28 32
29 33 ## Manual (without mld)
30 34
modified docs/troubleshooting.md +1 −0
@@ -15,6 +15,7 @@
15 15 | Download fails 429 | Downloads history | HF rate limit: retry later |
16 16 | `INSUFFICIENT_DISK` | System → Storage | `MIN_FREE_DISK_GB` reserve; delete unused models |
17 17 | Dashboard 503 "not reachable" | `pm2 logs llm-api-web`, port 8301 | Next not running / build failed (`pnpm build`) |
18 +| Public site 502 while `localhost:8300` is fine | `ssh BHS64 curl http://10.67.0.40:8300/health`, `pm2 ls` (`llm-api-wgfwd`) | WireGuard relay down: restart `llm-api-wgfwd` (managed firewall blocks non-Apple listeners on the wg address, see deployment.md) |
18 19 | Login loop | cookies | behind HTTPS set `SECURE_COOKIES=true`; ensure the proxy passes `X-Forwarded-Proto` |
19 20 | Model shows loaded after a crash | restart server | stale state is cleared at startup (`workers.json`) |
20 21 | GPU shows `—` | `ioreg -r -c IOAccelerator` | metric unavailable on this macOS; degrades gracefully |
modified server/llm_api/harvester.py +8 −5
@@ -44,7 +44,7 @@ def _base_key(repo: str, tags: list[str]) -> str:
44 44 if not bm:
45 45 bm = next((t.split(":", 1)[1] for t in tags if t.startswith("base_model:") and "finetune:" not in t), None)
46 46 if bm:
47 return bm.lower()
47 + return bm.split("/")[-1].lower()
48 48 name = repo.split("/")[-1].lower()
49 49 name = re.sub(r"-(\d+bit|q\d[_a-z0-9]*|iq\d[_a-z0-9]*|mxfp\d|nvfp4|bf16|fp16|f16|fp8|dwq|optiq|gguf|mlx|4bit|8bit)+$", "", name)
50 50 name = re.sub(r"-(gguf|mlx)$", "", name)
@@ -89,8 +89,10 @@ class Harvester:
89 89 max_ram = opt["max_ram_gb"] or budget
90 90 installed = {r["repository"] for r in await self.db.fetchall("SELECT repository FROM models WHERE installed=1 AND repository IS NOT NULL")}
91 91 installed_keys = set()
92 for r in await self.db.fetchall("SELECT name, tags FROM models WHERE installed=1"):
92 + for r in await self.db.fetchall("SELECT name, tags, repository FROM models WHERE installed=1"):
93 93 installed_keys.add(_base_key(r["name"], json.loads(r["tags"] or "[]")))
94 + if r["repository"]:
95 + installed_keys.add(_base_key(r["repository"], []))
94 96 sources: list[tuple[str, str]] = []
95 97 for rt in opt["runtimes"]:
96 98 authors = opt["authors"] or DEFAULT_SOURCES.get(rt, [])
@@ -220,10 +222,11 @@ class Harvester:
220 222 total_params = getattr(st, "total", None) or (st.get("total") if isinstance(st, dict) else None)
221 223 params_by_dtype = getattr(st, "parameters", None) or (st.get("parameters") if isinstance(st, dict) else None)
222 224 if params_by_dtype and isinstance(params_by_dtype, dict):
223 # packed uint32 quantized weights: count * 32 / bits
225 + # Hugging Face reports quantized uint32 tensors by *unpacked* parameter count
226 + # (an 8B 4-bit repo shows U32 ≈ 8.0e9): bytes = params × bits / 8, plus ~12 % for scales/biases.
224 227 for dt, cnt in params_by_dtype.items():
225 228 if dt in ("U32", "I32") and bits:
226 weights_bytes += int(cnt * 4)
229 + weights_bytes += int(cnt * bits / 8 * 1.12)
227 230 elif dt in ("F16", "BF16"):
228 231 weights_bytes += int(cnt * 2)
229 232 elif dt in ("F32",):
@@ -242,7 +245,7 @@ class Harvester:
242 245 if bits and total_params and bits < 16 and weights_bytes:
243 246 # 'total' counts packed uint32 elements as 1 param; approximate real param count
244 247 pass
245 param_count = pc or (int(weights_bytes * 8 / bits) if bits and weights_bytes else None)
248 + param_count = pc or (int(total_params) if total_params else None) or (int(weights_bytes * 8 / bits) if bits and weights_bytes else None)
246 249 download_bytes = int(weights_bytes * 1.02)
247 250 if not weights_bytes:
248 251 return None
249 252