wiki: add troubleshooting articles from MajorTwin v8 cycle
Two articles surfaced during the v8 deploy + eval on 2026-04-25: - Ollama: `ollama run` with piped stdin bypasses the chat template and SYSTEM prompt — output looks like raw base-model completion. Caught during initial v8 smoke test. Fix: use /api/chat HTTP endpoint. - rsync over Tailscale can hang in TCP teardown after the data has fully transferred. Verify with md5sum, then kill the hung pipeline. Includes a watcher-threshold gotcha (set below true file size, not above) and prevention tips. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
parent
ef91865f88
commit
0996861512
4 changed files with 181 additions and 0 deletions
|
|
@ -14,6 +14,7 @@ Practical fixes for common Linux, networking, and application problems.
|
|||
- [Mail Client Stops Receiving: Fail2ban IMAP Self-Ban](networking/fail2ban-imap-self-ban-mail-client.md)
|
||||
- [firewalld: Mail Ports Wiped After Reload](networking/firewalld-mail-ports-reset.md)
|
||||
- [Tailscale SSH: Unexpected Re-Authentication Prompt](networking/tailscale-ssh-reauth-prompt.md)
|
||||
- [rsync over Tailscale: Hung in TCP Teardown After Transfer Completes](networking/rsync-tailscale-teardown-stall.md)
|
||||
- [Windows OpenSSH: WSL Default Shell Breaks Remote Commands](networking/windows-openssh-wsl-default-shell-breaks-remote-commands.md)
|
||||
- [Pi-hole AI Blocklist Blocks Claude Desktop (ERR_CONNECTION_REFUSED)](networking/pihole-blocks-claude-desktop.md)
|
||||
- [ISP SNI Filtering & Caddy](isp-sni-filtering-caddy.md)
|
||||
|
|
@ -44,5 +45,6 @@ Practical fixes for common Linux, networking, and application problems.
|
|||
|
||||
## 🤖 AI / Local LLM
|
||||
- [Ollama Drops Off Tailscale When Mac Sleeps](ollama-macos-sleep-tailscale-disconnect.md)
|
||||
- [Ollama: `ollama run` with Piped Stdin Bypasses Chat Template + SYSTEM Prompt](ollama-chat-template-pipe-stdin-bypass.md)
|
||||
- [Windows OpenSSH Server (sshd) Stops After Reboot](networking/windows-sshd-stops-after-reboot.md)
|
||||
- [claude-mem Silently Fails with Claude Code 2.1+ (Empty `--setting-sources`)](claude-mem-setting-sources-empty-arg.md)
|
||||
|
|
|
|||
|
|
@ -0,0 +1,89 @@
|
|||
---
|
||||
title: "rsync over Tailscale: Hung in TCP Teardown After Transfer Completes"
|
||||
domain: troubleshooting
|
||||
category: networking
|
||||
tags: [rsync, ssh, tailscale, hang, tcp-fin, hash-mismatch]
|
||||
status: published
|
||||
created: 2026-04-25
|
||||
updated: 2026-04-25
|
||||
---
|
||||
|
||||
# rsync over Tailscale: Hung in TCP Teardown After Transfer Completes
|
||||
|
||||
A long rsync transfer over Tailscale finishes — the destination file is at full size, rsync's own summary line is in the log — but the rsync, ssh client, and parent bash processes never exit. The `&&` chain that should run after rsync (e.g. `&& echo DONE`) never fires. Watcher scripts polling for completion can stall indefinitely.
|
||||
|
||||
## The Short Answer
|
||||
|
||||
The data is fine. Verify with `md5sum` (or `md5 -q` on macOS) against the source, then kill the hung pipeline.
|
||||
|
||||
```bash
|
||||
# 1. confirm size matches rsync's reported total_size
|
||||
ls -lh ~/your-file.gguf
|
||||
tail ~/rsync.log # look for "total size is N" line
|
||||
|
||||
# 2. checksum end-to-end
|
||||
md5 -q ~/your-file.gguf # macOS
|
||||
ssh majorlinux@100.x.x.x 'md5sum /source/path/your-file.gguf' # Linux source
|
||||
|
||||
# 3. if hashes match, kill the hung pipeline by name
|
||||
pkill -f 'rsync.*your-file' || true
|
||||
pkill -f 'ssh .*rsync --server' || true
|
||||
```
|
||||
|
||||
## How to Notice
|
||||
|
||||
`ps aux | grep rsync` shows the rsync client, the spawned ssh, and the wrapping bash all in `S` state with **0 CPU activity** and timestamps from minutes-to-hours ago. The destination file already exists at the final (non-`.partial` / non-dotfile) path at full size. The trailing summary in the rsync log reads:
|
||||
|
||||
```
|
||||
sent N bytes received M bytes ... bytes/sec
|
||||
total size is X speedup is Y
|
||||
```
|
||||
|
||||
…but the bash `&&` followup that depends on rsync's exit code never runs.
|
||||
|
||||
## Why This Happens
|
||||
|
||||
rsync's exit waits for the underlying ssh transport to close cleanly. Over Tailscale (especially after a long-running connection that bridged a sleep, reconnect, or NAT shuffle), the TCP FIN/ACK handshake from the remote sshd can be lost or delayed indefinitely. The local end has all the data, has finalized the file, has printed its summary — but it's still blocked in `read()` on a socket that will never close on its own.
|
||||
|
||||
This is amplified when:
|
||||
- The transfer hits a hash-mismatch retry mid-flight (rsync re-pulls the temp file). Each retry re-establishes connection state that's more vulnerable to teardown weirdness.
|
||||
- The link briefly drops and reconnects via DERP relay during the transfer.
|
||||
- The source machine is on WSL2 — Windows network stack rewrites can defer FINs.
|
||||
|
||||
The upshot: the data was transferred correctly long before the pipeline reports done. Don't wait — verify and move on.
|
||||
|
||||
## Don't Just Kill — Verify First
|
||||
|
||||
Killing a hung rsync **before the file is complete** can leave a partial file that looks complete by size alone. Always:
|
||||
|
||||
1. Compare the on-disk size to the `total size is N` line in the rsync log
|
||||
2. md5 (or sha256) against the source to confirm bit-for-bit equality
|
||||
3. Only then kill the hung processes
|
||||
|
||||
Skipping the checksum step risks silently corrupting downstream consumers of the file (Ollama blobs, archive pipelines, etc.).
|
||||
|
||||
## Watcher Threshold Gotcha
|
||||
|
||||
If you have a polling watcher script that fires a notification when the file reaches some threshold size, **set the threshold below the actual file size**, not above it. Example: a 4.68 GB GGUF transferred fine but the watcher's threshold was set to 4.7 GB (`4_700_000_000` bytes), so the threshold never triggered even though the transfer completed.
|
||||
|
||||
```bash
|
||||
# bad — threshold above true size
|
||||
TARGET=4700000000 # 4.7 GB
|
||||
|
||||
# good — threshold below true size
|
||||
TARGET=4600000000 # 4.6 GB, fires at ~98% complete
|
||||
```
|
||||
|
||||
Or better: trust the rsync exit code / the `RSYNC_DONE` marker line your wrapper writes after `&&`, not file size.
|
||||
|
||||
## Prevention
|
||||
|
||||
- Wrap rsync in a watchdog. If rsync hasn't exited within `expected_runtime + 2 minutes`, snapshot status, md5-verify, and kill.
|
||||
- For very large files, use `rsync --partial-dir` so a fresh re-run resumes from the temp file instead of redoing the transfer.
|
||||
- Consider `rsync --inplace` for files that consumers will copy out of the destination anyway (Ollama blob copy step).
|
||||
- Add `ServerAliveInterval=30` / `ServerAliveCountMax=3` to your ssh config for the source host — kills the ssh transport if the remote stops responding to keepalives.
|
||||
|
||||
## Related
|
||||
|
||||
- [[tailscale-ssh-reauth-prompt]] — different Tailscale-over-ssh gotcha
|
||||
- [[../../02-selfhosting/storage-backup/rsync-backup-patterns|rsync backup patterns]] — general rsync usage in MajorInfrastructure
|
||||
88
05-troubleshooting/ollama-chat-template-pipe-stdin-bypass.md
Normal file
88
05-troubleshooting/ollama-chat-template-pipe-stdin-bypass.md
Normal file
|
|
@ -0,0 +1,88 @@
|
|||
---
|
||||
title: "Ollama: `ollama run` with Piped Stdin Bypasses Chat Template + SYSTEM Prompt"
|
||||
domain: troubleshooting
|
||||
category: ai-inference
|
||||
tags: [ollama, eval, chat-template, system-prompt, majortwin, gotcha]
|
||||
status: published
|
||||
created: 2026-04-25
|
||||
updated: 2026-04-25
|
||||
---
|
||||
|
||||
# Ollama: `ollama run` with Piped Stdin Bypasses Chat Template + SYSTEM Prompt
|
||||
|
||||
When eval'ing or smoke-testing an Ollama model, piping a prompt via stdin to `ollama run` skips the model's chat template **and** the SYSTEM prompt baked into the Modelfile. Output looks like raw base-model completion (often Mastodon-shaped or training-data-shaped), and you'll think the model is broken when it isn't.
|
||||
|
||||
## The Short Answer
|
||||
|
||||
For evals and any test where you want the model's actual chat behavior, **use the HTTP API at `/api/chat`** — never `ollama run` with `echo "..." | ollama run model`.
|
||||
|
||||
```python
|
||||
import json, urllib.request
|
||||
body = json.dumps({
|
||||
"model": "majortwin-v8",
|
||||
"messages": [{"role": "user", "content": "What's your name?"}],
|
||||
"stream": False,
|
||||
}).encode()
|
||||
req = urllib.request.Request(
|
||||
"http://localhost:11434/api/chat",
|
||||
data=body, headers={"Content-Type": "application/json"}, method="POST",
|
||||
)
|
||||
r = json.loads(urllib.request.urlopen(req).read())
|
||||
print(r["message"]["content"])
|
||||
```
|
||||
|
||||
Or with curl piped through jq:
|
||||
|
||||
```bash
|
||||
curl -s http://localhost:11434/api/chat -d '{
|
||||
"model": "majortwin-v8",
|
||||
"messages": [{"role": "user", "content": "What is your name?"}],
|
||||
"stream": false
|
||||
}' | jq -r .message.content
|
||||
```
|
||||
|
||||
## How to Notice
|
||||
|
||||
Symptom: model responses are weirdly raw — Mastodon-style hashtag rants, news headlines, multiple unrelated thoughts strung together — even though the same model behaves normally in Open WebUI or via the chat API. This is the canonical fingerprint of a chat-template-bypassed call.
|
||||
|
||||
## Why This Happens
|
||||
|
||||
`ollama run` is the CLI's interactive REPL. When stdin is a TTY, it reads input as user turns and applies the chat template. When stdin is a **pipe** (`echo "..." | ollama run model`), the CLI treats stdin as raw text and forwards it to `/api/generate` (the completion endpoint), not `/api/chat`. `/api/generate` does **not** apply the chat template, and the SYSTEM prompt only takes effect when the chat template wraps it.
|
||||
|
||||
The two endpoints serve different purposes:
|
||||
- `/api/generate` — raw completion, good for fill-in-the-blank or non-instruct base models
|
||||
- `/api/chat` — applies the model's chat template, includes SYSTEM, handles multi-turn message arrays
|
||||
|
||||
For an instruct-tuned model (Qwen2.5-Instruct, Llama-3.1-Instruct, etc.), bypassing the chat template means the model never sees the `<|im_start|>system ... <|im_end|>` framing it was trained to expect, and its responses regress toward base-model behavior.
|
||||
|
||||
## When You Actually Want `/api/generate`
|
||||
|
||||
Almost never, for instruct models. The legitimate use case is base models without a chat template, or specific completion-style prompts where you want the model to continue a string verbatim. For evals of a fine-tuned Modelfile, always use `/api/chat`.
|
||||
|
||||
## Reusable Eval Pattern
|
||||
|
||||
A minimal stdlib-only eval harness used for MajorTwin evals lives at `~/MajorTwin/scripts/eval_v8.py`. The key call is the `chat()` helper:
|
||||
|
||||
```python
|
||||
def chat(host, model, prompt, timeout=180):
|
||||
body = json.dumps({
|
||||
"model": model,
|
||||
"messages": [{"role": "user", "content": prompt}],
|
||||
"stream": False,
|
||||
}).encode()
|
||||
req = urllib.request.Request(
|
||||
f"{host}/api/chat",
|
||||
data=body,
|
||||
headers={"Content-Type": "application/json"},
|
||||
method="POST",
|
||||
)
|
||||
with urllib.request.urlopen(req, timeout=timeout) as r:
|
||||
return json.loads(r.read())["message"]["content"].strip()
|
||||
```
|
||||
|
||||
This applies the chat template and the SYSTEM prompt baked into the Modelfile. No need to re-specify SYSTEM per-call.
|
||||
|
||||
## Related
|
||||
|
||||
- [[ollama-macos-sleep-tailscale-disconnect]] — different Ollama gotcha (sleep + Tailscale)
|
||||
- [[20-Projects/MajorTwin/majortwin-v8-eval-report|MajorTwin v8 eval report]] — caught this issue during initial smoke test on 2026-04-25
|
||||
|
|
@ -88,6 +88,8 @@ updated: 2026-04-22T19:58
|
|||
* [Windows OpenSSH: WSL Default Shell Breaks Remote Commands](05-troubleshooting/networking/windows-openssh-wsl-default-shell-breaks-remote-commands.md)
|
||||
* [Pi-hole AI Blocklist Blocks Claude Desktop (ERR_CONNECTION_REFUSED)](05-troubleshooting/networking/pihole-blocks-claude-desktop.md)
|
||||
* [Ollama Drops Off Tailscale When Mac Sleeps](05-troubleshooting/ollama-macos-sleep-tailscale-disconnect.md)
|
||||
* [Ollama: `ollama run` with Piped Stdin Bypasses Chat Template + SYSTEM Prompt](05-troubleshooting/ollama-chat-template-pipe-stdin-bypass.md)
|
||||
* [rsync over Tailscale: Hung in TCP Teardown After Transfer Completes](05-troubleshooting/networking/rsync-tailscale-teardown-stall.md)
|
||||
* [macOS: Repeating Alert Tone from Mirrored iPhone Notification](05-troubleshooting/macos-mirrored-notification-alert-loop.md)
|
||||
* [ClamAV CPU Spike: Safe Scheduling with nice/ionice](05-troubleshooting/security/clamscan-cpu-spike-nice-ionice.md)
|
||||
* [Ansible: Vault Password File Not Found](05-troubleshooting/ansible-vault-password-file-missing.md)
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue