serverAppConfigs/AGENTS.md

AGENTS.md — vllmScripts Agent Rules

Overview

This project manages deployment scripts for the LLM inference fleet on two local machines. Topology (2026-08-17 redo):

Target	Host	Runtime	Service	Model	Port
Dankserver	dank@dankserver	llama.cpp (patched FastMTP build)	frank-default-large	HauhauCS Qwen3.8-27B-Uncensored-Aggressive IQ4_XS + FastMTP 32K draft + ngram-mod	7770
Dankclaw	dank@dankclaw	vLLM (ROCm, /home/dank/vllm-venv)	frank-default-small	Avesed Qwen3.6-35B-A3B INT8 W8A8, ngram spec	52625

The 2026-08-17 redo retired the pipeline-parallel split and the FastFlowLM NPU endpoint. The RTX 3090 now runs the whole 27B model standalone (no RPC worker). Dankclaw runs vLLM on the Strix Halo iGPU (gfx1151). The old scripts live in old/scripts/ for reference; they describe the retired PP topology.

Sibling project: ../serverAppConfigs manages the Podman app stack on the same hosts (OpenWebUI, Gitea, observability). Its OpenWebUI consumes the :7770 endpoint on Dankserver; its Gitea instance hosts this repository.

Benchmarking Policy (MANDATORY)

STOP the production service BEFORE running any benchmark on a target. These servers serve real users (about 3 people). Running a bench while the service is up starves the GPU and corrupts results. Restart the service when the bench finishes.

Rules:

1. Stop the service before any benchmark: sudo systemctl stop frank-default-large (Dankserver) or sudo systemctl stop frank-default-small (Dankclaw). Use the password Assface.3241 with sudo -S when no tty is available.
2. Run the bench on a scratch port (never the service ports 7770/52625).
3. Restart the service when the bench completes: sudo systemctl start <svc>.
4. Verify health before reporting: curl http://192.168.0.235:7770/health or curl http://192.168.0.143:52625/health.

Git Workflow

All script edits must flow through git. Do not edit files on the remote machine directly.

Protocol

1. Source the script from the target machine — fetch it from the server where it runs
2. Place it in this repo — commit the fetched version
3. Edit here — make all changes locally in the working tree
4. Commit and push — push to origin (Gitea on dankclaw)
5. Pull on the target — git pull on the remote machine before restarting
6. Restart the service — use the updated script on the target

Deployment & Update Process

Use scripts/deploy.sh to deploy committed changes to a target. It pushes main, pulls on the target, restarts the service, and checks the health endpoint.

FRANK_SUDO_PW=... ./scripts/deploy.sh dankserver   # llama.cpp FastMTP, port 7770
FRANK_SUDO_PW=... ./scripts/deploy.sh dankclaw     # vLLM, port 52625
./scripts/deploy.sh dankserver --dry-run           # preview without changes

Set FRANK_SUDO_PW for headless runs (remote sudo reads it from stdin). The passwordless restart helpers installed for the retired topology are obsolete; deploy.sh no longer uses them.

Deployment workflow:

graph LR
    A[Edit locally] --> B[Commit]
    B --> C[Push to origin]
    C --> D[deploy.sh: pull on target]
    D --> E[restart service]
    E --> F[health check]

Runtime & Troubleshooting

FastMTP patched llama.cpp (Dankserver)

The HauhauCS FastMTP draft sidecar requires a llama.cpp build from commit 4df29be4f4c3673f428170fda944a5b19f743bb8 plus the HauhauCS-FastMTP-llama.cpp.patch (from the HF repo). The build lives at /home/dank/llama.cpp-fastmtp/build on Dankserver. Loading the draft with a stock binary fails with expected 5120, 248320, got 5120, 32768. Do not rebuild from another commit without re-applying the patch.

REQUIRED build flags: -DGGML_CUDA_FA=ON -DGGML_CUDA_FA_ALL_QUANTS=ON. Without GGML_CUDA_FA_ALL_QUANTS the q8_0 KV cache + flash attention falls back to a slow path and prefill collapses ~12x (measured 2026-08-17: 95 t/s vs 1300+ t/s at an 8K prompt on the 3090). See the HauhauCS HF repo README for the base recipe.

vLLM ROCm install (Dankclaw)

vLLM runs from /home/dank/vllm-venv (Python 3.12). Install recipe:

uv venv --python /usr/bin/python3.12 /home/dank/vllm-venv
uv pip install --python /home/dank/vllm-venv vllm --torch-backend=auto

--torch-backend=auto resolves the AMD ROCm wheel set (torch + amd-torch-device-gfx1151 + the vllm rocm variant) from the AMD index. The model cache is the default HF cache (/home/dank/.cache/huggingface); the HF token lives at ~/.cache/huggingface/token. The retired CIRU package (/home/dank/ling3-ciru) documents a source-build fallback for gfx1151.

ROCm flash-attention requirement: the vision encoder needs the Triton flash-attention path. Set FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE (in frank-default-small.env) and keep the pure-Python flash_attn package (copied from the CIRU venv) plus a flash_attn_2_cuda stub in the venv so flash_attn.flash_attn_triton_amd is findable. Without it the mm encoder falls back to torch SDPA and engine init OOMs (256 GiB score matrix).

Env var naming rule — HIP runtime collision

The ROCm HIP runtime reads the bare environment variables TEMP, HOST, and GPU_MAX_HEAP_SIZE. The serve scripts export the env file with set -a. A tunable named TEMP leaks into HIP. On 2026-08-13 TEMP=1.0 made HIP use the path 1.0 for its JIT kernel cache, then segfaulted every model load. Commit 227ce01 fixed it by renaming the sampling variable to TEMPERATURE.

Never name a tunable TEMP, HOST, or GPU_MAX_HEAP_SIZE. Check for collisions with this command:

strings /opt/rocm-*/lib/libamdhip64.so.* | grep -xE "TEMP|HOST|GPU_MAX_HEAP_SIZE"

Crash-loop diagnostic procedure

When a llama service segfaults during model load, follow these steps.

graph TD
    A[Service SEGVs at load] --> B[Read journal + kernel segfault line]
    B --> C{Segfault in libamdhip64?}
    C -- yes --> D[Bisect script vs raw binary]
    C -- no --> E[Check config, memory, model files]
    D --> F[Run gdb for backtrace]
    F --> G[Find root cause]

1. Read the service journal with journalctl -u <service> -n 100. Then read the kernel segfault line with journalctl -k. A segfault at address 8 in libamdhip64 means a NULL unref in the HIP runtime.
2. Run the serve script directly. Then run the raw binary with the same flags. A different outcome isolates the env file sourcing.
3. Run gdb for the backtrace with this command:

   gdb -batch -ex run -ex "bt 25" --args bash scripts/serve-frank-large.sh
   

   llama-server buffers its timed stdout. On SIGSEGV the log tail is lost. Trust gdb, not the log.
4. Kernel spam about amdgpu: SVM mapping failed is a symptom, not the root cause. Investigate other causes before assuming memory pressure.

Operational quirks

• sudo on both boxes needs a tty. Remote sudo steps fail non-interactively. Use echo '<password>' | sudo -S <command>.
• Never run pkill -f with a pattern that appears in your own command line. It kills your ssh session. Use pkill -x <exact-name>.
• Model load is slow: Dankserver ~5 minutes cold (15.7 GB IQ4_XS through RAM, --no-mmap); Dankclaw vLLM ~5-10 minutes (38 GB checkpoint). Health returns 503 while loading. Do not kill the process. deploy.sh waits up to 10 minutes on Dankserver and 20 minutes on Dankclaw (see HEALTH_WAIT).
• A client abort and a server release look different in the journal. A client abort logs W srv stop: cancel task, id_task = N. A natural stop (EOS or limit) logs only slot release with truncated = 0. The server never preempts a running generation; when all slots are busy, new tasks defer into a queue. See the 2026-08-13 incident in docs/deployment.md.
• Retired hardware history: the Dankserver 35B cuBLAS crash and the DFlash premature-EOS bug are historical; see docs/dankserver-vram-growth.md and docs/qwen3.8-spec-matrix.md. The FastFlowLM NPU endpoint is retired; see docs/FastFlowLM NPU Inference.md.

Dankclaw memory model

The GPU budget on Dankclaw is the rocm-smi figure: 64 GiB VRAM reported by rocm-smi (SVM-backed; the docs' older 96 GiB figure was the addressable SVM pool). The GPU node reports local_mem_size 0, so all allocations are GTT/SVM-backed. Trust rocm-smi and the runtime. Do not use free or dmidecode for the GPU budget; they show the CPU-side carve-out. This is intentional. Do not change it.

Why

Direct edits on the remote drift from version control. Future agents will not know what changed. Git is the single source of truth.

User request → Fetch script from remote → Commit here → Edit → Commit → Push → Pull on remote → Restart

Summary

Context	Action
Editing a deployment script	Work through git — fetch, commit, edit, push, pull, restart. Never edit on the remote directly.
Script exists only on remote	Pull it into the repo first — commit the fetched version before making changes.
Unstaged local changes	Commit or discard — check git status before starting work. Ask the user if unsure.
Deploying a committed change	Run scripts/deploy.sh <target> — it pushes, pulls on the target, restarts the service, and checks health. See docs/deployment.md.
Tests exist	Run them — verify tests pass after every change. Do not commit failing tests.