Frank
55f87fcdce
The 2026-10-09 wire aliases become the service names everywhere: systemd units, serve scripts, env files and FRANK_* variables, deploy-plan/deploy-remote service keys, container names, Prometheus job names + alert rules, Grafana dashboard queries, bats seams, runbook filenames, homepage/AGENTS lore, and historical prose. Rollback tags renamed in lockstep. Residue in the tree: zero. Retired frank-medium/small/npu names are history and stay. Known casualty: scripts/patches/*.patch context lines were rewritten; those retired-era vLLM-fork patches no longer apply byte-exact.
242 lines
11 KiB
Markdown
242 lines
11 KiB
Markdown
---
|
|
type: reference
|
|
description: "**Status (historical, 2026-09-13):** was to deploy on Dankclaw."
|
|
tags:
|
|
- llm
|
|
- rocm
|
|
- frank
|
|
timestamp: 2026-09-15
|
|
---
|
|
|
|
# frank-dragoon: kingjones Qwen3.8-27B-ROCmFPX-Q8_0-AGENT
|
|
|
|
> **SUPERSEDED 2026-09-15.** The Q8_0_ROCMFPX-AGENT checkpoint (with the
|
|
> separate `draft-mtp` head) now serves Dankserver as **frank-medium** on the
|
|
> fork's fixed build — see [[frank-medium-rocmfpx]]. The failure below was
|
|
> specific to the gfx1151 legacy-read path and the Q6_0 single-file era.
|
|
|
|
**Status (historical, 2026-09-13):** was to deploy on Dankclaw. Serving is
|
|
**BLOCKED by
|
|
the `set_rows` build bug** — the model loads but crashes in
|
|
`ggml_compute_forward_set_rows` before it can serve. All config goals are met;
|
|
a build fix is required for the model to run.
|
|
|
|
**Alias:** `frank-dragoon`, port **7770**.
|
|
|
|
The large service serves the kingjones `Qwen3.8-27B-ROCmFPX-Q8_0-AGENT` GGUF
|
|
from `/home/dank/models/kingjones-Qwen3.8-27B-ROCMFPX-Q8_0-AGENT/` with a
|
|
charlie12345/ROCmFPX `llama-server` (build `build-kairic`). It runs a
|
|
**standalone MTP draft head** (loaded separately, on ROCm0), **TurboQuant** V
|
|
KV cache, the **Sharp** chat template with **reasoning on**, and a **full
|
|
262144 context / 1 slot** profile. The whole payload plus the 262k KV cache
|
|
fit in the lifted **114 GiB TTM/GTT** budget (see the 114 GiB change in
|
|
[[Dankclaw Improvements]] and [[qwen3.8-strix-optimizations]]).
|
|
|
|
| Figure | Value |
|
|
| ------ | ----- |
|
|
| Model | kingjones Qwen3.8-27B-ROCmFPX-Q8_0-AGENT (`Q8_0_ROCMFPX`) |
|
|
| HF repo | `kingjones777/Qwen3.8-27B-ROCmFPX-Q8_0-AGENT-GGUF` (public) |
|
|
| Runtime | charlie12345/ROCmFPX `llama-server`, `build-kairic` |
|
|
| Draft head | standalone `mtp-Qwen3.8-27B-Q4_0.gguf` on ROCm0 |
|
|
| KV cache | f16 K, TurboQuant (`turbo4`) V, full 262144 context, 1 slot |
|
|
| Reasoning | on, `deepseek`, budget `-1`, `--reasoning-format none` |
|
|
| Chat template | Sharp (repo-tracked), `enable_thinking` + `preserve_thinking` on |
|
|
| Host path | `/home/dank/models/kingjones-Qwen3.8-27B-ROCMFPX-Q8_0-AGENT/` |
|
|
| Tunables | `scripts/frank-dragoon.env` |
|
|
| Serve script | `scripts/serve-frank-dragoon.sh` |
|
|
|
|
## Serving status: the `set_rows` build bug (blocks serving)
|
|
|
|
As of 2026-08-25 the Q8_0 AGENT model will not serve on Dankclaw. It loads,
|
|
reaches `initializing slots, n_slots = 1`, but never prints `start serving`.
|
|
Health stays `503` until the process exits. The crash is in
|
|
`ggml_compute_forward_set_rows` in `libggml-cpu.so.0`, a NULL function-pointer
|
|
jump (address `0x00000000`).
|
|
|
|
The backtrace (gdb). The main thread and one OpenMP worker crash in the same
|
|
path:
|
|
|
|
```mermaid
|
|
graph TD
|
|
A[llama_context::graph_compute] --> B[ggml_backend_cpu_graph_compute]
|
|
B --> C[ggml_graph_compute]
|
|
C --> D[GOMP_parallel]
|
|
D --> E[ggml_compute_forward_set_rows]
|
|
E --> F[NULL dispatch -> address 0x0]
|
|
D --> G[Thread 23 worker: same path]
|
|
```
|
|
|
|
Test matrix. Five ROCmFPX builds and the all-CPU path were tested on a scratch
|
|
port (`/tmp/g80bench/run.sh`), `gl=45`, `ctx=131072`. All crash in `set_rows`.
|
|
|
|
| Build | Binary | Version | Result |
|
|
| ----- | ------ | ------- | ------ |
|
|
| build-kairic | `ciru-qwen38/ROCmFPX/build-kairic/bin/llama-server` | v81 (205a3e5f4) | segfault in set_rows |
|
|
| build-promptforge | `ciru-qwen38/ROCmFPX/build-promptforge/bin/llama-server` | PromptForge/ActiveFPX lineage | segfault in set_rows |
|
|
| build-rocmi4-w4a4 | `rocmi4-qwen38/ROCmFPX/build-rocmi4-w4a4/bin/llama-server` | v257 (c49ebdb) | segfault in set_rows |
|
|
| build-strix-rocmfp4 | `GitPackages/ROCmFPX/build-strix-rocmfp4/bin/llama-server` | v209 (00d5452) | segfault in set_rows |
|
|
| all-CPU | build-kairic, `-ngl 0` | v81 | segfault in set_rows |
|
|
|
|
The crash is not GPU-specific (the all-CPU path fails too), not MTP-specific
|
|
(`--spec-type none` fails), and not fixed by `LLAMA_SET_ROWS=0`. So it is a
|
|
build bug: the Q8_0 AGENT GGUF graph needs a `set_rows` dispatch the ROCmFPX
|
|
backends do not register. The older caf-eemin ROCmI4 and Kairic Edge models
|
|
served fine on related ROCmFPX builds. The caf-eemin payload was about 14.4
|
|
GiB; the separate
|
|
PromptForge/ActiveFPX deployment used a much larger multi-file payload of about
|
|
36.7 GiB. These are different deployments. The Q8_0 AGENT graph is the trigger
|
|
for the `set_rows` failure.
|
|
|
|
**Fix path:** rebuild charlie12345/ROCmFPX from source with a working
|
|
`set_rows` for the ROCmFPX backend, or find a build where it is registered.
|
|
This unblocks the model. See the context budget reality below once `set_rows`
|
|
is fixed.
|
|
|
|
## Model files
|
|
|
|
Three files, from the HF repo above:
|
|
|
|
| File | Size | Purpose |
|
|
| ---- | ---- | ------- |
|
|
| `Qwen3.8-27B-Q8_0_ROCMFPX_AGENT.gguf` | 28.21 GB | main weights |
|
|
| `mtp-Qwen3.8-27B-Q4_0.gguf` | 1.68 GB | standalone MTP draft head (loaded by the script) |
|
|
| `mmproj-Qwen3.8-27B-BF16.gguf` | 0.93 GB | vision projector (`--mmproj`) |
|
|
|
|
The repo also has a `mtp-Qwen3.8-27B-Q8_0.gguf` (3.16 GB) draft; the env uses
|
|
the Q4_0 draft. Download them from the resolve URLs below (they are large —
|
|
use the token in `~/.cache/huggingface/token`).
|
|
|
|
## Download
|
|
|
|
```bash
|
|
# on Dankclaw
|
|
TOKEN=$(cat ~/.cache/huggingface/token)
|
|
REPO=kingjones777/Qwen3.8-27B-ROCmFPX-Q8_0-AGENT-GGUF
|
|
DIR=/home/dank/models/kingjones-Qwen3.8-27B-ROCMFPX-Q8_0-AGENT
|
|
mkdir -p "$DIR"; cd "$DIR"
|
|
dl(){ curl -fsSL -C - -o "$1" --retry 5 --retry-delay 3 --retry-max-time 3600 \
|
|
-H "Authorization: Bearer $TOKEN" \
|
|
"https://huggingface.co/$REPO/resolve/main/$1"; echo "done: $1"; }
|
|
dl Qwen3.8-27B-Q8_0_ROCMFPX_AGENT.gguf
|
|
dl mmproj-Qwen3.8-27B-BF16.gguf
|
|
dl mtp-Qwen3.8-27B-Q4_0.gguf
|
|
```
|
|
|
|
The serve script checks each artifact is present and readable before launch
|
|
(`scripts/frank-dragoon.env` -> the four `MODEL_*` files plus the chat
|
|
template). A missing file fails the artifact check with `missing or
|
|
unreadable required artifact`.
|
|
|
|
## Serving profile
|
|
|
|
`scripts/serve-frank-dragoon.sh` builds the `llama-server` call from the env:
|
|
|
|
```
|
|
/dev/llama-server -m <main> --alias frank-dragoon
|
|
--host 0.0.0.0 --port 7770 --jinja
|
|
-dev ROCm0 -ngl 999 -np 1 -c 262144
|
|
-b 512 -ub 512 -fa on
|
|
-ctk f16 -ctv turbo4 -t 16 -tb 32
|
|
-ctxcp 32 --cache-ram 8192 --cache-prompt --cache-idle-slots --metrics
|
|
--spec-type draft-mtp --model-draft <mtp-Q4_0>
|
|
--mmproj <mmproj-BF16>
|
|
--spec-draft-device ROCm0 --spec-draft-ngl 999
|
|
--spec-draft-type-k f16 --spec-draft-type-v turbo4
|
|
--spec-draft-n-max 4 --spec-draft-n-min 0
|
|
--spec-draft-p-min 0.0 --spec-draft-p-split 0.10
|
|
--spec-draft-backend-sampling
|
|
--temp 0.7 --top-p 0.8 --top-k 20 --min-p 0.0
|
|
--presence-penalty 1.5
|
|
--reasoning on --reasoning-format none --reasoning-budget -1
|
|
```
|
|
|
|
Key choices:
|
|
|
|
- **Standalone MTP draft** (`--model-draft`, `--spec-type draft-mtp`). The
|
|
AGENT GGUF has no embedded draft head; the draft head is a separate file on
|
|
ROCm0. This matches the Qwen3.8-27B AGENT shape (`--spec-draft-type-k
|
|
f16 --spec-draft-type-v turbo4`).
|
|
- **f16 K + TurboQuant V** cache. f16 K is the safe long-ctx default; TurboQuant
|
|
halves the V cache at 262k while keeping quality (see the ubatch/threads-batch
|
|
notes in [[qwen3.8-strix-optimizations]]).
|
|
- **262144 context, 1 slot — the target.** The whole context KV cache plus the
|
|
~30 GB payload (main + draft + mmproj) are meant to fit in the 114 GiB
|
|
budget. Measured reality: `gl=999` (all layers on ROCm0) at 262144 **OOMs**
|
|
on device-local (`cudaMalloc failed`, the KV cache needs ~10-18 GiB that does
|
|
not fit once the payload is on the card). `gl=45` (partial offload) fits a
|
|
131072 context. The env ships `CTX_SIZE=262144`; the proven serving context
|
|
today is **131072 at `gl=45`** (before the `set_rows` fix). `deploy.sh` waits
|
|
up to 20 minutes (see `LLM_WAIT`).
|
|
- **Reasoning on.** The Sharp template drives thinking (medium effort);
|
|
`--reasoning-format none` lets the template project the reasoning content into
|
|
`reasoning_content`. `--reasoning-budget -1` = no budget cap.
|
|
|
|
## Reasoning + chat template
|
|
|
|
The Sharp template (`scripts/chat-template.jinja`, repo-tracked) is a
|
|
Qwen-AGENT shape. Env `CHAT_TEMPLATE_KWARGS` keeps the reasoning trail in the
|
|
conversation:
|
|
|
|
```json
|
|
{"enable_thinking": true, "preserve_thinking": true, "preserve_reasoning": true}
|
|
```
|
|
|
|
`preserve_thinking: true` keeps the model's prior thinking in the next prompt
|
|
(the user's "preserve thinking" request). `preserve_reasoning: true` keeps the
|
|
model's prior reasoning trace across requests too (OpenWebUI sends
|
|
`preserve_reasoning`/`preserve_thinking: true` when the conversation needs the
|
|
full trail). For multi-turn reasoning, the template stays in the reasoning
|
|
shape and `enable_thinking: true` per request.
|
|
|
|
## Memory: the 114 GiB TTM/GTT requirement
|
|
|
|
The 262k f16 KV cache is large (~a few GB) and the payload is ~30 GB. Both live
|
|
in the Strix Halo unified-memory GTT. The TTM `pages_limit` was raised to
|
|
114 GiB on 2026-08-25 (written live to
|
|
`/sys/module/ttm/parameters/pages_limit`, then persisted in
|
|
`/etc/modprobe.d/ttm.conf`). Without the raise the effective ceiling is ~15 GiB
|
|
and a 28 GB + 262k-KV load stalls. The GTT page-limit change is per the
|
|
`qwen3.8-strix-optimizations` table item 5.
|
|
|
|
## Health + troubleshooting
|
|
|
|
- Health: `curl http://192.168.0.143:7770/health` → `{"status":"ok"}`.
|
|
- `/api/models` reports `n_ctx:262144, n_ctx_train:262144, n_params:27.3B`.
|
|
- Model load is slow (a few minutes cold for the ~30 GB payload with a fresh
|
|
KV cache). Health returns 503 while loading. Do not kill the process.
|
|
`deploy.sh` waits up to 20 minutes (see `LLM_WAIT`).
|
|
- Kernel spam about `amdgpu: SVM mapping failed` is the known documented
|
|
symptom, not root cause (see the crash-loop procedure in the project
|
|
AGENTS.md). Investigate other causes before assuming memory pressure.
|
|
- A client abort logs `W srv stop: cancel task, id_task = N`; a natural stop
|
|
logs only `slot release` with `truncated = 0`.
|
|
|
|
## Env tunables
|
|
|
|
`scripts/frank-dragoon.env`. Serving: `CTX_SIZE=262144`,
|
|
`PARALLEL=1`, `BATCH_SIZE=512`, `UBATCH_SIZE=512`, `NUM_THREADS=16`,
|
|
`THREAD_BATCH=32`, `CACHE_TYPE_K=f16`, `CACHE_TYPE_V=turbo4`,
|
|
`CTX_CHECKPOINTS=32`, `CACHE_RAM=8192`. Reasoning: `REASONING=on`,
|
|
`REASONING_FORMAT=none`. Sampling: `TEMPERATURE=0.7`, `TOP_K=20`,
|
|
`TOP_P=0.8`. MTP: `MTP_N_MAX=4`, `MTP_N_MIN=0`, `MTP_P_MIN=0.0`,
|
|
`MTP_P_SPLIT=0.10`. Strix Halo: `HSA_OVERRIDE_GFX_VERSION=11.5.1`,
|
|
`GGML_HIP_ENABLE_UNIFIED_MEMORY=1`.
|
|
|
|
**Never name a tunable `TEMP`, `HOST`, or `GPU_MAX_HEAP_SIZE`** — the ROCm HIP
|
|
runtime reads those bare vars (see the 2026-08-13 TEMP incident). llama.cpp
|
|
reads `--host/--port/--temp` as CLI flags, never as env vars.
|
|
|
|
## History
|
|
|
|
- **2026-08-25 (today):** switched the large service to the kingjones
|
|
Q8_0-AGENT model with 262144 context and reasoning on. Downloaded the three
|
|
payload files from the HF repo, raised the TTM/GTT to 114 GiB for the full
|
|
context, and persisted the env/serve script in this repo. Measured that the
|
|
model loads but crashes in `ggml_compute_forward_set_rows` on all five
|
|
available ROCmFPX builds; serving is BLOCKED pending a rebuild. The config
|
|
goals (GTT 114 GiB, reasoning on, preserve_thinking/preserve_reasoning on,
|
|
262144 env context) are in place. Proven serving context today is 131072 at
|
|
`gl=45`. See the `set_rows` build bug section above.
|
|
- Preceding model: Kairic Edge Qwen3.8-27B IU4 (deployed 2026-08-25 ~05:42,
|
|
retired the same day for the AGENT model). See [[kairic-edge]].
|