Frank
f998a2d9f2
DFlash2 (100.0 t/s decode, 287.1 t/s burst-4) beats MTP self-speculation (85.9 t/s, 179.7 t/s) despite a lower accept length (4.24 vs 5.66): the small separate drafter steps faster than the target's own MTP layer applied autoregressively. DFlash2 is the deployed shape; both swap-script paths now default to dflash/7. Docs updated.
186 lines
8.7 KiB
Markdown
186 lines
8.7 KiB
Markdown
---
|
||
type: runbook
|
||
title: frank-medium Model Swap
|
||
description: How to swap the checkpoint behind the frank-medium endpoint (Dankserver :7770, vllm-radiance TP=2) — the streamlining toolchain (hf-manifest.py, hf-fetch.sh, swap-frank-medium.sh), the quantization-aware entrypoint contract, the quadlet edit, the deploy, the capability battery, and the rollback path.
|
||
tags:
|
||
- frank-medium
|
||
- dankserver
|
||
- model-swap
|
||
- vllm-radiance
|
||
- huggingface
|
||
- runbook
|
||
timestamp: 2026-09-20
|
||
---
|
||
|
||
# frank-medium Model Swap
|
||
|
||
Swapping the checkpoint behind **frank-medium** (Dankserver `:7770`,
|
||
vllm-radiance TP=2) is a four-step process, and the toolchain below drives all
|
||
of it. The design goal (2026-09-20, after the MXFP4 → FP8 swap) is that a swap
|
||
is **one command**: generate a manifest, download it verified, point the
|
||
quadlet at it, deploy. The entrypoint is quantization-aware, so the only repo
|
||
change a swap makes is the quadlet's four env vars.
|
||
|
||
## Why this exists
|
||
|
||
The 2026-09-20 MXFP4 → FP8 swap took most of a day, and the long pole was not
|
||
the GPU — it was the **download** (30.9 GB across 7 shards, hand-verified) and
|
||
the **serving-path decision** (which attention backend, which drafter, which
|
||
quantization flag). Both are now mechanical:
|
||
|
||
- The download is `hf-fetch.sh` — resumable, byte-count + sha256-verified,
|
||
idempotent (a re-run only fetches what is missing or corrupt).
|
||
- The serving path is the **quantization-aware entrypoint**
|
||
(`dankserver/frank-medium/radiance-entrypoint.sh`) — it reads
|
||
`quantization_config.quant_method` from the checkpoint's `config.json` and
|
||
picks the path. The swap operator never chooses it by hand.
|
||
|
||
## The toolchain
|
||
|
||
| Tool | Role |
|
||
|---|---|
|
||
| `scripts/hf-manifest.py` | Generate a fetch manifest from a live HF repo tree: `<filename> <size> <sha256-or-blank>` per line. LFS sha256s come from the resolve endpoint's `X-Linked-ETag`; non-LFS files get a blank hash (size-only check). |
|
||
| `scripts/hf-fetch.sh` | Fetch a manifest to a dir, resumable (`curl -C -`), verified (byte count + sha256), idempotent. Token out-of-band via a mode-600 `curl -K` file. |
|
||
| `scripts/swap-frank-medium.sh` | The orchestrator: generate manifest → download → detect quantization → edit the quadlet → deploy. |
|
||
|
||
The HF token is resolved on the **host** (dankserver) from `$HF_TOKEN`,
|
||
`~/.cache/huggingface/token`, or the llamaswap conf. Gated repos need a token
|
||
that has been granted access (the 2026-09-20 FP8 swap needed a one-time grant
|
||
for `orcarouter/Qwen3.8-27B-Uncensored-FP8`).
|
||
|
||
## The swap (one command)
|
||
|
||
```
|
||
scripts/swap-frank-medium.sh \
|
||
--repo <owner/repo> \
|
||
--target /home/dank/models/<name> \
|
||
--generate-manifest \
|
||
--apply \
|
||
--deploy
|
||
```
|
||
|
||
What it does, in order:
|
||
|
||
1. **Manifest** — `hf-manifest.py` lists the repo tree and writes
|
||
`scripts/frank-medium-<name>.manifest` (committed, so the payload is
|
||
reproducible). Pass `--manifest <path>` to reuse an existing one.
|
||
2. **Download** — stages the manifest + `hf-fetch.sh` on the host and fetches
|
||
to `--target`. Resumable: an interrupted swap re-runs and skips what is
|
||
already verified.
|
||
3. **Detect** — reads `quantization_config.quant_method` from the downloaded
|
||
`config.json` and maps it to the serving shape (below).
|
||
4. **Apply** — edits the quadlet's four env vars (`SNAP`, `SPEC_METHOD`,
|
||
`SPEC`, `KV_MEM`).
|
||
5. **Deploy** — `deploy.sh dankserver --service frank-medium --skip-e2e`
|
||
(pull, restart, health gate, up to 1800 s — the first start compiles
|
||
kernels).
|
||
|
||
Without `--apply` the script stops after the download + detection and prints
|
||
the quadlet lines to set, so you can eyeball them before committing.
|
||
|
||
## Quantization → serving path
|
||
|
||
The entrypoint (`radiance-entrypoint.sh`) reads `quant_method` and branches:
|
||
|
||
| `quant_method` | Path | Attention | Speculation | Drafter |
|
||
|---|---|---|---|---|
|
||
| `fp8` | plain FP8 | `ROCM_AITER_UNIFIED_ATTN` | **DFlash2 external drafter** (A/B winner) | `/models/Qwen3.8-27B-DFlash2-FP8` |
|
||
| `mxfp4` / `quark` / empty | MXFP4 | `R4D` (if `R4D_ATTN=1`) | DFlash2 external drafter | `/models/Qwen3.8-27B-DFlash2-FP8` |
|
||
|
||
The quadlet `SPEC_METHOD`/`SPEC` the swap script sets: **both paths use
|
||
`dflash/7`** (the DFlash2 drafter is quantization-agnostic — it proposes
|
||
tokens, the target verifies, so it works with any Qwen3.8-27B target). The
|
||
FP8 model also carries its own single-layer MTP head (22 tensors), so MTP
|
||
self-speculation (`mtp/8`) is available as an alternative — but A/B tested on
|
||
this checkpoint, **DFlash2 wins on decode** (100.0 vs 85.9 t/s geomean, 287.1
|
||
vs 179.7 t/s burst-4) despite a lower accept length (4.24 vs 5.66), because
|
||
the small separate drafter steps faster than the target's own MTP layer
|
||
applied autoregressively. The MXFP4 path additionally applies the
|
||
`radiance-vllm-mxfp4` patches at every container start.
|
||
|
||
**R4D allreduce is independent of the weight quantization.** The FP8 path
|
||
uses `ROCM_AITER_UNIFIED_ATTN` (the documented plain-FP8 default); R4D can be
|
||
tried later for a speed bump (`R4D_ATTN=1`), but the MXFP4 path is the one
|
||
tested with it.
|
||
|
||
## The quadlet contract
|
||
|
||
A swap edits exactly four env vars in
|
||
`dankserver/frank-medium/quadlet/frank-medium.container`:
|
||
|
||
- `SNAP=/models/<name>` — the container path (the host dir
|
||
`/home/dank/models/<name>` is mounted at `/models`).
|
||
- `SPEC_METHOD=mtp|dflash` — the speculative-decoding method.
|
||
- `SPEC=8|7` — the speculative depth.
|
||
- `KV_MEM=0` — the KV pool pin in MiB; `0` = auto-size (the documented
|
||
plain-FP8 default). After the first healthy start, recalibrate with
|
||
`calibrate-kv.sh` to use the extra VRAM on the 34.2 GiB cards.
|
||
|
||
The deploy handler (`deploy-remote.sh deploy_llm_medium`) reads `SNAP` and
|
||
`SPEC_METHOD` from the quadlet (the source of truth), maps `/models/` →
|
||
`/home/dank/models/`, and gates on the right assets: the drafter only for
|
||
`dflash`, `r4d.so` only for `mxfp4`/`quark`. So a swap that points the quadlet
|
||
at a new checkpoint is enough — the handler follows.
|
||
|
||
## Verify (capability battery)
|
||
|
||
After the deploy is green, run the battery (read-only — it does not stop the
|
||
service):
|
||
|
||
```
|
||
python3 scripts/bench-conversation.py http://192.168.0.235:7770 \
|
||
--model frank-medium --salt <name> --max-tokens 512 --burst 4
|
||
```
|
||
|
||
Then the capability checks (all must pass):
|
||
|
||
- `/health` → 200, `/v1/models` → `frank-medium`.
|
||
- A chat completion returns `content` **and** reasoning (vLLM emits the
|
||
`reasoning` field; the harness auto-detects it).
|
||
- MTP speculation: the boot log's `SpecDecoding metrics` line shows a mean
|
||
acceptance length > 1 (the 2026-09-20 FP8 swap measured 5.66/8, 66.2%
|
||
draft acceptance).
|
||
- Tool calls: a `tools` request returns `tool_calls`.
|
||
- Vision: an image request is answered (the FP8 checkpoint serves vision
|
||
despite `language_model_only: true` in `config.json`).
|
||
|
||
## Rollback
|
||
|
||
The previous checkpoint stays on disk (a swap never deletes it). To roll
|
||
back:
|
||
|
||
1. Point the quadlet at the old checkpoint (`SNAP`, `SPEC_METHOD`, `SPEC` —
|
||
the old values are in git history).
|
||
2. `deploy.sh dankserver --service frank-medium --skip-e2e`.
|
||
3. Re-run the capability battery.
|
||
|
||
The 2026-09-20 rollback target is the MXFP4 checkpoint at
|
||
`/home/dank/models/Qwen3.8-27B-Uncensored-MXFP4-awq` (19.37 GiB single
|
||
safetensors) with `SPEC_METHOD=dflash`, `SPEC=7`.
|
||
|
||
## 2026-09-20 swap record (MXFP4 → FP8)
|
||
|
||
- **Model**: `just1moremodel/Qwen3.8-27B-Uncensored-MXFP4-awq` (MXFP4,
|
||
19.37 GiB) → `orcarouter/Qwen3.8-27B-Uncensored-FP8` (FP8 E4M3 block-wise
|
||
128×128, dynamic activation quant, 30.9 GB / 7 shards).
|
||
- **Serving path**: DFlash2 external drafter → DFlash2 external drafter
|
||
(kept; `R4D` → `ROCM_AITER_UNIFIED_ATTN`). After the swap, MTP
|
||
self-speculation (`mtp/8`) was A/B tested against the DFlash2 drafter
|
||
(`dflash/7`) on the FP8 target.
|
||
- **DFlash2 vs MTP A/B** (the deployed shape is DFlash2):
|
||
- DFlash2: mean accept 4.24/7, geomean decode **100.0 t/s**, TTFT 0.19 s,
|
||
prefill 923.0 t/s, burst-4 **287.1 t/s**, KV pool 384,238 (1.47×).
|
||
- MTP: mean accept 5.66/8, geomean decode 85.9 t/s, TTFT 0.06 s, prefill
|
||
1096.9 t/s, burst-4 179.7 t/s, KV pool 461,540 (1.76×).
|
||
- **DFlash2 wins on decode (+16 %) and burst-4 (+60 %)**; MTP wins on TTFT
|
||
and prefill. Accept length is a proxy, not the metric — the small separate
|
||
drafter steps faster than the target's own MTP layer applied
|
||
autoregressively, and the faster step more than makes up the lower accept
|
||
length. FP8 + DFlash2 (100.0 t/s) also beats the original MXFP4 + DFlash2
|
||
(92.4 t/s) — the higher-precision FP8 target verifies more of the drafter's
|
||
proposals.
|
||
- **KV pool**: auto-sized to 384,238 tokens (1.47× full-window) — the
|
||
DFlash2 drafter's 2.1 GiB costs the pool vs MTP's 461,540 (1.76×); both
|
||
smaller than MXFP4's 943,581 (3.59×) because FP8 weights are bigger (30.9
|
||
vs 19.4 GiB).
|