feat(nh3-ml1): LFM2.5-VL-3B (llama.cpp) + VibeVoice-ASR-Streaming-1.5B (audio.cpp) utility seats

For brokkr's dataset foundry (operator-approved 2026-09-26, relayed).
- stacks/lfm-vl-seat: llama.cpp server-cuda b11176 (digest-pinned), Q5_K_M +
  mmproj Q8_0, :8030; gateway alias lfm25-vl-3b (LiteLLM restarted, 36 s).
  Positive control exact; null control shows it describes a missing image.
- stacks/vibevoice-asr-seat: audio.cpp v0.8.2-audio8-perf-hotfix (the GGUF's
  own runtime, not vibevoice.cpp) on cuda 12.8 runtime + libgomp + libsoxr,
  sha256-pinned; :8031 direct. LibriSpeech WER 3/69, RTF 0.07-0.14; ~31 s
  cold first request.
This commit is contained in:
vh
2026-09-26 00:41:01 -07:00
parent 45484a0007
commit f2792183d4
11 changed files with 235 additions and 0 deletions
+1
View File
@@ -203,6 +203,7 @@ in `servers/pfi-gx10/README.md`. → `persistent-memory.d/2026-09-24-gx10-ac-res
## Recent decisions
- `[2026-09-26]` **Two dataset-foundry utility seats LIVE on nh3-ml1 (brokkr; operator approval relayed):** LFM2.5-VL-3B on llama.cpp `:8030` (gateway `lfm25-vl-3b`, LiteLLM restarted 36 s at 0039) and VibeVoice-ASR-Streaming-1.5B on **audio.cpp** `:8031`, not vibevoice.cpp as specced; the GGUF card names audio.cpp. Controls: VL read a synthetic image exactly, but hallucinates when no image is sent; ASR WER 3/69 on the bundled LibriSpeech clips. libsoxr added (linear resampling misheard a word). First ASR request ~31 s cold. Answered brokkr: the old 27B image seat (10.250.50.54:8015) was retired 09-14; baseline = gateway `image-judge` (Flash-Next).
- `[2026-09-26]` **Coder seat STAYS on fv-ml1 (Prime).** The nh3-ml1 copy gave the same quality (teacher-forced true-code logprob diff +0.008 ± 0.019) but ran ~5× slower (64-tok FIM ~1.0 s vs 0.2 s; 63 vs 338 tok/s), and freeing 6.3 GB on fv-ml1 GPU 1 (20 GB spare) bought little. Copy removed; recipe kept in `stacks/coder-seat/`. The RTX 2000E suits embed/rerank/classify, not latency-sensitive generation. Only coder was a candidate: parakeet stays (earlier ruling), voices is a generation seat.
- `[2026-09-25]` **nh3-ml1 LIVE after the NH3 visit (SB off, IGFX restored): driver + CT 109 + TEI, parity vs esh-ml1 indistinguishable (cos min 0.999993 = own floor; overlap@10 1.000 vs MRL-256 control 0.684), same speed; monitoring/DNS wired. Autonomous: lxc-pve 6.0.0-2 upgrade on nh3-pve (Docker-in-LXC fix #7006). Gateway routing + AMT follow-up are Prime's.** → `persistent-memory.d/2026-09-25-nh3-ml1-live.md`
- `[2026-09-25]` **nh3-ml1 = CT 109 @ 10.100.50.80, second TEI embed/rerank backend on nh3-pve's RTX 2000E; GPU playbooks made host-generic — BUILD DEFERRED: blocked on nh3-pve Secure Boot until the NH3 visit; SB handling and gateway routing are Prime's calls.** Tracked: `7ddd116`, Current state. → `persistent-memory.d/2026-09-25-nh3-ml1-standup.md`
+8
View File
@@ -33,6 +33,14 @@ Prime's call (recommendation: load-share; see below).
VRAM ~2.7 GB for both, so ~13 GB is free.
**Dataset-foundry utility seats (2026-09-26, brokkr; operator-approved):**
- `stacks/lfm-vl-seat`: LFM2.5-VL-3B on llama.cpp, `:8030`, gateway `lfm25-vl-3b`.
- `stacks/vibevoice-asr-seat`: VibeVoice-ASR-Streaming-1.5B on audio.cpp,
`:8031`, direct only.
GPU total with TEI is ~7.5 of 16 GB. Both are batch workloads, which suits this
card.
**Tried and removed (2026-09-25):** a copy of the code-completion seat
(Qwen2.5-Coder-1.5B). It gave the same quality but ran about 5× slower than on
fv-ml1 (a 64-token FIM completion took ~1.0 s vs ~0.2 s), so it stays on fv-ml1
+14
View File
@@ -0,0 +1,14 @@
# lfm-vl-seat tunables (nh3-ml1). Copy to `.env` on the server.
# llama.cpp CUDA server, pinned by digest: build b11176 (commit f805c57a2),
# published 2026-09-25. Ada sm_89 is in the CUDA build's target set.
LLAMACPP_IMAGE=ghcr.io/ggml-org/llama.cpp@sha256:1f4b9cf58982dd4d7cc497aea31b1a456ca9a3a1f94f527d317d3fdee0d60ab6
VL_PORT=8030
VL_ALIAS=lfm25-vl-3b
VL_MODEL_FILE=LFM2.5-VL-3B-Q5_K_M.gguf
VL_MMPROJ_FILE=mmproj-LFM2.5-VL-3B-Q8_0.gguf
# 2 parallel slots sharing 16k context (8k each): room for a few images plus text.
VL_CTX=16384
VL_PARALLEL=2
+23
View File
@@ -0,0 +1,23 @@
# lfm-vl-seat
**LiquidAI LFM2.5-VL-3B**, a small vision-language model, on **nh3-ml1**, served
by llama.cpp (`ghcr.io/ggml-org/llama.cpp` server-cuda, build b11176, pinned by
digest) on `:8030`. A utility seat for brokkr's dataset foundry (image
understanding). Operator-approved 2026-09-26, relayed by brokkr-smithy-dev.
| | |
|---|---|
| direct | `http://10.100.50.80:8030/v1/chat/completions`, model `lfm25-vl-3b` |
| gateway | `lfm25-vl-3b` (ana-docker:4000, `hosted_vllm/`, `supports_vision`) |
| files | `LFM2.5-VL-3B-Q5_K_M.gguf` + **`mmproj-LFM2.5-VL-3B-Q8_0.gguf` (required for images)** from `LiquidAI/LFM2.5-VL-3B-GGUF` @ `6f730e9a2c45` |
| VRAM | ~2.9 GB (2 slots × 8k ctx) |
**Checks, 2026-09-26.**
- Positive control: a synthetic image (red square, blue circle, the text
"NH3 42") was described exactly at T=0, both direct and through the gateway.
- Speed: prefill ~800 tok/s, decode ~92 tok/s.
- ⚠ **Null control: with no image attached it confidently describes one anyway**
("a solid black background"). Callers must make sure the image actually went in.
Not wired yet: `LiquidAI/LFM2.5-VL-3B-DSpark-GGUF`, a speculative-decoding
drafter (brokkr's "optional later").
+62
View File
@@ -0,0 +1,62 @@
# lfm-vl-seat — LiquidAI LFM2.5-VL-3B (vision-language) on nh3-ml1 (CT 109 on
# nh3-pve, RTX 2000E Ada 16 GB), llama.cpp server. Utility seat for brokkr's
# dataset foundry (image understanding), operator-approved 2026-09-26 (relayed
# by brokkr-smithy-dev).
#
# llama-server LFM2.5-VL-3B Q5_K_M + mmproj Q8_0 → /v1/chat/completions
# (OpenAI, with image_url content parts) :8030
#
# ⚠ --mmproj is REQUIRED for image input; without it the model is text-only and
# says nothing about it.
# Batch work, not latency-critical: this card decodes ~5x slower than an fv-ml1
# Blackwell (servers/nh3-ml1/README.md), which is fine for a foundry.
#
# Files (not in git): /opt/aimodels/gguf/lfm25-vl-3b/ from
# LiquidAI/LFM2.5-VL-3B-GGUF @ 6f730e9a2c45.
name: lfm-vl-seat
services:
llama-server:
image: ${LLAMACPP_IMAGE}
container_name: lfm-vl
restart: unless-stopped
ports:
- "${VL_PORT}:8080"
volumes:
- /opt/aimodels/gguf/lfm25-vl-3b:/models:ro
command:
- -m
- /models/${VL_MODEL_FILE}
- --mmproj
- /models/${VL_MMPROJ_FILE}
- --alias
- ${VL_ALIAS}
- --host
- 0.0.0.0
- --port
- "8080"
- -ngl
- "999"
- -c
- "${VL_CTX}"
- -np
- "${VL_PARALLEL}"
- --jinja
deploy:
resources:
reservations:
devices:
- driver: nvidia
device_ids: ["0"]
capabilities: [gpu]
healthcheck:
test: ["CMD", "curl", "-fsS", "http://localhost:8080/health"]
interval: 30s
timeout: 10s
retries: 3
start_period: 120s
labels:
- homepage.group=AI - Eval & Retrieval
- homepage.name=VL — LFM2.5-VL-3B (llama.cpp, nh3-ml1)
- homepage.icon=mdi-image-search
- homepage.description=Image understanding for the dataset foundry (gateway lfm25-vl-3b)
- homepage.href=http://10.100.50.80:${VL_PORT}
+14
View File
@@ -512,6 +512,20 @@ model_list:
model_info:
mode: rerank
# --- lfm25-vl-3b → LiquidAI LFM2.5-VL-3B (Q5_K_M + mmproj Q8_0), llama.cpp server on
# nh3-ml1 :8030 (stacks/lfm-vl-seat, 2026-09-26). Image understanding for
# brokkr's dataset foundry; send images as OpenAI image_url parts. A small VLM:
# with NO image attached it confidently describes one anyway (measured), so
# callers must check the image actually went in. Batch-grade speed (~90 tok/s). ---
- model_name: lfm25-vl-3b
litellm_params:
model: hosted_vllm/lfm25-vl-3b
api_base: http://10.100.50.80:8030/v1
api_key: os.environ/VLLM_API_KEY
model_info:
mode: chat
supports_vision: true
# --- coder-fast → Qwen2.5-Coder-1.5B (BASE), FIM code-completion seat (ana-ml2
# GPU1 :8020, vLLM; deep-research pick 2026-07-27). For Zed editor inline
# edit-predictions via the LEGACY /v1/completions endpoint with Qwen FIM
+2
View File
@@ -0,0 +1,2 @@
# vibevoice-asr-seat tunables (nh3-ml1). Copy to `.env` on the server.
ASR_PORT=8031
+16
View File
@@ -0,0 +1,16 @@
# audio.cpp server (0xShug0/audio.cpp) for VibeVoice-ASR-Streaming on nh3-ml1.
# The release's Linux CUDA build ("cuda12.8-colab") is a bare binary that needs the
# CUDA 12 runtime libs (cudart, cublas, cufft, nccl) plus libgomp. This image
# supplies them, plus libsoxr (without it audio.cpp falls back to linear
# resampling for the 16k→24k step). Built locally on the host; not pushed.
FROM nvidia/cuda:12.8.1-runtime-ubuntu22.04
ARG AUDIOCPP_TAG=v0.8.2-audio8-perf-hotfix
ARG AUDIOCPP_SHA256=ccfb35869d67e520801981392189a1ab8b6c829151093e0816ecd6a2d828d8c8
RUN apt-get update && apt-get install -y --no-install-recommends libgomp1 libsoxr0 curl ca-certificates \
&& rm -rf /var/lib/apt/lists/*
RUN curl -fsSL -o /tmp/a.tgz "https://github.com/0xShug0/audio.cpp/releases/download/${AUDIOCPP_TAG}/audio-${AUDIOCPP_TAG}-bin-ubuntu-x64-cuda12.8-colab.tar.gz" \
&& echo "${AUDIOCPP_SHA256} /tmp/a.tgz" | sha256sum -c - \
&& mkdir -p /app && tar xzf /tmp/a.tgz -C /app --no-same-owner && rm /tmp/a.tgz \
&& chmod +x /app/audiocpp_server
WORKDIR /app
ENTRYPOINT ["/app/audiocpp_server"]
+30
View File
@@ -0,0 +1,30 @@
# vibevoice-asr-seat
**Microsoft VibeVoice-ASR-Streaming-1.5B** (Q4_K) on **nh3-ml1**, served by
**audio.cpp** (`0xShug0/audio.cpp` v0.8.2-audio8-perf-hotfix, Linux CUDA 12.8
build) on `:8031`, with direct access only. A utility seat for brokkr's dataset
foundry (speech → text with inline speaker labels). Operator-approved 2026-09-26,
relayed by brokkr-smithy-dev.
- `POST /v1/audio/transcriptions`: multipart `file=@x.wav`, `model=vibevoice-asr-streaming-1.5b`.
Returns `{"text": " \n Speaker 0:...", "timing": {"rtf": ...}}`.
- `POST /v1/audio/transcriptions/live?model=…&sample_rate=16000&channels=1&sample_format=s16le`
takes chunked raw PCM.
- `GET /v1/models`, `GET /health`.
**Why audio.cpp and not vibevoice.cpp or llama.cpp.** The
`christopherthompson81/VibeVoice-ASR-Streaming-1.5B-GGUF` files are audio.cpp
packages with their sidecars embedded, per the model card, so no separate
tokenizer is needed. The release binary expects the CUDA 12 runtime libraries,
so `Dockerfile` puts it on `nvidia/cuda:12.8.1-runtime` with a SHA-256-pinned
download, plus `libgomp1` and **`libsoxr0`**. Without libsoxr it falls back to
linear resampling (16k→24k), and on one clip that turned "cutter" into "country".
**Checks, 2026-09-26.** On the 4 LibriSpeech clips shipped with audio.cpp, WER
is 3/69 = 4.35%, identical across 3 reps. Two of the three errors are "I'm" vs
"I am" normalization. RTF is 0.07–0.14. ⚠ **The first request after a start
takes ~31 s** (CUDA graph warmup); later ones take 0.3–1 s. VRAM ~2.0 GB.
The model card recommends Q8_0 (3.3 GB; its CUDA WER is 5.80% vs Q4_K's 7.25%,
on 69 words). To switch, change `path` in `conf/server.json`, download the file,
and recreate the container.
+50
View File
@@ -0,0 +1,50 @@
# vibevoice-asr-seat — Microsoft VibeVoice-ASR-Streaming-1.5B on nh3-ml1 (CT 109,
# RTX 2000E Ada 16 GB), served by audio.cpp. Utility seat for brokkr's dataset
# foundry (speech → text, speaker-attributed), operator-approved 2026-09-26
# (relayed by brokkr-smithy-dev).
#
# audiocpp_server vibevoice-asr-streaming-1.5b (Q4_K) :8031
# POST /v1/audio/transcriptions (OpenAI-style, whole file)
# POST /v1/audio/transcriptions/live?... (chunked 16 kHz s16le PCM)
# GET /v1/models, /health
#
# Runtime is audio.cpp, per the GGUF's own model card (the GGUF is an audio.cpp
# package with its sidecars embedded, so no separate tokenizer file). It is NOT
# vanilla llama.cpp and NOT mudler's vibevoice.cpp.
# Direct access only; no gateway entry.
#
# Files (not in git): /opt/aimodels/gguf/vibevoice-asr-streaming-1.5b/ from
# christopherthompson81/VibeVoice-ASR-Streaming-1.5B-GGUF @ ff9615110299.
name: vibevoice-asr-seat
services:
audiocpp:
build:
context: .
image: local/audiocpp:v0.8.2-audio8-perf-hotfix
container_name: vibevoice-asr
restart: unless-stopped
command: ["--config", "/config/server.json", "--no-ui", "--log"]
ports:
- "${ASR_PORT}:8080"
volumes:
- /opt/aimodels/gguf/vibevoice-asr-streaming-1.5b:/models:ro
- /opt/docker/conf/vibevoice-asr-seat/server.json:/config/server.json:ro
deploy:
resources:
reservations:
devices:
- driver: nvidia
device_ids: ["0"]
capabilities: [gpu]
healthcheck:
test: ["CMD", "curl", "-fsS", "http://localhost:8080/health"]
interval: 30s
timeout: 10s
retries: 3
start_period: 120s
labels:
- homepage.group=AI - Eval & Retrieval
- homepage.name=ASR — VibeVoice Streaming 1.5B (audio.cpp, nh3-ml1)
- homepage.icon=mdi-microphone-message
- homepage.description=Streaming speech-to-text for the dataset foundry (direct, no gateway)
- homepage.href=http://10.100.50.80:${ASR_PORT}/v1/models
@@ -0,0 +1,15 @@
{
"host": "0.0.0.0",
"port": 8080,
"backend": "cuda",
"threads": 6,
"models": [
{
"id": "vibevoice-asr-streaming-1.5b",
"family": "vibevoice_asr_streaming",
"path": "/models/vibevoice-asr-streaming-1.5b-q4_k.gguf",
"task": "asr",
"mode": "streaming"
}
]
}