feat(selene+mistral): restore Selene judge (FP8, GPU1) + push Mistral to 256K

selene: AtlaAI Selene-1-Mini-Llama-3.1-8B judge restored on vLLM after the
llama-swap teardown took its Q6_K GGUF offline. FP8 (dynamic --quantization
fp8; FP8 >= the validated Q6_K fidelity, and text-only Llama so no vision-
tower-noise risk; NVFP4's W4A4 too aggressive for a precision judge). GPU1
util 0.13 (8.51 GiB weights + 2.53 GiB KV, 32K ctx, 1.27x concurrency),
~11 GB GPU1 buffer left. Gateway selene-1-mini-8b → :8011 (shadows the *
wildcard that used to reach it via llama-swap). Judge smoke: scored an
unfaithful claim 1/5 correctly.

mistral-small-4: max-model-len 131072 → 262144 (full native 256K) for
novel-length consistency-checking. KV pool is util-bound (~862K tokens), so
256K costs no extra VRAM — max concurrency just drops to 3.29x at full length.
max-num-seqs 64 → 32 keeps the warmup transient flat (scales with seqs × len),
so it fits the tight GPU0 (free unchanged at 5.2 GB). Verified loaded + healthy.
This commit is contained in:
2026-06-15 18:52:26 -07:00
parent 9a49963d07
commit c985ede07b
5 changed files with 142 additions and 14 deletions
+12
View File
@@ -95,6 +95,18 @@ model_list:
model_info:
mode: chat
# --- Selene 1 Mini 8B (AtlaAI judge, FP8) — restored on GPU1 after the
# llama-swap teardown (was the Q6_K GGUF in the swap zoo). vLLM dynamic fp8,
# :8011. Explicit entry shadows the "*" wildcard (which used to reach it via
# llama-swap). Hallucination/RAG-faithfulness judge; callers set temp ~0.01. ---
- model_name: selene-1-mini-8b
litellm_params:
model: hosted_vllm/selene-1-mini-8b
api_base: http://10.250.50.54:8011/v1
api_key: os.environ/VLLM_API_KEY
model_info:
mode: chat
# --- Qwen3 embeddings ---
- model_name: qwen3-embedding
litellm_params:
+10 -7
View File
@@ -22,14 +22,17 @@ MISTRAL_PORT=8010
# the embed/rerank/reward trio.
MISTRAL_GPU_ID=0
# util 0.93 (~89 GB budget) — 74.4 GB weights + ~5 GB CUDA/graph leaves ~10 GB
# KV. MLA keeps KV compressed so 131072 ctx fits; raise toward native 256K only
# after measuring real KV headroom. Dedicated card, so 0.93 is safe.
# util 0.93 (~89 GB budget) — measured 66.1 GiB weights + 0.8 GiB graph + ~18.5 GiB
# KV. MLA keeps KV compressed, so the FULL native 256K context fits (verified).
MISTRAL_GPU_MEM_UTIL=0.93
MISTRAL_MAX_MODEL_LEN=131072
# Single-card KV is tighter than the official TP=2 setup → cap concurrency at 64
# (official used 128 across two cards).
MISTRAL_MAX_NUM_SEQS=64
# 262144 = native 256K (for novel-length consistency-checking). The KV pool is
# util-bound (~862K tokens) regardless of max-len, so 256K costs no extra VRAM —
# it just lets one request use up to 256K (→ max concurrency 3.29x at full length).
MISTRAL_MAX_MODEL_LEN=262144
# 32 (halved from 64 when going to 256K): the warmup transient scales with
# max-num-seqs × max-model-len, so halving seqs while doubling len keeps it flat
# and fits the tight card. 32 is ample — this is a low-concurrency creative model.
MISTRAL_MAX_NUM_SEQS=32
# Optional — model is ungated (Apache-2.0), no token needed.
HF_TOKEN=
+8 -7
View File
@@ -19,16 +19,17 @@
# reference 80 GB cards can't fit 74.4 GB + context on one). The 96 GB Blackwell
# flips that to single-card: 74.4 GB weights + ~5 GB overhead leaves ~17 GB for
# KV. Mistral Small 4 uses MLA attention (TRITON_MLA) so KV is compressed/cheap —
# big context stays affordable even on a constrained KV pool. max-model-len is
# capped to 131072 on first bring-up (raise toward the native 256K once real KV
# headroom is measured).
# big context stays affordable even on a constrained KV pool. We serve the FULL
# native 256K (max-model-len 262144) — the KV pool is util-bound (~862K tokens)
# so 256K costs no extra VRAM, it just lets one request use up to 256K (max
# concurrency 3.29x at full length). For novel-length consistency-checking.
#
# vLLM FLOOR: needs >= 0.20 (Mistral Small 4 day-0 support); validated on 0.23.0.
# Do NOT reuse the qwen36-vl 0.19.1 image — it predates this model.
# vLLM PIN: v0.22.0 (in .env) — the last release with WORKING Mistral vision
# (#44911 fetch_images regression hit 0.22.1+/0.23.0). See the .env header.
#
# Serve flags mirror Mistral's official command (cited in README), adapted for
# single-card: TP 2->1, util 0.8->0.93, max-len 262144->131072, max-num-seqs
# 128->64. All tunables live in .env — edit that, not this file.
# single-card: TP 2->1, util 0.8->0.93, max-num-seqs 128->32 (32 keeps the
# 256K warmup transient flat). All tunables live in .env — edit that, not this file.
name: mistral-small-4
+26
View File
@@ -0,0 +1,26 @@
# Selene 1 Mini 8B judge (FP8) on ana-ml2 GPU 1 — copy to .env on the host.
# Real .env lives on ana-ml2 at /opt/docker/compose/selene/.env (gitignored).
#
# See compose.yaml header for the FP8-over-NVFP4 (judge-fidelity) rationale.
# 0.23.0 — Llama 3.1 + dynamic fp8 is rock-solid here (same digest as qwen36).
SELENE_IMAGE=vllm/vllm-openai@sha256:6d8429e38e3747723ca07ee1b17972e09bb9c51c4032b266f24fb1cc3b22ed8f
SELENE_CONTAINER_NAME=vllm-selene
SELENE_MODEL=AtlaAI/Selene-1-Mini-Llama-3.1-8B
SELENE_PORT=8011
# GPU 1 = shared with qwen36 (NVFP4) + granite + embed/rerank/reward.
SELENE_GPU_ID=1
# util 0.13 (~12.5 GB) — measured: 8.51 GiB FP8 weights + ~1.5 GiB graph +
# 2.53 GiB fp8 KV (41,456-token pool, 1.27x concurrency at full 32K). util 0.12
# was too thin (1.86 GiB KV < the 2.0 GiB a single 32K request needs → crash).
# Keeps GPU 1 total ~0.95 → ~5 GB buffer; ctx 32768 mirrors the old judge config.
SELENE_GPU_MEM_UTIL=0.13
SELENE_MAX_MODEL_LEN=32768
SELENE_MAX_NUM_SEQS=16
# Optional — checkpoint is ungated (Apache-2.0).
HF_TOKEN=
API_KEY=
+86
View File
@@ -0,0 +1,86 @@
# selene — AtlaAI Selene 1 Mini (Llama 3.1 8B) judge/eval model on ana-ml2 GPU 1.
#
# Restores the judge that went offline when llama-swap was downed (it was the
# Q6_K GGUF `selene-1-mini-8b` in the llama-swap zoo). Re-served on vLLM at the
# operator's request, FP8 (NVFP4 had no pre-made checkpoint and W4A4 is too
# aggressive for a precision judge validated at Q6_K — FP8 ≥ Q6_K fidelity).
#
# FP8 = vLLM DYNAMIC --quantization fp8 (W8A8) of the bf16 AtlaAI checkpoint —
# no offline quant needed, near-lossless, and Selene is text-only Llama 3.1 so
# there's NO vision tower for dynamic fp8 to noise-quantize (the qwen35-VL
# footgun doesn't apply here). ~8 GiB weights on GPU 1's headroom.
#
# Co-tenant on GPU 1 with qwen36 (NVFP4) + granite + embed/rerank/reward. Sized
# to fit the ~24 GB headroom while leaving GPU 1 a safe buffer (see .env).
# Served ONLY as `selene-1-mini-8b` (the name its consumers know). All tunables
# live in .env.
name: selene
services:
vllm-selene:
image: ${SELENE_IMAGE}
container_name: ${SELENE_CONTAINER_NAME}
restart: unless-stopped
ipc: host
ports:
- "${SELENE_PORT}:8000"
volumes:
- /tank/aimodels/huggingface:/hfcache
environment:
- HF_HOME=/hfcache
- HF_HUB_CACHE=/hfcache/hub
- HUGGING_FACE_HUB_TOKEN=${HF_TOKEN:-}
- VLLM_API_KEY=${API_KEY:-}
command:
- ${SELENE_MODEL}
# Dynamic FP8 (W8A8) from the bf16 checkpoint — no pre-quant needed.
- --quantization
- fp8
- --served-model-name
- selene-1-mini-8b
- --host
- 0.0.0.0
- --port
- "8000"
- --gpu-memory-utilization
- ${SELENE_GPU_MEM_UTIL}
- --max-model-len
- ${SELENE_MAX_MODEL_LEN}
- --max-num-seqs
- ${SELENE_MAX_NUM_SEQS}
# fp8 KV — matches the judge's old q8 KV posture + keeps the pool compact
# on the shared card.
- --kv-cache-dtype
- fp8
- --dtype
- auto
- --enable-prefix-caching
deploy:
resources:
reservations:
devices:
- driver: nvidia
device_ids:
- "${SELENE_GPU_ID}"
capabilities:
- gpu
healthcheck:
test: ["CMD", "curl", "-f", "http://localhost:8000/health"]
interval: 30s
timeout: 10s
retries: 3
start_period: 180s
networks:
- tnet
labels:
- homepage.group=AI Systems
- homepage.name=Selene 1 Mini 8B (judge, FP8)
- homepage.icon=mdi-gavel
- homepage.description=AtlaAI Selene 1 Mini Llama-3.1-8B judge (FP8) via vLLM (ana-ml2 GPU1)
- homepage.href=http://10.250.50.54:${SELENE_PORT}/docs
networks:
tnet:
name: traefik-net
external: true