feat(cyberprev-seat): mixed-NVFP4 quant of the abliterated cyber-preview, displacing sentinel-r3

Third sec-seat candidate: hotdogs/Qwen3.8-27B-abliterated-cyber-preview, an
abliteration (refusal-direction weight edit) of Qwen3.8-27B aimed at the
cyber-offense refusal surface -- distinct from mog-sec (persona on stock
weights) and sentinel-r3 (SFT finetune). Operator instruction: quant it, take
sentinel down, serve it with mtp or dflash.

Quantized to the house mixed recipe via services/gen-seat-mixed-quant/ (NVFP4
W4A4 on MLP layers 0-55 + FP8 W8A8 on attn/linear_attn/lm_head/MLP 56-63, FP8
KV). The prior attempt (2026-09-11/14) died with "Cannot determine
num_attention_heads" because it ran from a bare .venv whose newer
compressed-tensors reads that field at top level; quant_mixed_nvfp4.py already
promotes text_config attention fields for exactly this reason, and the run
through the canonical vLLM-image + llmcompressor 0.13.0 / compressed-tensors
0.18.0 path (versions recorded from the container) completed clean.

post_quant.py did its job: grafted the 15 MTP tensors verbatim (BF16), and
re-injected re:^mtp.* into the ignore list after llm-compressor pruned it for
matching no module at quant time -- without which vLLM loads the grafted head
uninitialised and speculative decoding runs at 0% acceptance. It also caught a
missing preprocessor_config.json (absent from the abliterated source AND its
hotdogs upstream); restored from Qwen/Qwen3.8-27B, verified byte-identical to
the working sentinel-r3 build, so the vision tower preprocesses.

Verified from the tensors, not the config: NVFP4 covers MLP 0-55, FP8 covers
56-63, no overlap, 168 weight_packed tensors (56x3), 15 BF16 MTP, 333 BF16
vision. 51.0 GiB bf16 -> 21.0 GiB.

Served under its own name (cyberprev-27b / -thinking), NOT sentinel-r3 --
serving different weights under a retired name is silent substitution. Takes
over :8025 and GPU 0 (co-resident with mog-sec). dflash k=7 is the default,
carried from the sentinel measurement, but is re-measured against MTP on this
ABLITERATED body before cutover, since abliteration is exactly what can desync
an MTP head. Drafter (qwen38-27b-dflash2-drafter) shared with mog-sec.

Context capped at 163840 not native 262K, inherited from mog-sec's hard-won
depth lesson; raise only after a non-repeating deep-prefill probe on this
checkpoint.

stacks/cyberprev-seat/.env.example

@@ -0,0 +1,54 @@
+# cyberprev-seat — hotdogs/Qwen3.8-27B-abliterated-cyber-preview, fv-ml1 GPU 0, :8025.
+# Copy to .env on the host at /opt/docker/compose/cyberprev-seat/.env.
+
+# ── Image ───────────────────────────────────────────────────────────────────
+# Same pinned nightly mog-sec and sentinel-r3 run. Not :latest — pin it, so a seat
+# restart cannot silently change the engine under a measured configuration.
+CYBER_IMAGE=vllm/vllm-openai:nightly-e9d1398d9edfd90fcc1cf783805240e3effec013
+API_KEY=replace-me
+
+# ── Placement ───────────────────────────────────────────────────────────────
+# GPU 0, co-resident with mog-sec. Took over :8025 from the retired sentinel-r3 seat.
+CYBER_GPU_ID=0
+CYBER_PORT=8025
+CYBER_CONTAINER_NAME=vllm-cyberprev
+
+# ── Model ───────────────────────────────────────────────────────────────────
+# In-house mixed quant of /tank/aimodels/cyberprev-bf16 (51.0 GiB bf16 source),
+# built with services/gen-seat-mixed-quant/. compressed-tensors, NOT modelopt_fp4.
+CYBER_MODEL=/tank/aimodels/cyberprev-nvfp4-mixed
+CYBER_QUANT=compressed-tensors
+# ⚠ Its OWN name. Do not reuse `sentinel-r3` — that seat is retired and its gateway
+# aliases are deliberately left to 404 rather than repointed at different weights.
+CYBER_SERVED_NAME=cyberprev-27b
+CYBER_SERVED_NAME_THINK=cyberprev-27b-thinking
+
+# ── Memory ──────────────────────────────────────────────────────────────────
+# 0.40 of the card, sharing GPU 0 with mog-sec. KV pinned in bytes (8 GiB) so the
+# figure is reproducible regardless of what else is resident at start time.
+CYBER_GPU_MEM_UTIL=0.40
+CYBER_KV_CACHE_MEMORY=8589934592
+
+# ⚠ 163840, NOT native 262144. Inherited from mog-sec, which crashed five times
+# learning that the KV pool's capacity and the card's processing depth are different
+# numbers. Raising this requires a deep-prefill probe with a NON-REPEATING prompt on
+# THIS checkpoint — a repeated prompt hashes to cached blocks and never prefills deep.
+CYBER_MAX_MODEL_LEN=163840
+CYBER_MAX_NUM_SEQS=16
+CYBER_MAX_NUM_BATCHED_TOKENS=4096
+CYBER_KV_CACHE_DTYPE=fp8
+
+# ── Speculative decoding ────────────────────────────────────────────────────
+# One var, whole JSON — the two shapes are not interchangeable (dflash needs a
+# "model", MTP must not have one).
+#   DFlash2 : {"method": "dflash", "model": "/drafter", "num_speculative_tokens": 7}
+#   MTP     : {"method": "qwen3_5_mtp", "num_speculative_tokens": 3}
+# dflash k=7 is the default, carried from the sentinel measurement (2.40 vs 2.18).
+# ⚠ Re-measure on THIS body before trusting it: cyberprev is ABLITERATED, and
+# abliteration is precisely what can desync an MTP head. Gate MTP on a measured
+# acceptance of >=~40%, never on an assumption.
+CYBER_SPEC_CONFIG={"method": "dflash", "model": "/drafter", "num_speculative_tokens": 7}
+
+CYBER_REASONING_PARSER=qwen3
+CYBER_REASONING_EFFORT=medium
+CYBER_ALLOC_CONF=

stacks/cyberprev-seat/compose.yaml

@@ -0,0 +1,141 @@
+# cyberprev-seat — hotdogs/Qwen3.8-27B-abliterated-cyber-preview, fv-ml1 GPU 0, :8025.
+#
+# The THIRD sec-seat candidate, and the one that displaced sentinel-r3 (operator instruction
+# 2026-09-14). Lineage of the three, because the distinction is the whole point of the A/B:
+#   mog-sec     — persona system prompt on STOCK Qwen3.8-27B weights
+#   sentinel-r3 — a REAL SFT pentest finetune (proprietary licence); RETIRED for this seat
+#   cyberprev   — an ABLITERATION of Qwen3.8-27B (refusal-direction weight edit, no finetune)
+#                 aimed at the cyber-offense refusal surface specifically
+#
+# Quantized in-house to the same mixed recipe as mog-sec / sentinel / gen:
+# NVFP4 W4A4 on MLP layers 0-55 + FP8 W8A8 on attn / linear_attn / lm_head / MLP 56-63,
+# FP8 KV, vision tower + linear_attn norms + `re:^mtp.*` ignored.
+# Pipeline: services/gen-seat-mixed-quant/. Source bf16: /tank/aimodels/cyberprev-bf16.
+#
+# ⚠ `re:^mtp.*` MUST be present in config.json quantization_config.ignore, or vLLM loads the
+#   grafted BF16 MTP head UNINITIALISED and speculative decoding silently runs at 0%
+#   acceptance. llm-compressor PRUNES ignore entries that matched no module at quant time,
+#   and the wrapper class never loads the MTP head — so the entry is dropped unless
+#   post_quant.py re-injects it. post_quant.py verifies rather than assumes; do not skip it.
+#
+# ⚠ max-model-len 163840, NOT the native 262K — inherited from mog-sec the hard way. Same
+#   base, same hybrid Qwen3_5 arch: mog-sec crashed five times (420k -> 384k -> 320k -> ...)
+#   because every cut sized the KV POOL while the crashes were governed by PROCESSING DEPTH.
+#   163840 buys a clean 400-refusal above the measured ceiling instead of an engine death.
+#   Do NOT raise without re-running a deep-prefill probe with a NON-REPEATING prompt on THIS
+#   checkpoint — a repeated prompt hashes to cached blocks and never prefills deep.
+#
+# ⚠ Speculative decoding: dflash k=7 is the DEFAULT here, carried over from the sentinel
+#   measurement (dflash 2.40 vs MTP 2.18 mean acceptance length on a 27B body, +11%). But
+#   cyberprev is an ABLITERATED body, not an SFT one, and abliteration is exactly what can
+#   desync an MTP head — so the two are re-measured on THIS checkpoint before believing either.
+#   See the CYBER_SPEC_CONFIG note below.
+name: cyberprev-seat
+
+services:
+  vllm-cyberprev:
+    image: ${CYBER_IMAGE:-vllm/vllm-openai:nightly-e9d1398d9edfd90fcc1cf783805240e3effec013}
+    container_name: ${CYBER_CONTAINER_NAME:-vllm-cyberprev}
+    restart: unless-stopped
+    ipc: host
+    ports:
+      - "${CYBER_PORT:-8025}:8000"
+    volumes:
+      - /tank/aimodels/huggingface:/hfcache
+      - ${CYBER_MODEL:-/tank/aimodels/cyberprev-nvfp4-mixed}:/model:ro
+      # DFlash2 drafter, mounted unconditionally — inert if CYBER_SPEC_CONFIG selects an MTP
+      # method, which does not reference /drafter. Shared with mog-sec and (formerly)
+      # sentinel-r3: all three share the same 64-layer Qwen3.8-27B base and vocab, so one
+      # drafter serves them all. Costs +3.6 GB of VRAM when dflash is active.
+      - ${CYBER_DRAFT_MODEL:-/tank/aimodels/qwen38-27b-dflash2-drafter}:/drafter:ro
+    environment:
+      - VLLM_API_KEY=${API_KEY:-}
+      - PYTORCH_CUDA_ALLOC_CONF=${CYBER_ALLOC_CONF:-}
+    command:
+      - /model
+      # ⚠ Served under its OWN name. It does NOT inherit `sentinel-r3` — serving different
+      # weights under a retired model's name is silent substitution. The sentinel-r3 gateway
+      # aliases are left to 404 deliberately; a `cyberprev` alias is added instead.
+      - --served-model-name
+      - ${CYBER_SERVED_NAME:-cyberprev-27b}
+      - ${CYBER_SERVED_NAME_THINK:-cyberprev-27b-thinking}
+      - --host
+      - 0.0.0.0
+      - --port
+      - "8000"
+      - --quantization
+      - ${CYBER_QUANT:-compressed-tensors}
+      - --gpu-memory-utilization
+      - "${CYBER_GPU_MEM_UTIL:-0.40}"
+      # KV pinned in BYTES, not a ratio — a ratio silently yields a different cache depending
+      # on what else is resident on GPU 0 at start time (mog-sec is co-resident here).
+      - --kv-cache-memory
+      - "${CYBER_KV_CACHE_MEMORY:-8589934592}"
+      - --max-model-len
+      - "${CYBER_MAX_MODEL_LEN:-163840}"
+      - --max-num-seqs
+      - "${CYBER_MAX_NUM_SEQS:-16}"
+      - --max-num-batched-tokens
+      - "${CYBER_MAX_NUM_BATCHED_TOKENS:-4096}"
+      - --trust-remote-code
+      - --dtype
+      - auto
+      # Hybrid backbone: linear_attention + full_attention layers; the model's own config
+      # asks for float32 mamba/SSM state.
+      - --mamba-cache-dtype
+      - float32
+      - --kv-cache-dtype
+      - ${CYBER_KV_CACHE_DTYPE:-fp8}
+      - --enable-prefix-caching
+      - --enable-chunked-prefill
+      - --limit-mm-per-prompt
+      - '${CYBER_LIMIT_MM:-{"image": 4}}'
+      - --mm-processor-kwargs
+      - '${CYBER_MM_PROCESSOR_KWARGS:-{"size": {"longest_edge": 4194304, "shortest_edge": 65536}}}'
+      - --reasoning-parser
+      - ${CYBER_REASONING_PARSER:-qwen3}
+      - --default-chat-template-kwargs
+      - '{"reasoning_effort": "${CYBER_REASONING_EFFORT:-medium}"}'
+      - --enable-auto-tool-choice
+      - --tool-call-parser
+      - qwen3_coder
+      # ONE env var carrying the whole JSON, because the two speculative shapes are not
+      # interchangeable: dflash needs a "model" pointing at the drafter, MTP must NOT have
+      # one. A method+tokens template cannot express both.
+      #   DFlash2 : {"method": "dflash", "model": "/drafter", "num_speculative_tokens": 7}
+      #   MTP     : {"method": "qwen3_5_mtp", "num_speculative_tokens": 3}
+      # ⚠ Comparing the two requires MATCHING num_speculative_tokens, or you are comparing
+      # draft width as well as method — quant playbook §5.1: MTP runs a single-module head
+      # autoregressively, so deeper k improves acceptance and destroys throughput.
+      # ⚠ To turn speculation OFF there is no honest placeholder value — delete these two
+      # lines rather than passing an empty string.
+      - --speculative-config
+      - '${CYBER_SPEC_CONFIG:-{"method": "dflash", "model": "/drafter", "num_speculative_tokens": 7}}'
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              device_ids:
+                - "${CYBER_GPU_ID:-0}"
+              capabilities:
+                - gpu
+    healthcheck:
+      test: ["CMD", "curl", "-f", "http://localhost:8000/health"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      start_period: 900s
+    networks:
+      - tnet
+    labels:
+      - homepage.group=AI - Inference
+      - homepage.name=CyberPreview (abliterated sec)
+      - homepage.icon=mdi-shield-bug
+      - homepage.description=Qwen3.8-27B abliterated for cyber-offense refusals, sec-seat candidate (fv-ml1 GPU 0)
+      - homepage.href=http://10.251.50.54:${CYBER_PORT:-8025}/docs
+
+networks:
+  tnet:
+    name: traefik-net
+    external: true