feat(erp-seat): NVFP4A16 quant pipeline for the Gemma-4 26B-A4B MoE ERP tune + ana-ml2 GPU1 serve stack

- services/erp-seat-quant/quant_nvfp4a16_gemma4_moe.py: linearize_moe first (playbook §3.15),
  asserts the expert Linear count, routers/vision/audio/norms/lm_head ignored, W4A16 for RP
  long-session fidelity, post-steps restore processor configs + template and reset the
  tokenizer truncation cap (§3.14); --dry-run proves targets before GPU time
- services/erp-seat-quant/run_quant_erp_v6.sh: detached container on GPU1 (vllm-llmcompressor)
- stacks/erp-seat: serve recipe copied from gemma4-charrp, true served name only, port 8021

services/erp-seat-quant/quant_nvfp4a16_gemma4_moe.py

@@ -0,0 +1,145 @@
+#!/usr/bin/env python3
+"""NVFP4A16 (weight-only) quant of a Gemma-4 26B-A4B **MoE** checkpoint for vLLM (compressed-tensors).
+
+Built for the ERP-seat tunes (merged LoRA on Gemma-4-26B-A4B-it or its abliteration). Replicates
+the published `prithivMLmods/gemma-4-26B-A4B-it-NVFP4A16` recipe (targets=Linear, NVFP4A16,
+routers + vision + lm_head ignored) with the fleet's own calibration corpus and chat template.
+
+Playbook rules honoured (docs/pfi/model-quantization-playbook.md):
+  §3.15  fused 3-D MoE experts are INVISIBLE to targets=["Linear"] -> linearize_moe() FIRST,
+         then ASSERT the expert Linear count (layers x experts x 3) before any GPU time.
+  §3.15  routers stay BF16 (a 4-bit router picks different experts).
+  §3.5   vision/audio towers + projector ignored (BF16); processor configs restored after save.
+  §3.6   CPU-resident load (device_map=None); llm-compressor onloads one layer at a time.
+  §3.10  do NOT set PYTORCH_CUDA_ALLOC_CONF=expandable_segments (corrupts retained tensors).
+  §3.14  calibration bakes a truncation cap into tokenizer.json -> reset to null after save.
+  §1     W4A16 chosen over mixed W4A4: RP long-session fidelity > prefill speed (gate-judged seat).
+"""
+import argparse, json, os, re, shutil, sys, hashlib
+
+IGNORE = [
+    "lm_head",
+    "re:.*embed_tokens.*",
+    "re:.*embed_vision.*",
+    "re:.*vision_tower.*",
+    "re:.*audio_tower.*",
+    "re:.*audio.*",
+    "re:.*multi_modal_projector.*",
+    "re:.*mm_projector.*",
+    "re:.*patch_embedder.*",
+    "re:.*norm.*",
+    "re:.*router.*",          # MoE routers stay BF16 (§3.15)
+    "re:.*layer_scalar.*",
+]
+
+def _ignored(name):
+    for pat in IGNORE:
+        if pat.startswith("re:"):
+            if re.fullmatch(pat[3:], name): return True
+        elif name == pat or name.endswith("." + pat): return True
+    return False
+
+def enumerate_targets(model):
+    import torch.nn as nn
+    lin = [n for n, m in model.named_modules() if isinstance(m, nn.Linear)]
+    will = [n for n in lin if not _ignored(n)]
+    experts = [n for n in will if ".experts." in n]
+    skipped = [n for n in lin if _ignored(n)]
+    return lin, will, experts, skipped
+
+def build_calib(path, tok, seqlen, n):
+    from datasets import Dataset
+    rows = [json.loads(l) for l in open(path) if l.strip()][:n]
+    out = []
+    for r in rows:
+        msgs = []
+        for m in r.get("messages", []):
+            c = m.get("content")
+            if isinstance(c, list):
+                c = " ".join(p.get("text", "") for p in c if isinstance(p, dict))
+            if c: msgs.append({"role": m.get("role", "user"), "content": c})
+        if not msgs: continue
+        try:
+            text = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=False)
+        except Exception:
+            text = "\n".join(f"{m['role']}: {m['content']}" for m in msgs)
+        out.append(tok(text, truncation=True, max_length=seqlen))
+    return Dataset.from_list(out)
+
+def sha256(p):
+    h = hashlib.sha256()
+    with open(p, "rb") as f:
+        for chunk in iter(lambda: f.read(1 << 20), b""): h.update(chunk)
+    return h.hexdigest()
+
+def post_steps(src, out):
+    """§4.3-style post-steps for a Gemma-4 (no MTP head): processor configs, template, tokenizer cap, ignore check."""
+    for f in ("processor_config.json", "preprocessor_config.json", "video_preprocessor_config.json", "generation_config.json"):
+        s = os.path.join(src, f)
+        if os.path.exists(s) and not os.path.exists(os.path.join(out, f)):
+            shutil.copy(s, os.path.join(out, f)); print(f"[post] restored {f}")
+    pc = os.path.join(out, "processor_config.json"); pp = os.path.join(out, "preprocessor_config.json")
+    if os.path.exists(pc) and not os.path.exists(pp):
+        d = json.load(open(pc))
+        if "image_processor" in d:
+            json.dump(dict(d["image_processor"]), open(pp, "w"), indent=1); print("[post] materialized preprocessor_config.json from processor_config.image_processor")
+    # chat template: ship the SOURCE's (the one training/serving used), byte-identical
+    st = os.path.join(src, "chat_template.jinja"); ot = os.path.join(out, "chat_template.jinja")
+    shutil.copy(st, ot); print(f"[post] chat_template.jinja <- source, sha256 {sha256(ot)[:16]}")
+    # tokenizer truncation cap (§3.14)
+    tj = os.path.join(out, "tokenizer.json"); t = json.load(open(tj))
+    if t.get("truncation") is not None:
+        print(f"[post] ⚠ tokenizer.json had truncation={t['truncation']} baked in -> reset to null")
+        t["truncation"] = None; json.dump(t, open(tj, "w"), ensure_ascii=False)
+    else:
+        print("[post] tokenizer.json truncation: null (clean)")
+    # quantization_config ignore must still carry the routers + vision
+    cfg = json.load(open(os.path.join(out, "config.json"))); ig = cfg["quantization_config"].get("ignore", [])
+    has_router = any("router" in x for x in ig); has_vision = any("vision" in x for x in ig)
+    print(f"[post] quantization_config.ignore: {len(ig)} entries; routers={has_router} vision={has_vision}")
+    if not (has_router and has_vision):
+        print("[post] ⚠ REFUSING: ignore list lost routers or vision — llm-compressor pruned unmatched entries; investigate before serving", file=sys.stderr)
+        return 3
+    return 0
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--model", required=True); ap.add_argument("--out", required=True)
+    ap.add_argument("--calib", default="/tank/aimodels/heretic2-nvfp4-work/production_calib_512.jsonl")
+    ap.add_argument("--num-samples", type=int, default=256); ap.add_argument("--seqlen", type=int, default=8192)
+    ap.add_argument("--scheme", default="NVFP4A16")
+    ap.add_argument("--expect-experts", type=int, default=30 * 128 * 3, help="layers x experts x projections; assert before GPU time")
+    ap.add_argument("--dry-run", action="store_true", help="load + linearize + enumerate targets only (no GPU, no save)")
+    a = ap.parse_args()
+
+    from transformers import AutoTokenizer, Gemma4ForConditionalGeneration
+    print(f"[load] {a.model} (CPU-resident)", flush=True)
+    tok = AutoTokenizer.from_pretrained(a.model)
+    model = Gemma4ForConditionalGeneration.from_pretrained(a.model, torch_dtype="auto", device_map=None)
+    from llmcompressor.modeling.moe.linearize import linearize_moe
+    linearize_moe(model)
+    lin, will, experts, skipped = enumerate_targets(model)
+    print(f"[targets] Linear modules {len(lin)}  WILL quantize {len(will)}  (experts: {len(experts)})  ignored {len(skipped)}", flush=True)
+    print("[targets] ignored sample:", sorted({re.sub(r'\.\d+\.', '.N.', n) for n in skipped})[:20], flush=True)
+    print("[targets] quantized sample:", sorted({re.sub(r'\.\d+\.', '.N.', n) for n in will})[:12], flush=True)
+    if len(experts) != a.expect_experts:
+        print(f"[targets] ⚠ REFUSING: expert Linear count {len(experts)} != expected {a.expect_experts} (§3.15)", file=sys.stderr); return 2
+    if any("router" in n or "vision" in n or "audio" in n for n in will):
+        print("[targets] ⚠ REFUSING: a router/vision/audio Linear is in the quantize set", file=sys.stderr); return 2
+    if a.dry_run:
+        print("[dry-run] OK — targets proven; exiting before calibration"); return 0
+
+    print(f"[calib] <= {a.num_samples} samples @ seq {a.seqlen} from {a.calib}", flush=True)
+    calib = build_calib(a.calib, tok, a.seqlen, a.num_samples); print(f"[calib] {len(calib)} rows", flush=True)
+    from llmcompressor import oneshot
+    from llmcompressor.modifiers.quantization import QuantizationModifier
+    recipe = QuantizationModifier(targets="Linear", scheme=a.scheme, ignore=IGNORE)
+    print(f"[quant] oneshot scheme={a.scheme} targets=Linear (post-linearize)", flush=True)
+    oneshot(model=model, dataset=calib, recipe=recipe, num_calibration_samples=len(calib), max_seq_length=a.seqlen)
+    print(f"[save] -> {a.out}", flush=True)
+    model.save_pretrained(a.out, save_compressed=True); tok.save_pretrained(a.out)
+    rc = post_steps(a.model, a.out)
+    print("DONE" if rc == 0 else f"DONE WITH POST-STEP FAILURE rc={rc}", flush=True); return rc
+
+if __name__ == "__main__":
+    sys.exit(main())

services/erp-seat-quant/run_quant_erp_v6.sh

@@ -0,0 +1,18 @@
+#!/usr/bin/env bash
+# NVFP4A16 quant of the ERP run-6 merged model on ana-ml2 GPU1 (co-resident with the GPU1 seats;
+# CPU-resident load, per-layer onload). Detached container; watch with `docker logs -f erp-v6-quant`.
+# NOTE: no PYTORCH_CUDA_ALLOC_CONF=expandable_segments (playbook §3.10).
+set -euo pipefail
+WORK=/tank/aimodels/erp-tune-v6-quant-work
+SRC="${1:-/tank/aimodels/erp-tune-v6-bf16}"
+OUT="${2:-/tank/aimodels/erp-tune-v6-nvfp4a16}"
+MODE="${3:-full}"          # full | dry-run
+EXTRA=""; [ "$MODE" = "dry-run" ] && EXTRA="--dry-run"
+NAME=erp-v6-quant; [ "$MODE" = "dry-run" ] && NAME=erp-v6-quant-dry
+docker rm -f "$NAME" 2>/dev/null || true
+docker run -d --name "$NAME" --gpus '"device=1"' --ipc host \
+  -v /tank/aimodels:/tank/aimodels \
+  --entrypoint python3 vllm-llmcompressor:latest \
+  "$WORK/quant_nvfp4a16_gemma4_moe.py" --model "$SRC" --out "$OUT" \
+  --num-samples "${NUM_SAMPLES:-256}" --seqlen "${SEQLEN:-8192}" $EXTRA
+echo "launched $NAME: $(docker ps --filter name=$NAME --format '{{.Status}}')"

stacks/erp-seat/.env.example

@@ -0,0 +1,12 @@
+# erp-seat — ana-ml2 GPU1. Real .env lives on the host at /opt/docker/compose/erp-seat/.env.
+ERP_IMAGE=vllm/vllm-openai:v0.26.0
+ERP_MODEL=/tank/aimodels/erp-tune-v6-nvfp4a16
+ERP_SERVED_NAME=erp-tune-v6-nvfp4a16
+ERP_CHAT_TEMPLATE=/tank/aimodels/erp-tune-v6-nvfp4a16/chat_template.jinja
+ERP_PORT=8021
+ERP_GPU_ID=1
+# 0.35 x 97.9 GiB = 34 GiB. GPU1 had ~47 GiB free on 2026-09-08 (scriberr/embed/rerank/coder/reward resident).
+ERP_GPU_MEM_UTIL=0.35
+ERP_MAX_MODEL_LEN=32768
+ERP_MAX_NUM_SEQS=8
+API_KEY=

stacks/erp-seat/README.md

@@ -0,0 +1,18 @@
+# erp-seat — ERP-tune seat on ana-ml2 (GPU1, `:8021`)
+
+Serves the latest gated ERP LoRA merge as an **NVFP4A16** (weight-only) compressed-tensors
+checkpoint so the GX10 is free to train the next run. First occupant: **run 6** —
+`erp-tune-v6-nvfp4a16` = merged-run06 (jenerallee78 ARA-abliterated Gemma-4-26B-A4B-it, index
+`33c59654…`, + R47 SFT r6) quantized by `services/erp-seat-quant/`.
+
+- **True name only.** `--served-model-name erp-tune-v6-nvfp4a16`. Gateway aliases (`trial`) are
+  set in LiteLLM on the operator's word, never here (no silent substitution — the bf16 arm on the
+  GX10 and this NVFP4 arm are different artifacts).
+- **Recipe** = `stacks/gemma4-charrp` (same arch + format, proven on this box): `gemma4` tool
+  and reasoning parsers, `enable_thinking` pinned false, the model's own stock template
+  (`ae53464b…`, the one it trained through). Without the reasoning parser the post-tool turn leaks
+  `<|channel>` markers; without the kwargs pin all prose lands in `reasoning_content`.
+- **GPU1 is shared** — check real usage (`nvidia-smi --query-compute-apps=pid,used_memory`) before
+  raising `ERP_GPU_MEM_UTIL`; the flag sizes KV, not CUDA context.
+- **Rollback / next run:** point `ERP_MODEL` + `ERP_SERVED_NAME` at the next quant dir, keep the
+  previous on disk. Deploy with `scripts/deploy-stack.sh ana-ml2 erp-seat`.

stacks/erp-seat/compose.yaml

@@ -0,0 +1,76 @@
+# erp-seat — the ERP-tune seat on ana-ml2 GPU1: NVFP4A16 quant of the latest gated ERP LoRA merge
+# (run 6 = jenerallee78 ARA-abliterated Gemma-4-26B-A4B + R47 SFT), served under its TRUE name.
+# Routing aliases (e.g. LiteLLM `trial`) are the operator's call and live in the gateway, not here.
+#
+# Serve recipe copied from stacks/gemma4-charrp (same architecture + quant format, proven on this
+# box): gemma4 tool + reasoning parsers, enable_thinking pinned false, model's own stock template.
+# GPU1 is SHARED (charrp-MoE moved? no — scriberr, embed, rerank, coder, reward live there):
+# ~47 GiB was free on 2026-09-08; 0.35 x 97.9 GiB = 34 GiB keeps ~13 GiB of real margin.
+# Quant pipeline: services/erp-seat-quant/. Tunables in .env.
+
+name: erp-seat
+
+services:
+  vllm-erp-seat:
+    image: ${ERP_IMAGE:-vllm/vllm-openai:v0.26.0}
+    container_name: ${ERP_CONTAINER:-vllm-erp-seat}
+    restart: unless-stopped
+    ipc: host
+    ports:
+      - "${ERP_PORT:-8021}:8000"
+    volumes:
+      - /tank/aimodels:/tank/aimodels
+    environment:
+      - VLLM_API_KEY=${API_KEY:-}
+    command:
+      - ${ERP_MODEL:-/tank/aimodels/erp-tune-v6-nvfp4a16}
+      - --quantization
+      - compressed-tensors
+      - --served-model-name
+      - ${ERP_SERVED_NAME:-erp-tune-v6-nvfp4a16}
+      - --tool-call-parser
+      - gemma4
+      - --enable-auto-tool-choice
+      - --reasoning-parser
+      - gemma4
+      - --default-chat-template-kwargs
+      - '{"enable_thinking": false}'
+      - --chat-template
+      - ${ERP_CHAT_TEMPLATE:-/tank/aimodels/erp-tune-v6-nvfp4a16/chat_template.jinja}
+      - --max-model-len
+      - "${ERP_MAX_MODEL_LEN:-32768}"
+      - --max-num-seqs
+      - "${ERP_MAX_NUM_SEQS:-8}"
+      - --gpu-memory-utilization
+      - "${ERP_GPU_MEM_UTIL:-0.35}"
+      - --kv-cache-dtype
+      - fp8
+      - --trust-remote-code
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              device_ids:
+                - "${ERP_GPU_ID:-1}"
+              capabilities:
+                - gpu
+    healthcheck:
+      test: ["CMD", "curl", "-f", "http://localhost:8000/health"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      start_period: 600s
+    networks:
+      - tnet
+    labels:
+      - homepage.group=AI - Inference
+      - homepage.name=erp-tune-v6 (Gemma-4 26B-A4B ARA, NVFP4A16)
+      - homepage.icon=mdi-fire
+      - homepage.description=ERP-seat run-6 LoRA merge on the jenerallee78 abliteration, NVFP4A16 MoE (ana-ml2 GPU1)
+      - homepage.href=http://10.250.50.54:${ERP_PORT:-8021}/docs
+
+networks:
+  tnet:
+    name: traefik-net
+    external: true