feat(fv-ml1): rebalance GPUs 0/1 by usage; contexts to published maximums; gen-small default output cap

Layout: GPU 0 = cyberprev (KV 16 -> 15 GiB, 1.29x), erp-seat (moved from GPU 1, KV 8.49 -> 5.5 GiB, 1.32x), voices (ctx 8,192 -> 40,960, its native maximum, KV 6 GiB, 1.07x), coder (moved from GPU 1, ctx 16,384 -> 32,768, the card maximum, 2.35x), parakeet-nemo. GPU 1 = gen-small (moved from GPU 0, 5.5 GiB, 1.76x), meromero-rp (22 GiB, 1.32x), intern-decision. Sizing rule: KV >= 1.25x max context rounded up to 0.5 GiB, checked against 7 days of LiteLLM traffic and the engines' own KV-usage logs. Each card keeps spare memory for its bursty tenant (parakeet's cap on GPU 0, intern-decision's 32k call on GPU 1). YaRN extensions are not enabled: they do not fit, and the Qwen cards warn static YaRN degrades short prompts. gen-small gets --override-generation-config max_new_tokens 32768. It applies only to requests that omit max_tokens (vLLM get_max_tokens) and answers 9 runaway ~260k-token replies in 7 days; the longest normal reply was 6,357. Applied with playbooks/fv-ml1-rebalance-20261005.yaml (boot-check guard and health/KV verification per seat); all nine seats healthy, gateway smoke tests pass.
2026-10-05 22:51:17 -07:00
parent 6851ecbd0b
commit c35f6510bf
15 changed files with 319 additions and 25 deletions
@@ -31,13 +31,14 @@ CYBER_SERVED_NAME_THINK=cyberprev-27b-thinking
 # 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
+# 2026-10-05 rebalance: 15 GiB (~337k tokens, 1.29x at 262,144). Live max use seen: 57% of 16 GiB.
+CYBER_KV_CACHE_MEMORY=16106127360

 # ⚠ 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_MODEL_LEN=262144
 CYBER_MAX_NUM_SEQS=16
 CYBER_MAX_NUM_BATCHED_TOKENS=4096
 CYBER_KV_CACHE_DTYPE=fp8
@@ -4,12 +4,13 @@ ERP_MODEL=/tank/aimodels/G4-MeroMero-26B-A4B-it-uncensored-heretic-NVFP4A16
 ERP_SERVED_NAME=G4-MeroMero-26B-A4B-it-uncensored-heretic-NVFP4A16
 ERP_CHAT_TEMPLATE=/tank/aimodels/G4-MeroMero-26B-A4B-it-uncensored-heretic-NVFP4A16/chat_template.jinja
 ERP_PORT=8021
-ERP_GPU_ID=1
+# 2026-10-05 rebalance: GPU 0, KV pinned 5.5 GiB (~346k tokens, 1.32x); util = boot check only.
+ERP_GPU_ID=0
 # 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=0.30
+ERP_GPU_MEM_UTIL=0.25
 ERP_MAX_MODEL_LEN=262144
 ERP_MAX_NUM_SEQS=32
 API_KEY=
 # 8.49 GiB -> 534,649 KV tokens -> 2.04x a 262,144 context (operator's KV = 2x rule).
-ERP_KV_CACHE_MEMORY=9114000000
+ERP_KV_CACHE_MEMORY=5905580032
 ERP_MOE_BACKEND=auto
@@ -1,4 +1,4 @@
-# erp-seat — ERP-tune seat on ana-ml2 (GPU1, `:8021`)
+# erp-seat — ERP-tune seat on fv-ml1 (ex ana-ml2; GPU0 since 2026-10-05, `: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.
@@ -38,7 +38,7 @@ flip plus `docker compose up -d`.*
  json_schema` (guided decoding) 18/18 on this seat** — that is the deterministic path for a
  forced call; `tool_choice: auto` works normally. A parser plugin that re-enables guided JSON
  would also need JSON extraction in the engine-parser path; not attempted.
- **GPU1 is shared** — check real usage (`nvidia-smi --query-compute-apps=pid,used_memory`) before
+- **GPU0 is shared** (since 2026-10-05: cyberprev, voices, coder, parakeet-nemo; KV pinned 5.5 GiB = 346,436 tokens, 1.32×) — 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`.
@@ -151,7 +151,7 @@ services:
      - homepage.group=AI - Inference
      - homepage.name=char-rp-fast (MeroMero 26B-A4B, NVFP4A16 MoE)
      - homepage.icon=mdi-fire
-      - homepage.description=MeroMero A4B abliterated RP seat, NVFP4A16 weight-only, vision intact (fv-ml1 GPU1)
+      - homepage.description=MeroMero A4B abliterated RP seat, NVFP4A16 weight-only, vision intact (fv-ml1 GPU0)
      - homepage.href=http://10.251.50.54:${ERP_PORT:-8021}/docs

 networks:
@@ -1,12 +1,13 @@
-# gen-small-seat — Qwen3.6-35B-A3B Heretic (llmfan46), fv-ml1 GPU 0, :8026.
+# gen-small-seat — Qwen3.6-35B-A3B Heretic (llmfan46), fv-ml1 GPU 1 (since 2026-10-05), :8026.
 # Copy to .env on the host at /opt/docker/compose/gen-small-seat/.env.

 GEN_SMALL_IMAGE=vllm/vllm-openai:nightly-e9d1398d9edfd90fcc1cf783805240e3effec013
 API_KEY=

-# ── Placement ── shares GPU 0 with cyberprev (sec). KV pinned in bytes so the two
+# ── Placement ── shares its card with other seats. KV pinned in bytes so the two
 # co-resident seats don't fight over a util fraction.
-GEN_SMALL_GPU_ID=0
+# 2026-10-05 rebalance: moved to GPU 1 (with meromero-rp and intern-decision).
+GEN_SMALL_GPU_ID=1
 GEN_SMALL_PORT=8026
 GEN_SMALL_CONTAINER_NAME=vllm-gen-small

@@ -1,5 +1,5 @@
 # gen-small-seat — a fast, cheap "gen-small" seat: Qwen3.6-35B-A3B (3B active MoE),
-# Heretic-abliterated, on fv-ml1 GPU 0, :8026.
+# Heretic-abliterated, on fv-ml1 GPU 1, :8026.
 #
 # WHY: a tiny-active (A3B = ~3B active of 35B total) gen tier for high-volume / low-latency
 # work, complementing the big gen seat (flash-next :8022). Operator-directed 2026-09-14.
@@ -59,6 +59,13 @@ services:
      # attention footprint make this cheap; tuned to a sane concurrency band, not max.
      - --kv-cache-memory
      - "${GEN_SMALL_KV_CACHE_MEMORY:-5905580032}"
+      # Default output cap (2026-10-05, Prime). A request that OMITS max_tokens stops at 32,768
+      # instead of max_model_len - prompt; an explicit max_tokens still wins (vLLM get_max_tokens:
+      # request value first, generation-config max_new_tokens only as the fallback). Reason: 9
+      # runaway replies in 7 days, each ~260k tokens over 20+ min (albok-nemi, all-agents-local),
+      # while the longest normal reply was 6,357 tokens.
+      - --override-generation-config
+      - '{"max_new_tokens": 32768}'
      - --max-model-len
      - "${GEN_SMALL_MAX_MODEL_LEN:-262144}"
      - --max-num-seqs
@@ -92,6 +92,9 @@ service. This container may therefore use its rest (8,812 MiB) plus GPU 1's nvid
 > 3,981 MiB, about 2.4 GB short of a 32k call. Fixed 2131–2135 the same night: meromero and coder now pin KV in
 > bytes (`playbooks/fv-ml1-kv-pin-20261005.yaml`), and `Free` reads **11,801 MiB**, a 20,613 MiB budget, so
 > ~5.4 GB is spare at the 15,220 peak. The static seats can no longer drift across restarts.
+> **Later that night (2226–2249), the GPU 0/1 rebalance** (`playbooks/fv-ml1-rebalance-20261005.yaml`) left this
+> card with gen-small, meromero-rp and this service only. `Free` read 10,864 MiB right after, and should settle
+> near ~9.7 GB once gen-small's post-boot growth levels off. That is a ~18.5 GB budget, so ~3.3 GB spare at the 32k peak.

 Measured on the live service (per-process nvidia-smi every 0.1 s; single `noul` question unless noted; n=3, deterministic):

@@ -83,12 +83,13 @@ HF_TOKEN=
 # and the keyless zed-fim-proxy (stacks/zed-fim-proxy). util 0.06 (~5.7 GB) holds
 # the 1.5B fp16 + fp8 KV; 8192 ctx is ample for FIM (KV = 13.75x concurrency).
 CODER_PORT=8020
-CODER_GPU_ID=1
+# 2026-10-05 rebalance: GPU 0; context raised to the card-published maximum (32,768).
+CODER_GPU_ID=0
 CODER_MODEL=Qwen/Qwen2.5-Coder-1.5B
 CODER_SERVED_NAME=qwen2.5-coder-1.5b
-CODER_MAX_MODEL_LEN=8192
+CODER_MAX_MODEL_LEN=32768
 CODER_KV_CACHE_DTYPE=fp8
-CODER_GPU_MEM_UTIL=0.055
+CODER_GPU_MEM_UTIL=0.06
 # KV pinned 2026-10-05 (77,056 tokens; see compose.yaml). Util now only gates the boot check.
 CODER_KV_CACHE_MEMORY=1104674816
 CODER_MAX_NUM_SEQS=32
@@ -16,7 +16,7 @@ model on the box, currently three services:

 **Server:** ana-ml2
 **Ports:** `8001` (embed), `8002` (rerank), `8003` (reward) — all configurable via `.env`
-**GPU:** all three services share GPU 1 by default (configurable)
+**GPU:** only `vllm-coder` is live in this stack; since 2026-10-05 it runs on GPU 0 (`CODER_GPU_ID`) at the card-published maximum context of 32,768 (KV pinned 77,056 tokens, 2.35×). The embed/rerank/reward services here are retired

 ## Why one process per service

@@ -69,7 +69,13 @@ separately.

 ## Placement warnings

-⚠ **GPU 0 is now at 96.0 of 97.9 GB.** This seat's 10.7 GB went into the last real gap on
+> **2026-10-05 rebalance:** context raised 8,192 → **40,960** (Qwen3-4B's native maximum: the card's 32,768
+> plus its 8,192 prompt reserve; 131,072 would need YaRN, which the card says hurts short prompts). KV pinned
+> 6 GiB bf16 = 43,680 tokens (1.07×), util 0.16, resident ~15.4 GB. Still GPU 0, now beside cyberprev,
+> erp-seat, coder and parakeet-nemo. The LoRA voices were trained on short prose; long-context behaviour
+> with an adapter is unmeasured.
+
+⚠ *(historical, pre-2026-10-05)* **GPU 0 is now at 96.0 of 97.9 GB.** This seat's 10.7 GB went into the last real gap on
 fv-ml1: GPU 1 has ~5.7 GB, GPU 2 has ~2.4 GB, and GPU 3 is a held reserve for a future
 full-card seat (`flash-next` alone needs 93 of 96 GiB). **There is no room for another seat
 here without a placement decision.**
@@ -87,11 +87,11 @@ services:
      - --port
      - "8000"
      - --gpu-memory-utilization
-      - "${VOICES_GPU_MEM_UTIL:-0.12}"
+      - "${VOICES_GPU_MEM_UTIL:-0.16}"
      - --kv-cache-memory
-      - "${VOICES_KV_CACHE_MEMORY:-2147483648}"
+      - "${VOICES_KV_CACHE_MEMORY:-6442450944}"
      - --max-model-len
-      - "${VOICES_MAX_MODEL_LEN:-8192}"
+      - "${VOICES_MAX_MODEL_LEN:-40960}"
      - --max-num-seqs
      - "${VOICES_MAX_NUM_SEQS:-8}"
      - --dtype