scriberr to fv-ml1 GPU 3 (on-demand, steps aside to irv-ml1 A6000); intern-decision 32k-token calls (cap 14.4 GiB)

Prime 2026-09-30: move scriberr to GPU 3 and extend the Jev endpoint to 32k tokens. Scriberr holds 0 VRAM idle; verified a 20-min job on GPU 3 at 5,496 MiB. With GPU 1 freed, intern-decision's measured card peak at MAX_TOKENS=32768 is 15,220 MiB against a 15,437 MiB budget (n=3, 1 and 16 questions); 32,769 tokens is refused 422 up front. JevBench v1.2.16 via /v1/systemone unchanged: 202/231, 0 diffs vs the bench.
2026-09-30 13:35:32 -07:00
parent 92501a29c1
commit 6b201e1d4a
8 changed files with 79 additions and 29 deletions
@@ -1,18 +1,19 @@
 # intern-decision — copy to /opt/docker/compose/intern-decision/.env on fv-ml1 (mode 0600).
 # Built on fv-ml1 from services/intern-decision-serve (see README "Building").
-IMAGE=intern-decision-serve:0.1.0
+IMAGE=intern-decision-serve:0.1.2
 PORT=8033
 HOST_IP=10.251.50.54
-# fv-ml1 GPU 1 = the utility card (vllm-coder, erp, meromero, scriberr).
+# fv-ml1 GPU 1 = the utility card (vllm-coder, erp, meromero). scriberr moved to GPU 3 (2026-09-30).
 GPU_ID=1
 # HARD torch-allocator cap: the single knob that holds the container's WHOLE nvidia-smi footprint
-# (CUDA context included) inside GPU 1's budget next to scriberr (infra-ops, 2026-09-30):
-#   GPU 1 nvidia-smi Free >= 15,400 MiB  =  our card peak 9,876 (cap 9.0 GiB + 660 MiB outside the
-#   allocator, measured) + scriberr's peak 5,496, rounded up. README "VRAM".
-VRAM_CAP_GIB=9.0
+# (CUDA context included, ~660 MiB outside the allocator) inside GPU 1's free memory (infra-ops,
+# 2026-09-30 1330): with the vLLM seats static, this container may use its rest (8,812) + GPU 1
+# nvidia-smi Free (6,625) = 15,437 MiB. 14.4 GiB cap -> card ceiling ~15,408. README "VRAM".
+VRAM_CAP_GIB=14.4
 # Tokens per CALL (state + up to 16 questions), checked BEFORE the forward pass: a longer call is a
-# clear 422. 7,168 is the largest call measured to fit under VRAM_CAP_GIB=9.0. Change the two TOGETHER,
-# and re-measure (README "VRAM"): a larger value would let a call reach the cap and return 503.
-MAX_TOKENS=7168
+# clear 422. 32,768 = Jev's "32k for state plus the longest question"; measured to fit under 14.4 GiB
+# at a card peak of 15,220 MiB (1 question and 16 questions, n=3 each). Change the two TOGETHER, and
+# re-measure (README "VRAM"): a larger value would let a call reach the cap and return 503.
+MAX_TOKENS=32768
 # >= 32 characters; source of truth: secret get intern-decision/api-token
 INTERN_DECISION_API_TOKEN=
@@ -80,7 +80,36 @@ curl -s -H "Authorization: Bearer $T" http://intern-decision.fv.internal:8033/de
     never truncated.
   - A request with more decisions is split into more calls, and each call must fit.

-## VRAM: fits beside scriberr's peak, whatever the request
+## VRAM: 32k-token calls since 2026-09-30 1330 (scriberr moved to GPU 3)
+
+**Current setting: `VRAM_CAP_GIB=14.4`, `MAX_TOKENS=32768`** (Prime: move scriberr to GPU 3 and "extend the jev
+endpoint to hit 32k tokens if possible"). With scriberr gone, GPU 1 holds only the static vLLM seats and this
+service. This container may therefore use its rest (8,812 MiB) plus GPU 1's nvidia-smi `Free` (6,625 MiB), which is
+**15,437 MiB**. The 14.4 GiB cap plus the ~660 MiB outside the allocator puts the card ceiling at ~15,408 MiB.
+
+Measured on the live service (per-process nvidia-smi every 0.1 s; single `noul` question unless noted; n=3, deterministic):
+
+| call tokens | card peak MiB | wall (warm) |
+|---|---|---|
+| 3,187 | 9,306 | 0.15 s |
+| 12,187 | 11,206 | 0.68 s |
+| 24,187 | 13,552 | 1.49 s |
+| 29,987 | 14,692 | 1.92 s |
+| **32,768 (limit)** | **15,220** | 2.12 s |
+| 32,765, 16 questions | 15,220 | 2.15 s |
+| 32,769 | refused 422 before the forward | 0.30 s |
+
+Spare at the limit: 217 MiB. JevBench v1.2.16 through `/v1/systemone` after the change: 202/231, with 0 answer and 0
+probability diffs against the bench's r1..r4. `/decide` is unchanged.
+
+⚠ **Cold-shape latency:** the first call in a new length bucket after a (re)start costs ~6.5 s extra
+(3,001 and 4,000 words were slow; 3,002–3,500 and 4,097–5,000 were not). The fast kernels autotune per shape
+bucket, and the result is cached in-process. Warm calls are as tabled. A startup warm-up across the buckets would
+remove it; that is not done yet.
+
+The table below is the ORIGINAL budget (9.0 GiB / 7,168 tokens, beside scriberr) and is kept for history.
+
+## VRAM (history): fits beside scriberr's peak, whatever the request

 **Budget (infra-ops, 2026-09-30):** GPU 1 needs nvidia-smi `Free` ≥ **15,400 MiB** before this
 service starts. That is our card peak of 9,876 MiB plus scriberr's peak of 5,496 MiB (with its
@@ -1,5 +1,6 @@
 # intern-decision: Intern-Decision-4B (internlm, Apache-2.0) behind intern-decision-serve, on
-# fv-ml1 GPU 1 (the utility card, beside vllm-coder, the erp/meromero seats and scriberr).
+# fv-ml1 GPU 1 (the utility card, beside vllm-coder and the erp/meromero seats; scriberr moved to
+# GPU 3 on 2026-09-30 1322, Prime, to free this card's headroom for 32k-token calls).
 # Replaces semif (Prime, 2026-09-30: "replace semif with intern-decision now").
 #
 # One forward pass per call, scored by the checkpoint's OWN inference.py (sha256-pinned); the
@@ -8,8 +9,8 @@
 # fv-ml1 from that dir.
 #
 # ⚠ VRAM_CAP_GIB is a HARD cap on torch's allocator (per-process memory fraction), set so the
-#   container's WHOLE nvidia-smi footprint, CUDA context included, fits beside scriberr's peak
-#   (infra-ops budget, 2026-09-30); MAX_TOKENS keeps every accepted call under the cap. A request that needs more
+#   container's WHOLE nvidia-smi footprint, CUDA context included, fits GPU 1's free memory beside
+#   the static vLLM seats (infra-ops budget, 2026-09-30); MAX_TOKENS keeps every accepted call under the cap. A request that needs more
 #   gets 503 out_of_memory and the service stays up. See the README before changing it.
 #
 # .env (tunables): IMAGE, PORT, GPU_ID, VRAM_CAP_GIB, MAX_TOKENS, HOST_IP, INTERN_DECISION_API_TOKEN