frank-primary: 5 x full 262144 contexts at fp8 (final shape)

Frank's call after the night ladder: five contexts at the model's full native
262,144 window beat six shortened ones. Six full-length contexts do NOT fit at
any KV dtype (6 x 241664 fp8 asked 10.70 GiB against 9.18 GiB free — the
6-seq graph/activation footprint leaves too little for the pool); 5 x 262144
at fp8 PASSED on the first rung: KV_BYTES 10,388,766,720/rank, measured
'GPU KV cache size: 1,350,182 tokens, Maximum concurrency for 262,144 tokens
per request: 5.15x' (39,462 tokens of margin).

Also documented: the SEQS=3 bf16 max (241,664, KV 732,359 / 3.03x) measured
earlier the same night, and the root-cause read that the 7.2 GiB KV ceiling
law was the MTP4 workspace, not VRAM. Seam and harness (contextWindow 262144)
re-pinned.

AGENTS.md

@@ -24,7 +24,7 @@ merge absorbed the former `vllmScripts` repo, history preserved).
 
 | Target | Service | Runtime | Model | Port |
 |--------|---------|---------|-------|------|
-| Dankserver | `frank-primary` | tcclaviger RDNA4 vLLM fork (`docker.io/tcclaviger/vllm@sha256:da4029666…`, digest-pinned, plain `podman run`, root systemd unit) | `orcarouter/Qwen3.8-Flash-Next-Uncensored-FP8` (~172.8 GiB, 131 shards), TP=2, expert offload automatic (fork plans it), prefetch OFF, **bf16 KV** (`--kv-cache-dtype auto`), **no speculative decode**, MNBT 1024, MoE gather 32×64, **241664 ctx × 3 seqs** at KV_BYTES 10,184,000,000/rank (measured pool 732,359 tokens / 3.03× concurrency), util 0.98, and the client contract wired: `--reasoning-parser qwen3` (thinking in `reasoning`, never inline), `--enable-auto-tool-choice --tool-call-parser qwen3_xml`, default chat-template kwargs `enable_thinking` + `preserve_thinking`. 241,664 is **max possible** on this box (Frank's subtract-16k ladder: 262144/245760 OOM at graph capture, 237568 passed, +4096 bonus step won). fp8+MTP4 shape restorable from git tag `frank-primary-fp8-mtp-20261001` | 7770 |
+| Dankserver | `frank-primary` | tcclaviger RDNA4 vLLM fork (`docker.io/tcclaviger/vllm@sha256:da4029666…`, digest-pinned, plain `podman run`, root systemd unit) | `orcarouter/Qwen3.8-Flash-Next-Uncensored-FP8` (~172.8 GiB, 131 shards), TP=2, expert offload automatic (fork plans it), prefetch OFF, **fp8 KV** (`--kv-cache-dtype fp8`), **no speculative decode**, MNBT 1024, MoE gather 32×64, **5 seqs × 262144 ctx (the model's full native window)** at KV_BYTES 10,388,766,720/rank (measured pool 1,350,182 tokens / 5.15× concurrency), util 0.98, and the client contract wired: `--reasoning-parser qwen3` (thinking in `reasoning`, never inline), `--enable-auto-tool-choice --tool-call-parser qwen3_xml`, default chat-template kwargs `enable_thinking` + `preserve_thinking`. Six full-length contexts do **not** fit (6×241664 asked 10.70 GiB vs 9.18 GiB free); 5×262144 was the first rung that passed (Frank's subtract-16k ladder). fp8+MTP4 shape restorable from git tag `frank-primary-fp8-mtp-20261001` | 7770 |
 | Dankclaw | `frank-large` | gufo `gufo-org/gufo` v0.1.1, the Strix Halo engine, BUILT FROM SOURCE (root systemd unit; links the host ROCm 7.2.4) | Swift-1.5-Qwen3.8-27B-Uncensored UD-Q8_K_XL (27B dense, ABLITERATED/uncensored, 29.3 GiB, SWIFT OPEN LICENSE 1.0) with the z-lab DFlash2 Q4_K_M drafter (adaptive block policy, cap 7) on the gfx1151 iGPU (Strix Halo LPDDR, 114 GiB GTT), 262144 ctx single session, greedy, gufo's compiled-in Qwen renderer (NO Qwen-Sharp), text only | 7770 |
 | Dankclaw | `frank-small` | mainline llama.cpp master 37b53fd (root systemd unit) | HauhauCS Gemma4-12B-QAT-Uncensored-Balanced Q4_K_M (BF16 mmproj + MTP assistant sidecar, `draft-mtp,ngram-mod`) on the RTX 3090, 1048576 pool / 4 slots of 262,144, q8_0 KV, the fleet's fastest vision endpoint | 7771 |
 

scripts/frank-primary.env

@@ -60,14 +60,19 @@ FRANK_PRIMARY_PUBLISH=7770:7770
 FRANK_PRIMARY_LEGACY_CONTAINERS="frank-primary-r9v qwen-fp8"
 
 # --- qualified engine shape (2026-10-01 night ladder, Frank's method) ---------
-# MAXLEN ladder per Frank: 262144 OOMed in graph capture (76 MiB short),
+# SEQS 5 x MAXLEN 262144 at fp8 KV is the DEPLOYED shape (Frank's call: five
-# 245760 OOMed (0 bytes free at graph capture), 237568 PASSED (718,606
+# contexts of the model's FULL native window beats six shortened ones).
-# tokens / 3.0x), then the +4096 bonus attempt 241664 PASSED and is
+# Measured: "GPU KV cache size: 1,350,182 tokens, Maximum concurrency for
-# PRODUCTION (KV 732,359 tokens / 3.03x at 241,664-token requests).
+# 262,144 tokens per request: 5.15x" — 39,462 tokens of margin over the
-FRANK_PRIMARY_MAXLEN=241664
+# 1,310,720-token contract.
-FRANK_PRIMARY_SEQS=3
+# The night's ladder, both contracts: at SEQS=3 bf16, 262144/245760 OOMed at
-FRANK_PRIMARY_KV_BYTES=10184000000
+# graph capture and 241664 was max possible (KV 732,359 / 3.03x). At SEQS=6
-FRANK_PRIMARY_KV_DTYPE=auto
+# fp8, 6 x 241664 OOMed (tried 10.70 GiB with 9.18 GiB free) — six
+# full-length contexts do NOT fit on this box; five do.
+FRANK_PRIMARY_MAXLEN=262144
+FRANK_PRIMARY_SEQS=5
+FRANK_PRIMARY_KV_BYTES=10388766720
+FRANK_PRIMARY_KV_DTYPE=fp8
 FRANK_PRIMARY_MNBT=1024
 FRANK_PRIMARY_GATHER_CHUNKS=32
 FRANK_PRIMARY_GATHER_LANES=64

serverAppConfigs-docs/LLM/dankserver-qwen38-fp8-expert-offload-tuning.md

@@ -275,39 +275,46 @@ Depth caveat: every number above is at a near-empty KV. Deep-context
 behavior (the gather win growing with depth, bf16's decode cost at depth)
 is only sampled at 3 depths; a 64k-192k-depth probe is the flagged gap.
 
-### The bf16 no-spec incumbent (production since 2026-10-01)
+### The deployed shape (2026-10-01: 5 x 262144, fp8, no spec)
 
 ```
-FRANK_PRIMARY_MAXLEN=241664    x SEQS 3 (724,992 tokens of contract)
+FRANK_PRIMARY_MAXLEN=262144    x SEQS 5 (1,310,720 tokens of contract)
-FRANK_PRIMARY_KV_BYTES=10184000000  PER RANK (bf16, pool scales with window)
+FRANK_PRIMARY_KV_BYTES=10388766720  PER RANK (fp8, pool scales with windows)
-FRANK_PRIMARY_KV_DTYPE=auto   --kv-cache-dtype auto (bf16, unquantized)
+FRANK_PRIMARY_KV_DTYPE=fp8    --kv-cache-dtype fp8 (K and V together)
 FRANK_PRIMARY_MNBT=1024       --max-num-batched-tokens 1024
 R4D_LRU_GATHER_CHUNKS=32 / R4D_LRU_GATHER_LANES=64  (container env)
 NO --speculative-config, NO --expert-prefetch
 ```
 
-Boot-probed 2026-10-01 (transient systemd units, full-log capture): 262144
+Boot-probed 2026-10-01 (transient systemd units at the production uid +
-and 245760 both died at graph capture (`Tried to allocate 76.00 MiB ...
+memlock, full-log capture; the rootless bench probe had masked the production
-0 bytes free` — the pool+profile+graph budget exceeds by a hair at util
+memory environment). The night searched two concurrency contracts:
-0.98); 237568 passed (KV 718,606 tokens) and the +4096 bonus attempt
-**241664 PASSED and is production: KV 732,359 tokens / Maximum concurrency
-for 241,664 tokens per request: 3.03x** — all three full windows fit with
-a 7,367-token margin. Decode measured at the 262k-era shape before the
-ladder: 39.38/38.79/36.42 tok/s (166/1087/4064 ctx).
-The old "KV_BYTES is PER RANK, 9/16 GiB OOM" law was the MTP4 WORKSPACE
-reservation, not a VRAM wall — no-spec bf16 takes ~9.5-10.2 GB/rank and
-boots clean. The failed 11.3 GB attempt OOMed at the engine's free-memory
-check (`RuntimeError: Worker failed with error 'CUDA out of memory'`,
-captured from the journald json stream — the fork's giant ERROR lines are
-split by journald, so grep the `-o json` output, not `-o cat`).
 
-Rollback to the fp8+MTP4 shape: git tag `frank-primary-fp8-mtp-20261001`
+| rung | result |
-(env trio + seam re-pinned there). The bf16 trade is deliberate: ~14%
+|---|---|
-decode cost at 4064 ctx vs the fp8+gather arm (49.25 -> 36.42) bought
+| 3 × 262144, bf16, KV 11,047,010,304 | **OOM at graph capture** — `Tried to allocate 76.00 MiB ... 0 bytes free`, 30.43 GiB of 31.86 GiB already allocated by PyTorch |
-clean cache + the largest window this box boots. The model's native
+| 3 × 245760, bf16 | **OOM identically** |
-262,144 exceeds what graph capture grants at util 0.98 here: the ladder
+| 3 × 237568, bf16 | PASS — KV 718,606 tokens |
-measured 262144/245760 OOM and 241,664 as max possible (Frank's
+| **3 × 241664, bf16** | **PASS — KV 732,359 tokens, 3.03x** (max possible at 3 contexts) |
-subtract-16k method, then one +4k bonus step).
+| 6 × 241664, fp8 | **OOM** — tried 10.70 GiB against 9.18 GiB free: six full-length contexts do NOT fit |
+| **5 × 262144, fp8** | **PASS on the first rung — KV 1,350,182 tokens, 5.15x; DEPLOYED** (39,462-token margin over the 1,310,720-token contract) |
+
+The final shape is Frank's call: **five contexts of the model's full native
+262,144 window** beat six shortened ones. fp8 KV is what buys the capacity —
+7,926 B/token against bf16's ~13,900-14,047 — at the cost of bf16's ~4 %
+decode edge. Six full-length contexts cannot fit at any KV dtype: the 6-seq
+graph/activation footprint leaves only 9.18 GiB for the pool, and 6 × 262144
+needs 12.47 GB even at fp8.
+
+The old "KV_BYTES is PER RANK, 9/16 GiB OOM" law was the MTP4 WORKSPACE
+reservation, not a VRAM wall. Failures are clean and diagnostic: the engine
+reports `torch.OutOfMemoryError: Tried to allocate X GiB. GPU N has ... of
+which Y GiB is free`, and the fork's giant ERROR lines are split by journald —
+read them via `journalctl -o json`, not `-o cat` (and `podman logs`, which is
+only available while the container lives).
+
+Rollback to the 2026-09-29 fp8+MTP4 shape: git tag
+`frank-primary-fp8-mtp-20261001` (env trio + seam re-pinned there).
 
 ## See Also
 

serverAppConfigs-docs/LLM/index.md

@@ -4,9 +4,9 @@ Documentation for the LLM inference fleet on the two hosts (2026-10-01):
 **Dankserver runs the tcclaviger RDNA4 vLLM fork on frank-primary** (the
 digest-pinned `tcclaviger/vllm` image driving `podman run` directly, serving
 orcarouter/Qwen3.8-Flash-Next-Uncensored-FP8 across both Radeon AI PRO R9700s
-on :7770, TP=2 with automatic tiered-expert offload, 241664 ctx x 3 seqs
+on :7770, TP=2 with automatic tiered-expert offload, **5 seqs x 262144 ctx
-(max possible — 262144 does not fit at a 3-seq contract on this box),
+(the model's full native window)**, fp8 KV (6 full-length contexts do not
-bf16 KV, no speculative decode, MNBT 1024, MoE gather 32x64, util 0.98,
+fit), no speculative decode, MNBT 1024, MoE gather 32x64, util 0.98,
 reasoning/tool parsers wired (qwen3 / qwen3_xml) with preserve_thinking on,
 measured 36.4-39.4 tok/s decode — the 2026-10-01 bf16 no-spec cutover plus
 the night ladder; rollback frank-primary-fp8-mtp-20261001), and Dankclaw
@@ -32,8 +32,8 @@ gfx1201, 32 GiB each]
         B2[tcclaviger RDNA4 vLLM fork, digest-pinned
 podman run, frank-primary, port 7770]
         C[Qwen3.8-Flash-Next FP8 bf16 no-spec
-TP=2, expert offload automatic, 241664 ctx x 3,
+TP=2, expert offload automatic, 5 x 262144 ctx,
-bf16 KV, MNBT 1024, gather 32x64, qwen3 parsers]
+fp8 KV, MNBT 1024, gather 32x64, qwen3 parsers]
     end
 
     subgraph Dankclaw[AMD Strix Halo - Ubuntu 26.04]

serverAppConfigs-docs/log.md

@@ -709,8 +709,17 @@
   --tool-call-parser qwen3_xml`, `--default-chat-template-kwargs
   '{"enable_thinking": true, "preserve_thinking": true}'`. The seam's
   "NO tool dialect" test is flipped to a PRESENT test.
-- Harness `/home/blank/.omp/agent/models.yml` — contextWindow 262144 ->
+- **Final shape (Frank's call): 5 seqs x the full native 262144 window at
-  241664 (compaction gap recomputed: 169,164 + 36,864 <= 241,664) and
+  fp8 KV.** 6 x 241664 at fp8 died (`Tried to allocate 10.70 GiB`, 9.18 GiB
+  free) — six full-length contexts do not fit at any KV dtype — so the
+  contract moved to five FULL windows: 5 x 262144 fp8 PASSED on the first
+  rung, KV_BYTES 10,388,766,720/rank, measured **1,350,182 tokens /
+  5.15x concurrency** (39,462 tokens of margin). Seam re-pinned; the bf16
+  three-context shape (241,664, KV 732,359 / 3.03x, measured earlier the
+  same night) is the documented max at 3 contexts.
+- Harness `/home/blank/.omp/agent/models.yml` — contextWindow back to 262144
+  with the five-window fp8 shape (compaction gap 183,501 + 36,864 = 220,365
+  <= 262,144) and
   `preserve_thinking` moved from a bare top-level body field to
   `chat_template_kwargs` (the form vLLM actually reads).
 - Production bring-up: the winner (241,664) boots and serves on :7770 —

tests/frank_primary_unit.bats

@@ -75,20 +75,20 @@ DOWNLOAD="$REPO_ROOT/scripts/download-frank-primary.sh"
 }
 
 @test "the qualified engine shape is in the env file" {
-    # 2026-10-01 night ladder (Frank's method: subtract until it boots, then
+    # 2026-10-01 night ladder (Frank's method). Two contracts were searched:
-    # one bonus step): 262144 OOMed at graph capture (76 MiB short, 0 bytes
+    # SEQS=3 bf16 topped out at 241664 (262144/245760 OOMed at graph capture;
-    # free), 245760 OOMed identically, 237568 PASSED (KV 718,606 tokens /
+    # KV 732,359 tokens / 3.03x). SEQS=6 fp8 did NOT fit — 6 x 241664 tried
-    # 3.0x), and the +4096 bonus attempt 241664 PASSED and IS production
+    # 10.70 GiB against 9.18 GiB free, so six full-length contexts are
-    # (KV 732,359 tokens / 3.03x at 241,664-token requests, 7,367-token
+    # impossible here. SEQS=5 fp8 at the model's FULL native 262144 window
-    # margin over the 3-window contract). UTIL stays 0.98 — the production
+    # PASSED on the first rung and is DEPLOYED: KV 1,350,182 tokens,
-    # unit's pinned-expert path needs the workspace the bench probe hid;
+    # Maximum concurrency 5.15x, 39,462 tokens of margin over the
-    # 0.98 with the smaller window fits. KV_BYTES 10,184,000,000 scales the
+    # 1,310,720-token contract. UTIL stays 0.98 (lowering it shrinks the
-    # pool with the window.
+    # engine's own budget instead of buying slack).
-    grep -q '^FRANK_PRIMARY_MAXLEN=241664$' "$ENV_FILE"
+    grep -q '^FRANK_PRIMARY_MAXLEN=262144$' "$ENV_FILE"
-    grep -q '^FRANK_PRIMARY_SEQS=3$' "$ENV_FILE"
+    grep -q '^FRANK_PRIMARY_SEQS=5$' "$ENV_FILE"
     grep -q '^FRANK_PRIMARY_UTIL=0.98$' "$ENV_FILE"
-    grep -q '^FRANK_PRIMARY_KV_BYTES=10184000000$' "$ENV_FILE"
+    grep -q '^FRANK_PRIMARY_KV_BYTES=10388766720$' "$ENV_FILE"
-    grep -q '^FRANK_PRIMARY_KV_DTYPE=auto$' "$ENV_FILE"
+    grep -q '^FRANK_PRIMARY_KV_DTYPE=fp8$' "$ENV_FILE"
     grep -q '^FRANK_PRIMARY_MNBT=1024$' "$ENV_FILE"
     grep -q '^FRANK_PRIMARY_GATHER_CHUNKS=32$' "$ENV_FILE"
     grep -q '^FRANK_PRIMARY_GATHER_LANES=64$' "$ENV_FILE"
@@ -151,13 +151,14 @@ DOWNLOAD="$REPO_ROOT/scripts/download-frank-primary.sh"
     ! grep -q -- '--speculative-config' "$SERVE"
 }
 
-@test "bf16 KV is PRESENT — clean cache won the trade" {
+@test "fp8 KV is PRESENT — capacity won over the bf16 decode edge" {
-    # --kv-cache-dtype auto = bf16 attention KV. Doubles attention-KV
+    # --kv-cache-dtype fp8 (the stock flag quantizes K and V together).
-    # bytes/token (measured 14,047 vs 7,926) and costs ~14% decode vs fp8
+    # Capacity is what buys the five full 262,144-token contexts: fp8 costs
-    # at 4064 ctx (49.25 -> 36.42 with the gather win) — traded deliberately
+    # 7,926 B/token against bf16's ~13,900-14,047, so 5x262144 fits in
-    # for clean cache and 262k windows (Frank's call). The fp8 shape is
+    # 10,388,766,720 B/rank (KV 1,350,182 tokens / 5.15x) where bf16 topped
-    # restorable from git tag frank-primary-fp8-mtp-20261001.
+    # out at three 241,664-token contexts. The bf16 ~4% decode edge was
-    grep -q '^FRANK_PRIMARY_KV_DTYPE=auto$' "$ENV_FILE"
+    # traded away for it (Frank's call, 2026-10-01).
+    grep -q '^FRANK_PRIMARY_KV_DTYPE=fp8$' "$ENV_FILE"
     grep -q -- '--kv-cache-dtype "$FRANK_PRIMARY_KV_DTYPE"' "$SERVE"
 }