This website requires JavaScript.
a02850203d
frank-small: three contexts, vision on, 2-3 GiB of the 3090 left free
Frank
2026-10-02 03:08:00 -07:00
8f129180d8
docs+scripts: reconcile payload inventory after both swaps
Frank
2026-10-02 02:31:39 -07:00
72bd4c967e
docs: record the 2026-10-02 swaps — frank-large on ROCmFPX, frank-small on Q5_K_M
Frank
2026-10-02 01:51:11 -07:00
125d04967f
frank-large: ROCmFPX llama.cpp on the Q8_0_ROCMFPX_AGENT file, MTP drafting
Frank
2026-10-02 00:47:16 -07:00
d82c5c03b7
frank-small: Q5_K_M GGUF on llama.cpp, two full native windows
Frank
2026-10-02 00:00:17 -07:00
4bb10cd20e
fetch: kingjones Q8_0_ROCMFPX_AGENT payload, frank-small Q5_K_M GGUF
Frank
2026-10-01 23:50:52 -07:00
a7c68db9d0
docs: record the frank-small llama.cpp Q6_K cutover and the frank-large gufo rollback
Frank
2026-10-01 23:50:52 -07:00
8551bb006b
frank-large: roll back to gufo Qwen3.8-27B Q4_K_XL, one session x 262144 f16 KV
Frank
2026-10-01 22:40:59 -07:00
e452a1afbf
harness: tolerate missing container in ASI capture; llama cmdline fallback
Frank
2026-10-01 22:11:34 -07:00
01ba41cab4
harness: llama-server lifecycle in pre-clean and health guard
Frank
2026-10-01 21:58:51 -07:00
cde91c1a00
llama.cpp runtime arm: serve-script runtime switch, harness containerless support, master build
Frank
2026-10-01 21:56:30 -07:00
6513672f31
SPEC= (no draft): decode +46%, prefill +52%, ttft -35% vs MTP-2 baseline. Spec decode is NET-NEGATIVE on this endpoint for real chat+reasoning traffic at both depths; the bench-window win did not transfer.
Frank
2026-10-01 21:00:19 -07:00
5df6ef3166
autoresearch: harness setup
Frank
2026-10-01 20:41:51 -07:00
3c339eeba2
frank-large: record deployed boot facts (pool 819,332 tokens, 3.13x, 230 s)
Frank
2026-10-01 19:48:56 -07:00
780c307354
frank-large: hand vLLM the CONTAINER model path; link the new runbook
Frank
2026-10-01 19:39:55 -07:00
de383fe53e
frank-large: deployer spec gate must accept the sharded MTP layout
Frank
2026-10-01 18:52:17 -07:00
1127a52078
docs: frank-large vLLM MiMo-BF16 runbook; retire the gufo runbook
Frank
2026-10-01 18:48:51 -07:00
dab483c79f
frank-large: cutover to vLLM TheRock gfx1151, MiMo-V2.6 BF16 + cross-model MTP
Frank
2026-10-01 18:48:32 -07:00
b3313340fa
frank-large: fetch scripts for MiMo BF16, Qwen3.5-9B MTP source, DFlash drafter
Frank
2026-10-01 18:48:05 -07:00
e703b0c15c
frank-primary: restore fp8 3x262144 (bf16 boot fails on a dynamo constraint)
Frank
2026-10-01 14:40:17 -07:00
d3d180fd3f
frank-primary: drop --expert-precise-memory (boot crash loop with bf16)
Frank
2026-10-01 14:16:47 -07:00
db54fa3af1
frank-small: record 2x262144 verification numbers
Frank
2026-10-01 14:08:51 -07:00
8029119b19
frank-large: Q4_K_XL payload + 4 x 163840 sessions (f16 KV)
Frank
2026-10-01 14:08:50 -07:00
35e24db7e6
frank-small: reslice to 2 x 262,144 native slots (was 4 x 151,552)
Frank
2026-10-01 14:03:25 -07:00
389b605baf
frank-primary: 2 bf16 windows + --expert-precise-memory
Frank
2026-10-01 13:52:13 -07:00
3f1860aa5f
docs(log): record e2e 12-crawl flakiness (Plex redirect + transient ERR_ABORTED)
Frank
2026-10-01 13:40:24 -07:00
e3c3fad0f1
frank-primary: TimeoutStopSec=180, drop duplicate StartLimitIntervalSec
Frank
2026-10-01 13:30:29 -07:00
a1b650aabc
frank-primary: 3 full windows, pool at the contract floor, MNBT 2048
Frank
2026-10-01 13:25:27 -07:00
6bbcaf3999
docs(frank-primary): parallelism survey — DCP rejected on 2 GPUs (GQA constraint)
Frank
2026-10-01 13:17:45 -07:00
148a5f2575
frank-primary: wire --decode-context-parallel-size (opt-in)
Frank
2026-10-01 13:05:16 -07:00
667ec2987c
frank-small: correct the unit description (no MTP, 4x151k)
Frank
2026-10-01 07:06:15 -07:00
0c5e054fb8
frank-small: SGLang MiMo-V2.6-Distill-Qwen-9B INT8 (qwen3_5 hybrid, RTX 3090)
Frank
2026-10-01 06:55:48 -07:00
7da504a651
frank-primary: four full windows + Restart=always (runtime survivability)
Frank
2026-10-01 04:14:25 -07:00
bd311bf591
frank-primary: 5 x full 262144 contexts at fp8 (final shape)
Frank
2026-10-01 03:52:10 -07:00
7cab15ac39
frank-primary: max-possible window (241664) + reasoning/tool contract
Frank
2026-10-01 03:08:52 -07:00
e8524d4046
frank-primary: util 0.90 — production bf16 boot OOM fix
Frank
2026-10-01 02:12:54 -07:00
c1248d4a5d
frank-primary: bf16 no-spec cutover (262144x3, gather 32x64, MNBT 1024)
Frank
2026-10-01 01:54:27 -07:00
26f6f82c98
frank-primary: qwen cutover docs + deployer identity + seam re-pin
Frank
2026-09-30 20:24:49 -07:00
d91379cff2
frank-primary: roll back to qwen FP8 + MTP4 (2339fae
Frank
2026-09-30 20:13:23 -07:00
caee26e3c3
llm: patch 04b v2 - plan-time signal is the spec-config METHOD
Frank
2026-09-30 19:29:18 -07:00
9198da375f
docs: dflash KV journey — 03 verified (drafter graphs capture), 04 capacity/aliasing wall
Frank
2026-09-30 19:22:45 -07:00
d56db43aa3
llm: patch 04 - glm5 KV pool carries draft + hidden-state groups
Frank
2026-09-30 19:19:59 -07:00
2bc2f8c833
llm: patch 03 - pin the kpool tail block through the unifier
Frank
2026-09-30 18:38:53 -07:00
a1f14a1c0f
docs: dflash patch v5 clears the EAGLE3 gate — wrapper is the real target
Frank
2026-09-30 17:59:35 -07:00
b877d87588
patches: glm53-dflash-01 v5 -- patch the CONDITIONAL wrapper (the real eagle3 target)
Frank
2026-09-30 17:41:19 -07:00
7560d40543
patches: glm53-dflash-01 v4 -- all three SupportsEagle3 members on the wrapper
Frank
2026-09-30 17:24:56 -07:00
e044277d8b
patches: glm53-dflash-01 v3 -- declare supports_eagle3 as a class attribute
Frank
2026-09-30 17:08:50 -07:00
9c5c2494da
patches: glm53-dflash-01 v2 -- satisfy SupportsEagle3 by inheriting the Protocol
Frank
2026-09-30 16:53:24 -07:00
ce280e7295
fleet docs: AGENTS.md fork-work paragraph corrected and the DFlash port recorded
Comrade Chief Designer
2026-09-30 16:29:19 -07:00
d2e059262f
frank-primary: patch 05 is not part of the confirmed blocker
Comrade Chief Designer
2026-09-30 16:27:24 -07:00
3cf543ead6
frank-primary: fork-bomb hazard, patch 06 documented, retired probe
Comrade Chief Designer
2026-09-30 16:24:16 -07:00
c7db6d94e9
frank-primary: DFlash2 enabling patch — EAGLE3 aux-hidden-state interface for the GLM trunk
Comrade Chief Designer
2026-09-30 15:57:27 -07:00
0465b9b5a7
frank-primary: pin down the compile blocker — a name-resolution KeyError in an inlined callee
Comrade Chief Designer
2026-09-30 15:39:10 -07:00
62779aa237
frank-primary: document the experiment-window trap that faked two successful restores
Comrade Chief Designer
2026-09-30 14:54:07 -07:00
71db0debcc
frank-primary: patch 05, five verified artifacts, and correct the wrong confound theory
Comrade Chief Designer
2026-09-30 14:19:25 -07:00
fd864b5a42
frank-primary: compile changelist — real diffs, break diagnostics, patch 04, corrected overclaim
Comrade Chief Designer
2026-09-30 13:59:09 -07:00
1c18b9c808
frank-primary: DFlash2 load attempt + fp8 verdict + compile blocker narrowed
Comrade Chief Designer
2026-09-30 12:49:26 -07:00
bd8c72ff47
frank-primary: profile the GLM decode step and stage the compile changelist
Comrade Chief Designer
2026-09-30 09:36:48 -07:00
979a4308f4
frank-primary runbook: record the allocator env and the e2e/FrankBook-Pro coupling
Frank
2026-09-30 05:40:57 -07:00
a3e1c97408
frank-primary: pin PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
Frank
2026-09-30 05:13:50 -07:00
f2bcf84e88
frank-primary cutover: GLM-5.3-Flash W4A16 int4 (Intel AutoRound)
Frank
2026-09-30 04:22:11 -07:00
2339faeaa4
frank-primary: fix the SIGPIPE race in both image-digest gates
Frank
2026-09-30 01:05:05 -07:00
01fedb0081
frank-primary.env: quote the space-separated LEGACY_CONTAINERS list
Frank
2026-09-30 00:58:32 -07:00
3bb0e82ccf
frank-primary cutover: tcclaviger RDNA4 stack, 600k ctx / 3 sessions / q8-q8 KV
Frank
2026-09-30 00:46:30 -07:00
9cf4b1de82
docs: qwen-fp8 endpoint published on 7773; fix log bullet placement
Frank
2026-09-29 23:51:53 -07:00
d28ab8ef8c
docs: record the qwen-fp8-tuned-20260929 tag + digest-pinned image
Frank
2026-09-29 23:31:13 -07:00
f532e03bf4
docs: GLM-5.3-Flash readiness revised — tcclaviger fork has native MXFP4/DSFP4
Frank
2026-09-29 23:28:04 -07:00
57072a44df
docs: final 256k verdict — boots at 4.03-4.5 GiB KV, band 34.1-36.5 straddles target
Frank
2026-09-29 21:55:58 -07:00
0ccb01708b
docs: qwen38-fp8 expert-offload tuning — prefetch off hits 37.4-40.6 tok/s (target met)
Frank
2026-09-29 21:22:21 -07:00
67672dc794
docs(llm): record a measured NEGATIVE tuning result for frank-large
Frank
2026-09-29 05:06:06 -07:00
3a931ce2c1
chore(dankclaw): reclaim 510G of retired payloads; fix a stale fallback
Frank
2026-09-29 00:00:49 -07:00
7f9424554c
frank-large: serve the abliterated Swift-1.5 UD-Q8_K_XL
Frank
2026-09-28 23:29:05 -07:00
32abc8d543
chore(llm): add dankstatus.sh — one verified parse path for Dankserver state
Frank
2026-09-28 23:27:50 -07:00
b4b6777e49
fetch: add the Swift-1.5 frank-large candidate download
Frank
2026-09-28 22:48:53 -07:00
9d9d28cda6
docs(llm): record the gufo cutover verification
Frank
2026-09-28 22:28:44 -07:00
a196b0fa45
test: scope the halogen-quadlet check to tracked files
Frank
2026-09-28 22:12:16 -07:00
aaef419a6d
frank-large: replace stock llama.cpp with gufo (gufo-org/gufo) on Dankclaw
Frank
2026-09-28 22:08:05 -07:00
d29d104f01
docs: frontier-MoE feasibility on Dankserver (V4.1 / MiMo / GLM-5.3)
Frank
2026-09-28 14:58:59 -07:00
564a781539
log: frank-primary deploy-state reconciled to ba2100e
Frank
2026-09-28 13:22:14 -07:00
ba2100e22b
docs: prefix caching deployed on frank-primary (E8-qualified, 9x/63x reuse TTFT)
Frank
2026-09-28 12:33:28 -07:00
41e918929f
frank-primary: enable prefix caching (E8 qualified: 9x-63x reuse TTFT, zero VRAM cost, cold path unchanged)
Frank
2026-09-28 12:03:56 -07:00
92579b0674
log: frank-primary restore verified (f187165
Frank
2026-09-28 11:21:43 -07:00
f18716563a
frank-primary: PLE ready timeout 1800s + bench matrix runbook; index/doc-link fixes
Frank
2026-09-28 10:32:00 -07:00
671f784ca4
docs: frank-primary q8 cutover — new runbook, uncensored retired as rollback, fleet table/index/log synced
Frank
2026-09-27 22:53:19 -07:00
ca2810a319
frank-primary: unit description/notes retuned for the Q8_0 payload
Frank
2026-09-27 22:27:32 -07:00
b06895d725
frank-primary: cutover state to Qwen3.8-Flash-Next Q8_0 (6 shards, PLE pinned, MTP4, sharp template)
Frank
2026-09-27 22:25:08 -07:00
928fd81dbd
docs(log): record Q8_0 bring-up qualification and uncensored restore posture
Frank
2026-09-27 18:19:56 -07:00
efff47a75d
AGENTS.md: frank-primary section rewritten for the R9V cutover
Frank
2026-09-27 04:39:30 -07:00
ca6ba6d0e9
frank-primary: R9V runbook + docs cutover
Frank
2026-09-27 04:35:05 -07:00
7562e41ba4
frank-primary: retire the FP8 fetch path with the vLLM 0.30 wiring
Frank
2026-09-27 04:21:47 -07:00
095128b669
frank-primary: R9V radiance cutover wiring (launcher-driven serve + state gates)
Frank
2026-09-27 04:05:46 -07:00
baba84fc90
frank-primary vllm: orphan-container self-heal + HOME + stop/rm order
Frank
2026-09-27 01:08:21 -07:00
ecaad36654
frank-primary: vLLM 0.30 UVA-offload shape (pre-deploy commit)
Frank
2026-09-27 00:16:42 -07:00
82960c77db
frank-primary: add FP8 payload manifest + fetch wrapper
Frank
2026-09-26 23:50:58 -07:00
7a8d992fce
searxng: wire paid Kagi engine with deploy-time key rendering
Frank
2026-09-26 23:25:55 -07:00
4992e5a010
hf-manifest.py: follow the 307 resolve-cache redirect for non-LFS files
Frank
2026-09-26 22:50:17 -07:00
5bd12d3193
hf-manifest.py: read X-Linked-* headers, never follow the resolve redirect
Frank
2026-09-26 22:46:24 -07:00
f41e0384dd
frank-primary: measure expert routing; block choice is immaterial
Frank
2026-09-26 21:58:03 -07:00
d67768e1a5
frank-primary runbook: 5/3 block fill is the deployed shape
Frank
2026-09-26 21:40:25 -07:00
d317d6b246
frank-primary: 5th expert block onto ROCm0 (operator-directed fill)
Frank
2026-09-26 21:38:05 -07:00
Frank
Comrade Chief Designer