This website requires JavaScript.
f68bda56c7
frank-large: serve script feeds the launcher its runtime root from the SDK venv
Frank
2026-09-19 03:49:04 -07:00
b6ce3afb1f
frank-large: CIRU v4.4.1 Orca cutover — root unit, one stream, full context
Frank
2026-09-19 02:59:01 -07:00
d374ab963f
frank-medium: orcarouter Q8_0 + DFlash2 Q4_K_M, 2x160k, balanced 32:33
Frank
2026-09-19 02:10:38 -07:00
374041cd7a
frank-small: revert vLLM cutover — back to llama.cpp Gemma4-12B Q4_K_M
Frank
2026-09-19 01:42:37 -07:00
bab8f7f839
frank-small: revert to bf16 KV — the only dtype that boots Gemma4
Frank
2026-09-19 00:26:11 -07:00
7e6ef1ac99
frank-small: --attention-backend flag (env var removed in this build)
Frank
2026-09-19 00:17:41 -07:00
966b70db1a
frank-small: pin VLLM_ATTENTION_BACKEND=TURBOQUANT
Frank
2026-09-19 00:07:57 -07:00
5e581666f2
frank-small: turboquant_k8v4 KV — int8 cannot boot this model
Frank
2026-09-18 23:50:19 -07:00
8d88b3a0a9
frank-small: int8 KV (per-token, per-head) — doubles the pool
Frank
2026-09-18 23:24:37 -07:00
ae137317ab
frank-small quadlet: Entry -> Entrypoint (actual v1 generator key)
Frank
2026-09-18 21:33:39 -07:00
f55189e880
frank-small quadlet: ContainerEntrypoint -> Entry (v1 generator key)
Frank
2026-09-18 21:27:52 -07:00
a2b39b179a
frank-small: llama.cpp -> vLLM (gemma-4-12B-it-qat-w4a16-ct, w4a16 Marlin)
Frank
2026-09-18 21:06:46 -07:00
cf06c4eab8
docs: frank-medium prompt cache -- why the cap is load-bearing, and the vision correction
Frank
2026-09-18 15:57:25 -07:00
3364e9afad
log: frank-medium prompt cache doubled to 16 GiB
Frank
2026-09-18 15:26:25 -07:00
1acca430a7
frank-medium: double the host-RAM prompt cache to 16 GiB
Frank
2026-09-18 15:25:55 -07:00
cfabe07959
log: frank-medium restored after orphaned merge-build server held both R9700s
Frank
2026-09-18 12:29:36 -07:00
59c76035d3
log: frank-medium left down after bench round 04:25, restored and verified
Frank
2026-09-18 08:15:14 -07:00
6c6464a31a
frank-medium: record 2026-09-18 rebalance measurement, keep 35:30
Frank
2026-09-18 00:28:29 -07:00
9764e86986
log: homepage model docs, fleet sweep, guard audit, sightsound runner lockout fix
Frank
2026-09-17 23:37:53 -07:00
bd96906aee
tests: migrate pending-recognition seam to the pruning contract
Frank
2026-09-17 23:34:09 -07:00
02a787683c
deploy-remote: prune stale pending state for retired services
Frank
2026-09-17 23:32:02 -07:00
32751c02d5
homepage: live model info on the fleet cards, user docs page at /docs
Frank
2026-09-17 23:25:50 -07:00
2fc0689858
frank-medium runbook: record the verified 35:30 post-deploy state
Frank
2026-09-17 20:14:10 -07:00
6569c63ec4
frank-medium: shift the layer split to 35:30, leaning work onto the x16 card
Frank
2026-09-17 18:58:06 -07:00
d6bc245934
relay: restore SIGNAL_CLI constant dropped by a boundary-repaired edit
Frank
2026-09-17 18:38:54 -07:00
698759c61e
relay: address the account by phone number and pass the message via -m
Frank
2026-09-17 18:28:20 -07:00
f73a7e80a1
relay: fix UnboundLocalError in dedupe pruning (rebind made the dict local)
Frank
2026-09-17 15:48:34 -07:00
d251dbfe62
relay: admit loopback and the host itself before the DROP
Frank
2026-09-17 14:37:38 -07:00
f73fd43027
relay: bind the host LAN IP, admit the podman bridge subnet only
Frank
2026-09-17 13:52:20 -07:00
8f747a2104
relay: bind 169.254.1.2 (host.containers.internal), not loopback
Frank
2026-09-17 13:26:24 -07:00
5f55efcb9d
alertmanager: drop --web.enable-lifecycle (not a flag in this binary; lifecycle is always on)
Frank
2026-09-17 13:18:30 -07:00
49ddce9d00
docs: halogen 400-zone note (compaction gap vs fit check); chore: gitignore third_party/ fork tree against peer stash sweeps
Frank
2026-09-17 13:14:51 -07:00
d856cdd9cc
deploy: retry the relay health probe after restart (raced python startup)
Frank
2026-09-17 13:11:25 -07:00
590853c133
alerting: Alertmanager + Signal relay deliver firing rules to Frank's phone
Frank
2026-09-17 13:09:59 -07:00
17450dd172
docs: the max-length sweep tabled; frank-small.env NGL/FLASH_ATTN restored
Frank
2026-09-17 04:56:59 -07:00
24eaba1078
frank-small: batch 8192/1024 — the measured max-length winner
Frank
2026-09-17 04:28:51 -07:00
74dda30fc7
frank-medium: seam recalibrated to measured MTP charge; docs to MTP/393216/34:31
Frank
2026-09-17 03:35:46 -07:00
4d13333da7
frank-small: wire SPEC_DRAFT_N_MAX and POLL_LEVEL knobs
Frank
2026-09-17 03:03:46 -07:00
23f09788e9
frank-medium: MTP drafter, 393216 pool (2x196k), split 34:31
Frank
2026-09-17 02:56:36 -07:00
dbd764a493
frank-small: four FULL 262144-token slots — the max that fits the card
Frank
2026-09-17 02:41:11 -07:00
87caca4894
docs: frank-small is Gemma4-12B QAT now — new runbook, Qwen retired as rollback target
Frank
2026-09-17 02:18:44 -07:00
0a25351caa
docs: the 2026-09-17 frank-medium shape -- 36:29, 327680 pool, vision
Frank
2026-09-17 02:08:14 -07:00
4b7de59dc8
frank-small: Gemma4-12B Q4_K_M, MTP sidecar + ngram-mod, q8 KV, --fit ctx
Frank
2026-09-17 01:59:10 -07:00
b6c083f522
tests: un-interleave the mmproj seam comment
Frank
2026-09-17 01:54:51 -07:00
ddf69c4107
frank-medium: pass --mmproj explicitly; this build has no autoload
Frank
2026-09-17 01:54:26 -07:00
de12a54df9
On main: peer frank-small WIP held (test file)
Frank
2026-09-17 01:47:30 -07:00
72fbcfa907
index on main: d95f55b
Frank
2026-09-17 01:47:30 -07:00
3b696107ad
On main: peer frank-small WIP held for frank-medium deploy
Frank
2026-09-17 01:47:22 -07:00
98912aec89
index on main: d95f55b
Frank
2026-09-17 01:47:22 -07:00
d95f55b822
frank-medium: 36:29 split, 327680 context, vision enabled
Frank
2026-09-17 01:37:55 -07:00
24cdbb0e45
docs: the VRAM figures settle 0.2 GiB below the post-load reading
Frank
2026-09-17 00:29:21 -07:00
6472d706be
docs: the ubatch optimum, the rebalance, and the live atomics chain
Frank
2026-09-17 00:27:59 -07:00
a11fae07d8
frank-medium: ubatch 512, and rebalance the split to 35:30
Frank
2026-09-17 00:21:48 -07:00
2f012fba39
the x4 link, measured: 6.72 GB/s — which decides the streaming question
Frank
2026-09-16 23:25:06 -07:00
262ac53c14
docs: how to bench this endpoint, and the llama-bench properties that bite
Frank
2026-09-16 22:11:14 -07:00
a40a1aad92
model cache cleared, re-fetched, and the update turned out to be real
Frank
2026-09-16 22:07:04 -07:00
77a07ffc30
scripts: fix the unbound local that killed the sweep at its first cell
Frank
2026-09-16 21:43:09 -07:00
87845fca22
scripts: a batch x ubatch sweep harness for frank-medium's model
Frank
2026-09-16 21:41:51 -07:00
a04a2c27a9
AGENTS.md: a token on a curl command line is world-readable
Frank
2026-09-16 21:10:45 -07:00
f917a2ef55
docs: the runbook records the DSpark excursion as an excursion
Frank
2026-09-16 21:01:45 -07:00
91546dbb95
frank-medium: back to the DFlash2 drafter, and rebalance to 36:29
Frank
2026-09-16 20:59:20 -07:00
718a283afc
docs: the frank-medium runbook catches up with DSpark, 2 slots and the rebalance
Frank
2026-09-16 20:21:31 -07:00
b312042706
frank-medium: 3x batch over 2 slots, and rebalance the split to pay for it
Frank
2026-09-16 20:17:27 -07:00
88f0d79021
docs: how the Uncensored payload is really fetched (it is shipped in parts)
Frank
2026-09-16 20:13:44 -07:00
ad18d2ddd8
frank-medium: step 3 -- the skew is the balance, and equality is off-limits
Frank
2026-09-16 20:00:17 -07:00
169eba0734
frank-medium: checkpoints back to the build's default (0)
Frank
2026-09-16 19:51:06 -07:00
b6d4914fcc
frank-medium: pay for DSpark's draft KV with a 38:27 split
Frank
2026-09-16 19:48:29 -07:00
46b56ab5e8
x4 note: the pause process died mid-run (18:58:30), not after the runs
Frank
2026-09-16 19:42:23 -07:00
4d6511d8b3
docs: the pause-process trigger is unproven — say so, everywhere
Frank
2026-09-16 19:34:40 -07:00
7e11d247a8
frank-medium: swap the drafter to magnitudedev/Qwen3.8-27B-DSpark
Frank
2026-09-16 19:34:00 -07:00
96062b627e
x4 note: post-reboot headroom (12.3 GiB, 240 blocks vs 475) is part of the record
Frank
2026-09-16 19:33:36 -07:00
803398051a
x4 pair round two: pure Vulkan changes nothing, and the first verdict was wrong
Frank
2026-09-16 19:32:53 -07:00
d14e0a94eb
hermes vhost: dial netavark's host address, not the LAN IP (fixes the 502)
Frank
2026-09-16 19:28:34 -07:00
65fbef63dc
docs: what the x4 pair can and cannot do — measured, not argued
Frank
2026-09-16 17:40:37 -07:00
e2dc2b67ce
AGENTS.md: what can be chained, and what cannot
Frank
2026-09-16 16:52:05 -07:00
9902267f85
deploy: stop the frank-medium unit before waiting for port 7770
Frank
2026-09-16 16:47:24 -07:00
a4f4429a99
frank-medium: chain DFlash2 with ngram-mod -- measured free where it cannot match
Frank
2026-09-16 16:44:45 -07:00
95936bbc81
bench-spec-ab: rewrite (write_env was lost), restore the unit in a finally
Frank
2026-09-16 16:38:12 -07:00
605e32dfc3
bench-spec-ab: pep8 spacing
Frank
2026-09-16 16:31:55 -07:00
9c59513cf5
bench-spec-ab: sweep an impl's own window, not just the spec type
Frank
2026-09-16 16:31:48 -07:00
e767810c9a
bench-spec-ab: think-free echo/structured shapes, and prove the echo is an echo
Frank
2026-09-16 16:27:57 -07:00
6e6e7187ca
bench-spec-ab: absolute launcher paths
Frank
2026-09-16 16:23:14 -07:00
21daf07713
frank-medium: ngram-mod is a legal chain member -- wire it and measure it
Frank
2026-09-16 16:19:15 -07:00
34253a9903
frank-medium: the drafter beats MTP and MTP beats nothing -- measured, not assumed
Frank
2026-09-16 16:15:59 -07:00
9a83dae73a
docs: blank line before the throughput heading
Frank
2026-09-16 15:52:45 -07:00
ff52543d8a
docs: frank-medium splits weights and KV across both cards, it does not replicate them
Frank
2026-09-16 15:52:18 -07:00
c97e359bc2
docs: the LLM index and the cooperation page catch up with layer split
Frank
2026-09-16 15:39:47 -07:00
87492c7b10
frank-medium: the agent-loop shape on one slot, and a harness header that matches the runtime
Frank
2026-09-16 15:38:20 -07:00
dc76dc5dc4
frank-medium: serve the whole 262144-token window, and repoint every signal to llama.cpp
Frank
2026-09-16 15:29:40 -07:00
ab51d61198
docs: frank-medium is llama.cpp with DFlash2 on both cards
Frank
2026-09-16 15:15:02 -07:00
c9c42e7489
build: bundle the rocBLAS device libraries at the depth the runtime looks for
Frank
2026-09-16 15:15:02 -07:00
05c0d3a00a
frank-medium: llama.cpp layer split across both R9700s, DFlash2 drafting
Frank
2026-09-16 14:59:23 -07:00
49e8dfff66
docs: the vault and AGENTS.md describe two Dankclaw endpoints
Frank
2026-09-16 14:52:51 -07:00
40850bbc8f
frank-medium: add the RDNA4 llama.cpp build and the GGUF fetch
Frank
2026-09-16 14:52:11 -07:00
f367a5a856
deploy: gate the port wait after stopping this service's own quadlet
Frank
2026-09-16 14:06:51 -07:00
7403bd3d50
deploy: never start the replacement while the predecessor holds the port
Frank
2026-09-16 14:06:03 -07:00
79cca19919
observability: the LLM surfaces follow the two-endpoint split
Frank
2026-09-16 13:59:49 -07:00
43ec77fee6
inference: frank-large returns to halogen, frank-small takes the 3090
Frank
2026-09-16 13:58:00 -07:00
cc29875842
research: how to make the two R9700s work together (and one correction to my own claim)
Frank
2026-09-16 08:59:12 -07:00
f134cc8876
docs+units: fix every claim my own changes made false
Frank
2026-09-16 07:46:23 -07:00
Frank