memory: snapshot — 2026-06-13 ana-ml2 Ada→Blackwell + NVFP4-infeasible + Qwen3.5-VL FP8 + comfyui→arbo + GPU-1 rebalance + prefix-caching; archive 8 (2026-06-03 cluster)
This commit is contained in:
+119
-83
@@ -1,6 +1,6 @@
|
||||
# Persistent memory — eshpfi-management
|
||||
|
||||
_Last updated: 2026-06-11_
|
||||
_Last updated: 2026-06-13_
|
||||
|
||||
## Repo purpose
|
||||
|
||||
@@ -91,43 +91,100 @@ no longer deployed sidecars here. See Recent decisions.)
|
||||
|
||||
## Current state / in-flight
|
||||
|
||||
_As of 2026-06-11:_
|
||||
_As of 2026-06-13:_
|
||||
|
||||
- **Mac Pro migration is the big open project** — `migration-plan.md` (repo root) frames a
|
||||
WORKSTATION-ONLY migration of nh3-dev's dev env to an M2 Ultra Mac Pro Rack, racked in NH3 on
|
||||
the 10.100 subnet. nh3-dev wears two hats; only **Hat 1** (the ~40 repos, all of `~/.claude`,
|
||||
dotfiles, toolchain) moves. **Hat 2** fleet sidecars (egress SOCKS5 proxy, mead-hall, ttyd
|
||||
seats, `althing-forseti`) STAY on the Linux VM. **Phase 0 is the only time-sensitive step**:
|
||||
push the pushable repos + confirm which of ~6 no-upstream repos are local-only. Cutover =
|
||||
`rsync` the working trees, **NOT re-clone** (re-clone would lose unsaved work in ~15 repos).
|
||||
macOS gotcha pre-flagged: `/home/lkraven` → `/Users/lkraven` repath. Everything else waits on
|
||||
the hardware landing in the rack.
|
||||
- **ana-ml2 went Ada → Blackwell** (dual NVIDIA RTX PRO 6000 Blackwell Max-Q, 96 GB each, cc 12.0 /
|
||||
sm_120 — was dual RTX 6000 Ada 48 GB / cc 8.9, confirmed live via `nvidia-smi`). GPU layout now:
|
||||
**GPU 0** held free for large-model hot-loads (llama-swap pinned there, `edf0f91`); **GPU 1** is the
|
||||
steady-tenant card — granite (131k ctx) + qwen3.5-vl (65k) + the embed/rerank/reward trio, ~3.5 GB
|
||||
free after the rebalance. CLAUDE.md's server table + the granite/qwen "Ada (cc 8.9)" compose comments
|
||||
are STALE — doc-fix offered, **pending operator go-ahead** (/snapshot doesn't auto-edit CLAUDE.md).
|
||||
|
||||
- **Two commits/branches committed-not-pushed** (operator's push call): `95b2701` (GLM thinking
|
||||
config) on main; and the R17 v2 results branch `r17-v2-characterization` on the irv-ml1
|
||||
checkout (push HELD — see decisions).
|
||||
- **R16 "vmoan" TTS LoRA POC — awaiting the operator's ear** (Brokkr's designated gate). Two auditions
|
||||
on NFS: `/mnt/smithy/scratch/r16-vmoan-pilot/vmoan_audition_v1.wav` (42 s, run-on) vs `_v2.wav` (10 s,
|
||||
tight). v2 fixed the run-on; Brokkr holds the pilot verdict on the operator's listen. Grab:
|
||||
`scp irv-ml1:/mnt/smithy/scratch/r16-vmoan-pilot/vmoan_audition_v{1,2}.wav .`. LOCAL-ONLY corpus
|
||||
(soundgasm-derived) — distribution barred; never echo transcripts to the bus.
|
||||
|
||||
- **R17 v2 corpus characterization DONE on irv-ml1** (brokkr, ~8.5h): 119 clips → **18 MULTI /
|
||||
101 ELIGIBLE** (Stage-0 drop list) + 19936 candidates + ASR speech + blend verdicts. Outputs
|
||||
on the local branch; **push HELD** — no-push rule AND the `speech.json` are ASR transcripts of
|
||||
- **R17 v2 corpus characterization still on the local irv-ml1 branch, push HELD** (`r17-v2-character
|
||||
ization`, 18 MULTI / 101 ELIGIBLE + 19936 candidates + `speech.json` ASR). No-push rule AND ASR of
|
||||
the intimate-audio batch (content-exposure call). Awaiting brokkr-collect or explicit push approval.
|
||||
|
||||
- **pi coding agent + GLM 5.1 harness live on nh3-dev** — `glm` launcher runs pi against `glm-5.1`
|
||||
(thinking-off) via the LiteLLM gateway; `glm-5.1-reasoning` for opt-in thinking.
|
||||
- **Mac Pro migration is still the big open project** — `migration-plan.md` (repo root): WORKSTATION-
|
||||
ONLY move of nh3-dev's Hat-1 dev env (~40 repos, all of `~/.claude`, dotfiles, toolchain) to an M2
|
||||
Ultra Mac Pro Rack on the 10.100 subnet. Hat-2 fleet sidecars (egress SOCKS5, mead-hall, ttyd seats,
|
||||
`althing-forseti`) STAY on the Linux VM. **Phase 0** (push pushables + confirm the ~6 local-only repos)
|
||||
is the only time-sensitive step. Cutover = `rsync` working trees, NOT re-clone (re-clone loses unsaved
|
||||
work in ~15 repos). macOS `/home/lkraven`→`/Users/lkraven` repath. Everything else waits on hardware.
|
||||
|
||||
- **Worldtree config-propagation lane is mature + humming** — pre-merge delta-ping → sync-on-merge.
|
||||
All of v0.33.x→v0.34.1 reconciled to the demo+personal bind-mounts; demo/personal both on v0.34.x,
|
||||
pinned stays pre-cutover (full migration at its next re-image).
|
||||
- **sglang-vs-vLLM bench stack staged but parked** (`5f049cb`) — the originating question (SGLang
|
||||
RadixAttention caching) resolved without it: vLLM v1 already defaults prefix-caching ON, now pinned
|
||||
explicit on granite+qwen. The NVFP4 chase is abandoned. Bench stack stays staged if ever revisited;
|
||||
the MAX_JOBS-on-shared-prod foot-gun (see Tried) caps any future from-source build on ana-ml2.
|
||||
|
||||
- **Disclosed-keys hygiene queue** carries forward (rotate at convenience): HF token, the wt-personal
|
||||
keys, Gitea runner reg token, MINIFLUX_PASSWORD. `ZAI_API_KEY` (z.ai MCP web-tools key) now also
|
||||
lives in `~/.config/secrets/env.sh` (shared from Worldtree env for the Brokkr panel).
|
||||
- **pi + GLM 5.1 harness live on nh3-dev** — `glm` launcher runs pi against `glm-5.1` (thinking-off)
|
||||
via the LiteLLM gateway; `glm-5.1-reasoning` for opt-in thinking.
|
||||
|
||||
- **Still open from prior:** clean legacy `news-digest` on ana-docker; pin llama-swap to GPU 0; the
|
||||
`docker push 60s ceiling` mystery uninstrumented.
|
||||
- **Worldtree config-propagation lane is mature + humming** — pre-merge delta-ping → sync-on-merge to
|
||||
the demo+personal bind-mounts; pinned stays pre-cutover (full migration at its next re-image).
|
||||
|
||||
- **Disclosed-keys hygiene queue** (rotate at convenience): HF token, the wt-personal keys, Gitea
|
||||
runner reg token, MINIFLUX_PASSWORD, ZAI_API_KEY. (The shared all-agents LiteLLM key is intentional,
|
||||
not hygiene-debt — see decisions; rotatable via infra-ops only if it leaks.)
|
||||
|
||||
- **Still open from prior:** clean legacy `news-digest` on ana-docker; the `docker push 60s ceiling`
|
||||
mystery uninstrumented. (llama-swap GPU-0 pin — DONE this session, `edf0f91`.)
|
||||
|
||||
## Recent decisions
|
||||
|
||||
- `[2026-06-13]` **ana-ml2 upgraded Ada → dual RTX PRO 6000 Blackwell Max-Q** (96 GB each, cc 12.0 /
|
||||
sm_120; was dual RTX 6000 Ada 48 GB / cc 8.9 — confirmed live via `nvidia-smi`). Unlocks NVFP4 (FP4
|
||||
tensor cores) and doubles VRAM headroom. CLAUDE.md's server table + the granite/qwen compose
|
||||
"Ada (cc 8.9)" comments are now stale; doc-fix **offered, pending operator go-ahead** (snapshot doesn't
|
||||
auto-edit CLAUDE.md). Tracking surface: this snapshot + commits `19a07b9`/`1e2a3a1` ("Blackwell 96GB").
|
||||
|
||||
- `[2026-06-13]` **NVFP4-W4A4 is infeasible for Granite — FP8 stays the Granite-on-Blackwell format.**
|
||||
W4A4 (4-bit weights + activations) collapses at 30k context, proven **producer-independent** (modelopt
|
||||
AND llm-compressor both clean-NONE from the same BF16 base + wikitext-2k calib). No 4-bit wins both
|
||||
axes: W4A4 = quality collapse; W4A16-NVFP4/AWQ = weight-only dequant → bf16 (no FP4-core speedup), so
|
||||
the tempting ~40% uplift doesn't materialize without quality loss. **30B retired** (not enough quality
|
||||
for the VRAM/perf hit for our use case). (auto-memory `reference_nvfp4_w4a4_granite_infeasible`)
|
||||
|
||||
- `[2026-06-13]` **Qwen3.5-9B VL (FP8) deployed on ana-ml2 GPU 1** — `qwen35-vl` stack, :8007, gateway
|
||||
alias `qwen3.5-9b-fp8`. Clean FP8 (no NVFP4 for vision). **Pinned nightly digest, not `:latest`**: the
|
||||
stable release quantizes the Qwen3.5-VL *vision tower* under `--quantization fp8` → garbage vision (LM
|
||||
fine, "sees" noise); the nightly correctly excludes the vision tower. Re-pin to `:latest` + drop the
|
||||
pin once that exclusion lands stable. (`2e3dcc2`)
|
||||
|
||||
- `[2026-06-13]` **comfyui 325 G model tree migrated worktank → `/storetank/arbo`** (worktank 97% → 26%).
|
||||
`arbo` is the consuming app (ComfyUI-backed prompt/image pipeline); models live under its name on the
|
||||
bigger pool. Overlay bind-mount via `COMFYUI_MODELS_DIR` in the comfyui compose; inventory at
|
||||
`docs/arbo-comfyui-model-catalog.md` (retain-large-portion-for-future-use decision). (`38186be`)
|
||||
|
||||
- `[2026-06-13]` **GPU layout settled on the Blackwell box.** GPU 0 held free for large-model hot-loads
|
||||
(llama-swap pinned there, `edf0f91`); GPU 1 is the steady-tenant card — granite 131k ctx (was 51k),
|
||||
qwen 65k, embed/rerank/reward trio, ~3.5 GB free after rebalance (`1e2a3a1`, `19a07b9`; trio re-floored
|
||||
for 96 GB, 20×-parallel-stable). embed/rerank left at floor — long docs are chunked BEFORE embedding,
|
||||
so a longer embed ctx buys nothing. PagedAttention note: max-model-len is a ceiling, not a reservation,
|
||||
so small requests aren't blocked by the big ceiling; concurrency = KV-pool-tokens / actual-request-size.
|
||||
|
||||
- `[2026-06-13]` **Prefix caching pinned explicit on granite + qwen** — benched ~6.5× faster TTFT (45 ms
|
||||
vs 292 ms) on a shared ~4.5k-token summarizer template; soft/evictable KV, neutral when prefixes don't
|
||||
repeat. vLLM v1 `:latest` defaults it ON (granite) but the qwen nightly defaults it OFF — pin both so a
|
||||
version flip can't silently disable it. (`a9a2be7`)
|
||||
|
||||
- `[2026-06-13]` **granite-4.1-8b listed as the always-available summarizer/classifier endpoint + a
|
||||
shared all-agents key minted** (operator-directed). Added to the global `~/.claude/CLAUDE.md` Global-
|
||||
tools section; key alias `all-agents-local`, scoped to the FREE local models only (granite + qwen-
|
||||
vision + embed/rerank, NOT the paid GLM), internal-gateway-only, rotatable. The standing "reach for
|
||||
this before spending premium tokens on low-caliber high-volume work" lever. (auto-memory
|
||||
`reference_litellm_gateway`)
|
||||
|
||||
- `[2026-06-13]` **arbo engine + frontend stack stood up** (ADR-0001) — irv-ml1 co-located inference
|
||||
engine (`ee57e69`), python-based healthcheck (the slim image ships no curl/wget, `bdb3312`), frontend
|
||||
ro-mounted from the v0.11.2 checkout (`922e8ad`, ADR-0001 D2). arbo = the ComfyUI-consuming app whose
|
||||
models now live at `/storetank/arbo`.
|
||||
|
||||
- `[2026-06-11]` **GLM thinking inverted at the LiteLLM gateway** (operator call): `glm-5.1` defaults
|
||||
thinking-OFF; new `glm-5.1-reasoning` alias = same z.ai upstream with thinking ON (opt-in).
|
||||
Mechanism: `litellm_params.extra_body:{thinking:{type:disabled}}` — LiteLLM `drop_params` STRIPS a
|
||||
@@ -172,12 +229,10 @@ _As of 2026-06-11:_
|
||||
|
||||
- `[2026-06-05]` **Granite 4.1 8B FP8 replaced phi4-mini as the production summarizer** (supersedes
|
||||
the 2026-06-04 phi4 decision below). Beat phi4 on precision in brokkr's R15 P03. **Staying FP8, not
|
||||
Q4/AWQ** — primary workload (agent memory + summarization) is high-concurrency, where FP8-on-Ada
|
||||
scales ~linearly (profiled 2010 tok/s @ C=32; single-stream 67.5 is batch-1 GEMV physics, not a
|
||||
config bug — placement/kernel/contention all ruled out). vLLM `vllm-granite` :8004 GPU 1, official
|
||||
IBM compressed-tensors FP8, CUDA graphs. **GPU-1 retune** (trio utils 0.2/0.2/0.3→0.07/0.07/0.18,
|
||||
granite 0.36) freed ~10 GB → CUDA graphs + a Granite-text-LoRA hedge. nevermore repointed. (`34a43a0`,
|
||||
auto-memory `reference_ana_ml2_vllm_granite`)
|
||||
Q4/AWQ** — primary workload (agent memory + summarization) is high-concurrency, where FP8 scales
|
||||
~linearly (profiled 2010 tok/s @ C=32; single-stream 67.5 is batch-1 GEMV physics, not a config bug).
|
||||
vLLM `vllm-granite` :8004 GPU 1, official IBM compressed-tensors FP8, CUDA graphs. (Then on Ada cc 8.9;
|
||||
the box has since gone Blackwell — see 2026-06-13.) (`34a43a0`, auto-memory `reference_ana_ml2_vllm_granite`)
|
||||
|
||||
- `[2026-06-05]` **Langfuse v3 stood up on ana-docker (:3001) as the gateway trace UI**; LiteLLM
|
||||
`success_callback:[langfuse]` live (project `gateway`). Pretty prompt/completion/reasoning traces +
|
||||
@@ -225,34 +280,38 @@ _As of 2026-06-11:_
|
||||
- `[2026-06-04]` **`brokkr` user (no-sudo) on irv-ml1; R14/R15/R16 substrate moved to /home/brokkr.**
|
||||
Persistent box services there need SYSTEM systemd units (see Tried).
|
||||
|
||||
- `[2026-06-03]` **yt-voice-clipper bot-gate fix = route yt-dlp through NH3 residential
|
||||
egress, NOT cookies/PO-token.** YouTube hard-flags the Irvine colo IP (LOGIN_REQUIRED on a
|
||||
public video even with no cookies). Cookies + the bgutil PO-token + deno JS-runtime all
|
||||
loaded fine — the gate is pure IP reputation. Operator chose proxy-via-nh3-dev → durable
|
||||
dante proxy → proven. The egress proxy is a reusable fleet lever for any datacenter-IP-gated
|
||||
service.
|
||||
|
||||
- `[2026-06-03]` **yt-voice-clipper push-to-deploy via gitea webhook** (operator-directed,
|
||||
after 6 manual rebuilds in ~40 min). Webhook (not poll) — gitea CAN reach the WG IP per the
|
||||
operator. The proxy env + Homepage labels live in the **host-specific override** (untracked
|
||||
→ survive the auto-deploy's `git reset --hard`), NOT yt-voice-clipper-dev's image. Runbook
|
||||
`d4f180d`.
|
||||
|
||||
- `[2026-06-03]` **R14 scope = (a) provision-only.** infra-ops provides box + CUDA env +
|
||||
engines + weights + NFS; brokkr/dev wires `arms.py` + runs — keeps infra-ops OFF the
|
||||
VIVAE-processing path (VIVAE = Variably Intense Vocalizations of Affect/Emotion, CHARTER §4
|
||||
highest-liability; operator authorized R&D-eval-only, quarantined). Box = irv-ml1 (A6000
|
||||
free; ana-ml2 GPU-saturated). Per-engine venvs (divergent torch stacks); A6000 = `cuda:0`
|
||||
NATIVE (≠ docker `=1`).
|
||||
|
||||
- `[2026-06-03]` **Declined worldtree v0.30.4 staging deploy** — that's worldtree-team's
|
||||
CI/CD lane (a developer `staging/vX.Y.Z` git-tag promote), not infra-ops. They self-corrected
|
||||
to the same conclusion independently.
|
||||
|
||||
_40 older entries archived to archival-memory.md._
|
||||
_44 older entries archived to archival-memory.md._
|
||||
|
||||
## Tried and abandoned
|
||||
|
||||
- `[2026-06-13]` **Heavy from-source compile (`MAX_JOBS=128` sglang fork build) on the shared PROD GPU
|
||||
box PINS it** — load hit 187, prod vLLM services restarted, killed an in-flight quant. ana-ml2 hosts
|
||||
live inference; never run a big build there at full parallelism. Cap `MAX_JOBS≤32`, build off-box, or
|
||||
cgroup-constrain. (Operator ran the kill; infra-ops NOPASSWD-sudo confirmed working on ana-ml2 — retry
|
||||
infra-ops on an ssh-255 before concluding "no access.")
|
||||
|
||||
- `[2026-06-13]` **`--quantization fp8` on a VL model can quantize the VISION TOWER → garbage vision**
|
||||
(Qwen3.5-VL on the stable vLLM: gray-grid output; the LM answers text fine, so it "looks" healthy
|
||||
until you actually feed it an image). The nightly excludes the vision tower. Lesson: validate the
|
||||
VISION path on a quantized VLM, not just text — and pin the engine digest that has the exclusion.
|
||||
|
||||
- `[2026-06-13]` **vLLM's `--gpu-memory-utilization` is checked against FREE VRAM at startup
|
||||
(`free >= util*total`), not total** — on a shared card, growing one service before trimming a
|
||||
co-tenant OOMs ("free 48.56 < desired 80.72" at util 0.85 on a half-occupied 96 GB card). Start-order
|
||||
matters: trim the shrinking service FIRST, then grow the other. Size to the FREE budget, not the total.
|
||||
|
||||
- `[2026-06-13]` **The `vllm/vllm-openai` entrypoint is already `["vllm","serve"]`** — the compose
|
||||
`command:` supplies the model as the first POSITIONAL arg + flags; a second `serve` (or `--model X`)
|
||||
yields "unrecognized arguments". Same-class gotchas this session: `tee` masks the real exit code (use
|
||||
a `>` redirect to keep rc); HF `datasets` rejects bare `wikitext` (needs `Salesforce/wikitext`).
|
||||
|
||||
- `[2026-06-13]` **Chatterbox-Turbo LoRA finetune: the repo's `setup.py` loads the WRONG tokenizer** —
|
||||
`merge_and_save_turbo_tokenizer()` pulls gpt2-medium + a grapheme merge file (len mismatch) instead of
|
||||
the chatterbox-turbo GPT2 tokenizer (vocab.json+merges.txt, len 50276). Fix = override with the correct
|
||||
tokenizer + delete the grapheme `tokenizer.json`; `[vmoan]` token → new_vocab_size 50277 (1-row resize),
|
||||
lora_r 64 / alpha 128, modules_to_save=[text_emb,text_head]. Also: a unique-stem corpus collision (53
|
||||
rows, 32 wavs) needs `{index}_{stem}` IDs. (irv-ml1 `~/r16-vmoan-harness`)
|
||||
|
||||
- `[2026-06-11]` **A completion-poll `while pgrep -f <scriptname>` SELF-MATCHES its own remote shell
|
||||
argv** — the poll's command line contains the script name, so its own `pgrep -f` always finds itself
|
||||
→ the loop never exits, the poll never fires. Use a match pattern ABSENT from the poll command (pgrep
|
||||
@@ -261,7 +320,7 @@ _40 older entries archived to archival-memory.md._
|
||||
|
||||
- `[2026-06-08]` **Demucs `uv pip install demucs` pulls torch 2.12/torchaudio 2.11 → `ta.save()`
|
||||
requires torchcodec → dies AFTER separating (0 stems written, rc=1).** Same class as the torch-2.12
|
||||
torchcodec foot-gun below. Fix = pin `torch==torchaudio==2.4.1` (pre-torchcodec save backend) +
|
||||
torchcodec foot-gun. Fix = pin `torch==torchaudio==2.4.1` (pre-torchcodec save backend) +
|
||||
`UV_LINK_MODE=copy` for the EPERM-hardlink quirk. Lesson restated: validate the SAVE path, not just
|
||||
import + GPU inference, on a bleeding-edge torch.
|
||||
|
||||
@@ -317,27 +376,4 @@ _40 older entries archived to archival-memory.md._
|
||||
widget targets (retired ESH firewall IPs) hanging the node event loop into `exit_mmap`, needing a
|
||||
host reboot. Check homepage's siteMonitors against retired hosts. (`incident_esh_docker_nfs_boot_race`)
|
||||
|
||||
- `[2026-06-03]` **gitea webhook to a private IP is denied by `webhook.ALLOWED_HOST_LIST`**
|
||||
(anti-SSRF; default `external` blocks private/loopback). Symptom: delivery shows
|
||||
`dial tcp ...: webhook can only call allowed HTTP servers`. Fix = APPEND the target net to
|
||||
ALLOWED_HOST_LIST in gitea's app.ini (keep `external`; scope tight, never `*`/`private`) +
|
||||
restart gitea (act_runner job containers survive a restart). gitea runs as a container on
|
||||
ana-docker (`gitea_gitea_data` volume, `/data/gitea/conf/app.ini`).
|
||||
|
||||
- `[2026-06-03]` **torch-2.12 venvs need `uv pip install torchcodec`** — torchaudio 2.12
|
||||
defaults to the TorchCodec backend for `.load`; without it, real audio I/O throws "TorchCodec
|
||||
is required" — and it ONLY surfaces at actual conversion, NOT at import/model-load. Lesson:
|
||||
validate real I/O, not just import, when provisioning ML engine envs. (seed-vc on torch 2.4
|
||||
uses the legacy backend, exempt.)
|
||||
|
||||
- `[2026-06-03]` **Backgrounding `althing-cli monitor` with an inline shell `&` (instead of
|
||||
the Bash-tool `run_in_background`) orphans it** — it survives the shell exit, holds the
|
||||
per-handle flock UNTRACKED (won't notify the session), and `stop-monitor` doesn't detect it.
|
||||
Fix: find + kill the orphan PID (verify cwd=this repo / handle first — nh3-dev is shared, other
|
||||
agents' monitors run there too), then re-arm via run_in_background. Always re-arm tracked.
|
||||
|
||||
- `[2026-06-03]` **`uv pip install .` fails on SmoothKen/knn-svc** (and similar script-repos)
|
||||
— it's analysis scripts + a poetry pyproject, no buildable package (setuptools
|
||||
package-discovery error). Install the pyproject deps directly, don't build the "package".
|
||||
|
||||
_44 older entries archived to archival-memory.md._
|
||||
_48 older entries archived to archival-memory.md._
|
||||
|
||||
Reference in New Issue
Block a user