Frank
c3f30bdbd0
Purge live-state claims that ~/.pi is our harness; keep general pi documentation in omp-docs/ and dated historical records. - README.md: single config source; Legacy Notes reworded - AGENTS.md: drop 'like ~/.pi did' framing - APPEND_SYSTEM.md: okf-spec -> omp-docs/references/okf-spec.md - dotfiles/zsh/.zshrc: systemprompt alias -> APPEND_SYSTEM.md - managed skill llm-fleet-model-swap + model_config_sync skill: single-source models.yml; validation script YAML-only - memories (MEMORY.md, memory_summary.md, learned.md): single source - omp-docs: .okf root -> ~/.omp/omp-docs; de-live pi-era claims in pi/index.md, pi-web-access.md, engineering/agents-files.md
27 lines
2.2 KiB
Markdown
27 lines
2.2 KiB
Markdown
## System Topology & Hardware
|
|
- **dankclaw**: AMD Strix Halo APU (gfx1151) + RTX 3090 in M.2 dock (PCIe x4, ~29 GB/s sequential). Combined GPU budget ~88 GiB. OS: Ubuntu 24.04.
|
|
- **dankserver**: ASUS TUF GAMING B650-E WIFI. Dual AMD R9700 GPUs.
|
|
- **RCCL Constraint**: Chipset-attached PCIe lanes fail RCCL atomics (broadcast/all-reduce/send-recv). Fix: Move card #2 to CPU-attached M.2_2 slot via PCIe x4 riser to restore atomic-op routing.
|
|
|
|
## LLM Model Portfolio
|
|
- **frank-large (halogen)**: Qwen3.8-27B-ROCmFPX-Q8_0. Runs on Strix Halo (gfx1151).
|
|
- *KV Pool*: 524288 positions (~14.4 GiB). Admission is pool-bound, not slot-bound (caps at 3 simultaneous full-length requests).
|
|
- *Behavior*: Defaults to greedy. Stalls (full-CPU, no disk activity, no output) under co-tenancy with pinned trunk are documented behavior, not crashes.
|
|
- *Constraints*: maxTokens <= 65536. First request after restart may be silent ~45-60s while first-touching KV pool.
|
|
- **frank-small**: Qwen3.8-27B GSQ-RCO IQ3_S. Runs on RTX 3090 (M.2 dock). Defaults to sampling (temp 1.0, top_k 20, top_p 0.95).
|
|
- **ik_llamacpp**: mradermacher/Qwen3.8-Flash-Next-Uncensored-i1-GGUF (i1-Q4_K_M).
|
|
- *Backend*: CUDA (3090) + Vulkan (Strix Halo) split.
|
|
- *Performance*: ~3.8-4.0 t/s decode, ~192 t/s prefill.
|
|
- *Bottleneck*: PCIe x4 launch/sync overhead; CUDA graphs currently abort on hc ops (ggml_repeat_4d).
|
|
|
|
## Software Infrastructure
|
|
- **Podman**: Rootless deployment using Quadlet units for boot persistence. Note: `restart: unless-stopped` survives crashes, but not explicit stops (kernel power-off).
|
|
- **Authentik**: Anubis guard chain. Marreta app connected via ProxyProvider (fixed null flow crash).
|
|
- **AGENTS.md**: Structured as (1) Active-topology table, (2) Runbook pointers, (3) Key Lessons Learned.
|
|
|
|
## Durable Lessons
|
|
- **ROCm**: Use layer-split-not-tensor; mmap/PLE constraints apply; HIP env collisions (`TEMP`/`HOST`).
|
|
- **Memory**: `free` command lies about available memory (doesn't account for pinned/reserved weights).
|
|
- **PCIe**: M.2 x4 dock is ~29 GB/s; do not assume x16 bandwidth.
|
|
- **Skill Discovery**: `omp` scans `~/.omp/agent/skills/` only. New skills surface on the next session.
|