DeepReinforce Ornith-1.0 (dense 9B on Qwen 3.5, Q5_K_M, MIT), an
agentic-coding model. Tool-calls + <think> work under --jinja; native
256k so no YaRN. Loads at ~7.7GB VRAM @ 64k.
Benchmark (docs/2026-06-27-ornith-9b-benchmark.md): quality ties
gpt-oss-20b but gen is ~2.5-3x slower (dense 9B active vs gpt-oss MoE
3.6B active on the compute-bound P100). Default stays gpt-oss-20b;
ornith kept as a coding specialist in the menu.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Portainer's git checkout auto-creates a relative repo-file bind as a directory,
breaking the /app/config.yaml mount. Use the absolute host path like the share
stack; repo copy stays canonical, mirrored to /config/llm on deploy.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replace single llama-server with llama-swap so all benchmarked models are
selectable from Hermes' menu and hot-swapped on the one P100. Menu: gpt-oss-20b
(default, ~45s cold start), gemma-4-26b-a4b (MoE), gemma-4-12b, gemma-4-e4b.
qwen3-30b-a3b excluded (OOMs at 64k in 16GB). All 64k, q8/q8 KV, --parallel 1.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The q4_0 V-cache + flash-attention path is pathological on the GP100: 1.28
tok/s generation at 5-8% GPU util. q8_0 V-cache gives 9.2 tok/s and still fits
64k context in 16GB (15.3GB used, ~950MB free).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
llama-server caps the slot to the GGUF training context (32768) and ignores the
YaRN-extended size, leaving per-seq context at 32k. Raise qwen2.context_length
metadata to 65536 so the full window is served per request.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
With the default 4 slots, llama-server splits ctx into 32k per sequence, which
fails Hermes' 64K minimum. One slot serves the full 65536 per request (serial
agent use; concurrent calls queue).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Hermes Agent rejects models with <64K context. Qwen2.5-14B is 32k native, so
enable YaRN rope-scaling (2x → 65536) and drop the V-cache to q4_0 for VRAM
headroom on the 16GB P100.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The server-cuda image parses -fa as --flash-attn [on|off|auto], so a bare -fa
swallowed the following --cache-type-k as its value and crash-looped.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
New stacks/llm/ serves Qwen2.5-14B-Instruct (Q4_K_M GGUF) via llama.cpp's
OpenAI-compatible server on the Tesla P100 (CDI nvidia.com/gpu=0), published on
172.20.0.1:8090 for the host-side Hermes agent. vLLM was rejected: the P100
(cc 6.0) lacks the DP4A INT8 instructions its AWQ/GPTQ kernels need.
Includes design spec and implementation plan under docs/superpowers/.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
