Files
serverAppConfigs/dankclaw/hermes/config.yaml.example
T
Frank 4fca739bf8 frank-medium: deploy two single-GPU replicas — TP=2 is blocked by PCIe atomics
The radiance MXFP4 stack cannot run TP=2 on this host. RCCL's collective
kernels carry a hidden hostcall-buffer parameter, and ROCr refuses to
dispatch them on a device without PCIe atomics: the AQL dispatch fails with
hipErrorIllegalState ("the operation cannot be performed in the present
state"), which is what every two-rank collective hit. The second R9700 sits
behind the chipset bridges 05:00.0/06:09.0, which advertise
`AtomicOpsCap: Routing-` — a read-only capability, so no software can enable
atomic routing on that path and the driver's
pci_enable_atomic_ops_to_root() fails for that card. The first card, on CPU
root port 00:01.1, has atomics enabled.

Ruled out: kernel version (7.0.0-28/-30/-31 identical, after a full upgrade),
iommu=pt, P2P hardware (45.3 GB/s between the cards), the shipped librccl (a
from-source rebuild reproduces the failure at the same call path), vLLM and
pynccl (a bare two-rank torch.distributed allreduce fails the same way), all
NCCL transport knobs, and single-rank tests (NCCL short-circuits at
world_size=1, so one-card success proves nothing).

Deployment: three rootless Quadlets — frank-medium-a (:7771, card 1),
frank-medium-b (:7772, card 2), both TP=1 at 114688 context (262144 and
131072 do not fit one card's KV) with vLLM-profiled KV — and
frank-medium-proxy (:7770), a round-robin that preserves the fleet contract
of one endpoint and one served model name. frank-medium.container stays as
the TP=2 target for a host whose second card is atomic-capable; the deploy
handler does not install it and trashes any stale copy.

Measured through the proxy (vllm bench serve, random dataset, --ignore-eos):
conc 1 42.8 tok/s, conc 4 147.6 tok/s, conc 8 230.8 tok/s, 32k prefill at
~2.8k tok/s per card, TPOT p50 23.6-28.8 ms.

Also: healthcheck reads PORT so one probe serves all three units, the hermes
example and the harness models.yml drop to the real 114688 context, and the
runbook records the full evidence chain.
2026-09-16 03:47:11 -07:00

101 lines
3.4 KiB
Plaintext

# Hermes configuration backup / non-secret baseline.
#
# Source on Dankclaw: /home/dank/Containers/hermes/data/config.yaml
# This file is a sanitized Git backup, not a complete replacement for the
# live Hermes state. Do not add dashboard password hashes, phone numbers,
# OAuth credentials, API keys, sessions, memories, or generated state here.
# The live directory is bind-mounted into the containers as /data.
model:
provider: dankclaw-large
base_url: http://192.168.0.143:7770/v1
api_key: not-needed
default: frank-large
providers:
dankclaw-large:
api: http://192.168.0.143:7770/v1
discover_models: false
models:
frank-large:
context_length: 114688
dankserver-small:
api: http://192.168.0.235:7770/v1
discover_models: false
models:
frank-medium:
context_length: 114688
# Auxiliary routing from the live configuration. Named providers route via
# their `api:` URL; `custom` entries route by their explicit `base_url`.
auxiliary:
vision: {provider: dankclaw-large, model: frank-large}
web_extract: {provider: dankserver-small, model: frank-medium}
compression: {provider: dankserver-small, model: frank-medium}
skills_hub: {provider: dankserver-small, model: frank-medium}
approval: {provider: dankserver-small, model: frank-medium}
mcp: {provider: dankserver-small, model: frank-medium}
# Title generation goes out over the Responses transport on purpose: Hermes
# hardcodes response_format=json_schema on this one call, no local back-end
# implements constrained decoding (it is a 400), and this transport is the
# only one that drops the field. reasoning.effort=none is required with it —
# the call caps output at 64 tokens, which a thinking pass would consume
# entirely, leaving an empty title.
title_generation:
provider: dankserver-small
model: frank-medium
api_mode: codex_responses
extra_body:
reasoning:
effort: none
memory_query_rewrite: {provider: custom, model: frank-medium, base_url: http://192.168.0.235:7770/v1}
tts_audio_tags: {provider: custom, model: frank-medium, base_url: http://192.168.0.235:7770/v1}
triage_specifier: {provider: dankclaw-large, model: frank-large}
kanban_decomposer: {provider: dankclaw-large, model: frank-large}
profile_describer: {provider: dankserver-small, model: frank-medium}
curator: {provider: dankserver-small, model: frank-medium}
dashboard:
# basic_auth intentionally omitted; its password_hash remains live-only.
font: atkinson-hyperlegible
moa:
presets:
default:
reference_models:
- provider: openrouter
model: deepseek/deepseek-v4-pro
enabled: false
degraded_reference_policy: loud
fanout: user_turn
reference_models:
- provider: openrouter
model: deepseek/deepseek-v4-pro
enabled: false
aggregator:
provider: openrouter
model: anthropic/claude-opus-4.8
degraded_reference_policy: loud
max_tokens: 4096
fanout: user_turn
enabled: true
skills:
disabled: []
onboarding:
seen:
busy_input_prompt: true
_config_version: 34
platforms:
signal:
# Live config has the Signal adapter disabled; http_url points at the
# host signal-cli daemon. The account phone number stays live-only.
enabled: false
extra:
http_url: http://192.168.0.143:8080
mcp_servers:
chrome-devtools:
command: /usr/local/bin/npx
args:
- -y
- chrome-devtools-mcp@latest
- --browserUrl=http://127.0.0.1:9222