Frank
4fca739bf8
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.
99 lines
4.9 KiB
INI
99 lines
4.9 KiB
INI
# Quadlet unit: frank-medium replica B (rootless, boot-persistent).
|
|
#
|
|
# Install path on dankserver: ~/.config/containers/systemd/frank-medium-b.container
|
|
# Unit name: frank-medium-b.service (container name: frank-medium-a)
|
|
#
|
|
# WHY TWO REPLICAS INSTEAD OF TP=2. The intended shape for frank-medium is one
|
|
# tensor-parallel process across both R9700s (see frank-medium.container, kept
|
|
# as the target). That shape cannot run on this host: RCCL's collective kernels
|
|
# carry a hidden hostcall-buffer parameter, and the ROCr runtime refuses to
|
|
# dispatch them on a device without PCIe atomics ("Pcie atomics not enabled,
|
|
# hostcall not supported" -> AQL dispatch failed -> hipErrorIllegalState). 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
|
|
# atomics on that path; the amdgpu driver's pci_enable_atomic_ops_to_root()
|
|
# therefore fails for that GPU and KFD marks the CPU<->GPU2 link NO_ATOMICS.
|
|
# The first R9700 sits on CPU root port 00:01.1 (Routing+) and does have
|
|
# atomics. Result: every collective that touches GPU 2 dies at launch.
|
|
# Two independent single-GPU replicas use both cards instead; the proxy unit
|
|
# keeps the single :7770 endpoint. Full evidence: the frank-medium runbook.
|
|
#
|
|
# SHAPE. vLLM binds 0.0.0.0:7772 on the host network directly; no PublishPort.
|
|
# MAXLEN is 114688, not the TP=2 deployment's 262144: one card carries the
|
|
# whole model (~20 GiB of weights + drafter), leaving ~5.2 GiB of KV, and
|
|
# 131072 does not fit (measured: needs 5.58 GiB). KV_MEM=0 lets vLLM profile
|
|
# this shape instead of using the 2x7551-32624 pin, which is only valid at
|
|
# TP=2/262144.
|
|
#
|
|
# DEVICES. /dev/kfd + /dev/dri; HIP_VISIBLE_DEVICES=1 (the second R9700). The
|
|
# 2 GiB Raphael iGPU enumerates as index 2 and is excluded by the pin.
|
|
# Only HIP_VISIBLE_DEVICES is set, never ROCR_VISIBLE_DEVICES: ROCR filters
|
|
# first and HIP indexes into the filtered list, so a single-valued pair
|
|
# (ROCR=1,HIP=1) leaves an empty device set.
|
|
#
|
|
# DO NOT name an environment variable TEMP, HOST or GPU_MAX_HEAP_SIZE: the
|
|
# HIP runtime reads bare TEMP for its JIT cache path (2026-08-13 incident).
|
|
[Unit]
|
|
Description=frank-medium replica B — vllm-radiance MXFP4 Qwen3.8-27B on R9700 #2
|
|
After=network-online.target
|
|
|
|
[Container]
|
|
ContainerName=frank-medium-b
|
|
Image=localhost/radiance-vllm:0.9.3
|
|
AddDevice=/dev/kfd
|
|
AddDevice=/dev/dri
|
|
GroupAdd=keep-groups
|
|
# The upstream launcher's exact podman posture (rootless).
|
|
PodmanArgs=--privileged --ipc=host --security-opt seccomp=unconfined --cap-add SYS_PTRACE --network=host
|
|
Environment=HIP_VISIBLE_DEVICES=1
|
|
Environment=SNAP=/models/Qwen3.8-27B-Uncensored-MXFP4-awq
|
|
Environment=DRAFTER=/models/Qwen3.8-27B-DFlash2-FP8
|
|
Environment=CHAT_TEMPLATE=/patches/qwen-fixed-v22.3.jinja
|
|
Environment=SERVED_NAMES=frank-medium
|
|
Environment=PORT=7772
|
|
Environment=TP=1
|
|
Environment=MAXSEQS=4
|
|
Environment=CHUNK=8192
|
|
Environment=MAXLEN=114688
|
|
Environment=GPU_UTIL=0.98
|
|
Environment=KV_MEM=0
|
|
Environment=SPEC_METHOD=dflash
|
|
Environment=SPEC=7
|
|
# Checkpoints (rw: nothing downloads at runtime, but vLLM mmaps the shards).
|
|
Volume=/home/dank/models:/models
|
|
# Tokenizer files ride the HF cache; HF_HUB_OFFLINE=1 is set in the entrypoint.
|
|
Volume=/home/dank/.cache/huggingface:/root/.cache/huggingface
|
|
# Compile cache. Shared with the other replica and with the TP=2 target: the
|
|
# entries are content-addressed, so shapes simply add subdirectories.
|
|
Volume=/home/dank/.radiance-cache-w4a8-093-gdnm-nqft-fp8s-gnq:/cache
|
|
# The pinned radiance clone: patches applied at every container start.
|
|
Volume=/home/dank/GitPackages/radiance-vllm-mxfp4:/patches:z
|
|
# The patched libr4d (pin b9e42ab + repo extras, key b9e42ab-rx6). The image's
|
|
# shipped r4d.so predates the GDN overflow fix and NaNs this model.
|
|
Volume=/home/dank/.cache/radiance-libr4d/b9e42ab-rx6:/r4d:z
|
|
# Entrypoint + healthcheck from this checkout, mounted read-only.
|
|
Volume=/home/dank/serverAppConfigs/dankserver/frank-medium/radiance-entrypoint.sh:/usr/local/bin/frank-medium-entrypoint.sh:ro
|
|
Volume=/home/dank/serverAppConfigs/dankserver/frank-medium/frank-medium-healthcheck.py:/usr/local/bin/frank-medium-healthcheck.py:ro
|
|
# The entrypoint must be the container command; the image's own entrypoint
|
|
# (which the patched script execs at the end) must not run unpatched.
|
|
PodmanArgs=--entrypoint /usr/local/bin/frank-medium-entrypoint.sh
|
|
# 200 only once the engine serves; during first-start compilation the
|
|
# container is legitimately unhealthy. HealthCmd is split on spaces into an
|
|
# argv array by the podman 5.x v1 generator, hence the file-based probe.
|
|
HealthCmd=python3 /usr/local/bin/frank-medium-healthcheck.py
|
|
HealthInterval=60s
|
|
HealthTimeout=30s
|
|
HealthRetries=3
|
|
HealthStartPeriod=1800s
|
|
Pull=never
|
|
|
|
[Service]
|
|
# First start compiles Triton/inductor kernels (several minutes) before the
|
|
# engine loads ~20 GiB of weights.
|
|
TimeoutStartSec=1800
|
|
Restart=on-failure
|
|
RestartSec=10
|
|
|
|
[Install]
|
|
WantedBy=default.target
|