Fresh install doesn't work

[yundddd]

Hello. I'm following the installation guide and there are couple issues I'm working through.
1: It looks like the guard only accepts two pins:

Genesis pin-gate] STRICT FAIL: vllm pin '0.20.2' is NOT on the Genesis known-good list (2 entries). Allowed pins: ['0.20.1rc1.dev16+g7a1eb8ac2', '0.20.2rc1.dev9+g01d4d1ad3']. To accept this pin, add it to KNOWN_GOOD_VLLM_PINS in vllm/_genesis/guards.py and document the validation in CHANGELOG.

Am I supposed to not use nightly as it's way ahead? I'm confused how we reproduce results with a nightly image.

2: The mounting point for _genesis in startup script is at

v /home/tim/genesis-vllm-patches/vllm/_genesis:/usr/local/lib/python3.12/dist-packages/vllm/_genesis:ro \

Looking at vllm/vllm-openai:v0.20.2 (which I assume is close to the image you pinned), it actually uses /site-packages/ instead of /dist-packages.

3: I guess this -network genesis-vllm-patches_default can be removed, aside from other changes (scripts/start_27b_int4_TQ_k8v4.sh)?

I'm also a bit confused for the oneline install script.

curl -sSL https://raw.githubusercontent.com/Sandermage/genesis-vllm-patches/main/install.sh | bash

Genesis vLLM Patches — installer
https://github.com/Sandermage/genesis-vllm-patches


» Pre-flight checks
  ok: OS: Linux
  ok: Python: 3.14 (/usr/bin/python3)
  ok: git + curl available

» GPU detection
  ok: GPU: NVIDIA GeForce RTX 5090 × 1
      matched preset GPU class: rtx 5090

» vLLM detection
  ok: vllm: ?

» Container-runtime caveat probe (Proxmox VE)
  ok: no Proxmox VE markers — docker path should be safe

» Pick workload
  ok: workload: balanced (non-interactive default — re-run with --workload to change)

» Resolve Genesis pin
  ok: pin: v7.51-stable-2026-04-27 (latest stable tag)

» Genesis source (/home/tim/.genesis)
info: found existing clone — updating
  ok: Genesis at a8c202d (ref: v7.51-stable-2026-04-27)
 err: Genesis tree at v7.51-stable-2026-04-27 is missing vllm/_genesis/compat/cli.py — wrong pin?

It looks like v7.51 tag was behind some refactor on master and is missing the cli.py.
Is there any reason to clone the repo again to ./genesis if we could just checkout different branches locally?

I'm interested in providing some data points on Ubuntu and in WSL but there are some rough edges here I need to overcome. Thanks.

[Sandermage]

@yundddd — thanks for the report, and sorry for the slow turnaround. This is now fixed.

The repo has been updated to the v12 (sndr-platform) tree, and the specific things you hit are resolved.

The pin gate (your STRICT FAIL). That allowlist was 2 stale pins; it now tracks the live nightly — the current canonical pin is 0.23.1rc1.dev424+g3f5a1e173 (image vllm/vllm-openai:nightly-3f5a1e1733200760169ff31ebe60a271072b199e), with 0.23.1rc1.dev301+g04c2a8dea kept as the rollback pin. Also worth knowing: the STRICT FAIL … exit(2) you saw is not the default — the default policy is warn (it logs a warning and boots; patches whose anchors drifted on an unknown pin self-SKIP, visible in the startup log). The hard exit only fires when something sets GENESIS_VLLM_PIN_POLICY=strict (our prod start scripts set it on purpose, to fail fast on accidental drift). So a newer/unknown nightly just warns by default — you don't have to be on an allowlisted pin to try one.

The installer — you were right it was broken. Two issues, both now fixed:

• The curl … | bash install.sh still referenced the retired vllm/sndr_core/ (and older vllm/_genesis/) namespace — the code moved to the top-level sndr/ package in v12, so the sanity check died with the exact missing … cli.py — wrong pin? you saw. install.sh now checks the correct sndr/ paths, installs via pip install --no-deps -e . (which registers the vllm.general_plugins entry-point and puts sndr on sys.path), and calls python3 -m sndr.compat.cli.
• --pin stable resolved to the old pre-refactor v7.51-stable-2026-04-27 tag (and there was no v12 tag). --pin stable is now v12-aware, and a v12.0.0 release tag is published — so the default one-liner resolves to a real, reproducible v12 target.

Fresh install (works now)

One-liner:

curl -fsSL https://raw.githubusercontent.com/Sandermage/genesis-vllm-patches/main/install.sh | bash
# pinned release:  ... | bash -s -- --pin v12.0.0
# bleeding tip:    ... | bash -s -- --pin main

Manual (also in docs/INSTALL.md § "Step-by-step setup"):

python3.12 -m venv .venv && source .venv/bin/activate
pip install --upgrade pip wheel setuptools
pip install --pre vllm==0.23.1rc1.dev424+g3f5a1e173 \
  --extra-index-url https://wheels.vllm.ai/nightly
git clone https://github.com/Sandermage/genesis-vllm-patches.git
cd genesis-vllm-patches
pip install --no-deps -e .
pip install pandas scipy xxhash         # runtime extras
python3 -c "import sndr; print(sndr.__version__)"
sndr doctor
sndr launch prod-qwen3.6-35b-balanced --dry-run   # inspect, then drop --dry-run to boot

Thanks again for flagging it — the namespace migration genuinely hadn't been carried into install.sh. If the one-liner still trips on your setup, paste the output and I'll dig in.