flare-2026-11-16-upgrade.md 9.6 KB

FLARE: fleet upgrade to main@6b68b9b4 and the corrections it forces (2026-11-16)

Written after rebuilding both lanes and brain-checking the old claims against the running system. This document supersedes specific statements in flare-bringup.md, flare-storage-nas.md and flare-roce-poison.md; those files stay as history (they describe the October stand) but where they disagree with this page, this page was measured on 2026-11-16.

1. What is deployed

prod lane sparks lane
nodes dgx1→dgx4, cluster sparks, group deepseek1017 dgx5→dgx8, cluster dgx58, group ds41-tp4
controller gx10-dgx1 gx10-dgx5
API http://dgx1.lan:30000/v1, alias deepseek-v4.1-flash http://dgx5.lan:30000/v1, same alias
core db20209d361f473b8452f24ad517fa0d-33966105 (ABI 10) edb5231cb9b847b7914a6475a5c02b3d-24324121 (ABI 14)
model checkpoint-expanded-scales (local-inference-lab) JMNI-Labs/DeepSeek-V4.1-Flash-1017-FLARE@gb10-tp4

Both lanes run one identical binary, sha256 3cd3f442ccb9b499f8469be1f878ee49541f6edfab41db5c140495aa257aa178 (26,120,544 bytes) at /usr/bin/flare and ~/.local/lib/flare-agent/flare.

Source: NVIDIA/hyperflare-inference main 6b68b9b4 (2026-11-15T18:14:40Z, the v0.2.1 line). Built on dgx5:

git clone --depth 1 https://github.com/NVIDIA/hyperflare-inference ~/hyperflare-src
cd ~/hyperflare-src
cargo build --release --locked --manifest-path flare-cli/Cargo.toml --features roce
# -> flare-cli/target/release/flare

2. Corrections to our earlier claims

Each row was re-tested on 2026-11-16; "before" is what our docs/config said.

# earlier claim measured now
1 build with --manifest-path flare/Cargo.toml, binary flare/target/release/flare, "~6 min, 20 MB" the serving binary is the flare-cli crate: --manifest-path flare-cli/Cargo.toml, output flare-cli/target/release/flare. Pass --features roce (harmless on aarch64 — it is on by default there — required on x86_64). ~21 min cold with 344 crates
2 flare --version prints 0.1.0 for every build a source build prints flare 0.0.0 (the crate carries no version; tags stamp release builds only). Never use --version as the identity test — use sha256sum
3 PR #184 detection: strings <bin> \| grep -c "embedding read worker panicked" → 1 obsolete: current main contains no such string (0 in both PPA and source builds), because the disk-Engram reader was rewritten. The marker cannot distinguish builds any more
4 --no-embedding-table-prefetch is part of the publisher's TP4 recipe removed from the CLI. Disk prefetch is now native on Linux and there is no switch; --embedding-table-workers exists but is Windows-only. Launching with the publisher's launch.sh verbatim fails on this flag
5 cache layout is packages/ and paths.package_cache installed cores live in cores/<core_id>-<bytes>/, prepared models and peer shares in models/<model>/<version>/, hub transport in transport/hub/v1/. package_cache is still accepted as an alias for core_cache, so an old config.toml silently redirects the store
6 FLARE rejects reasoning_effort / chat_template_kwargs; thinking is fixed by the package false from this build on. reasoning_effort: high returns reasoning_content; chat_template_kwargs.enable_thinking=false returns content with no reasoning. Verified on both lanes
7 a forced named tool_choice with thinking off returns HTTP 500 (TC-45) fixed: returns 200 with a correct tool_calls[0] = get_weather {"city":"Berlin"}
8 packages record "proxy ABI 9/10/11" and the runtime must match runtime PROXY_ABI = 14 and it loads cores recording 9, 10, 11 or 14. Proven live: the prod core (ABI 10, prepared by the October ignition) serves on the ABI-14 runtime
9 rank processes survive an agent restart ranks are children of the local agent: restarting the unit kills them. What survives is controller loss (control plane only). Both remain true simultaneously — do not conflate them
10 version directories are stable identities every acquisition re-stamps the version directory (gb10-tp4-202611152102 → …2357 on re-acquire). Never hard-code a version path in a group definition; re-point it after a re-acquire

3. Operational facts worth keeping

  • Mixed builds cannot pair. A new CLI against old agents fails immediately with handshake read failed: failed to fill whole buffer. The swap must therefore be one window: stop the group → swap both paths on every node → restart the agents → relaunch with the new CLI.
  • --gpus 0 on orbit run is not "the node's GPU 0"; it is read as a GPU count and fails with core has no tensor-parallel size 1 deployment (available: [4]). Omit the flag and let the cluster's node count set tensor parallelism.
  • Hub transport keeps the payload one time. models/ entries are hardlinks into transport/hub/v1/blobs (inode-shared: 196 of 197 files, 480.1 GiB of 480.1 GiB shared). A re-acquire after a broken tree re-hashes the local blobs (~950 MB/s, NIC ~0) instead of re-downloading. Copy such trees with -aH.
  • Installing a core is a copy, not a hardlink: cores/<id>-<bytes> held its own inodes (hardlinks=1) next to the download store, so a controller that keeps both needs the payload twice.
  • Both lanes are --persistent (since 2026-11-16). The definition lives in the controller's orbit.toml ([[group]]: core, serving flags, rank-zero flags, rank-ordered nodes, transfer streams, publisher keys) and survives stop, history expiry and reboot. Restart is then one short command — no flags:

    ssh dgx1.lan 'flare orbit run --name deepseek1017 --detach'   # prod, dgx1-4
    ssh dgx5.lan 'flare orbit run --name ds41-tp4 --detach'       # sparks, dgx5-8
    

Persistence does not start anything after a reboot; it only removes the need to remember the flag set. --dangerously-approve-key / --dangerously-accept-publisher are not stored in the definition (trust lives in trust.toml), which is why the saved group starts without them. An override on a later run applies to that launch only unless --persistent is passed again.

4. The 2026-11-16 prod window (record)

dgx1–dgx4 were moved from e3fe697eab1f… (build of 2026-11-04) to 3cd3f442… without changing flags, core or publisher trust:

event UTC
orbit stop deepseek1017 (drained, not killed) 10:44:05
binaries swapped on all four (host-scripts/flare-swap.sh, FORCE=1) ~10:44:2x
group relaunched with the identical flag set 10:44:50
/health 200 10:45:23

Total outage ≈ 78 s. Verification after: /v1/models (deepseek-v4.1-flash, 1,048,576 ctx), pi completion via dgx1/deepseek-v4.1-flash, forced tool call 200, 8 concurrent requests all 200 (1.0–2.5 s). Rollback binaries kept as flare.orig-20261116 at both paths on each node.

Follow-up: both lanes made persistent (same day)

deepseek1017 was relaunched with --persistent (10:49:54→10:50:28Z, ~34 s), then stopped and restored with the short form orbit run --name deepseek1017 --detach to prove the restore path (10:50:33→10:51:27Z, ~54 s). The restored rank command line is flag-for-flag identical to the pre-window one. ds41-tp4 was launched persistent from the start.

5. Apt decoupling (2026-11-16)

The upstream PPA (https://ppa.local-inference-lab.ai noble/main, keyring /etc/apt/keyrings/lil-flare-archive.asc, package flare-inference) shipped exactly one payload: /usr/bin/flare. We overwrite that file with our source build, so any apt upgrade would have replaced the CLI while the agent kept the old inode — the mixed-build handshake failure, delivered by apt. There is also no way for apt-get install -y --only-upgrade flare-inference to be "harmless": the package owns the path.

All eight nodes were decoupled the same way:

cp -a /usr/bin/flare /tmp/flare-keep                  # our build
sudo apt-get purge -y flare-inference                # removes the file it owns
sudo install -m 0755 /tmp/flare-keep /usr/bin/flare   # ours, now unowned
sudo rm -f /etc/apt/sources.list.d/flare.sources \
           /etc/apt/keyrings/lil-flare-archive.asc

Result on every node: dpkg -S /usr/bin/flare → no path found, no flare apt source, no LIL key, apt-get update clean, /usr/bin/flare and ~/.local/lib/flare-agent/flare both 3cd3f442…. Unattended upgrades are disabled anyway (APT::Periodic::Unattended-Upgrade "0"), so the fleet now has no path by which an upstream build arrives uninvited: upgrades are manual, /usr/local/share/hyperflare/v0.2.1 (staged on all eight nodes, with the bundled LICENSE/NOTICE/THIRD_PARTY_NOTICES replacing the purged package docs), and rollback is the single .orig-20261116 generation.

Rebuilds are now entirely ours. Bumping upstream means: watch the repo, rebuild, stage a new vX.Y.Z, swap with host-scripts/flare-swap.sh, restart the agents together, relaunch both persistent groups.

6. Not fixed by this upgrade

  • dgx4 still has the KingSpec XG7000-2TB 2242 whose stop-and-go prefill poisoned RoCE on the October package — unchanged by a runtime upgrade; see flare-roce-poison.md. The disk-bound Engram behaviour it causes is a hardware property, not a build property.
  • TP4 disk Engram still reads ~189 GiB of tables through a 1 GiB row cache; the sparks-lane measurement (--embedding-table-resident-scales, disk) fits ~10.5 GiB of GPU headroom at c32, i.e. no change in the KV budget.