Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2608.05944

Operating Multi-Node Full Fine-Tuning on NVIDIA B300: A Field Report on Telemetry-Based Triage, Negative Results, and Operational Hardening

Seon Ho Kim·Ui Jeong Jeon·Su Hyeon Kim·Min Tae Hwang
View original source ↗Read source document ↗Export metadata ↓

Abstract

We report operational experience full-fine-tuning a 32.76B-parameter dense model (Qwen3-32B) on 16 x NVIDIA B300 (two nodes, FSDP / ZeRO-3) -- among the first published field accounts on this accelerator. We claim no new algorithm. The individual mechanisms we use are established practice; our contribution is the integrated field experience and a set of calibrated measurements on new hardware. Concretely we offer four practitioner artifacts. (1) A B300-calibrated power-draw triage table that distinguishes compute / communication / data-starvation / checkpoint-or-deadlock / idle by board wattage (utilization% reads 100% during an NCCL hang). (2) A set of honest negative results that dispel common optimization folklore at this scale: a controlled A/B in which per-step NFS reading matches a pretokenized local cache (~53k tok/s) because the corpus fits in page cache and the job is compute-bound; and a reconstruction of an earlier "throughput collapse" as NFS/CPU contention rather than a storage-medium limit. (3) Calibrated 4/8/16-GPU strong-scaling and GPU-hour numbers on B300 (near-linear, as expected in this regime; we report absolute values as reference data). (4) A worked failure case -- an epoch-end NCCL deadlock from per-rank token-packing imbalance -- together with a 2.7-second pre-run invariant gate and an external watcher that turn multi-hour silent failures into instant rejections. This deadlock and its remedy correspond to PyTorch's documented Join / equalize-to-minimum practice; we position our instantiation against that prior art and report the GPU-hours the failure cost and the gate saves. The transferable takeaway is operational, not algorithmic: for data-dependent data-parallel jobs, watch power rather than utilization, and verify invariants before launch -- a passing smoke test is not evidence of a safe full run.

Explore related subjects

cs.DCcs.LG

Keep this discovery

BibTeXRIS

Seon Ho Kim, Ui Jeong Jeon, Su Hyeon Kim, Min Tae Hwang. 2026-08-06. Operating Multi-Node Full Fine-Tuning on NVIDIA B300: A Field Report on Telemetry-Based Triage, Negative Results, and Operational Hardening. https://arxiv.org/abs/2608.05944

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Online Treasure Hunt in Vertex-Permuted Dynamic Rings

We study the problem of treasure hunt by a group of $k \geq 1$ agents in vertex-permuted dynamic rings (VP). In this model, the $n$ vertices remain on a ring but are permuted at each time step. We first show that treasure hunt is impossible for any $k \leq n-3$ agents, if there are no restrictions on the sequence of permutations used in the dynamic ring. We then study the $VP(\delta)$ setting, in which for every pair $i, j$ of vertices, the edge $(i, j)$ is guaranteed to appear within $\delta$ steps. We show that the class $VP(\delta)$ is feasible only for $\delta \geq \left\lceil \frac{n-1}{2}\right\rceil$. For the one-agent case, we show a tight bound of $\Theta(\delta n)$ on the worst-case search time as well as competitive ratio of any online algorithm for treasure hunt, provided $\delta \geq 2n$. We then give an optimal algorithm for $k$ agents, thereby showing that $k$ agents can obtain a speedup of $k$ on the worst-case search time. Finally, in the R-VP setting, in which in every step, the vertices are arranged as a ring according to a random permutation, we show that treasure hunt takes expected $\Theta(n)$ steps against an oblivious adversary and $\Theta(n \log n)$ steps against an adaptive adversary.

cs.DC↗
arXiv2026

The Computing Channel: How Modulation Programs the Airwaves

Distributed computing and distributed artificial intelligence require frequent exchanges of intermediate results, although many applications need only an aggregate rather than messages from individual devices. Conventional systems recover each message before computing the aggregate, whereas over-the-air computation (OAC) exploits simultaneous transmission to obtain it directly. However, dominant OAC implementations rely on analog signaling, creating a mismatch with finite-precision data and digital communication procedures. This article presents digital function-oriented communication, in which finite-alphabet symbol representations and receiver decisions are jointly designed so that multiple-access superposition encodes the desired function without recovering individual inputs. We introduce its computational-constellation principle, main design approaches, extensions, and implementation challenges. Federated edge learning illustrates how the framework can reduce user-dependent data-bearing resources while operating directly on quantized model updates.

cs.DC↗
arXiv2026

Can AI Remediate Backend Failures Safely? GuardedAct with Blast-Radius-Aware Sandboxing

Large Language Models (LLMs) have shown promising capabilities in generating remediation actions for microservice failures. However, directly executing AI-generated repair actions in production risks cascading collateral damage. We propose GuardedAct, a sandbox-first remediation framework that interposes a blast-radius-aware verification layer between the LLM action generator and the production environment. GuardedAct operates in four phases: (1) ingesting a diagnosis report together with the live system topology and recent telemetry, (2) prompting an LLM to produce a ranked list of candidate remediation actions, (3) simulating each action in a lightweight digital-twin sandbox that estimates the blast radius and assigns a risk label, and (4) enforcing a rollback-confidence gate that auto-executes only low-risk actions while escalating high-risk ones for human review. We evaluate GuardedAct on five fault scenarios injected into the DeathStarBench social-network application. Experimental results show that GuardedAct achieves an overall recovery rate of 87.4% while reducing collateral damage by 79.7% relative to direct LLM execution (from 25.6% to 5.2%), at the cost of a modest sandbox-induced increase in mean time to recovery (approximately 8 s). Ablation studies confirm that each component contributes meaningfully to the safety-speed trade-off.

cs.DC↗