Spark Training Gotchas

DGX Spark’s GB10 chip (Grace Blackwell, SM121, 128GB unified memory, aarch64) has ten recurring failure modes across launch, memory, thermals, bandwidth, and precision. Each is named G1–G10 so it can be checked by number — the numbering is load-bearing for tooling that runs these checks. Read this before a long run, not after hour six.

When to Use This Skill

• A training run fails to start, with an import error or a segfault that doesn’t point at the real cause.
• A run OOMs while nvidia-smi still shows headroom.
• Throughput degrades partway through a run that started fine.
• Before any multi-hour or multi-epoch job on GB10.
• Wiring two Sparks together, before picking a parallelism strategy.
• Choosing between FP8 and NVFP4 for a Spark-hosted run.

Common Issues Quick Reference

The Ten Gotchas

G1: CUDA 12/13 ABI Mismatch

• SYMPTOM: ImportError: undefined symbol naming a CUDA function, or a segfault on the first .cuda() call.
• CAUSE: most PyPI wheels link libcudart.so.12; Spark ships CUDA 13. pip never checks CUDA ABI, so it surfaces only at import or first kernel launch.
• CHECK: references/gotcha-checks.md G1 — the wheel’s CUDA build tag.
• FIX: reinstall from download.pytorch.org/whl/cu130 or use a matched container.

G2: flash-attn — Skip the pip Build, Watch Unsloth’s Auto-Detect

• SYMPTOM: pip install flash-attn still fails/hangs. Unsloth may also silently train flash-attn over an explicitly requested SDPA.
• CAUSE: no aarch64/sm_121 wheel for bare pip — but NGC containers ship a working SM121 flash-attn, and Unsloth auto-prefers it, dropping attn_implementation="sdpa".
• CHECK: references/gotcha-checks.md G2 — is flash-attn already present and working.
• FIX: bare pip — skip flash-attn, use SDPA (unchanged). On NGC — the only reliable override is the monkeypatch in references/gotcha-checks.md G2.

G3: UMA OOM Below 128GB

• SYMPTOM: OOM during model load/training while nvidia-smi still reports free memory under the 128GB cap — or, on some setups, [N/A] outright instead of a number.
• CAUSE: mmap and the CUDA allocator double-count pages during safetensors load; QLoRA can OOM earlier than bf16 since dequantization adds transient allocs.
• CHECK: references/gotcha-checks.md G3 — read free -g and /proc/meminfo, not nvidia-smi.
• FIX: drop the page cache with sync; echo 3 > /proc/sys/vm/drop_caches — needs root, a between-run reset, not a mid-training step.

G4: Thermal Throttling

• SYMPTOM: throughput drops partway through a multi-hour run, or the box spontaneously reboots under sustained load.
• CAUSE: sustained power draw caps around 100W versus the 240W rated figure; long runs push into that ceiling and throttle or, sometimes, reboot.
• CHECK: references/gotcha-checks.md G4 — sample nvidia-smi --query-gpu=temperature.gpu,power.draw.
• FIX: if power plateaus under 240W while temperature climbs, treat throttling as the cause; improve cooling or cap run length.

G5: Bandwidth Ceiling

• SYMPTOM: memory-bound workloads, decode-heavy RL loops especially, plateau well below expected throughput.
• CAUSE: 273 GB/s is a spec ceiling, not sustained; measured bandwidth runs 180–192 GB/s.
• CHECK: references/gotcha-checks.md G5 — observed step time vs. the measured range, not spec.
• FIX: budget throughput from 180–192 GB/s; revise a plan built on the 273 GB/s figure.

G6: Global UMA Resource Contention

• SYMPTOM: a process’s KV cache/weights get evicted mid-run silently, no OOM in its own logs.
• CAUSE: unified memory is one global pool; an uncapped or near-capacity process competes with anything else and can evict it. A small, bounded workload doesn’t — a <4GB LoRA coexists fine alongside vLLM capped at gpu-memory-utilization<=0.5.
• CHECK: references/gotcha-checks.md G6 — other GPU-resident processes and whether capped.
• FIX: the one-heavy-job rule applies to uncapped or near-capacity workloads — cap or stop unrelated servers first. A small, capped workload need not stop.

G7: NVFP4 Slower Than FP8 on SM121

• SYMPTOM: switching an inference workload from FP8 to NVFP4 on Spark makes it slower, not faster.
• CAUSE: SM121 lacks cvt.e2m1x2 unless kernels target sm_121a; NVFP4 runs ~32% slower without it.
• CHECK: references/gotcha-checks.md G7 — capability reports (12, 1); does the build target sm_121a?
• FIX: stay on FP8 unless the build targets sm_121a.

G8: Stale Official Playbooks

• SYMPTOM: following an official DGX Spark playbook still fails, with no local misconfiguration explaining it.
• CAUSE: official playbooks have shipped broken before; the stack moves faster than the docs.
• CHECK: references/gotcha-checks.md G8 — the playbook repo’s recent issues.
• FIX: check github.com/NVIDIA/dgx-spark-playbooks issues before trusting a recipe for an expensive run.

G9: Container-First, Not Bare Pip

• SYMPTOM: a bare-pip environment that worked yesterday breaks after an unrelated pip install, or two “identical” environments behave differently.
• CAUSE: bare pip lets Triton, xformers, and transformers drift independently; nothing pins them to GB10’s SM121 target.
• CHECK: references/gotcha-checks.md G9 — container or bare pip?
• FIX: prefer an NGC container (see spark-environment-setup for tag guidance) or Unsloth’s container. If bare pip is unavoidable, follow the NVIDIA install order, including --no-deps on Unsloth.

G10: Dual-Spark Is DDP/FSDP Only

• SYMPTOM: a tensor-parallel launch across two Sparks hangs, runs far slower than single-Spark, or errors out.
• CAUSE: ConnectX-7 is fast enough for gradient/parameter sync (DDP, FSDP) but too thin for TP’s fine-grained traffic.
• CHECK: references/gotcha-checks.md G10 — the configured parallelism strategy.
• FIX: on a two-Spark setup, choose DDP or FSDP, never tensor parallelism — TP is single-node only here.

Fast Triage

The cheapest checks to run before anything else:

python3 -c "import torch; print(torch.version.cuda)"  # expect 13.x (G1); NGC builds have no +cu130 tag — that's not a failure

import torch; print(torch.cuda.get_device_capability())  # expect (12, 1) (G7)

{ [ -f /.dockerenv -o -f /run/.containerenv ] || grep -qE 'docker|containerd' /proc/1/cgroup; } 2>/dev/null && echo container || echo unknown  # G9

assets/preflight.sh runs G1, G3, G4, G7, G9 and produces one output line per gotcha in a fixed format: G-number first, then PASS/FAIL/WARN where automatable, SKIP when unavailable, or INFO: for a raw reading (G3, G4). Full commands: references/gotcha-checks.md. See also spark-environment-setup for the environment assumed working.