Spark Memory Thermal Ops

DGX Spark’s GB10 chip has one 128GB unified memory (UMA) pool shared by CPU and GPU, and a sustained power ceiling well below its rated figure. Both break discrete-GPU assumptions: headroom isn’t what nvidia-smi reports, and a run that starts fast will slow down mid-job with nothing misconfigured. This skill covers planning memory headroom, working an actual OOM, and watching thermals across a long job. For launch-time failure modes (ABI mismatches, flash-attn, playbook breakage), see spark-training-gotchas — this skill assumes the job starts.

Common Issues Quick Reference

When to Use This Skill

• Sizing a training run against the 128GB pool before launch — will this model, method, and batch/pack combination fit.
• A run OOMs mid-load or mid-step and the remediation order matters — what to try first, second, third.
• Watching temperature and power during a multi-hour job, deciding whether a slowdown is thermal throttling or something else.
• Planning to run a trainer alongside an inference server (vLLM, Ollama) on the same box.

UMA Memory Model

Spark has no separate GPU VRAM — the GPU and CPU share one 128GB pool. Two consequences:

• nvidia-smi and cudaMemGetInfo underreport pressure — or report nothing at all. Both report CUDA-allocator-visible memory, not the pool’s actual state — a box can show headroom in nvidia-smi and still OOM, because page-cache and mmap’d pages the allocator doesn’t see consume the same pool. On some driver/setups, the memory query returns [N/A], [N/A] outright instead of a number — a script grepping for a numeric value there gets nothing, not a misleading undercount (see spark-training-gotchas gotcha G3).

• Model load is a transient peak, not the steady state. Loading safetensors weights mmaps the file, then copies into CUDA tensors — for a window during load, both the mmap’d pages and the CUDA copy count against the pool at once. A model that fits while training can still OOM during load if headroom was sized for the post-load footprint instead of this doubled transient.

Plan and diagnose with free -g, not nvidia-smi:

free -g | awk 'NR==2 {print "free:", $4, "GB"}'

Rule of thumb: take that free figure, subtract a few GB for OS/driver overhead, and budget against the result — not the 128GB spec number. The worksheet in references/uma-accounting.md accepts parameter count, dtype, and method as input, and returns a memory estimate to compare against known anchors.

Planning Sequence

Before launch, work through these in order:

1. Read free -g; subtract OS/driver overhead for the budget.
2. Estimate weights + optimizer + gradients + activations from references/uma-accounting.md.
3. Compare against the closest anchor (70B QLoRA, 27B LoRA, 9B full FT), not the estimate alone.
4. If the estimate is close to the budget, start with shorter packing or a smaller batch — cheaper than hitting the OOM Ladder mid-run.

Example: Sizing a 70B QLoRA Run

A sanity check of the worksheet formula against the ≈40GB anchor:

params = 70e9
weights_gb = params * 0.5 / 1e9      # NF4, step 1
adapter_gb = 0.5                     # step 5, negligible
total_gb = weights_gb + adapter_gb   # + activations
print(f"{total_gb:.0f}GB before activations")

Weights alone land near the ≈40GB anchor — a plan estimating far above that for the same model class is a signal to recheck dtype and method.

The OOM Ladder

When a job OOMs on unified memory, work this ladder in order. Each step is more disruptive than the last — don’t skip ahead: reducing batch size is never step 1.

1. Flush the buffer cache. Page cache from a previous run or a large dataset read often accounts for GB of the “missing” headroom. This costs nothing but a rerun and doesn’t touch the job’s configuration:

   bashCopy

   sync; echo 3 > /proc/sys/vm/drop_caches

   Needs root; a between-run reset, not a mid-training step. See spark-training-gotchas (gotcha G3) for the full diagnostic behind this step.

2. Reduce batch size or packing length. Only after a flush fails to free enough headroom, cut batch size or packing length — the first step that changes what the run does. Prefer packing length first; it drives activation footprint more directly at long context.

3. Downgrade the method: bf16 LoRA before QLoRA. If flushing and shrinking batch/pack still OOM, drop the method a tier — bf16 LoRA is next, not the reverse. QLoRA’s bitsandbytes dequantization buffers are transient CUDA-side allocations that can OOM before an equivalent bf16 LoRA run would, even though QLoRA’s steady-state footprint is smaller. A QLoRA OOM is not proof the model doesn’t fit.

Fall back further (smaller model, multi-Spark) only after all three steps and the job still won’t fit.

Thermal Monitoring

Multi-hour runs push into Spark’s sustained power ceiling, well under the rated figure — expected platform behavior, not a symptom to explain away:

• Sample temperature and power alongside the training logs, not after a slowdown is noticed — every 30-60 seconds correlates a throughput drop with a thermal event. Keep the CSV output format assets/thermal-sample.sh writes, so timestamps line up against the log:

  bashCopy

  bash assets/thermal-sample.sh 30 thermal.log

• A sustained ~100W power draw is the platform cap, not a configuration bug. Don’t re-tune batch size or precision to “fix” a plateau that’s the box behaving normally under load. If temperature climbs while power stays flat under the rated 240W figure, that’s the signature to recognize.

• Log throttle events explicitly instead of letting a run silently slow down unrecorded. A run whose per-step time doubles two hours in should show that in the log, correlated against the thermal sample at that timestamp. Full throttling diagnostics: spark-training-gotchas (gotcha G4).

Concurrent Workloads

Because the 128GB pool is global, eviction happens without either process’s logs showing an OOM:

• The one-heavy-job rule applies to uncapped or near-capacity workloads — an uncapped trainer and inference server (vLLM, Ollama) compete for the same pool. A small, capped workload doesn’t: a <4GB LoRA fine-tune coexists fine alongside vLLM capped at gpu-memory-utilization<=0.5 — check the other process’s cap, not just its presence, before stopping it.

• Inference servers evict trainer pages silently under uncapped/near-capacity contention, and vice versa — neither logs an error, so a slow run or lost KV cache is a contention symptom to check for. Stop unrelated uncapped servers before a long or full-pool run.

Check for GPU-resident processes first:

ps aux | grep -E 'vllm|ollama|trl|axolotl' | grep -v grep

This procedure complements spark-training-gotchas (gotchas G3, G4, G6) — that skill covers launch-time failures; this one, the running job.

Memory math worksheets: references/uma-accounting.md.