csh
/
playbook
1
1
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
playbook/antigravity-awesome-skills/skills/remote-gpu-trainer/references/training/throughput-profiling.md

28 KiB
Raw Permalink Blame History

Throughput & profiling — make training FAST, find the one bottleneck

How to tell why a rented GPU is underfed (GPU-bound vs data-bound vs comms-bound), then apply the right speedup in cost order — from a free dataloader knob to torch.compile and fused attention. This layer owns making it RUN fast + locating the mechanical bottleneck; verifying-dl-experiments owns is the resulting number correct. Cross-link it (REQUIRED) wherever a speedup risks changing the science (a kernel that alters numerics, a precision swap, dropping samples to "go faster").

Size the run to the box — then PIN it for any comparison. Auto-sizing batch/num_workers to the measured GPU/VRAM/vCPU (Phase 0) to use the card well is fine for a STANDALONE job; but for an ablation or baseline-vs-variant comparison, pin the same batch across all cells — auto-maximizing per-box silently changes a variable and breaks comparability (verifying-dl-experiments, REQUIRED).

To jump: grep -in '<keyword>' references/training/throughput-profiling.md (e.g. bound, workers, compile, recompile, flash, sdpa, nsys, py-spy, channels_last, tf32, overlap).

Table of contents

  • Diagnose first — T1 the 3-way split (GPU/data/comms-bound) · T2 util%-is-a-liar pointer · T3 the cheap CPU/GPU-busy triage
  • Dataloader (the #1 cause of a starved GPU) — T4 num_workers · T5 persistent_workers · T6 pin_memory + non_blocking · T7 prefetch_factor · T8 IO-bound vs CPU-transform-bound
  • Free / near-free knobs — T9 TF32 + matmul precision · T10 cudnn.benchmark · T11 channels_last · T12 set_to_none + disable debug APIs
  • Mixed precision for speed — T13 bf16/fp16 throughput
  • Kernels — T14 SDPA / FlashAttention · T15 torch.compile gains · T16 torch.compile recompilation traps
  • Memory↔speed trades — T17 activation checkpointing speed cost · T18 batch size vs throughput
  • Profilers — T19 torch.profiler (is-it-data-bound) · T20 nsys / Nsight Systems · T21 py-spy (live, no restart) · T22 memory-snapshot pointer
  • Multi-GPU / multi-node comms — T23 DDP/FSDP compute-comm overlap
  • Pointers — gotchas_universal.md U8/U21/U24/U25/U38 · oom-memory.md · distributed-launch.md · multinode.md · verifying-dl-experiments (skill)

Diagnose first — do NOT tune blind

T1 — The 3-way split: GPU-bound vs data-bound vs comms-bound (decide before touching a knob)

Symptom: training is "slow" and the instinct is to change the model or batch size at random.

Root cause: throughput is gated by exactly one of three resources at a time; the fix for each is disjoint, so guessing wastes paid wall-clock (principle #1).

Fix — classify with one cheap reading each (heuristic: util consistently >90% ⇒ GPU-bound; low/fluctuating ⇒ elsewhere; both CPU+GPU low ⇒ I/O — https://apxml.com/courses/planning-optimizing-ai-infrastructure/chapter-5-strategies-for-performance-optimization/identifying-performance-bottlenecks):

  • GPU-bound (the good case): util high and SM clock/power high (T2); adding workers doesn't help. Only levers left: kernels (T14–T15), precision (T13), a bigger card.
  • Data-bound: util low-but-nonzero or sawtoothing, host CPU busy in DataLoader/transforms; a trace shows GPU-idle gaps lining up with CPU data work (T19). Go to T4–T8.
  • Comms-bound (multi-GPU/-node only): per-GPU util high, scaling efficiency poor; time in nccl:all_reduce/all_gather not overlapped with compute. Go to T23.

The highest-signal instrument is a profiler trace (T19) — read it before changing anything.

T2 — nvidia-smi GPU-Util % lies; correlate clock + power → gotchas_universal.md U21

A 100%-util tile can hide a starved GPU (a trickle of tiny kernels reads as 100%). The full diagnosis — correlate clocks.current.sm + mem-bandwidth util + power via nvidia-smi dmon -s pucvmet -d 1, and the thermal/power-throttle slowdown — lives in gotchas_universal.md U21/U23; read it before concluding a run is GPU-bound. The 0%-util-but-running (CPU-data-bound) inverse is U38, owned by verifying-dl-experiments.

T3 — Cheap triage when no profiler is wired yet: is the host CPU busy?

Symptom: need a 30-second answer to "GPU or data?" before instrumenting.

Fix: watch GPU and CPU at once for ~10 s —

nvidia-smi dmon -s pu -d 1 -c 10          # per-second SM% + power; sawtooth/low = starved
top -b -n 1 | grep -i python | head        # a worker pegged at ~100% CPU = CPU-transform-bound

GPU SM% high and steady ⇒ GPU-bound (stop here, go to kernels/precision). GPU SM% sawtoothing while a python worker is CPU-pegged ⇒ data-bound (T4–T8). Both idle ⇒ I/O-bound (stage to NVMe, U8). Then confirm with a real trace (T19) before investing in a fix. GPU SM% low while many python threads thrash a few cores (not one worker pegged) ⇒ intra-op thread oversubscription on a vCPU slice, not data-bound — cap OMP_NUM_THREADS to your cgroup quota (gotchas_universal.md U40), don't add dataloader workers.

Dataloader — the #1 reason a rented GPU sits idle

The partial-starve knob set (and its order) is gotchas_universal.md U24; this section is the per-knob why/when. Each helps a different failure, so apply by symptom, not as a blanket cargo-cult.

T4 — num_workers: 0 means the main process loads serially (the default starves the GPU)

Symptom: DataLoader(num_workers=0) (the default) — every batch is fetched on the main thread, GPU waits the whole fetch.

Root cause: with num_workers=0 "the data will be loaded in the main process" — no overlap between data prep and compute (https://docs.pytorch.org/docs/2.12/data.html).

Fix: set num_workers > 0 to load asynchronously and overlap fetch with the GPU step. Start at cores − 1, but size against per-worker RAM, not CPU count — each worker forks a full copy of any large in-dataset object; too many OOM the cgroup with a bare Killed (the quadratic trap + sizing rule are gotchas_universal.md U9). Not monotonic: past the point where the GPU is fed, extra workers only add RAM and startup cost.

T5 — persistent_workers=True: stop paying worker-startup every epoch

Symptom: a visible stall at the start of every epoch (especially short epochs / many epochs); GPU idle while workers respawn.

Root cause: default persistent_workers=False shuts down all workers after the dataset is consumed once and re-forks them next epoch — re-importing, re-opening files, rebuilding the dataset object each time (https://docs.pytorch.org/docs/2.12/data.html).

Fix: persistent_workers=True keeps the worker Dataset instances alive between epochs, removing the per-epoch respawn cost. Requires num_workers > 0. Biggest win when epochs are short or the dataset's __init__ is heavy (loads an index/manifest).

T6 — pin_memory=True + non_blocking=True: overlap the host→device copy

Symptom: the H2D copy (x.to('cuda')) sits on the critical path between fetch and forward.

Root cause: a pageable-memory tensor must be staged through a pinned buffer by the driver before DMA; a synchronous .to(device) blocks the step. "When using a GPU it's better to set pin_memory=True" (https://docs.pytorch.org/tutorials/recipes/recipes/tuning_guide.html).

Fix: DataLoader(pin_memory=True) allocates batches page-locked, then transfer x = x.to(device, non_blocking=True) so the copy runs async on a copy stream and overlaps compute. Both halves needed — pin_memory alone still blocks; non_blocking without pinned memory silently falls back to a blocking copy. Costs host RAM (pinned pages aren't swappable) — back off if it pressures the cgroup (U9).

T7 — prefetch_factor: deepen the queue when fetch time is bursty

Symptom: with workers on, the GPU still periodically stalls — every Nth step (N = num_workers) has a long idle gap because all workers were busy producing the next batch when the GPU asked (https://docs.pytorch.org/tutorials/intermediate/tensorboard_profiler_tutorial.html).

Root cause: prefetch_factor defaults to 2 when num_workers>0 (None when 0) — "2 means there will be a total of 2 * num_workers batches prefetched across all workers" (https://docs.pytorch.org/docs/2.12/data.html). A shallow queue can't absorb a variance spike in per-sample fetch/decode time.

Fix: raise prefetch_factor (3–4) so workers run ahead and bursts hide — at the cost of more resident batches in RAM (re-check U9). A smoothing knob, not a multiplier: if the average fetch rate is below the GPU's consume rate, no depth helps — fix the rate (workers T4, GPU transform T8, NVMe U8) instead.

T8 — IO-bound vs CPU-transform-bound are different data-bound cases (different fix)

Symptom: data-bound (T1), but adding workers barely helps.

Root cause — split the case:

  • IO-bound: bytes arrive slowly from network/HDD/object store; workers sit in read. Stage the working set to instance-local NVMe (HDD→NVMe gaps reach ~35×) = gotchas_universal.md U8; the many-tiny-files transaction death + shard-into-tar / WebDataset fix = U25.
  • CPU-transform-bound: a heavy per-sample augment (resize/decode/FFT) saturates CPU; workers CPU-pegged (T3), capping at core count. Move the transform to the GPU (NVIDIA DALI, torchvision.transforms.v2 on tensors, kornia) onto idle GPU cycles. The 0%-util serialized-transform variant is U38, owned by verifying-dl-experiments REQUIRED (which also owns whether a GPU-side transform shifted the data distribution).

Fix: read the trace (T19) — time in read/stat ⇒ U8/U25; time in a transform fn ⇒ move to GPU.

Free / near-free knobs (set these once at startup on any box)

T9 — TF32 / set_float32_matmul_precision("high") — the "why is my A100 slow" footgun

The biggest free speedup on Ampere+ for any fp32 matmul path; OFF by default since PyTorch 1.12. The decision and exact knobs (torch.set_float32_matmul_precision("high"), the legacy allow_tf32 flags, --tf32 1 in HF Trainer, convergence impact) are owned by references/training/precision-stability.md P2 (cross-link there; do NOT restate). If a fresh PyTorch 2.x rental's fp32-heavy run is 2–4× slow with no bug, this is the first suspect.

T10 — cudnn.benchmark=True: autotune conv algorithms (fixed input shapes only)

Symptom: a conv-heavy net (CNN/UNet) is slower than it should be; input shapes are constant.

Root cause: by default cuDNN picks a generic conv algorithm; the autotuner benchmarks variants on the first batch of each new shape and caches the fastest (https://docs.pytorch.org/tutorials/recipes/recipes/tuning_guide.html).

Fix: torch.backends.cudnn.benchmark = True once at startup. Only helps when input shapes are stable — with variable shapes (dynamic resolution, ragged batches) it re-benchmarks every new shape and loses time. Trade-off: it is nondeterministic (picks by first-batch timing), so it fights the determinism knobs — whether to enable it for a clean datapoint is owned by precision-stability P19 / verifying-dl-experiments (U36, REQUIRED).

T11 — channels_last: free Tensor-Core speedup for conv nets under AMP

Symptom: a CNN under mixed precision isn't hitting Tensor Cores; throughput below the card's potential.

Root cause: default NCHW contiguous layout forces layout transposes around Tensor-Core convolutions.

Fix: convert model and inputs to memory_format=torch.channels_last — model = model.to(memory_format=torch.channels_last) and x = x.to(memory_format=torch.channels_last). Optimizes convolutional networks with Tensor Cores + AMP (https://docs.pytorch.org/tutorials/recipes/recipes/tuning_guide.html). Marked experimental and CNN-specific (no benefit for pure transformers). No numerics change — purely a layout speedup.

T12 — set_to_none + disable debug APIs (two free per-step taxes to remove)

  • optimizer.zero_grad(set_to_none=True) (the default since PyTorch 2.0) over zero-filling — assigning None skips a memory-write kernel per param and lets the next backward write fresh (https://docs.pytorch.org/tutorials/recipes/recipes/tuning_guide.html). Edge case: code reading .grad between steps must tolerate None.
  • Turn OFF debug APIs for the real run — torch.autograd.set_detect_anomaly(True), torch.autograd.profiler.profile, gradcheck add per-op bookkeeping (anomaly detection is ~10× slower, precision-stability P9). Grep detect_anomaly / leftover with profile( wrappers before a long launch (https://docs.pytorch.org/tutorials/recipes/recipes/tuning_guide.html); easy to leave on after a NaN hunt.

Mixed precision for speed

T13 — bf16/fp16 is a throughput lever, not just a memory lever

Symptom: fp32 training under-uses Tensor Cores; the GPU has bf16/fp16 tensor cores.

Root cause: 16-bit matmuls run on Tensor Cores at much higher FLOP/s and halve activation read/write bandwidth — a speedup on top of the memory saving (oom-memory.md M6).

Fix: torch.autocast("cuda", dtype=torch.bfloat16) on Ampere+ (the modern default; no GradScaler — precision-stability P6) or bf16=True in HF TrainingArguments. The full precision decision (bf16 vs fp16 vs the V100/T4 fp16-only path, GradScaler mechanics, NaN/overflow) is owned by references/training/precision-stability.md P1–P10 (cross-link; do NOT restate). The memory angle and the activation-bucket math is oom-memory.md M6. A NaN/divergence after the swap is a numerics question → precision-stability / verifying-dl-experiments (REQUIRED).

Kernels — the levers left once the GPU is fed

T14 — SDPA / FlashAttention: stop materializing the O(seq²) attention matrix

Symptom: a transformer is attention-bound; long sequences are slow and memory-heavy; or flash_attn "installed" but the run is no faster.

Root cause: the eager/math attention path materializes the full seq×seq score matrix. The fused FlashAttention / memory-efficient backends never do, but PyTorch's scaled_dot_product_attention silently falls back to the slow math backend when the fused kernel's input constraints aren't met (wrong dtype, head dim, mask shape) — "if a fused implementation is not available, a warning will be raised" (https://docs.pytorch.org/docs/2.12/generated/torch.nn.functional.scaled_dot_product_attention.html).

Fix:

  • Use F.scaled_dot_product_attention(q,k,v) (or attn_implementation="sdpa", the HF default on 2.1.1+), which auto-picks FlashAttention / memory-efficient / cuDNN / math. Feed it fp16/bf16 inputs — the fused backends need 16-bit (the math fallback is what runs in fp32).
  • Force-verify the fast backend instead of trusting silence: 
    from torch.nn.attention import sdpa_kernel, SDPBackend
with sdpa_kernel(backends=[SDPBackend.FLASH_ATTENTION]):   # errors loudly if it can't be used
    out = F.scaled_dot_product_attention(q, k, v, is_causal=True)
  • Installing flash_attn from source is a trap: without ninja (pip install ninja) the CUDA extension compiles single-threaded ~2 h; with ninja ~3–5 min on a 64-core box. With many cores but <96 GB RAM ninja over-parallelizes and OOMs the build — cap MAX_JOBS=4 pip install flash-attn --no-build-isolation. Prefer a prebuilt wheel matching the cuXX/torchYY/cpZZ triple (https://github.com/Dao-AILab/flash-attention/issues/1038, https://pypi.org/project/flash-attn/). A torch/CUDA mismatch is gotchas_universal.md U28. Whether the fused kernel changes outputs (causal-mask edge cases) is a numerics check → verifying-dl-experiments (REQUIRED).

T15 — torch.compile: fuse kernels + cut launch overhead (one line, real gains)

Symptom: many small pointwise/elementwise ops; Python/launch overhead dominates between big matmuls.

Root cause: eager launches each op separately; Inductor fuses adjacent ops into Triton kernels and (in CUDA-graph modes) eliminates per-step launch overhead, reusing the execution plan across steps.

Fix: wrap the model — model = torch.compile(model). Modes (https://huggingface.co/docs/transformers/en/perf_torch_compile):

  • default — balanced speed/memory.
  • mode="reduce-overhead" — uses CUDA graphs to kill Python overhead (best for many tiny ops / small-batch / inference), at a little more memory.
  • mode="max-autotune" — longest compile, fastest steady-state.
  • HF TrainingArguments(torch_compile=True, torch_compile_mode="reduce-overhead").

Reported ~2.2× mean-inference speedups; training gains real but model-dependent. First step(s) are slow — compilation is lazy on first call (https://huggingface.co/docs/transformers/en/perf_torch_compile); exclude warm-up from any throughput measurement. Set fullgraph=True while developing to surface graph breaks loudly instead of silently losing speed. Whether the compiled numbers match eager → verifying-dl-experiments (REQUIRED).

T16 — torch.compile recompilation trap: variable shapes silently blow the cache → eager

Symptom: a compiled run is slower than eager, or stutters periodically; throughput never stabilizes. Common with variable batch/seq-len, dynamic padding, or per-step changing shapes.

Root cause: compile creates guards on traced shapes; a new shape violates a guard and triggers a recompile. Past the recompile cap (torch._dynamo.config.recompile_limit, default 8; legacy cache_size_limit) Dynamo stops compiling that function and runs it eagerly — paying all the compile cost and getting none of the benefit (https://docs.pytorch.org/docs/stable/compile/programming_model.recompilation.html, https://github.com/pytorch/pytorch/issues/93457).

Fix:

  • See it: TORCH_LOGS=recompiles python train.py logs which function recompiled and the failed guard; TORCH_LOGS=graph_breaks and torch._dynamo.explain(...) locate graph breaks (https://docs.pytorch.org/docs/stable/torch.compiler_troubleshooting.html).
  • Tame shapes: pad/bucket to a few fixed shapes so guards stop firing; or mark the varying dim dynamic — torch.compile(model, dynamic=True) (or mark_dynamic / TORCH_COMPILE_DYNAMIC_SOURCES) compiles one shape-generic graph instead of one per size. dynamic=False forces a fresh recompile per distinct size (use only with truly few shapes) (https://docs.pytorch.org/docs/stable/compile/programming_model.html).
  • Last resort: raise torch._dynamo.config.recompile_limit only if a handful of stable extra shapes legitimately exist — raising it to mask genuinely unbounded shapes just thrashes.

Memory ↔ speed trades

T17 — Activation checkpointing buys memory by spending ~20–30% compute (know the cost)

Symptom: gradient/activation checkpointing is on "to be safe" and training is slow — but the model actually fits without it.

Fix: checkpointing recomputes activations in backward instead of storing them — trading ~20–30% extra compute for a large memory cut (https://docs.pytorch.org/tutorials/recipes/recipes/tuning_guide.html, oom-memory.md M7). Enable it only when activations actually OOM (full rationale + use_reentrant=False / use_cache=False gotchas = oom-memory.md M7); if it fits without, turning it off is a free ~25% speedup. On the frontier, checkpoint only the fewest/heaviest blocks needed to fit, not the whole model.

T18 — Bigger micro-batch ≈ better GPU utilization (up to the memory wall)

Symptom: tiny batches under-feed the GPU; util and throughput both low though VRAM is mostly free (small batches under-fill Tensor Cores and amortize launch/sync overhead poorly).

Fix: raise micro-batch toward the VRAM limit; keep the effective batch fixed with grad-accum if the result depends on it (batch 4 × accum 16 beats batch 1 × accum 64 — oom-memory.md M5). Accuracy/effective- batch implications (LR scaling, accumulation loss-weighting) → verifying-dl-experiments (REQUIRED). Sizing alongside a concurrent job + expandable_segments = gotchas_universal.md U10 / oom-memory.md M8.

Profilers — measure the bottleneck, don't guess it

T19 — torch.profiler: the definitive data-bound vs compute-bound verdict

Symptom: need to prove where time goes (which T1 case), not infer from util%.

Fix — scheduled profile of a few steps (https://docs.pytorch.org/tutorials/recipes/recipes/profiler_recipe.html):

from torch.profiler import profile, schedule, ProfilerActivity, tensorboard_trace_handler
with profile(
    activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
    schedule=schedule(wait=1, warmup=1, active=3),     # skip warm-up; record 3 steps
    on_trace_ready=tensorboard_trace_handler("./tb_trace"),
    record_shapes=True, with_stack=True,
) as prof:
    for step, batch in enumerate(loader):
        train_step(batch); prof.step()
        if step >= 6: break
print(prof.key_averages().table(sort_by="self_cuda_time_total", row_limit=15))

Read it: large GPU-timeline gaps with CPU busy in DataLoader/transforms during the gap ⇒ data-bound (T4–T8); the TensorBoard "Performance Recommendation" panel names the DataLoader directly (https://docs.pytorch.org/tutorials/intermediate/tensorboard_profiler_tutorial.html). Densely-packed GPU timeline ⇒ GPU-bound; sort by self_cuda_time_total for the hottest kernel (T14/T15). Time in nccl:* not overlapped ⇒ comms-bound (T23). On a remote box write the trace and view locally — for raw export_chrome_trace("trace.json") open at chrome://tracing; scp it down (references/ssh_transport.md), never run a viewer over ssh.

T20 — nsys / Nsight Systems: system-wide timeline when the gap is below PyTorch's view

Symptom: torch.profiler shows GPU-idle gaps but not why (CPU launch latency, a hidden sync, a memcpy, a kernel-launch storm); or want CUDA-API + NVTX + OS-runtime on one timeline.

Root cause: torch.profiler sees PyTorch ops; nsys traces the whole system — CUDA API, kernels, memcpy, NVTX ranges, OS-runtime — so it exposes launch-bound stalls and CPU↔GPU sync that PyTorch can't. "Periodic gaps in the CUDA HW row are moments when the GPU is idle — a red flag" (https://docs.lxp.lu/howto/pytorch-profiling-with-nsight/).

Fix — profile a bounded window on the box, view locally (canonical PyTorch recipe, https://gist.github.com/mcarilli/376821aa1a7182dfcf59928a7cde3223):

nsys profile -w true -t cuda,nvtx,osrt,cudnn,cublas -s cpu \
  --capture-range=cudaProfilerApi -x true -o report python train.py

In the script, bound the window so the .nsys-rep stays small:

torch.cuda.profiler.cudart().cudaProfilerStart()   # after warm-up
# ... a handful of steps, optionally wrapped in torch.cuda.nvtx.range_push/pop ...
torch.cuda.profiler.cudart().cudaProfilerStop()

scp the .nsys-rep down, open in the Nsight Systems GUI. Nsight Systems finds which kernel is slow; Nsight Compute (ncu) finds why (occupancy, bandwidth, warp stalls) — but ncu is heavy, reserve it for one hot kernel (https://www.spheron.network/blog/gpu-profiling-ai-workloads-nsight-compute-pytorch-profiler-guide/).

T21 — py-spy: profile a LIVE training process with no restart, no code change

Symptom: a long run is mysteriously slow or apparently hung; restarting it to add a profiler would cost hours and might not reproduce.

Root cause: a Python-side bottleneck or deadlock (a slow transform, a lock, a blocking collective) that needs inspection in situ.

Fix — attach by PID, zero instrumentation (https://github.com/benfred/py-spy):

py-spy dump --pid <PID>            # one-shot stack of every thread → where it's hung RIGHT NOW
py-spy top  --pid <PID>            # live "which functions burn time" (Unix top-style)
py-spy record -o prof.svg --pid <PID>   # flame graph over a window

"The profiled program needs no import, no decorator, and no restart." On a rented box mid-run, py-spy dump instantly distinguishes a hung process (stuck in recv/lock/all_reduce) from a slow one (busy in a transform) — pairs with the "is it actually hung?" check (gotchas_universal.md U17, verifying-dl-experiments REQUIRED). May need --native for C-extension frames and sudo/SYS_PTRACE to attach.

T22 — CUDA memory snapshot/visualizer → oom-memory.md M19

For what allocated the memory (not time), the torch.cuda.memory._record_memory_history snapshot + https://pytorch.org/memory_viz timeline is owned by references/training/oom-memory.md M19/M18. It is a memory tool, not a throughput tool — listed here only so the profiler menu is complete. Do NOT restate.

Multi-GPU / multi-node communication

T23 — Compute-comms overlap: DDP overlaps by default; tune the bucket, watch for breakers

Symptom: scaling efficiency is poor — per-GPU util high, but N GPUs deliver far less than N× throughput; trace shows all_reduce/all_gather not overlapped with backward compute.

Root cause: DDP overlaps gradient all-reduce with backward by bucketing gradients and launching each bucket's reduce on a separate CUDA stream as soon as it's ready (https://github.com/pytorch/pytorch/issues/67570). Overlap breaks when something forces a sync: an unused-parameter recompute, an off-by-default find_unused_parameters=True, a .item()/print/.cpu() in the step, or too-small/too-large buckets.

Fix (single box, DDP/FSDP — the launch/sharding mechanics live in references/training/distributed-launch.md, REQUIRED):

  • Tune bucket_cap_mb (DDP) to batch gradient chunks into fewer, larger all-reduces; set gradient_as_bucket_view=True to cut a copy. Buckets too small = launch overhead; too large = late overlap.
  • FSDP: enable backward_prefetch (prefetch the next layer's all-gather during current backward) and forward_prefetch so comms hide under compute; limit_all_gathers if memory-pressured.
  • Remove per-step host syncs (loss.item() every step, prints, eager .cpu()) that serialize the stream.

Inter-node transport (NCCL picking the wrong NIC, fabric-manager hang, 1800 s timeout masking a straggler, MTU mismatch) is references/multinode.md (REQUIRED for ≥2 instances) — a comms "slowdown" across boxes is usually one of those, not a bucket-size tune. Whether a world-size change silently rescaled the effective batch/LR is a science question → verifying-dl-experiments (REQUIRED).

Pointers — throughput gotchas catalogued elsewhere (do NOT restate)

  • gotchas_universal.md — U8 stage hot data to local NVMe (IO-bound) · U21 nvidia-smi util% is a liar (+ U23 thermal/power throttle) · U24 dataloader-starvation knob order · U25 millions of small files → shard into tar/WebDataset · U38 GPU 0%-util CPU-data-bound (owned by verifying-dl).
  • references/training/oom-memory.md — M5 micro-batch/grad-accum · M6 bf16 activations · M7 activation checkpointing memory rationale · M8 expandable_segments · M19 memory snapshot/visualizer.
  • references/training/precision-stability.md — P1–P10 the precision decision + AMP mechanics · P2 the TF32-off footgun · P19 determinism-vs-cudnn.benchmark speed trade.
  • references/training/distributed-launch.md — torchrun/Accelerate/DeepSpeed launch, DDP/FSDP sharding, and the desync/hang toolkit (the launch substrate this file's T23 sits on).
  • references/multinode.md — inter-node NCCL/NIC/fabric/timeout/MTU (the wire between boxes). Single-box users skip.
  • verifying-dl-experiments (REQUIRED) — owns is-the-number-real: whether a kernel/precision/compile swap changed the result, whether dropping samples or a GPU-side transform shifted the distribution, the 0%-util diagnosis (U38), determinism (U36). This file makes training fast; that skill decides if the faster result is still true.