Memory Tuning#
Stable docs: docs/parallelisms.md
Card: card.yaml (co-located)
What It Is#
GPU OOM failures during training often stem from memory fragmentation rather than raw capacity. PyTorch’s default CUDA allocator can leave unusable gaps between allocations. The single most effective fix is:
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
This tells PyTorch to use expandable (non-fixed-size) memory segments, which dramatically reduces fragmentation and often eliminates borderline OOM without any model or parallelism changes.
Beyond fragmentation, actual peak memory is determined by:
Parameter + optimizer state memory — controlled by TP, PP, DP sharding (distributed optimizer, FSDP)
Activation memory — controlled by activation recompute, sequence length, micro-batch size
Temporary / workspace memory — CUDA kernels, NCCL buffers, CUDA graphs
Quick Decision#
When a training run OOMs or is close to the memory limit:
Set
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:Truefirst. This fixes fragmentation-induced OOM with zero performance cost. Most Slurm launch templates already include it.Add selective activation recompute (
recompute_modules=[core_attn]) if not already enabled. Seeskills/perf-techniques/activation-recompute/SKILL.md.Avoid increasing TP as a memory fix — doubling TP dramatically increases NVLink all-reduce volume and often kills throughput (-28% on Llama3 70B).
Avoid increasing PP at the cost of DP — halving DP doubles gradient accumulation steps and hurts throughput (~6%).
Consider
mlprecompute if still OOM. Saves ~3 GB but costs ~16% GPU utilization on large dense models (Llama3 70B).CPU offloading is blocked when PP > 1.
Enablement#
Expandable segments (recommended first step)#
Set in the job’s environment before launching:
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
In Slurm scripts this is typically placed alongside other env vars:
export CUDA_DEVICE_MAX_CONNECTIONS=1
export NVTE_ALLOW_NONDETERMINISTIC_ALGO=1
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
No model config changes needed. Zero throughput cost.
Parallelism resizing#
If the model genuinely does not fit (not fragmentation), adjust parallelism:
Strategy |
Memory effect |
Throughput cost |
Notes |
|---|---|---|---|
Increase PP (keeping DP) |
Fewer layers per stage |
Moderate (~6% if DP halved) |
Only if GPU count allows |
Increase TP |
Fewer params per GPU |
Severe (-28% on 70B) |
Last resort |
Distributed optimizer |
Shards optimizer state across DP ranks |
~1-2% |
Recommended for large models |
FSDP |
Shards params + grads + optimizer |
Varies |
See |
Activation recompute#
See skills/perf-techniques/activation-recompute/SKILL.md for full details.
CPU offloading#
cfg.model.cpu_offloading = True
Incompatible with PP > 1. Only usable when pipeline_model_parallel_size = 1.
A Note on VPP#
Virtual pipeline parallelism (VPP) is primarily a throughput optimization that reduces pipeline bubble overhead by interleaving smaller model chunks. Its effect on peak memory is minimal — changing VPP does not meaningfully change the total activation, parameter, or optimizer memory on a GPU.
In earlier experiments we incorrectly attributed an OOM fix to VPP tuning
(VPP 5→10). The actual fix was PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
which eliminated memory fragmentation. The VPP=10 run actually used slightly
more peak memory (60.2 GB vs 58.8 GB) but did not OOM because expandable
segments prevented fragmentation.
VPP should be tuned for pipeline bubble reduction (see docs/parallelisms.md),
not as a memory fix.
Compatibility and Constraints#
expandable_segments:Trueis incompatible with--use-nccl-ub(NCCL user-buffer registration). See Megatron-FSDP docs.When using CUDA graphs with
expandable_segments:True, setNCCL_GRAPH_REGISTER=0(required on pre-Blackwell GPUs, enforced by MCoreCudaGraphManager).CPU offloading requires
pipeline_model_parallel_size = 1.Distributed optimizer requires
use_distributed_optimizer = Truein the optimizer config.
Measured Results#
Llama3 70B SFT on 32x H100 80GB, FP8 (Current Scaling):
Baseline: TP=4, PP=4, VPP=5, DP=2, MBS=1, GBS=32, seq_len=4096
Golden GPU utilization: 709.93 TFLOP/s/GPU
Regression threshold: 5%
Strategy comparison: parallelism changes for memory reduction#
Experiment |
TP |
PP |
VPP |
DP |
TFLOP/s/GPU |
vs Golden |
Peak Mem (GB) |
Result |
|---|---|---|---|---|---|---|---|---|
Baseline |
4 |
4 |
5 |
2 |
~704 |
-0.8% |
58.8 |
OOM (fragmentation) |
More PP |
4 |
8 |
5 |
1 |
668.0 |
-5.9% |
53.2 |
Borderline perf |
More TP |
8 |
4 |
5 |
1 |
508.7 |
-28.4% |
50.2 |
Severe regression |
Baseline + expandable_segments |
4 |
4 |
5 |
2 |
~704 |
-0.8% |
~59 |
Passed |
Key takeaways:
expandable_segments:Trueis the winner. The baseline OOM was caused by memory fragmentation, not insufficient capacity. Setting this env var eliminated the OOM with zero throughput cost and no parallelism changes.PP=8 works for memory but loses DP (2→1), meaning 32 gradient accumulation steps per batch, which hurts throughput by ~6%.
TP=8 is catastrophic (-28%) because doubling TP increases all-reduce communication volume proportionally across NVLink, and DP=1 means no micro-batch overlap.
CPU offloading: blocked#
Experiment |
offload_layers |
Result |
|---|---|---|
Exp 4 |
2 |
Incompatible (PP > 1) |
Exp 5 |
4 |
Incompatible (PP > 1) |
Exp 6 |
6 |
Incompatible (PP > 1) |
ValueError: Currently there is no support for Pipeline parallelism with CPU offloading. This approach is blocked for any model using PP > 1.
Activation recompute: expensive alternative#
Selective activation recompute with mlp saved ~3 GB peak memory but cost
~16% GPU utilization on this workload. See
skills/perf-techniques/activation-recompute/SKILL.md for full results.
Code Anchors#
CPU offloading PP incompatibility (MCore)#
if self.cpu_offloading and self.pipeline_model_parallel_size > 1:
raise ValueError(
"Currently there is no support for Pipeline parallelism with CPU offloading"
)
VPP config and layer divisibility validation (MCore)#
if pipeline_parallel_size and self.virtual_pipeline_model_parallel_size is not None:
num_layers_per_middle_pipeline_rank = num_layers // pipeline_parallel_size
if (
not num_layers_per_middle_pipeline_rank
% self.virtual_pipeline_model_parallel_size
== 0
):
raise ValueError(
f"number of layers on each middle pipeline rank:"
f"{num_layers_per_middle_pipeline_rank} must be divisible by virtual"
f"pipeline parallel degree {self.virtual_pipeline_model_parallel_size}"
)
Parallelism docs on interleaved pipeline schedule#
To minimize the pipeline bubble, the computation on each GPU can be divided into multiple subsets of layers (referred to as model chunks), rather than a single contiguous block. Enable this by setting `virtual_pipeline_model_parallel_size`:
model_config = GPTModelProvider(
pipeline_model_parallel_size=4,
virtual_pipeline_model_parallel_size=2, # 2 model chunks per pipeline stage
# ... other model parameters
)
Failure Diagnosis#
Symptom |
Cause |
Confirm |
Fix |
|---|---|---|---|
OOM on a single rank despite headroom on others |
Memory fragmentation |
check if |
set |
OOM with |
Genuine capacity limit |
check |
increase PP, use distributed optimizer, or add recompute |
|
using cpu_offloading with PP > 1 |
check PP config |
disable CPU offloading or set PP=1 |
|
NCCL UB incompatible with expandable allocator |
check env vars |
remove |
Known Limitations#
expandable_segments:Trueis incompatible with NCCL user-buffer registrationCPU offloading is blocked when PP > 1
Parallelism resizing (TP/PP) often has significant throughput costs
No automatic memory profiling to recommend the optimal strategy
Verification#
Quick check that expandable_segments:True is active:
import os
assert "expandable_segments:True" in os.environ.get("PYTORCH_CUDA_ALLOC_CONF", "")
For Slurm jobs, verify the env var is exported before the training command in the launch script.