megatron-lm: bump to NGC pytorch:26.02 and add Llama 3 8B sbatch

[KeitaW]

Summary

Bumps the Megatron-LM test case to nvcr.io/nvidia/pytorch:26.02-py3 and adds
an sbatch tuned for Llama 3 8B on P6-B300. The base image ships CUDA 13 + a
recent NCCL with native sm_103 binaries, eliminating the PTX-JIT fallback
on Blackwell Ultra.

Stack delta

Component	Before	After
NGC PyTorch base	25.06-py3	26.02-py3
GDRCopy	v2.5.1	v2.5.2
EFA installer	1.43.2	1.48.0
NCCL (ARG)	(inherited)	v2.30.4-1 (declared explicit)
AWS_OFI_NCCL (ARG)	(inherited)	v1.19.0 (declared explicit)
transformers	4.52.4	4.57.6
Megatron-LM	core_v0.12.1	core_v0.17.0

NCCL_VERSION and AWS_OFI_NCCL_VERSION are declared explicitly so the
repo's CI version-gate (which greps the Dockerfile for nccl/efa lines
and parses versions) sees values at or above the enforced minimums
(EFA ≥ 1.47.0, NCCL ≥ 2.28).

New: slurm/llama3/pretrain-llama3-8b.sbatch

Drop-in companion to slurm/llama2/. Defaults tuned for 8× B300 per node:
TP=1, PP=1, CP=2, seq_length=8192, MBS=1 GBS=512, --bf16,
--use-flash-attn, --transformer-impl transformer_engine, Llama 3 RoPE
(--rotary-base 500000), HuggingFaceTokenizer.

Compatibility fixes (required for core_v0.17.0)

Two bugs that the new container surfaces, both fixed in this PR:

1. pretrain-llama2.sbatch missing --eval-interval:
   core_v0.17.0's data iterator builder dereferences args.eval_interval
   even when --eval-iters 0
   (megatron/training/training.py:1143):

   eval_iters = (args.train_iters // args.eval_interval + 1) * args.eval_iters
   TypeError: unsupported operand type(s) for //: 'int' and 'NoneType'
   

   Crashes before iter 1 with the new container. The shipped gpt3 sbatch
   already has --eval-interval 1000; this PR adds it to llama2 too.

2. helpers_cpp C++ extension built lazily by rank 0 (Dockerfile fix):
   core_v0.17.0 lazy-builds megatron.core.datasets.helpers_cpp the
   first time a dataset is accessed. The build runs on rank 0 only, but
   /workspace inside each Pyxis container is local — when training spans
   multiple nodes, ranks on other nodes never see the rank-0 build and crash
   with ModuleNotFoundError: No module named 'megatron.core.datasets.helpers_cpp'. Fixed by pre-building the
   extension during docker build (passes 1n; was the only barrier to
   multi-node post runs).

Performance

Hardware: SageMaker HyperPod Slurm cluster, p6-b300.48xlarge (B300, sm_103),
8 GPUs/node. Llama 3 8B-equivalent (pretrain_gpt.py with the new sbatch
shape: 32L × 4096H × 32A, GQA 8, seq_len=8192, MBS=1 GBS=512, CP=2,
mock data), --log-throughput. Steady-state averaged over iters 5–25.

Nodes	GPUs	Pre TFLOPS/GPU	Post TFLOPS/GPU	Δ
1	8	710	809	+13.9%
4	32	LOAD_FAIL (NCCL/OFI hang)	782	n/a
8	64	LOAD_FAIL	760	n/a
16	128	LOAD_FAIL	740	n/a
32	256	LOAD_FAIL	732	n/a

Pre baseline (1n): 707–713 TFLOPS/GPU stable across iters 20–29 of a
100-iter run.
Post 1n: 800–820 TFLOPS/GPU steady state (mean 809).
Post 4–32n: 732–782 TFLOPS/GPU; the per-rank floor drops ~10% from 1n→32n,
which is consistent with CP=2 communication overhead growing with cluster
size — there's no grad-accumulation buffer to hide the all-gather, so ~700+
TFLOPS/GPU at 256 GPUs is the right neighborhood for this shape on B300.

Multi-node pre = LOAD_FAIL: All four pre runs (4/8/16/32n) hung at
> building train, validation, and test datasets for GPT ... with
NCCL WARN NET/OFI Attempt to call recv_close_deferred with outstanding requests and never reached iter 1. The old NCCL 2.27.3 + bundled OFI
plugin in NGC pytorch:25.06 doesn't survive multi-rank dataset-builder
collectives at our scale. That is itself the evidence supporting the bump
— the pre stack literally cannot run multi-node on B300, while the post
stack does.

Test plan

• make all from slurm/ builds and imports successfully.
• CI version-gate passes against the new Dockerfile.
• pretrain-llama2.sbatch smoke run on pre image — passes 1n
  (~710 TFLOPS/GPU); multi-node hangs (LOAD_FAIL above).
• pretrain-llama2.sbatch smoke run on post image, after both
  compatibility fixes — passes (~809 TFLOPS/GPU 1n; ~732–782 4n–32n).
• pretrain-llama3-8b.sbatch smoke run on post image (1n).