Reference

Compatibility

The versions, hardware and models Meep² is tested with.

TL;DR

Python 3.12, PyTorch 2.11 with CUDA from your environment, and the exact versions below, which pip installs with Meep². An NVIDIA GPU of the Ampere generation or later; tested on NVIDIA GB10 and RTX 3090.

Packages

Package Versions Notes

PyTorch

2.11, from the environment

Not installed by Meep², so a PyTorch built for your GPU is never replaced: NVIDIA’s container 26.03, or pip install torch==2.11.0 --index-url https://download.pytorch.org/whl/cu130.

Transformers

5.20.2

TRL

1.16.0

PEFT

0.21.2

datasets

4.8.4

accelerate

1.15.0

Liger Kernel

0.8.5

Meep² sets the chunk size of Liger’s fused loss through its ops module, and runs Liger’s rotary embedding on its Triton kernel: the CuTe DSL kernel Liger picks where NVIDIA’s CUTLASS DSL is installed failed on a GB10 at step 136 of a Llama 3.1 8B training. LIGER_KERNEL_IMPL_ROPE set before loading overrides that.

bitsandbytes

0.50.2

4-bit models, 8-bit optimizers, fused LoRA.

flash-linear-attention, fla-core

0.6.0

Padding-free batches on Qwen3.5, Qwen3.6 and Qwen3-Next.

FlashAttention

2.7, optional

Not installed by pip; NVIDIA’s container has it. Without it, padding-free batches run on PyTorch’s memory-efficient kernel (meepsquared_varlen).

vLLM

0.22 or later (TRL’s floor)

Optional; GRPO generation in colocate mode. vLLM and PyTorch must be the same pair: on NVIDIA’s PyTorch container, the wheel of tk-vllm-wheels built for it (Installing vLLM).

Hardware

GPU Tested

NVIDIA GB10 (DGX Spark), 128 GB shared memory

SFT with LoRA and QLoRA, frozen weights in FP8, mixture of experts, GRPO generating with Transformers and with vLLM, shard-by-shard loading.

2 × NVIDIA RTX 3090, 24 GB each, PCIe without NVLink

One GPU, DDP, FSDP2, QLoRA with DDP and FSDP2, a model spread over both.

Models

The models of the benchmarks are tested end to end. Other decoder-only models of Transformers load and train; Liger’s kernels need a model type Liger patches, and liger=False trains any other.

Vision layers are not trained.