A cross-framework ragged tensor primitive for variable-length sequence data.

📚 Documentation · 📦 PyPI · 💬 Discussions · 🐛 Issues

import scree
import numpy as np

# Three sequences of different lengths.
seqs = [np.random.randn(n, 8).astype(np.float32) for n in [4, 2, 7]]

# Pack them into one scree.Array — no padding.
arr = scree.pack(seqs)
# arr.values: shape (13, 8), arr.offsets: [0, 4, 6, 13]

# Run varlen attention. Each sequence attends only to itself.
from scree.kernels.reference import varlen_attention
out = varlen_attention(arr, arr, arr, causal=True)

Why

Variable-length sequence data is everywhere in modern ML — transformer training, inference batching, multimodal interleaving, MoE routing — yet every team carries their own incompatible representation:

• torch.nested (PyTorch only, in beta since 2021)
• TF RaggedTensor (TensorFlow only)
• FlashAttention cu_seqlens (a convention, not a typed primitive)
• vLLM / SGLang packed batches (internal data structures)
• HuggingFace attention_mask (pads, then masks — wasting memory and FLOPs)

scree ships one primitive — a packed values + offsets + ragged_dim array — that bridges across frameworks and ships with reference varlen kernels for attention, layernorm, softmax, and scatter/gather.

What you get

Memory savings vs HF padded on realistic LLM length distributions (log-normal):

Workload	Mean savings	Min – Max
Training-style (batch 64, mean_len 256, σ=0.6)	71%	63% – 84%
Inference-style (batch 32, mean_len 1024, σ=1.2)	85%	75% – 94%

Reproduce: python benchmarks/bench_memory.py

CPU throughput vs a naive padded attention baseline on a real batch (16 seqs × log-normal lengths, 1980 real / 4464 padded tokens, 4 heads × head_dim 32, fp32, no mask optimization):

Operation	scree	padded baseline	Speedup
varlen_attention	34.7 ms	228.3 ms	6.6×
varlen_rmsnorm	0.13 ms	0.28 ms	2.1×

Reproduce: python benchmarks/bench_throughput.py

GPU forward + training step vs FlashAttention-2 on H100. Headline workload: 16 seqs × log-normal lengths, 12160 total tokens, 16 heads × head_dim 64, fp16, causal:

Operation	FA-2	scree-Triton	Ratio
forward only	0.165 ms	0.216 ms	1.30×
forward + backward (training step)	0.688 ms	1.106 ms	1.61×

Correctness: forward max abs diff 4.88e-04; backward dq 9.77e-04, dk 1.95e-03, dv 1.95e-03 vs FA-2 (all PASS within fp16 tolerance). Reproduce: modal run benchmarks/modal_bench.py + modal run benchmarks/modal_autograd_bench.py (~$0.40 of Modal credit total).

Across 27 shapes (head_dim × n_heads × mean_len): scree is closer to FA-2 on large workloads, slower on small ones (wrapper allocation overhead is per-call; it amortizes as the kernel grows). See benchmarks/modal_multishape_sweep.py.

Workload range	Forward ratio	Training-step ratio
Best (large: head_dim=64, n_heads=16, mean_len=2048)	1.21×	1.45×
Median across 27 shapes	1.95×	2.01×
Worst (toy: head_dim=32, n_heads=4, mean_len=256)	3.53×	1.77×

For production LLM training (head_dim ≥ 64, n_heads ≥ 8, mean_len ≥ 1024), expect 1.2–2.0× of FA-2.

Zero-copy bridges to the things you already use:

import scree.bridges as bridges

arr = scree.from_cu_seqlens(values, cu_seqlens)         # FlashAttention
arr = bridges.from_hf_padded(hidden_states, attn_mask)  # HuggingFace
arr = bridges.from_torch_nested(nt)                     # torch.nested

bridges.to_torch_nested(arr)   # → torch.NestedTensor
bridges.to_hf_padded(arr)      # → (hidden_states, attention_mask)
bridges.to_torch(arr)          # numpy values → torch tensors via DLPack

One primitive, every framework — values and offsets can be NumPy, PyTorch, MLX (Apple Silicon, via Metal), or JAX. All four backends pass the same correctness suite end-to-end.

The name

A scree is the irregular pile of rock fragments accumulated on a mountain slope. Variable-length sequences pack against each other the same way: irregular shapes, fitted by their irregularity, not despite it.

Status

v0.0.1 on PyPI. Reference Python kernels for all four backends and a full FA-2-style Triton kernel set (forward + backward + RMSNorm + LayerNorm) ship today. See the table below.

Component	Status
scree.Array type + invariants	✅
pack / unpack / to_padded / from_padded	✅
Reference varlen attention / layernorm / softmax	✅
Bridges: torch.nested, HF padded, FA cu_seqlens, DLPack	✅
NumPy + PyTorch + MLX + JAX backends	✅
Triton varlen_attention forward (H100)	✅ 1.30× of FA-2
Triton varlen_attention backward (FA-2 style: preprocess + dKV + dQ)	✅ 1.61× of FA-2 (full training step)
Triton varlen_rmsnorm (H100)	✅ 13.97× of PyTorch reference
Triton varlen_layernorm (H100)	✅ 1.31× of torch.nn.functional.layer_norm

Install

pip install scree              # numpy backend
pip install "scree[torch]"     # + PyTorch backend
pip install "scree[mlx]"       # + MLX backend (Apple Silicon, Metal)
pip install "scree[jax]"       # + JAX backend

Examples

• examples/01_quickstart.py — pack/unpack + varlen attention
• examples/02_no_pad_transformer.py — full transformer block, zero padding
• examples/03_train_step.py — training step with PyTorch autograd flowing through scree (loss drops 80× in 30 steps)
• examples/04_hf_compat.py — HuggingFace Transformers migration recipe (drop-in via bridges.from_hf_padded / to_hf_padded)
• examples/05_multimodal_interleaved.py — interleaved text + image-patch sequences packed into one scree.Array

Documentation

Full rendered site at https://jvoltci.github.io/scree/.

• Getting started — install, first program, common patterns
• Concepts — the mental model behind values + offsets + ragged_dim
• API reference — every public function and class
• Bridges & migration — moving from torch.nested, HuggingFace, FlashAttention
• Kernels — reference and Triton kernel design
• Architecture — internal layout for contributors
• Benchmarks — methodology and reproduction
• FAQ

Contributing

PRs welcome. See CONTRIBUTING.md for the workflow. Open a GitHub Discussion for anything beyond a small fix.

License

Apache-2.0