comfy-sparse-attn 0.0.3

Sparse / variable-length attention dispatch for ComfyUI custom nodes

comfy-sparse-attn

Sparse / variable-length attention dispatch for ComfyUI custom nodes, plus a filesystem-link utility for staging sparse primitives under the comfy namespace ahead of their upstream merge.

Quick Start

pip install comfy-sparse-attn

Varlen Attention

Handles the varlen case that ComfyUI's built-in attention doesn't cover: packed sequences of different lengths across a batch, as produced by sparse 3D models (point clouds, voxels, VarLenTensors). For ordinary dense attention use optimized_attention_for_device from ComfyUI directly.

from comfy_sparse_attn import dispatch_varlen_attention

# q, k, v: [T, H, D]  — total tokens packed across the batch
# cu_seqlens_q / cu_seqlens_kv: [B+1] int32 cumulative lengths
# max_seqlen_q / max_seqlen_kv: int, longest sequence in the batch
out = dispatch_varlen_attention(
    q, k, v,
    cu_seqlens_q, cu_seqlens_kv,
    max_seqlen_q, max_seqlen_kv,
)
# out: [T, H, D]

Backend Priority

GPU	Priority
SM ≥ 8.0 (Ampere / Ada / Hopper / Blackwell)	sage2 > flash > xformers > sdpa
Older / CPU	xformers > sdpa

comfy Namespace Injection

Sparse primitives (comfy.sparse, comfy.ops_sparse, comfy.attention_sparse) are written as if already merged into ComfyUI main. setup_link creates symlinks at those paths at node-load time so the imports work today. When ComfyUI ships the real files, setup_link detects them and becomes a no-op — no code changes needed.

Call it from your node's __init__.py before any model imports:

import pathlib
from comfy_sparse_attn import setup_link

_HERE = pathlib.Path(__file__).parent / "my_model"
setup_link(_HERE / "sparse.py",           "sparse.py")            # → comfy/sparse.py
setup_link(_HERE / "ops_sparse.py",       "ops_sparse.py")        # → comfy/ops_sparse.py
setup_link(_HERE / "attention_sparse.py", "attention_sparse.py")  # → comfy/attention_sparse.py

setup_link skips silently if:

• The target is a real file (ComfyUI shipped it natively).
• The target is already a symlink pointing to the same source.

What This Package Is Not

• Not a dense attention dispatcher. Use comfy.ldm.modules.attention.optimized_attention_for_device.
• Not a kernel implementation. Routes to ComfyUI's built-in backends or xformers/SDPA.

License

MIT