mchl-labs/custom-bidirectional-hip-smoke

Custom Bidirectional HiP Smoke Tests

Permanent repo for Modal smoke tests related to zero-dense bidirectional Block-HiP attention for diffusion LLMs.

Files

• zero_dense_bihip_modal.py: Modal app for a no-full-attention bidirectional Block-HiP smoke test.

Run

Install Modal and authenticate first:

pip install modal
modal setup

Download or copy the script, then run:

modal run zero_dense_bihip_modal.py

Or download directly:

wget https://huggingface.co/datasets/mchl-labs/custom-bidirectional-hip-smoke/resolve/main/zero_dense_bihip_modal.py
modal run zero_dense_bihip_modal.py

What this tests

The zero-dense path avoids full attention in the algorithmic path:

• no full Q @ K.T selector
• bidirectional hierarchical representative search over key blocks
• local-window anchors
• prefix anchors
• global sink/pivotal blocks estimated from K norms
• sparse-only refresh

An optional dense diagnostic is included only for small-sequence comparison and is not used by the zero-dense algorithm.

Expected pass indicators:

PASS=True
frac_queries_with_future > 0.05
frac_queries_with_past > 0.05
avg_unique_blocks_per_query reasonably above local-only baseline
no NaN/Inf outputs

Notes

This is a correctness smoke test, not a speed benchmark. The sparse attention executor uses Python loops and should later be replaced with a batched/Triton block-sparse kernel.