Glq

Highlights

Block-diagonal GLQ decode is now faster than the legacy power-of-2 fused path. On Blackwell RTX PRO 6000 with SmolLM2-135M 4bpw:

| Path | v0.2.9 | v0.2.10 |
|------|--------|---------|
| Block-diag eager | 51.1 tok/s | 53.6 tok/s |
| Block-diag + CUDA graph | 121.6 tok/s | 136.3 tok/s |
| Legacy pow2 + CUDA graph | 124.3 tok/s | 124.3 tok/s |

Block-diag graph now beats pow2 graph by ~10% — same butterfly work, 3 launches total, exact in_features (no padding).

Features

• Phase A — CUDA-graph-safe forward for block-diagonal E8RHTLinear. Eager _blocks_n/m_tensor construction in __init__, cached empty placeholders, explicit device="cpu" pin so HF's init_empty_weights meta-default-device doesn't silently promote bookkeeping tensors to meta.
• Phase B — Fused multi-block FHT kernel (glq_{input,output}_rht_multiblock_kernel). Collapses N per-sub-block launches into one. gridDim.y = num_blocks, blockIdx.y selects the sub-block via packed int4 device metadata. Dispatches via max_bs ≤ 8192 gate; legacy per-block loop retained for larger blocks.

Fixes

• fast_hadamard_transform falls back to the PyTorch implementation when the input is on CPU (previously raised even with the CUDA pkg installed).
• _process_model_before_weight_loading no longer crashes on models without .config (e.g. bare nn.Sequential).

Testing

• 30 new CUDA fast-path tests for block-diagonal E8RHTLinear covering multiblock↔legacy equivalence, B=1 matvec vs B≥2 TC consistency, CUDA graph bit-exactness, and large-block fallback.
• 8 stale tests updated for block-diag-default and nsamples=128-default semantics.
• Full suite: 260 passed, 3 skipped, 0 failed.

Install

pip install glq==0.2.10