Glq

New features

• Hessian-based sensitivity profiling for per-layer bit allocation
• --bpw 2.5 auto-allocates 2/3bpw per layer to hit target average
• --min-bpw 2 --max-bpw 4 enables full {2,3,4} bpw range per layer
• Greedy marginal-gain optimizer considers all upgrade jumps (2→3, 2→4, 3→4)
• New module: glq/sensitivity.py with allocate_bpw() + print_allocation_summary()

Results (SmolLM3-3B, WikiText-2, 128 nsamples, L40S)

| Model | Eff. BPW | Perplexity | vs bf16 |
|-------|----------|------------|---------|
| bf16 | 16.00 | 7.04 | 1.00x |
| GLQ 4bpw | 4.00 | 7.19 | 1.02x |
| GLQ 3.5bpw mixed | 3.50 | 7.20 | 1.02x |
| GLQ 3bpw | 3.00 | 7.64 | 1.09x |
| GLQ 3bpw mixed (2+4) | 3.00 | 7.65 | 1.09x |
| GLQ 2.5bpw mixed | 2.50 | 8.08 | 1.15x |
| GLQ 2bpw | 2.00 | 9.61 | 1.36x |

GLQ 3.5bpw mixed matches uniform 4bpw quality at 10% less storage. GLQ 3bpw mixed (2+4) matches uniform 3bpw at 20% less storage.

5-task lm-eval accuracy (SmolLM3-3B)

| Method | Avg | % of bf16 |
|--------|-----|-----------|
| bf16 | 0.709 | 100% |
| GLQ 4bpw | 0.699 | 98.6% |
| GLQ 2bpw | 0.623 | 87.9% |

Also in this release

• Fix KV cache bug (v0.2.1): 0.6 → 14.0 tok/s decode
• Remove B from TC kernel autotune keys (v0.2.1)
• Fair perplexity re-measurement on L40S with 128 nsamples