Qwen3.8-27B — EXL3 4.0 bpw (v2)

EXL3 trellis quantization of Qwen/Qwen3.8-27B at 4.0 bpw, 16 GB, with AWQ smoothing and calibrated FP8 KV-cache scales.

What changed in v2

v1 was calibrated on exllamav3's bundled generic text, because convert_model.py exposes only -cr/-cc and has no option to supply a corpus. v2 changes three things:

1. Calibrated on our own corpus — 2,353 documents / 3.7M tokens, self-distilled from the target model plus a domain blend, replacing the bundled wiki/c4/code text (exllamav3's random-token rows are kept, so full-vocabulary coverage is not lost).
2. 4x the calibration rows (-cr 1000). We measured the statistics the quantizers actually consume: at v1's 250 rows the Hessian XᵀX sits at ~0.66 of its converged value, and AWQ's per-channel means at ~0.94.
3. AWQ smoothing re-run on 8x the data (2.1M tokens vs 262K).

At 4.0 bpw that is worth a full bitrate tier — every disagreement bucket improved significantly over v1 (confident z=−11.99, certain z=−14.02, moderate z=−17.12, near-tie z=−7.30). The gain is concentrated at low bitrates: the identical change at 5.5 bpw moves only moderate significantly, because a quantizer with capacity to spare tolerates a worse Hessian.

Measured

Disagreement against the BF16 base, bucketed by how confident the base model was at each position (its top1−top2 logprob margin). 231 documents, 142,727 scored positions. confident is the pre-registered decider — near-tie is numerical noise and not a quality signal. divmed is the median first-divergence token under free greedy decode, higher is better.

Significance is paired McNemar over matched token positions, n = 33,812 in the confident bucket.

Throughput is deliberately omitted: the v2 builds were benchmarked while another quantization job held the second GPU, so their single-stream numbers understate them and are not comparable to the v1 rows. The architecture is identical to v1 at the same bitrate.

KV cache

Ships per-layer k_scale / v_scale in model-kv-scales.safetensors with kv_cache_scheme under quantization_config — the contract llm-compressor emits and vLLM reads. Measured on this checkpoint, post-RoPE, over the full 3.7M-token corpus, which is where the V absmax stops moving. Only 16 of 64 layers are full_attention and carry a cache, so this is 32 scalars in a 4 KB file.

--kv-cache-dtype fp8      # 32 KiB/token, from 64 KiB at fp16

Notes

• 3.5 bpw was built and rejected. It scores 2.63% confident for 1 GB less than 4.0 bpw's 1.47%. Only 70% of an EXL3 checkpoint is quantized layers — embeddings and head are 23% and fixed with respect to bitrate, because the 248,320-token vocabulary makes them ~1.3B parameters. Below 4.0 bpw the size curve flattens while the damage curve keeps climbing.
• Perplexity is not shown. It is anti-correlated with behavioural damage on this model, so it is excluded from decisions rather than printed beside numbers that drive them.