inference/research
verifiedhigh · 0.90updated 2026-10-03

INT4 per-channel affine WikiText-2 perplexity delta vs fp16, Llama2-7B (RTN)

0.79 perplexity_delta (WikiText-2-test ppl points, quantized minus FP16)

Fields

id
benchmarks/int4-perchannel-wikitext2-ppl-delta-llama2-7b-rtn
Metric
quality
Value
0.79
Unit
perplexity_delta (WikiText-2-test ppl points, quantized minus FP16)
Model
Llama2-7B
Methodology
PTQ-with-microscaling-formats paper, Table 1, row 'A:INT8, W:4, RTN' - plain round-to-nearest, no AWQ, no GPTQ, no SmoothQuant. Model Llama2-7B. Fixed-point per-channel affine quantization for both activations and weights; calibration 128 random WikiText-2-train sentences, MinMaxObserver for ranges. Metric: WikiText-2 TEST perplexity. FP16 baseline 5.12; INT8-activation / INT4-weight RTN scores 5.91 - the figure the paper's own text quotes directly as the naive-quantization cost.
Measured by
third_party
Reproducible
yes
notes
DELTA = +0.79 ppl (5.91 vs 5.12). This record is the control. In the same table and the same model: RTN 5.91, GPTQ 5.45, SmoothQuant 5.60, AWQ 5.43, AWQ+GPTQ at INT4 5.37, AWQ+GPTQ at MXINT4 5.53. So the honest spread for '4-bit Llama2-7B on WikiText-2' is 5.37 to 5.91 - 0.54 ppl - and every one of those six numbers is presented somewhere as 'the 4-bit perplexity'. Quoting any single one without naming the algorithm, the activation precision, the group/block size and the harness version is the most common unreproducible claim in quantization work. format_id int8-channelwise is the nearest existing record; the activation side here is W8A8 per-channel affine, so read this together with the smoothquant record rather than as a weight-only AWQ/GPTQ number.

Notes

DELTA = +0.79 ppl (5.91 vs 5.12). This record is the control. In the same table and the same model: RTN 5.91, GPTQ 5.45, SmoothQuant 5.60, AWQ 5.43, AWQ+GPTQ at INT4 5.37, AWQ+GPTQ at MXINT4 5.53. So the honest spread for '4-bit Llama2-7B on WikiText-2' is 5.37 to 5.91 - 0.54 ppl - and every one of those six numbers is presented somewhere as 'the 4-bit perplexity'. Quoting any single one without naming the algorithm, the activation precision, the group/block size and the harness version is the most common unreproducible claim in quantization work. format_id int8-channelwise is the nearest existing record; the activation side here is W8A8 per-channel affine, so read this together with the smoothquant record rather than as a weight-only AWQ/GPTQ number.

Cross-references

Engine (1)

Format (1)