SmoothQuant google/gemma-2-2b-it
Quantized derivative of google/gemma-2-2b-it.
Not Hub fp16. Weights are the packed/runtime artifact from job 20260921T132955Z-21a810.
WikiText-2 (test, 2048-token windows, 65504 tokens)
Measured on NVIDIA A10G (g5.2xlarge) against the original fp16 snapshot.
Numbers copied from jobs/20260921T132955Z-21a810/benchmark.json.
| Quantized | fp16 snapshot | |
|---|---|---|
| Perplexity | 15.128110 | 14.809264 |
| NLL loss | 2.716555 | 2.695253 |
| Tokens/s (2048 prefill, after warmup) | 6322.2 | 5413.0 |
| Peak VRAM (GB) | 8.090 | 9.973 |
ppl_ratio: 1.021530 (quality_ok=True)improved_throughput: Trueimproved_vram: True
Pull
from huggingface_hub import snapshot_download
path = snapshot_download("chris320211/smoothquant-gemma2-2b-g52xlarge")
This is not a drop-in AutoModelForCausalLM.from_pretrained checkpoint.
Reload with the included quant_agent_inference_adapter.py after the method
repo (and overlay, if any) is on QUANT_AGENT_METHOD_REPO. See
quantization_config.json.
License
Base model license (MIT for Phi-3) plus the method repository license.
Keep LICENSE / NOTICE.md from the snapshot when present.