Update with latest and polish

README.md

@@ -81,6 +81,20 @@ To evaluate models that are loaded via `AutoSeq2SeqLM` in Huggingface, you inste
 
 > **Warning**: Choosing the wrong model may result in erroneous outputs despite not erroring.
 
+### Neural Magic `deepsparse`
+
+Models from [SparseZoo](https://sparsezoo.neuralmagic.com/) can be evaluated directly in lm-evaluation-harness using [DeepSparse](https://github.com/neuralmagic/deepsparse):
+```bash
+pip install deepsparse-nightly[llm]
+python main.py --model deepsparse --model_args pretrained=zoo:mpt-7b-gsm8k_mpt_pretrain-pruned70_quantized --tasks gsm8k
+```
+
+Compatible models hosted on Hugging Face Hub can be used as well:
+```bash
+python main.py --model deepsparse --model_args pretrained=hf:mgoin/TinyLlama-1.1B-Chat-v0.3-ds --tasks hellaswag
+python main.py --model deepsparse --model_args pretrained=hf:neuralmagic/mpt-7b-gsm8k-pruned60-quant-ds --tasks gsm8k
+```
+
 ### Commercial APIs
 
 Our library also supports language models served via the OpenAI API:

setup.py

@@ -50,5 +50,6 @@ setuptools.setup(
         "sentencepiece": ["sentencepiece>=0.1.98", "protobuf>=4.22.1"],
         "auto-gptq": ["auto-gptq[triton] @ git+https://github.com/PanQiWei/AutoGPTQ"],
         "anthropic": ["anthropic"],
+        "deepsparse": ["deepsparse-nightly[llm]"],
     },
 )