Fix: Re-enable Sparse Compression for 2of4 Examples (#1153)

This PR restores sparse compression for our `2of4` examples, which was
previously disabled due to a bug in the vLLM Cutlass integration.

#### Background  
A bug in the Cutlass integration caused certain sparse-only compressed
models to produce gibberish results. To mitigate this issue, we
temporarily turned off sparse compression for our `2of4` examples.

The bug has since been fixed by @tlrmchlsmth in
[vllm-project/vllm#13198](vllm-project/vllm#13198).
With this fix in place, we can safely re-enable sparse compression for
these examples.

#### Changes  
- Re-enable sparse compression for `2of4` examples.  

#### Testing  
- Verified that sparse-only compressed models now produce expected
outputs.

---------

Signed-off-by: Rahul Tuli <rahul@neuralmagic.com>
Co-authored-by: Dipika Sikka <dipikasikka1@gmail.com>

Author: rahul-tuli

examples/sparse_2of4_quantization_fp8/llama3_8b_2of4.py

@@ -116,7 +116,5 @@ def get_recipe(fp8_enabled):
 print("==========================================\n")
 
 # Save compressed model and tokenizer
-model.save_pretrained(
-    save_dir, save_compressed=args.fp8, disable_sparse_compression=True
-)
+model.save_pretrained(save_dir, save_compressed=args.fp8)
 tokenizer.save_pretrained(save_dir)

tests/e2e/vLLM/configs/sparse_24.yaml

@@ -5,4 +5,4 @@ recipe: tests/e2e/vLLM/recipes/Sparse_2of4/recipe_sparse_2of4.yaml
 scheme: sparse2of4_only
 dataset_id: HuggingFaceH4/ultrachat_200k
 dataset_split: train_sft
-save_compressed: False
+save_compressed: True