Sparse 2:4 + FP8 Quantization e2e vLLM tests (#1073)

SUMMARY:
- Add 2:4 Sparsity + FP8 Quantization e2e tests

TEST PLAN:
- Models produced by the tests:
nm-testing/TinyLlama-1.1B-Chat-v1.0-sparse2of4_fp8_dynamic-e2e
nm-testing/TinyLlama-1.1B-Chat-v1.0-sparse2of4_only-e2e
- Verified to run e2e with vLLM

Signed-off-by: Kyle Sayers <kylesayrs@gmail.com>

tests/e2e/vLLM/configs/sparse2of4_fp8_dynamic.yaml

@@ -0,0 +1,7 @@
+cadence: "nightly"
+test_type: "regression"
+model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
+recipe: tests/e2e/vLLM/recipes/Sparse_2of4/recipe_sparse_2of4_fp8_dynamic.yaml
+scheme: sparse2of4_fp8_dynamic
+dataset_id: HuggingFaceH4/ultrachat_200k
+dataset_split: train_sft

tests/e2e/vLLM/configs/sparse_24.yaml

@@ -0,0 +1,8 @@
+cadence: "nightly"
+test_type: "regression"
+model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
+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

tests/e2e/vLLM/recipes/Sparse_2of4/recipe_sparse_2of4.yaml

@@ -0,0 +1,6 @@
+sparsity_stage:
+  sparsity_modifiers:
+    SparseGPTModifier:
+      sparsity: 0.5
+      mask_structure: "2:4"
+      sequential_update: false

tests/e2e/vLLM/recipes/Sparse_2of4/recipe_sparse_2of4_fp8_dynamic.yaml

@@ -0,0 +1,25 @@
+sparsity_stage:
+  run_type: oneshot
+  sparsity_modifiers:
+    SparseGPTModifier:
+      sparsity: 0.5
+      mask_structure: "2:4"
+      sequential_update: false
+quantization_stage:
+  run_type: oneshot
+  quantization_modifiers:
+    ConstantPruningModifier:
+      targets: [
+        're:.*q_proj.weight',
+        're:.*k_proj.weight', 
+        're:.*v_proj.weight',
+        're:.*o_proj.weight',
+        're:.*gate_proj.weight',
+        're:.*up_proj.weight',
+        're:.*down_proj.weight',
+      ]
+      start: 0
+    QuantizationModifier:
+      targets: ["Linear"]
+      ignore: ["lm_head"]
+      scheme: "FP8_DYNAMIC"

tests/e2e/vLLM/test_vllm.py

@@ -21,6 +21,7 @@
     vllm_installed = False
     logger.warning("vllm is not installed. This test will be skipped")
 
+
 HF_MODEL_HUB_NAME = "nm-testing"
 
 TEST_DATA_FILE = os.environ.get("TEST_DATA_FILE", "")
@@ -74,6 +75,7 @@ def set_up(self):
         self.recipe = eval_config.get("recipe")
         self.quant_type = eval_config.get("quant_type")
         self.save_dir = eval_config.get("save_dir")
+        self.save_compressed = eval_config.get("save_compressed", True)
 
         logger.info("========== RUNNING ==============")
         logger.info(self.scheme)
@@ -113,7 +115,9 @@ def test_vllm(self):
         self._check_session_contains_recipe()
 
         logger.info("================= SAVING TO DISK ======================")
-        oneshot_model.save_pretrained(self.save_dir)
+        oneshot_model.save_pretrained(
+            self.save_dir, save_compressed=self.save_compressed
+        )
         tokenizer.save_pretrained(self.save_dir)
         recipe_path = os.path.join(self.save_dir, "recipe.yaml")