[LM Eval] Update / fix testing (#1992)

SUMMARY:
- Bump LM Eval to support chartqa - #1953 landed without updating
lm-eval which is required to support chatqa for VL models
- Update recovery metrics for two cases to use different thresholds, not
95%

Author: dsikka

setup.py

@@ -151,7 +151,7 @@ def localversion_func(version: ScmVersion) -> str:
             "pytest>=6.0.0",
             "pytest-mock>=3.6.0",
             "pytest-rerunfailures>=13.0",
-            "lm_eval==0.4.5",
+            "lm_eval==0.4.9",
             # test dependencies
             "beautifulsoup4~=4.12.3",
             "cmarkgfm>=2024.1.14",

tests/lmeval/configs/w4a16_actorder_weight.yaml

@@ -5,6 +5,9 @@ recipe: tests/e2e/vLLM/recipes/actorder/recipe_w4a16_actorder_weight.yaml
 dataset_id: HuggingFaceH4/ultrachat_200k
 dataset_split: train_sft
 lmeval:
+  recovery_threshold:
+    exact_match,strict-match: 0.94
+    exact_match,flexible-extract: 0.94
   metrics:
     exact_match,flexible-extract: 0.72
     exact_match,strict-match: 0.72

tests/lmeval/configs/w4a16_awq_sym.yaml

@@ -5,6 +5,9 @@ recipe: tests/e2e/vLLM/recipes/WNA16/recipe_w4a16_awq_sym.yaml
 dataset_id: HuggingFaceH4/ultrachat_200k
 dataset_split: train_sft
 lmeval:
+  recovery_threshold:
+    exact_match,strict-match: 0.92
+    exact_match,flexible-extract: 0.93
   metrics:
     exact_match,flexible-extract: 0.70
     exact_match,strict-match: 0.70