From eeb21a039d22abce7b2b53a1f9b8a2008f5decfd Mon Sep 17 00:00:00 2001
From: Nikolay <nikolay.lyalyushkin@intel.com>
Date: Tue, 6 Aug 2024 17:54:35 +0200
Subject: [PATCH] New metrics for weight compression with dynamic quantization

---
 .../llm_compression/openvino/tiny_llama/main.py  |  2 +-
 .../openvino/tiny_llama_find_hyperparams/main.py |  1 -
 tests/post_training/data/wc_reference_data.yaml  | 16 ++++++++--------
 .../pipelines/lm_weight_compression.py           |  1 -
 4 files changed, 9 insertions(+), 11 deletions(-)

diff --git a/examples/llm_compression/openvino/tiny_llama/main.py b/examples/llm_compression/openvino/tiny_llama/main.py
index dd03a4361c6..f2be54ce1aa 100644
--- a/examples/llm_compression/openvino/tiny_llama/main.py
+++ b/examples/llm_compression/openvino/tiny_llama/main.py
@@ -67,7 +67,7 @@ def transform_fn(data, model, tokenizer):
     )
     model.save_pretrained(OUTPUT_DIR)
 
-    model = OVModelForCausalLM.from_pretrained(OUTPUT_DIR, ov_config={"DYNAMIC_QUANTIZATION_GROUP_SIZE": "0"})
+    model = OVModelForCausalLM.from_pretrained(OUTPUT_DIR)
     input_ids = tokenizer("What is PyTorch?", return_tensors="pt").to(device=model.device)
 
     start_t = time.time()
diff --git a/examples/llm_compression/openvino/tiny_llama_find_hyperparams/main.py b/examples/llm_compression/openvino/tiny_llama_find_hyperparams/main.py
index 7ab0176eb85..6b57b9481f2 100644
--- a/examples/llm_compression/openvino/tiny_llama_find_hyperparams/main.py
+++ b/examples/llm_compression/openvino/tiny_llama_find_hyperparams/main.py
@@ -245,7 +245,6 @@ def main():
         "PERFORMANCE_HINT": "LATENCY",
         "NUM_STREAMS": "1",
         "CACHE_DIR": "",
-        "DYNAMIC_QUANTIZATION_GROUP_SIZE": "0",
     }
     model = OVModelForCausalLM.from_pretrained(
         model_id,
diff --git a/tests/post_training/data/wc_reference_data.yaml b/tests/post_training/data/wc_reference_data.yaml
index 08a5b426746..d1d988e2a88 100644
--- a/tests/post_training/data/wc_reference_data.yaml
+++ b/tests/post_training/data/wc_reference_data.yaml
@@ -1,32 +1,32 @@
 tinyllama_data_free_backend_OV:
-  metric_value: 0.73873
+  metric_value: 0.72494
   num_int4: 114
   num_int8: 84
 tinyllama_data_aware_backend_OV:
-  metric_value: 0.85767
+  metric_value: 0.85635
   num_int4: 94
   num_int8: 124
 tinyllama_data_aware_awq_stateful_backend_OV:
-  metric_value: 0.85616
+  metric_value: 0.84951
   num_int4: 94
   num_int8: 124
 tinyllama_data_aware_awq_scale_estimation_backend_OV:
-  metric_value: 0.85502
+  metric_value: 0.85299
   num_int4: 94
   num_int8: 124
 tinyllama_data_aware_awq_scale_estimation_stateful_backend_OV:
-  metric_value: 0.85502
+  metric_value: 0.84125
   num_int4: 94
   num_int8: 124
 tinyllama_int8_data_free_backend_TORCH:
-  metric_value: 0.95624
+  metric_value: 0.94924
   num_int4: 0
   num_int8: 312
 tinyllama_data_aware_gptq_backend_OV:
-  metric_value: 0.87134
+  metric_value: 0.86621
   num_int4: 94
   num_int8: 124
 tinyllama_scale_estimation_per_channel_backend_OV:
-  metric_value: 0.81389
+  metric_value: 0.82356
   num_int4: 188
   num_int8: 124
diff --git a/tests/post_training/pipelines/lm_weight_compression.py b/tests/post_training/pipelines/lm_weight_compression.py
index 27479fe6a50..06074701b42 100644
--- a/tests/post_training/pipelines/lm_weight_compression.py
+++ b/tests/post_training/pipelines/lm_weight_compression.py
@@ -275,7 +275,6 @@ def _validate(self):
                 load_in_8bit=False,
                 compile=False,
                 stateful=is_stateful,
-                ov_config={"DYNAMIC_QUANTIZATION_GROUP_SIZE": "0"},
             )
         print("Evaluation of the target model")
         _, all_metrics = evaluator.score(compressed_model_hf)