set kv cache in f16 (#3101)

### Changes

explicitly disable kv cache compression to u8, f16 precision is used
instead.

### Reason for changes

PTWC nightly has a different metrics (ticket 157594). 
It happens, because since
openvinotoolkit/openvino#27454 KV Cache
compressed to u8 by default and it affects accuracy of fp32 models
(ticket 157571).

Propose using kv cache in the f16 in order to handle issues in nncf
rather than in ov (there's still an open issue with kv cache
compression, and it can be modified in the nearest future)

### Related tickets

157571
157594

### Tests

- [x] openvino-nightly/job/post_training_weight_compression/56

![image](https://github.com/user-attachments/assets/0772a8e5-0f92-4f53-8ac0-e16841bd8193)
- [x] https://github.com/openvinotoolkit/nncf/actions/runs/11934079602
- [x] job/weekly/job/openvino-nightly/job/test_examples/77

examples/llm_compression/openvino/tiny_llama/main.py

@@ -67,7 +67,9 @@ 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, ov_config={"DYNAMIC_QUANTIZATION_GROUP_SIZE": "0", "KV_CACHE_PRECISION": "f16"}
+    )
     input_ids = tokenizer("What is PyTorch?", return_tensors="pt").to(device=model.device)
 
     start_t = time.time()

examples/llm_compression/openvino/tiny_llama_find_hyperparams/main.py

@@ -249,6 +249,7 @@ def main():
         "NUM_STREAMS": "1",
         "CACHE_DIR": "",
         "DYNAMIC_QUANTIZATION_GROUP_SIZE": "0",
+        "KV_CACHE_PRECISION": "f16",
     }
     model = OVModelForCausalLM.from_pretrained(
         model_id,

tests/post_training/pipelines/lm_weight_compression.py

@@ -271,7 +271,12 @@ def _validate(self):
         if os.getenv("NNCF_TEST_REGEN_DOT") is not None:
             print("Collection ground-truth reference data")
             model_gold = OVModelForCausalLM.from_pretrained(
-                self.fp32_model_dir, trust_remote_code=True, load_in_8bit=False, compile=False, stateful=is_stateful
+                self.fp32_model_dir,
+                trust_remote_code=True,
+                load_in_8bit=False,
+                compile=False,
+                stateful=is_stateful,
+                ov_config={"KV_CACHE_PRECISION": "f16"},
             )
             evaluator = Evaluator(base_model=model_gold, tokenizer=self.preprocessor, metrics=("similarity",))
             evaluator.dump_gt(str(gt_data_path))
@@ -290,7 +295,7 @@ def _validate(self):
                 load_in_8bit=False,
                 compile=False,
                 stateful=is_stateful,
-                ov_config={"DYNAMIC_QUANTIZATION_GROUP_SIZE": "0"},
+                ov_config={"DYNAMIC_QUANTIZATION_GROUP_SIZE": "0", "KV_CACHE_PRECISION": "f16"},
             )
         print("Evaluation of the target model")
         _, all_metrics = evaluator.score(compressed_model_hf)