Added ability to compare results vs. llama.cpp (#1461)

Example:
```bash
rm -rf results/smollm2_N_FP16/gt.csv
mkdir -p results/smollm2_N_FP16

# References from PyTorch FP16
wwb --base-model HuggingFaceTB/SmolLM2-360M-Instruct --gt-data results/smollm2_N_FP16/gt.csv --hf --num-samples 4


#huggingface-cli download "bartowski/SmolLM2-360M-Instruct-GGUF" "SmolLM2-360M-Instruct-f16.gguf" 
wwb --target-model models/SmolLM2-360M-Instruct-f16.gguf --gt-data results/smollm2_N_FP16/gt.csv --llamacpp --output results/smollm2_N_L_FP16 --num-samples 
```

tools/who_what_benchmark/whowhatbench/model_loaders.py

@@ -41,8 +41,19 @@ def load_text_genai_pipeline(model_dir, device="CPU", ov_config=None):
     return GenAIModelWrapper(openvino_genai.LLMPipeline(model_dir, device=device, **ov_config), model_dir, "text")
 
 
+def load_text_llamacpp_pipeline(model_dir):
+    try:
+        from llama_cpp import Llama
+    except ImportError:
+        logger.error(
+            "Failed to import llama_cpp package. Please install llama-cpp-python.")
+        exit(-1)
+    model = Llama(model_dir)
+    return model
+
+
 def load_text_model(
-    model_id, device="CPU", ov_config=None, use_hf=False, use_genai=False
+    model_id, device="CPU", ov_config=None, use_hf=False, use_genai=False, use_llamacpp=False,
 ):
     if use_hf:
         logger.info("Using HF Transformers API")
@@ -53,6 +64,9 @@ def load_text_model(
     elif use_genai:
         logger.info("Using OpenVINO GenAI API")
         model = load_text_genai_pipeline(model_id, device, ov_config)
+    elif use_llamacpp:
+        logger.info("Using llama.cpp API")
+        model = load_text_llamacpp_pipeline(model_id)
     else:
         logger.info("Using Optimum API")
         from optimum.intel.openvino import OVModelForCausalLM
@@ -276,7 +290,7 @@ def load_inpainting_model(
 
 
 def load_model(
-    model_type, model_id, device="CPU", ov_config=None, use_hf=False, use_genai=False
+    model_type, model_id, device="CPU", ov_config=None, use_hf=False, use_genai=False, use_llamacpp=False
 ):
     if model_id is None:
         return None
@@ -288,7 +302,7 @@ def load_model(
         ov_options = {}
 
     if model_type == "text":
-        return load_text_model(model_id, device, ov_options, use_hf, use_genai)
+        return load_text_model(model_id, device, ov_options, use_hf, use_genai, use_llamacpp)
     elif model_type == "text-to-image":
         return load_text2image_model(
             model_id, device, ov_options, use_hf, use_genai

tools/who_what_benchmark/whowhatbench/text_evaluator.py

@@ -108,6 +108,7 @@ def __init__(
         generation_config=None,
         generation_config_base=None,
         seqs_per_request=None,
+        use_chat_template=None,
     ) -> None:
         assert (
             base_model is not None or gt_data is not None
@@ -123,6 +124,7 @@ def __init__(
         self.generation_config_base = generation_config
         self.seqs_per_request = seqs_per_request
         self.generation_fn = gen_answer_fn
+        self.use_chat_template = use_chat_template
         if self.generation_config is not None:
             assert self.seqs_per_request is not None
 
@@ -202,15 +204,21 @@ def worst_examples(self, top_k: int = 5, metric="similarity"):
         return res
 
     def _generate_data(self, model, gen_answer_fn=None, generation_config=None):
-        def default_gen_answer(model, tokenizer, prompt, max_new_tokens, crop_question):
-            inputs = self.tokenizer(prompt, return_tensors="pt")
-
-            tokens = model.generate(**inputs, do_sample=False, max_new_tokens=max_new_tokens)
-
-            if crop_question:
-                tokens = tokens[:, inputs["input_ids"].shape[-1] :]
-
-            return self.tokenizer.batch_decode(tokens, skip_special_tokens=True)[0]
+        def default_gen_answer(model, tokenizer, prompt, max_new_tokens, crop_question, use_chat_template=False):
+            if use_chat_template:
+                message = [{"role": "user", "content": prompt}]
+                inputs = tokenizer.apply_chat_template(message, tokenize=True, add_generation_prompt=True, return_tensors="pt")
+                tokens = model.generate(inputs, do_sample=False, max_new_tokens=max_new_tokens)
+                if crop_question:
+                    tokens = tokens[:, inputs.shape[-1]:]
+                res = self.tokenizer.decode(tokens[0], skip_special_tokens=True)
+                return res
+            else:
+                inputs = self.tokenizer(prompt, return_tensors="pt")
+                tokens = model.generate(**inputs, do_sample=False, max_new_tokens=max_new_tokens)
+                if crop_question:
+                    tokens = tokens[:, inputs["input_ids"].shape[-1] :]
+                return self.tokenizer.batch_decode(tokens, skip_special_tokens=True)[0]
 
         gen_answer_fn = gen_answer_fn or default_gen_answer
 
@@ -250,6 +258,7 @@ def default_gen_answer(model, tokenizer, prompt, max_new_tokens, crop_question):
                         p,
                         self.max_new_tokens,
                         self._crop_question,
+                        self.use_chat_template
                     )
                 )
         else:

tools/who_what_benchmark/whowhatbench/wwb.py

@@ -40,6 +40,11 @@ def parse_args():
         default=None,
         help="Tokenizer for divergency metric. If not provided, it will be load from base_model or target_model.",
     )
+    parser.add_argument(
+        "--chat-template",
+        action="store_true",
+        help="Whether apply the default chat template.",
+    )
     parser.add_argument(
         "--gt-data",
         default=None,
@@ -137,6 +142,11 @@ def parse_args():
         action="store_true",
         help="Use LLMPipeline from transformers library to instantiate the model.",
     )
+    parser.add_argument(
+        "--llamacpp",
+        action="store_true",
+        help="Use llama-cpp-python to instantiate the model.",
+    )
     parser.add_argument(
         "--image-size",
         type=int,
@@ -190,9 +200,13 @@ def load_prompts(args):
 def load_tokenizer(args):
     tokenizer = None
     if args.tokenizer is not None:
-        tokenizer = AutoTokenizer.from_pretrained(
-            args.tokenizer, trust_remote_code=True
-        )
+        if args.llamacpp:
+            from llama_cpp.llama_tokenizer import LlamaHFTokenizer
+            tokenizer = LlamaHFTokenizer.from_pretrained(args.tokenizer)
+        else:
+            tokenizer = AutoTokenizer.from_pretrained(
+                args.tokenizer, trust_remote_code=True
+            )
     elif args.base_model is not None:
         tokenizer = AutoTokenizer.from_pretrained(
             args.base_model, trust_remote_code=True
@@ -246,8 +260,29 @@ def diff_strings(a: str, b: str, *, use_loguru_colors: bool = False) -> str:
     return "".join(output)
 
 
-def genai_gen_text(model, tokenizer, question, max_new_tokens, skip_question):
-    return model.generate(question, do_sample=False, max_new_tokens=max_new_tokens)
+def genai_gen_text(model, tokenizer, question, max_new_tokens, skip_question, use_chat_template=False):
+    if use_chat_template:
+        model.start_chat()
+        result = model.generate(question, do_sample=False, max_new_tokens=max_new_tokens)
+        model.finish_chat()
+        return result
+    else:
+        return model.generate(question, do_sample=False, max_new_tokens=max_new_tokens)
+
+
+def llamacpp_gen_text(model, tokenizer, question, max_new_tokens, skip_question, use_chat_template=False):
+    if use_chat_template:
+        output = model.create_chat_completion(messages=[{"role": "user", "content": question}], max_tokens=max_new_tokens, temperature=0.0)
+        text = output["choices"][0]["message"]["content"]
+        if skip_question:
+            text = text[len(question):]
+        return text
+    else:
+        output = model(question, max_tokens=max_new_tokens, echo=True, temperature=0.0)
+        text = output["choices"][0]["text"]
+        if skip_question:
+            text = text[len(question):]
+        return text
 
 
 def genai_gen_image(model, prompt, num_inference_steps, generator=None):
@@ -322,7 +357,15 @@ def create_evaluator(base_model, args):
         prompts = load_prompts(args)
 
         if task == "text":
-            tokenizer = load_tokenizer(args)
+            tokenizer = load_tokenizer(args) if not args.llamacpp else None
+
+            if args.genai:
+                gen_answer_fn = genai_gen_text
+            elif args.llamacpp:
+                gen_answer_fn = llamacpp_gen_text
+            else:
+                gen_answer_fn = None
+
             return EvaluatorCLS(
                 base_model=base_model,
                 gt_data=args.gt_data,
@@ -331,7 +374,8 @@ def create_evaluator(base_model, args):
                 similarity_model_id=args.data_encoder,
                 num_samples=args.num_samples,
                 language=args.language,
-                gen_answer_fn=genai_gen_text if args.genai else None,
+                gen_answer_fn=gen_answer_fn,
+                use_chat_template=args.chat_template,
             )
         elif task == "text-to-image":
             return EvaluatorCLS(
@@ -467,10 +511,11 @@ def main():
                 args.ov_config,
                 args.hf,
                 args.genai,
+                args.llamacpp
             )
             all_metrics_per_question, all_metrics = evaluator.score(
                 target_model,
-                evaluator.get_generation_fn() if args.genai else None,
+                evaluator.get_generation_fn() if args.genai or args.llamacpp else None,
                 output_dir=args.output
             )
         logger.info("Metrics for model: %s", args.target_model)