[llm_bench] update optimum bench hook for transformer-based imagegen (openvinotoolkit#1525)

Co-authored-by: Ilya Lavrenov <ilya.lavrenov@intel.com>

Author: eaidova

tools/llm_bench/llm_bench_utils/hook_forward.py

@@ -9,6 +9,7 @@ def __init__(self):
         self.text_encoder_step_count = 0
         self.unet_step_count = 0
         self.vae_decoder_step_count = 0
+        self.main_model_name = "unet"
 
     def get_text_encoder_latency(self):
         return (self.text_encoder_time / self.text_encoder_step_count) * 1000 if self.text_encoder_step_count > 0 else 0
@@ -56,7 +57,9 @@ def my_text_encoder(inputs, share_inputs=True, **kwargs):
         pipe.text_encoder.request = my_text_encoder
 
     def new_unet(self, pipe):
-        old_unet = pipe.unet.request
+        main_model = pipe.unet if pipe.unet is not None else pipe.transformer
+        self.main_model_name = "unet" if pipe.unet is not None else "transformer"
+        old_unet = main_model.request
 
         def my_unet(inputs, share_inputs=True, **kwargs):
             t1 = time.time()
@@ -66,7 +69,7 @@ def my_unet(inputs, share_inputs=True, **kwargs):
             self.unet_time_list.append(unet_time)
             self.unet_step_count += 1
             return r
-        pipe.unet.request = my_unet
+        main_model.request = my_unet
 
     def new_vae_decoder(self, pipe):
         old_vae_decoder = pipe.vae_decoder.request

tools/llm_bench/llm_bench_utils/metrics_print.py

@@ -97,17 +97,17 @@ def print_stable_diffusion_infer_latency(iter_str, iter_data, stable_diffusion,
     iter_data['first_token_infer_latency'] = iter_data['first_token_latency']
     iter_data['other_tokens_infer_avg_latency'] = iter_data['other_tokens_avg_latency']
     prefix = f'[{iter_str}][P{prompt_idx}]'
-    log.info(f"{prefix} First step of unet latency: {iter_data['first_token_latency']:.2f} ms/step, "
-             f"other steps of unet latency: {iter_data['other_tokens_avg_latency']:.2f} ms/step",)
+    log.info(f"{prefix} First step of {stable_diffusion.main_model_name} latency: {iter_data['first_token_latency']:.2f} ms/step, "
+             f"other steps of {stable_diffusion.main_model_name} latency: {iter_data['other_tokens_avg_latency']:.2f} ms/step",)
     has_text_encoder_time = stable_diffusion.get_text_encoder_step_count() != -1
     log_str = (
         f"{prefix} Text encoder latency: {stable_diffusion.get_text_encoder_latency():.2f}" if has_text_encoder_time else f"{prefix} Text encoder latency: N/A "
-        f"unet latency: {stable_diffusion.get_unet_latency():.2f} ms/step, "
+        f"{stable_diffusion.main_model_name} latency: {stable_diffusion.get_unet_latency():.2f} ms/step, "
         f"vae decoder latency: {stable_diffusion.get_vae_decoder_latency():.2f} ms/step, ")
     if has_text_encoder_time:
         log_str += f"text encoder step count: {stable_diffusion.get_text_encoder_step_count()}, "
     log_str += (
-        f"unet step count: {stable_diffusion.get_unet_step_count()}, "
+        f"{stable_diffusion.main_model_name} step count: {stable_diffusion.get_unet_step_count()}, "
         f"vae decoder step count: {stable_diffusion.get_vae_decoder_step_count()}")
     log.info(log_str)
 

tools/llm_bench/llm_bench_utils/ov_utils.py

@@ -363,10 +363,11 @@ def create_genai_image_gen_model(model_path, device, ov_config, **kwargs):
     import openvino_genai
 
     class PerfCollector:
-        def __init__(self) -> types.NoneType:
+        def __init__(self, main_model_name="unet") -> types.NoneType:
             self.iteration_time = []
             self.start_time = time.perf_counter()
             self.duration = -1
+            self.main_model_name = main_model_name
 
         def __call__(self, step, num_steps, latents):
             self.iteration_time.append(time.perf_counter() - self.start_time)
@@ -405,8 +406,6 @@ def get_unet_step_count(self):
         def get_vae_decoder_step_count(self):
             return 1
 
-    callback = PerfCollector()
-
     adapter_config = get_lora_config(kwargs.get("lora", None), kwargs.get("lora_alphas", []))
     if adapter_config:
         ov_config['adapters'] = adapter_config
@@ -416,6 +415,11 @@ def get_vae_decoder_step_count(self):
         data = json.load(f)
 
     model_class_name = data.get("_class_name", "")
+    main_model_name = "unet" if "unet" in data else "transformer"
+    callback = PerfCollector(main_model_name)
+
+    orig_tokenizer = AutoTokenizer.from_pretrained(model_path, subfolder="tokenizer")
+    callback.orig_tokenizer = orig_tokenizer
 
     start = time.perf_counter()
 

tools/llm_bench/task/image_generation.py

@@ -123,6 +123,7 @@ def run_image_generation(image_param, num, image_id, pipe, args, iter_data_list,
 def run_image_generation_genai(image_param, num, image_id, pipe, args, iter_data_list, proc_id, mem_consumption, callback=None):
     set_seed(args['seed'])
     input_text = image_param['prompt']
+    input_token_size = callback.orig_tokenizer(input_text, return_tensors="pt").input_ids.numel()
     input_args = collects_input_args(image_param, args['model_type'], args['model_name'], args["num_steps"], args.get("height"), args.get("width"), callback)
     out_str = f"Input params: Batch_size={args['batch_size']}, " \
               f"steps={input_args['num_inference_steps']}, width={input_args['width']}, height={input_args['height']}"
@@ -157,6 +158,7 @@ def run_image_generation_genai(image_param, num, image_id, pipe, args, iter_data
     generation_time = end - start
     iter_data = gen_output_data.gen_iterate_data(
         iter_idx=num,
+        in_size=input_token_size * args['batch_size'],
         infer_count=input_args["num_inference_steps"],
         gen_time=generation_time,
         res_md5=result_md5_list,
@@ -230,8 +232,7 @@ def run_image_generation_benchmark(model_path, framework, device, args, num_iter
                 prefix = '[warm-up]' if num == 0 else '[{}]'.format(num)
                 log.info(f"{prefix}[P{p_idx}] start: {iter_timestamp[num][p_idx]['start']}, end: {iter_timestamp[num][p_idx]['end']}")
 
-    if not use_genai:
-        metrics_print.print_average(iter_data_list, prompt_idx_list, args['batch_size'], False)
+    metrics_print.print_average(iter_data_list, prompt_idx_list, args['batch_size'], False)
     return iter_data_list, pretrain_time, iter_timestamp