[ADD] add transformer_int4_fp16_loadlowbit_gpu_win api (#11511)

* [ADD] add transformer_int4_fp16_loadlowbit_gpu_win api

* [UPDATE] add int4_fp16_lowbit config and description

* [FIX] fix run.py mistake

* [FIX] fix run.py mistake

* [FIX] fix indent; change dtype=float16 to model.half()

python/llm/dev/benchmark/all-in-one/README.md

@@ -41,6 +41,7 @@ test_api:
   # - "transformer_int4_gpu"                # on Intel GPU, transformer-like API, (qtype=int4), (dtype=fp32)
   # - "transformer_int4_gpu_win"            # on Intel GPU for Windows, transformer-like API, (qtype=int4), (dtype=fp32)
   # - "transformer_int4_loadlowbit_gpu_win" # on Intel GPU for Windows, transformer-like API, (qtype=int4), use load_low_bit API. Please make sure you have used the save.py to save the converted low bit model
+  # - "transformer_int4_fp16_loadlowbit_gpu_win" # on Intel GPU for Windows, transformer-like API, (qtype=int4), (dtype=fp16), use load_low_bit API. Please make sure you have used the save.py to save the converted low bit model
   # - "bigdl_fp16_gpu"                      # on Intel GPU, use ipex-llm transformers API, (dtype=fp16), (qtype=fp16)
   # - "optimize_model_gpu"                  # on Intel GPU, can optimize any pytorch models include transformer model
   # - "deepspeed_optimize_model_gpu"        # on Intel GPU, deepspeed autotp inference
@@ -64,7 +65,7 @@ task: 'continuation' # task can be 'continuation', 'QA' and 'summarize'
 ```
 
 ## (Optional) Save model in low bit
-If you choose the `transformer_int4_loadlowbit_gpu_win` test API, you will need to save the model in low bit first.
+If you choose the `transformer_int4_loadlowbit_gpu_win` or `transformer_int4_fp16_loadlowbit_gpu_win` test API, you will need to save the model in low bit first.
 
 Run `python save.py` will save all models declared in `repo_id` list into low bit models under `local_model_hub` folder.
 

python/llm/dev/benchmark/all-in-one/config.yaml

@@ -17,6 +17,7 @@ test_api:
   # - "transformer_int4_gpu"                # on Intel GPU, transformer-like API, (qtype=int4), (dtype=fp32)
   # - "transformer_int4_gpu_win"            # on Intel GPU for Windows, transformer-like API, (qtype=int4), (dtype=fp32)
   # - "transformer_int4_loadlowbit_gpu_win" # on Intel GPU for Windows, transformer-like API, (qtype=int4), use load_low_bit API. Please make sure you have used the save.py to save the converted low bit model
+  # - "transformer_int4_fp16_loadlowbit_gpu_win" # on Intel GPU for Windows, transformer-like API, (qtype=int4), (dtype=fp16), use load_low_bit API. Please make sure you have used the save.py to save the converted low bit model
   # - "bigdl_fp16_gpu"                      # on Intel GPU, use ipex-llm transformers API, (dtype=fp16), (qtype=fp16)
   # - "optimize_model_gpu"                  # on Intel GPU, can optimize any pytorch models include transformer model
   # - "deepspeed_optimize_model_gpu"        # on Intel GPU, deepspeed autotp inference

python/llm/dev/benchmark/all-in-one/run.py

@@ -137,6 +137,10 @@ def run_model(repo_id, test_api, in_out_pairs, local_model_hub=None, warm_up=1,
         # drop the results of the first time for better performance
         run_transformer_int4_loadlowbit_gpu_win(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials, num_beams, low_bit, cpu_embedding, batch_size, streaming)
         result = run_transformer_int4_loadlowbit_gpu_win(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials, num_beams, low_bit, cpu_embedding, batch_size, streaming)
+    elif test_api == 'transformer_int4_fp16_loadlowbit_gpu_win':
+        # drop the results of the first time for better performance
+        run_transformer_int4_fp16_loadlowbit_gpu_win(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials, num_beams, low_bit, cpu_embedding, batch_size, streaming)
+        result = run_transformer_int4_fp16_loadlowbit_gpu_win(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials, num_beams, low_bit, cpu_embedding, batch_size, streaming)
     elif test_api == 'transformer_autocast_bf16':
         result = run_transformer_autocast_bf16(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials, num_beams, batch_size)
     elif test_api == 'bigdl_ipex_bf16':
@@ -170,7 +174,7 @@ def run_model(repo_id, test_api, in_out_pairs, local_model_hub=None, warm_up=1,
                             low_bit,
                             cpu_embedding,
                             round(result[in_out_pair][-1][5], 2),
-                            result[in_out_pair][-1][6] if any(keyword in test_api for keyword in ['int4_gpu', 'int4_fp16_gpu_win', 'int4_loadlowbit_gpu', 'fp16_gpu', 'deepspeed_optimize_model_gpu']) and not lookahead else 'N/A',
+                            result[in_out_pair][-1][6] if any(keyword in test_api for keyword in ['int4_gpu', 'int4_fp16_gpu_win', 'int4_loadlowbit_gpu', 'int4_fp16_loadlowbit_gpu', 'fp16_gpu', 'deepspeed_optimize_model_gpu']) and not lookahead else 'N/A',
                             streaming if 'win' in test_api else 'N/A',
                             use_fp16_torch_dtype if 'pipeline_parallel_gpu' in test_api else 'N/A'],
                             ) 
@@ -1191,6 +1195,109 @@ def run_transformer_int4_loadlowbit_gpu_win(repo_id,
     return result
 
 
+def run_transformer_int4_fp16_loadlowbit_gpu_win(repo_id,
+                                                 local_model_hub,
+                                                 in_out_pairs,
+                                                 warm_up,
+                                                 num_trials,
+                                                 num_beams,
+                                                 low_bit,
+                                                 cpu_embedding,
+                                                 batch_size,
+                                                 streaming):
+    from ipex_llm.transformers import AutoModel, AutoModelForCausalLM
+    from transformers import AutoTokenizer, GPTJForCausalLM, LlamaTokenizer, TextStreamer
+    model_path = get_model_path(repo_id, local_model_hub)
+    # Load BigDL-LLM optimized low bit model
+    st = time.perf_counter()
+    if repo_id in CHATGLM_IDS:
+        model = AutoModel.load_low_bit(model_path+'-'+low_bit, optimize_model=True, trust_remote_code=True,
+                                       use_cache=True, cpu_embedding=cpu_embedding).eval()
+        tokenizer = AutoTokenizer.from_pretrained(model_path+'-'+low_bit, trust_remote_code=True)
+        model = model.half().to('xpu')
+    elif repo_id in LLAMA_IDS:
+        model = AutoModelForCausalLM.load_low_bit(model_path+'-'+low_bit, optimize_model=True, trust_remote_code=True,
+                                                  use_cache=True, cpu_embedding=cpu_embedding).eval()
+        tokenizer = LlamaTokenizer.from_pretrained(model_path+'-'+low_bit, trust_remote_code=True)
+        model = model.half().to('xpu')
+    elif repo_id in LLAVA_IDS:
+        llava_repo_dir = os.environ.get('LLAVA_REPO_DIR')
+        sys.path.append(rf"{llava_repo_dir}")
+        from llava.model.language_model.llava_llama import LlavaLlamaForCausalLM
+        model = AutoModelForCausalLM.load_low_bit(model_path+'-'+low_bit, optimize_model=True, trust_remote_code=True,
+                                                  use_cache=True, cpu_embedding=cpu_embedding).eval()
+        tokenizer = AutoTokenizer.from_pretrained(model_path+'-'+low_bit, trust_remote_code=True)
+        model = model.half().to('xpu')
+    else:
+        model = AutoModelForCausalLM.load_low_bit(model_path+'-'+low_bit, optimize_model=True, trust_remote_code=True,
+                                                  use_cache=True, cpu_embedding=cpu_embedding).eval()
+        tokenizer = AutoTokenizer.from_pretrained(model_path+'-'+low_bit, trust_remote_code=True)
+        model = model.half().to('xpu')
+    end = time.perf_counter()
+    load_time = end - st
+    print(">> loading of model costs {}s and {}GB".format(load_time, torch.xpu.memory.memory_reserved()/(1024**3)))
+
+    model = BenchmarkWrapper(model)
+    streamer = TextStreamer(tokenizer, skip_prompt=True)
+
+    result = {}
+    with torch.inference_mode():
+        for in_out in in_out_pairs:
+            try:
+                in_out_len = in_out.split("-")
+                in_len = int(in_out_len[0])
+                out_len = int(in_out_len[1])
+                # As different tokenizer has different encodings,
+                # in_len.txt maybe shorter than we need,
+                # use much longer context to make sure input length
+                test_length = min(in_len*2, 8192)
+                while test_length not in [32, 256, 1024, 2048, 8192]:
+                    test_length = test_length * 2
+                input_str = open(f"prompt/continuation/{test_length}.txt", 'r').read()
+                # As different tokenizer has different encodings,
+                # slice the input_ids to ensure the prompt length is required length.
+                input_ids = tokenizer.encode(input_str, return_tensors="pt")
+                input_ids = input_ids[:, :in_len]
+                true_str = tokenizer.batch_decode(input_ids)[0]
+                input_list = [true_str] * batch_size
+                input_ids = tokenizer(input_list, return_tensors="pt").input_ids.to('xpu')
+                actual_in_len = input_ids.shape[1]
+                result[in_out] = []
+                for i in range(num_trials + warm_up):
+                    st = time.perf_counter()
+                    if streaming:
+                        output_ids = model.generate(input_ids, do_sample=False,
+                                                    max_new_tokens=out_len, min_new_tokens=out_len,
+                                                    num_beams=num_beams, streamer=streamer)
+                    else:
+                        output_ids = model.generate(input_ids, do_sample=False,
+                                                    max_new_tokens=out_len, min_new_tokens=out_len,
+                                                    num_beams=num_beams)
+                    torch.xpu.synchronize()
+                    end = time.perf_counter()
+                    output_ids = output_ids.cpu()
+                    print("model generate cost: " + str(end - st))
+                    output = tokenizer.batch_decode(output_ids)
+                    if not streaming:
+                        print(output[0])
+                    actual_out_len = output_ids.shape[1] - actual_in_len
+                    if i >= warm_up:
+                        result[in_out].append([model.first_cost, model.rest_cost_mean, model.encoder_time,
+                                               actual_in_len, actual_out_len, load_time, model.peak_memory])
+                    # torch.xpu.empty_cache() # this may make first token slower
+            except RuntimeError:
+                traceback.print_exc()
+                pass
+            torch.xpu.synchronize()
+            torch.xpu.empty_cache()
+    model.to('cpu')
+    torch.xpu.synchronize()
+    torch.xpu.empty_cache()
+    del model
+    gc.collect()
+    return result
+
+
 def run_transformer_autocast_bf16( repo_id,
                     local_model_hub,
                     in_out_pairs,