GEAK-agent_debug/src/agents/OptimAgent.py

from tqdm import tqdm
import os
import json
from concurrent.futures import ThreadPoolExecutor, as_completed
from agents.reflexion_oneshot import Reflexion_Oneshot
from utils.utils import clear_code, extract_function_signatures, clear_json
from memories.Memory import MemoryClassMeta
from prompts import prompt_for_generation, prompt_for_reflection
from loguru import logger
from tenacity import RetryError


class OptimAgent(Reflexion_Oneshot):
    def __init__(self, model, dataset, corpus_path, max_perf_debug_num=5, mem_file=None):
        super().__init__(model, dataset, corpus_path, mem_file)
        self.max_perf_debug_num = max_perf_debug_num

    def memory_init(self, mem_file=None):
        """
        Args:
            mem_file: previous stored memories, which can be loaded to continue run
        """
        class Memory(metaclass=MemoryClassMeta, field_names=["ps", 
                                                             "call_err_msg", 
                                                             "exe_err_msg",
                                                             "reflection", 
                                                             "function_signatures", 
                                                             "oneshot", 
                                                             "perf_candidates",
                                                             "perf_strategy",
                                                             "raw_code",
                                                             "call_candidate",
                                                             "exe_candidate",
                                                             "perf_debug_num",
                                                             "pass_call", 
                                                             "pass_exe",
                                                             "pass_perf"]):
            pass
        
        if mem_file is not None:
            assert mem_file.endswith(".json"), f"expect a json file, but got {mem_file} instead"
            with open(mem_file, "r") as f:
                input_mems = json.load(f)
            assert len(input_mems) == len(self.dataset), f"expect {len(self.dataset)} samples, but got {len(input_mems)} instead"

        for ps in self.dataset.problem_states:
            if ps.label:
                fs_mem = extract_function_signatures(ps.label)
            else:
                fs_mem = None
            raw_code = [ps.solution] if ps.solution else [""]
            if mem_file is None:
                os_mem = self.instruction_retriever.query(ps.instruction)[0]
                tmp_mem = Memory(ps=ps, 
                                call_err_msg=None,
                                exe_err_msg=None, 
                                reflection=None, 
                                function_signatures=fs_mem, 
                                oneshot=os_mem["code"], 
                                perf_candidates=[],
                                perf_strategy=None,
                                raw_code=raw_code,
                                call_candidate=None,
                                exe_candidate=None,
                                perf_debug_num=0,
                                pass_call=False,
                                pass_exe=False,
                                pass_perf=False,
                                )
            else:
                input_mem = input_mems[ps.filename]
                tmp_mem = Memory(
                    ps=ps,
                    call_err_msg=input_mem["call_err_msg"],
                    exe_err_msg=input_mem["exe_err_msg"], 
                    reflection=input_mem["reflection"], 
                    function_signatures=fs_mem, 
                    oneshot=input_mem["oneshot"], 
                    perf_candidates=input_mem["perf_candidates"],
                    perf_strategy=input_mem["perf_strategy"],
                    raw_code=raw_code,
                    call_candidate=input_mem["call_candidate"],
                    exe_candidate=input_mem["exe_candidate"],
                    perf_debug_num=input_mem["perf_debug_num"],
                    pass_call=input_mem["pass_call"],
                    pass_exe=input_mem["pass_exe"],
                    pass_perf=input_mem["pass_perf"],
                )

            self.memories.append(tmp_mem)
    
    def write_memories(self, file_path):
        output_dict = {}
        with open(file_path, "w") as f:
            for mem in self.memories:
                output = {
                    "call_err_msg": str(mem.call_err_msg),
                    "exe_err_msg": str(mem.exe_err_msg),
                    "reflection": mem.reflection, 
                    "oneshot": mem.oneshot, 
                    "perf_candidates": [list(cand) for cand in mem.perf_candidates],
                    "perf_strategy": mem.perf_strategy,
                    "call_candidate": mem.call_candidate,
                    "exe_candidate": mem.exe_candidate,
                    "perf_debug_num": mem.perf_debug_num,
                    "pass_call": mem.pass_call, 
                    "pass_exe": mem.pass_exe,
                    "pass_perf": mem.pass_perf
                }
                output_dict[mem.ps.filename] = output
            json.dump(output_dict, f)
    
    def run(self, output_path=None, multi_thread=True, datalen=None, iteration_num=0, temperature=0, ancestor_num=2, start_idx=0, gpu_id=0, start_iter=0):
        """
        Args:
            output_path: the folder to store the final result
            multi_thread: whether use multithreading for generating
            datalen: for debug, to specify how many data from the dataset you want to use
            iteration_num: how many iterations you want to run
            temperature: LLM temperature
            ancestor_num: how many samples you want to add in the prompt when optimize the code
            start_idx: start idx of the data rows
            gpu_id: which gpu you want to use when you test the scripts
            start_iter: which iteration you want to start with. useful when you load previous result and memory
        """
        assert ancestor_num >= 0, f"expect ancestor_num to be larger than 0, bug got {ancestor_num}"
        data_len = datalen if datalen else len(self.dataset)
        for iter in range(start_iter, start_iter + iteration_num):
            logger.info(f"\n=== Iteration {iter} ===")
            if output_path is not None:
                root, extension = os.path.splitext(output_path)
                iter_path = f"{root}_{iter}{extension}"
                mem_output_path = f"{root}_mem_{iter}.json"

            if multi_thread:
                thread_num = 3
            
            # generate solution
            logger.info(f"\ngenerate solution")
            with tqdm(total=data_len) as pbar:
                if multi_thread:
                    
                    with ThreadPoolExecutor(max_workers=thread_num) as executor:
                        futures = {executor.submit(self.generate_solution, mem, temperature): mem for mem in self.memories[start_idx:(start_idx + data_len)]}
                        for future in as_completed(futures):
                            pbar.update(1)
                else:
                    for mem in self.memories[start_idx:(start_idx + data_len)]:
                        self.generate_solution(mem, temperature=temperature)
                        pbar.update(1)
            
            # run scripts
            logger.info(f"\nrun scripts on gpu")
            if output_path is not None:
                root, extension = os.path.splitext(output_path)
                tmp_dir = f"{root}_tmp"
                exe_dir = f"{root}_pass_exe"
                perf_result_dir = f"{root}_perf_results"
                perf_log_dir = f"{root}_perf_logs"

            else:
                tmp_dir = "tmp"
                exe_dir = "pass_exe"
                perf_result_dir = "perf_results"
                perf_log_dir = "perf_logs"
            
            for mem in tqdm(self.memories[start_idx:(start_idx + data_len)]):
                try:
                    pass_call, pass_exe, call_stdout, call_stderr, exe_stdout, exe_stderr = self.dataset.test_opt_correctness(mem.raw_code[0], mem.ps.filename, tmp_dir, exe_dir=exe_dir)
                except Exception as e:
                    print(f"failed to test the code for {mem.ps.filename}")
                    mem.call_err_msg = f"failed to test the code due to: {e}"
                    mem.exe_err_msg = f"failed to test the code due to: {e}"
                    continue

                if not pass_call:
                    mem.call_err_msg = call_stderr
                    mem.exe_err_msg = exe_stderr
                elif pass_call and not pass_exe:
                    mem.pass_call = True
                    if exe_stderr == "None":
                        mem.exe_err_msg = None
                    else:
                        mem.exe_err_msg = exe_stderr
                    mem.call_candidate = mem.raw_code[0]
                else:
                    mem.pass_call = True
                    mem.pass_exe = True
                    mem.exe_candidate = mem.raw_code[0]
            
            script_dir = os.path.join(tmp_dir, "perf_gen")
            self.dataset.write_perf_file(input_folder_path=exe_dir, results_path=perf_result_dir, tmp_dir=script_dir) 
            # if multi_thread:
            #     self.dataset.run_perf_scripts_multithread(gpu_count=gpu_count, log_dir=perf_log_dir)
            # else:
            self.dataset.run_perf_scripts(gpu_id=gpu_id, script_dir=script_dir, log_dir=perf_log_dir)

            # get ms and efficiency
            print("===="*40)
            
            for mem in self.memories[start_idx:(start_idx + data_len)]:
                path_gen = os.path.join(perf_result_dir, mem.ps.filename[:-3] + ".json")
                path_ref = os.path.join(self.dataset.perf_ref_folder, mem.ps.filename[:-3] + ".json")
                if not os.path.exists(path_gen):
                    continue
                try:
                    _, efficiency, ms = self.dataset.calculate(path_gen, path_ref)

                    # if efficiency < 0.01:
                    #     print(f"{mem.ps.filename} failed due to inordinarily low efficiency")
                    #     mem.exe_err_msg = f"inordinarily low efficiency: {efficiency}"
                    #     continue

                    mem.pass_perf = True
                    print(f"{mem.ps.filename}: {ms}")
                    print(f"{mem.ps.filename}: {efficiency}\n")
                    
                    mem.raw_code.extend([ms, efficiency])
                    
                except Exception as e:
                    print(f"{mem.ps.filename} failed due to {e}")
                    mem.pass_perf = False
                    # mem.exe_err_msg = e

            # generate reflections
            logger.info(f"\ngenerate reflections")
            with tqdm(total=data_len) as pbar:
                if multi_thread:
                    with ThreadPoolExecutor(max_workers=thread_num) as executor:
                        futures = {executor.submit(self.generate_reflexion, mem, temperature): mem for mem in self.memories[start_idx:(start_idx + data_len)]}
                        for future in as_completed(futures):
                            pbar.update(1)
                else:
                    for mem in self.memories[start_idx:(start_idx + data_len)]:
                        self.generate_reflexion(mem, temperature=temperature)
                        pbar.update(1)

            # update perf_candidates
            for mem in self.memories[start_idx:(start_idx + data_len)]:
                if not mem.pass_perf:
                    continue

                if len(mem.perf_candidates) < ancestor_num:
                    mem.raw_code.append(mem.reflection)
                    if len(mem.raw_code) < 4:
                        logger.info(f"no latency and efficiency info in the raw code for {mem.ps.filename}")
                        mem.pass_perf = False
                        continue
                    mem.perf_candidates.append(tuple(mem.raw_code))
                    mem.perf_candidates = sorted(mem.perf_candidates, key=lambda x: x[1], reverse=True)

                elif mem.perf_candidates[0][1] > mem.raw_code[1]:
                    mem.raw_code.append(mem.reflection)
                    mem.perf_candidates[0] = tuple(mem.raw_code)
                    mem.perf_candidates = sorted(mem.perf_candidates, key=lambda x: x[1], reverse=True)

            for mem in self.memories[start_idx:(start_idx + data_len)]:
                if len(mem.perf_candidates) > 0:
                    mem.ps.solution = mem.perf_candidates[-1][0]
                elif mem.pass_exe:
                    mem.ps.solution = mem.exe_candidate
                elif mem.pass_call:
                    mem.ps.solution = mem.call_candidate
                else:
                    mem.ps.solution = mem.raw_code[0]

            if output_path is not None:
                self.dataset.write_file(iter_path)
                self.write_memories(mem_output_path)

            os.system(f'rm -rf {exe_dir}')
            os.system(f'rm -rf {perf_result_dir}')
            os.system(f'rm -rf {perf_log_dir}')
    
    def generate_solution(self, mem, temperature=0):

        tab = "\n"
        fss_text = "".join(f"* {sig}{tab}" for sig in mem.function_signatures)
        text = prompt_for_generation.prompt.format(
            instruction=mem.ps.instruction,
            function_signatures=fss_text
        )

        # for the one that has perf_candidates, and the code generated in this round pass_exe, we need to generate a new code
        # for the one that has perf_candidates, but the code generated in this round not pass_exe, if the debug_num has exceeds the man_debug_num, then generate a new code
        # otherwise, go to debug
        if len(mem.perf_candidates) > 0 and (mem.pass_exe or (not mem.pass_exe and mem.perf_debug_num >= self.max_perf_debug_num)):
            mem.perf_debug_num = 0

            text += """There are some reference codes(NO.1, NO.2 and so on). The reference codes are arranged in ascending order based on their performance, where lower latencies and higher efficiencies indicate better performance. According to their performance(latency in ms and efficiency in TFLOPS or GB/s) and the corresponding analysis, you need to generate a new code with better performance. You should maintain code correctness during optimization."""

            text +="\nYou can use optimization strategies such as Memory access efficiency, Hardware resource utilization, IR analysis, Assembly analysis, Kernel occupancy, TorchInductor with Triton tuning knobs and Auto-tunable kernel configurations and environment variables."

            for i, cand in enumerate(mem.perf_candidates):
                text += f"\nreference code: {cand[0]}"
                text += f"\nOriginal latency(ms): {cand[1]}"
                text += f"\noriginal efficiency(TFLOPS, GB/s): {cand[2]}"
                text += f"\nAnalysis: {cand[3]}"
            
            text += "\nAnalyze and compare all optimization strategies based on correct codes and give a better strategy motivated by them. Generate a better optimization code based on the better strategy ."
            text += "\nThink before writing the optimization and no more explanation is required after the thinking."
            text += "\nYou should not suggest changes to the name of the function and parameter names, counts, or order."
        else:
            if not mem.raw_code or mem.raw_code[0] == "":
                text += f"\nHere is an example snippet of code: {mem.oneshot}"
            else:
                one_shot = self.code_retriever.query(mem.raw_code[0])[0]["code"]
                text += f"\nHere is an example snippet of code: {one_shot}"
                text += f"\nPrevious attempt implementation:{mem.raw_code[0]}"
                
                if not mem.pass_call:
                    text += f"\nTest messages for previous attempt:{mem.call_err_msg}"
                    text += f"\nTest messages for correctness check of previous attempt:{mem.exe_err_msg}"
                
                elif not mem.pass_exe:
                    text += "\nThe previous attempt implementation can be run successfully."
                    text += f"\nTest messages for correctness check of previous attempt:{mem.exe_err_msg}"
                
                if len(mem.perf_candidates) > 0:
                    mem.perf_debug_num += 1
            
            
            if mem.reflection:
                text += f"\nReflection on previous attempt:{mem.reflection}"

        text += "\nOutput your answer in json format, with the format as follows: {\"thought\": \"\", \"code\": \"\"}. Please strictly output in JSON format."
        text += "\nGenerate the strategy that used to correct and optimized code in the \"strategy\" field."
        text += "\nGenerate the correct and optimized code without explanation, which we can run directly in the \"code\" field."

        msg = [
            {"role": "user", "content": text},
        ]

        try:
            response = self.model.generate(msg, temperature=temperature, max_tokens=30000)
        except:
            logger.info(f"failed to call LLM for {mem.ps.filename}")
            response = {"code": ""}
            
        try:
            mem.raw_code = [clear_code(clear_json(response)["code"])]
        except:
            print(f"failed to extract code for {mem.ps.filename}")
            fail_dir = "failed_to_extract"
            fail_path = os.path.join(fail_dir, mem.ps.filename)
            os.makedirs(fail_dir, exist_ok=True)

            with open(fail_path, "w") as f:
                f.write(response)

            raw_code = response.split("\"code\":")[1]
            raw_code = raw_code.split("}")[0]
            mem.raw_code = [clear_code(raw_code)]
        # finally:
        
        if mem.raw_code[0] is None or mem.raw_code is None:
            print(f"raw code for {mem.ps.filename} is None")
            mem.raw_code = [""]

        mem.pass_call = False
        mem.pass_exe = False
        mem.pass_perf = False

        return
    
    def generate_reflexion(self, mem, temperature):
        if mem.pass_perf:
            reflect_txt = prompt_for_reflection.prompt_ga.format(
                problem=mem.ps.instruction,
                code=mem.raw_code[0],
                latency=mem.raw_code[1],
                efficiency=mem.raw_code[2]
            )
        elif mem.pass_call and mem.pass_exe:
            reflect_txt = prompt_for_reflection.prompt_ga.format(
                problem=mem.ps.instruction,
                code=mem.raw_code[0],
                latency="",
                efficiency=""
            )
        elif mem.pass_call:
            reflect_txt = prompt_for_reflection.prompt_exe.format(
                problem=mem.ps.instruction,
                solution=mem.raw_code[0],
                call_test_result="succeed",
                exe_test_result=mem.exe_err_msg
            )
        else:
            reflect_txt = prompt_for_reflection.prompt.format(
                problem=mem.ps.instruction,
                solution=mem.raw_code[0],
                test_result=mem.call_err_msg
            )
        
        reflect_msg = [
            {
                "role": "user",
                "content": reflect_txt
            }
        ]
        mem.reflection = self.model.generate(reflect_msg, temperature=temperature)