train with 4gpu

llama-factory/config/mac_template_4gpu.yaml

@@ -0,0 +1,42 @@
+### model
+model_name_or_path: ORG_NAME/MODEL_NAME
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: alpaca_mac
+template: CHAT_TEMPLATE
+cutoff_len: 1024
+max_samples: 4528
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: saves/MODEL_NAME
+logging_steps: 5
+save_steps: 70
+plot_loss: true
+# overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 8
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 6.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
+
+### eval
+val_size: 0.01
+per_device_eval_batch_size: 1
+eval_strategy: steps
+eval_steps: 70
+
+report_to: wandb
+run_name: MODEL_NAME_lora_sft

llm_toolkit/eval_epochs.py

@@ -0,0 +1,149 @@
+import os
+import sys
+import subprocess
+from dotenv import find_dotenv, load_dotenv
+
+from llm_toolkit.llm_utils import *
+from llm_toolkit.translation_utils import *
+
+
+def evaluate_model_all_epochs(
+    model,
+    tokenizer,
+    model_name,
+    adapter_path_base,
+    num_of_entries=-1,
+    result_file=None,
+    start_epoch=0,
+    end_epoch=-1,
+):
+    new_env = os.environ.copy()
+    new_env["MODEL_NAME"] = model_name
+    model = model_name.split("/")[-1]
+
+    new_env["LOAD_IN_4BIT"] = "true" if load_in_4bit else "false"
+    if result_file is not None:
+        new_env["RESULTS_PATH"] = result_file
+
+    if adapter_path_base is None:
+        num_train_epochs = 0
+        print(f"No adapter path provided. Running with base model:{model_name}")
+    else:
+        if end_epoch >= 0:
+            num_train_epochs = end_epoch
+            print(f"Running from epoch {start_epoch} to {end_epoch}")
+        else:
+            # find subdirectories in adapter_path_base
+            # and sort them by epoch number
+            subdirs = [
+                d
+                for d in os.listdir(adapter_path_base)
+                if os.path.isdir(os.path.join(adapter_path_base, d))
+            ]
+
+            subdirs = sorted(subdirs, key=lambda x: int(x.split("-")[-1]))
+            num_train_epochs = len(subdirs)
+            print(f"found {num_train_epochs} checkpoints: {subdirs}")
+
+    for i in range(start_epoch, num_train_epochs + 1):
+        print(f"Epoch {i}")
+        if i == 0:
+            os.unsetenv("ADAPTER_NAME_OR_PATH")
+        else:
+            adapter_path = adapter_path_base + "/" + subdirs[i - 1]
+            new_env["ADAPTER_NAME_OR_PATH"] = adapter_path
+
+        print(f"adapter path: {new_env.get('ADAPTER_NAME_OR_PATH')}")
+
+        log_file = "./logs/{}_epoch_{}.txt".format(model, i)
+        with open(log_file, "w") as f_obj:
+            subprocess.run(
+                f"python llm_toolkit/eval_shots.py {num_of_entries}",
+                shell=True,
+                env=new_env,
+                stdout=f_obj,
+                text=True,
+            )
+
+
+if __name__ == "__main__":
+    found_dotenv = find_dotenv(".env")
+
+    if len(found_dotenv) == 0:
+        found_dotenv = find_dotenv(".env.example")
+    print(f"loading env vars from: {found_dotenv}")
+    load_dotenv(found_dotenv, override=False)
+
+    workding_dir = os.path.dirname(found_dotenv)
+    os.chdir(workding_dir)
+    sys.path.append(workding_dir)
+    print("workding dir:", workding_dir)
+    print(f"adding {workding_dir} to sys.path")
+    sys.path.append(workding_dir)
+
+    model_name = os.getenv("MODEL_NAME")
+    adapter_path_base = os.getenv("ADAPTER_PATH_BASE")
+    start_epoch = int(os.getenv("START_EPOCH", 0))
+    end_epoch = os.getenv("END_EPOCH", -1)
+    load_in_4bit = os.getenv("LOAD_IN_4BIT", "true").lower() == "true"
+    result_file = os.getenv("RESULTS_PATH", None)
+
+    num_of_entries = int(sys.argv[1]) if len(sys.argv) > 1 else -1
+
+    print(
+        model_name,
+        adapter_path_base,
+        load_in_4bit,
+        start_epoch,
+        result_file,
+    )
+
+    device = check_gpu()
+    is_cuda = torch.cuda.is_available()
+
+    print(f"Evaluating model: {model_name} on {device}")
+
+    if is_cuda:
+        torch.cuda.empty_cache()
+        gpu_stats = torch.cuda.get_device_properties(0)
+        start_gpu_memory = round(
+            torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3
+        )
+        max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)
+        print(f"(0) GPU = {gpu_stats.name}. Max memory = {max_memory} GB.")
+        print(f"{start_gpu_memory} GB of memory reserved.")
+
+    model, tokenizer = load_model(model_name, load_in_4bit=load_in_4bit)
+
+    datasets = load_translation_dataset(data_path, tokenizer, num_shots=0)
+    print_row_details(datasets["test"].to_pandas())
+
+    if is_cuda:
+        gpu_stats = torch.cuda.get_device_properties(0)
+        start_gpu_memory = round(
+            torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3
+        )
+        max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)
+        print(f"(1) GPU = {gpu_stats.name}. Max memory = {max_memory} GB.")
+        print(f"{start_gpu_memory} GB of memory reserved.")
+
+    evaluate_model_all_epochs(
+        model,
+        tokenizer,
+        model_name,
+        adapter_path_base,
+        start_epoch=start_epoch,
+        end_epoch=end_epoch,
+        load_in_4bit=load_in_4bit,
+        num_of_entries=num_of_entries,
+        result_file=result_file,
+    )
+
+    if is_cuda:
+        gpu_stats = torch.cuda.get_device_properties(0)
+        start_gpu_memory = round(
+            torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3
+        )
+        max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)
+        print(f"(3) GPU = {gpu_stats.name}. Max memory = {max_memory} GB.")
+        print(f"{start_gpu_memory} GB of memory reserved.")

notebooks/00b_Data Analysis_Few_Shots.ipynb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9659822963e465fcc321c5cc538b9e2eabefcc5492ddc1eaaafd510e854b7202
-size 835024
+oid sha256:ab12535095ccc1615c5379d154c81142686e4fa3dc33fc51808b679fd6fa7116
+size 1234635

results/mac-results_few_shots_metrics.csv

@@ -1,41 +1,47 @@
 model,shots,meteor,bleu_1,rouge_l,ews_score,repetition_score,total_repetitions,rap,num_max_output_tokens,eval_time
-Qwen/Qwen2-72B-Instruct,0,0.4003638205699929,0.12223832517678616,0.3847629463750985,0.0,0.19593998234774934,0.19593998234774934,0.3970180421898014,1,8.894969108561341
-Qwen/Qwen2-72B-Instruct,1,0.4068727655718769,0.13151008586303575,0.3944115167988717,0.0,0.15798764342453664,0.15798764342453664,0.4041216347207881,1,8.983230361871138
-Qwen/Qwen2-72B-Instruct,3,0.4086244766794449,0.13771788946915253,0.39794626590394355,0.0,0.12709620476610767,0.12709620476610767,0.4063954239173824,0,11.657546337157987
-Qwen/Qwen2-72B-Instruct,5,0.4132330811975005,0.1439773872150899,0.4036491307620504,0.0,0.11915269196822595,0.11915269196822595,0.41111822769434864,0,17.167696381288614
-Qwen/Qwen2-72B-Instruct,10,0.41598174489789025,0.14493475334416772,0.4065107334928856,0.0,0.09620476610767872,0.09620476610767872,0.4142591929807702,0,29.728155339805824
-Qwen/Qwen2-72B-Instruct,50,0.4401536409204816,0.1538634893900684,0.4172243423111057,0.0,0.10150044130626655,0.10150044130626655,0.43823160654983345,0,112.50397175639894
-Qwen/Qwen2-7B-Instruct,0,0.377477070949433,0.11783492823424507,0.3679428026833277,0.0,0.07149161518093557,0.07149161518093557,0.3763128359886437,0,0.9805825242718447
-Qwen/Qwen2-7B-Instruct,1,0.38000752971097884,0.11731917392837622,0.3716350861124088,0.0,0.07413945278022947,0.07413945278022947,0.37879237953430883,0,1.0529567519858782
-Qwen/Qwen2-7B-Instruct,3,0.38678180999660744,0.12368875746156333,0.3775321935602687,0.0,0.1412180052956752,0.1412180052956752,0.38444052153933106,0,1.6010591350397176
-Qwen/Qwen2-7B-Instruct,5,0.38784856371389564,0.1227725469820483,0.3821016566709109,0.0,0.09179170344218888,0.09179170344218888,0.38631555618548774,0,2.2894969108561343
-Qwen/Qwen2-7B-Instruct,10,0.38526484346757095,0.12535252418966952,0.3819282327854221,0.0,0.10326566637246248,0.10326566637246248,0.3835535147682633,0,4.006178287731686
-Qwen/Qwen2-7B-Instruct,50,0.3953455943001352,0.12949951844499932,0.3902417317381649,0.0,0.10061782877316858,0.10061782877316858,0.39363409715118836,0,17.46425419240953
-internlm/internlm2_5-7b-chat,0,0.36816799960793073,0.11360521358693174,0.3600976114330957,0.0,0.2144748455428067,0.2144748455428067,0.3648059323539847,0,1.2241835834068844
-internlm/internlm2_5-7b-chat,1,0.3719587471180722,0.1157707566176535,0.3636179646738983,0.0,0.14033539276257723,0.14033539276257723,0.36972107700643503,0,1.3124448367166814
-internlm/internlm2_5-7b-chat,3,0.3747105229822289,0.1154826016668525,0.3683821652101455,0.0,0.17740511915269197,0.17740511915269197,0.37187052462735126,0,1.8578993821712269
-internlm/internlm2_5-7b-chat,5,0.37285562384505977,0.11541534709366409,0.3683735908407485,0.0,0.14827890556045895,0.14827890556045895,0.37048732274065205,0,2.860547219770521
-internlm/internlm2_5-7b-chat,10,0.3750895095392996,0.11696492920010637,0.3679980919671262,0.0,0.13062665489849956,0.13062665489849956,0.37298723763770353,0,5.722859664607237
-internlm/internlm2_5-7b-chat,50,0.37213069871716603,0.11404688073207249,0.36275069101944124,0.0,0.16857899382171226,0.16857899382171226,0.3694484047441432,8,42.29214474845543
-shenzhi-wang/Llama3.1-70B-Chinese-Chat,0,0.3638704024273502,0.10874677881601094,0.3532404744913337,0.0,0.15445719329214475,0.15445719329214475,0.3614642386796342,0,7.8331862312444835
-shenzhi-wang/Llama3.1-70B-Chinese-Chat,1,0.37956764543783084,0.11805442002282653,0.3695558384587787,0.0,0.12533097969991175,0.12533097969991175,0.3775255236309064,0,8.307149161518094
-shenzhi-wang/Llama3.1-70B-Chinese-Chat,3,0.38622483411876246,0.12306660851355093,0.374537863679436,0.0,0.14386584289496912,0.14386584289496912,0.38384366117983154,0,11.681376875551633
-shenzhi-wang/Llama3.1-70B-Chinese-Chat,5,0.3895488616778815,0.12582029733797498,0.37834594185588616,0.0,0.14386584289496912,0.14386584289496912,0.38714719527562863,0,17.23389232127096
-shenzhi-wang/Mistral-7B-v0.3-Chinese-Chat,0,0.3284997502705771,0.08313795089297474,0.31822074756329843,0.0,0.12797881729920565,0.12797881729920565,0.3266954815790356,0,1.204766107678729
-shenzhi-wang/Mistral-7B-v0.3-Chinese-Chat,1,0.3342115436248988,0.08857909016110346,0.3235832621122292,0.0,0.6690203000882613,0.6690203000882613,0.3250691235789747,1,1.4819064430714917
-shenzhi-wang/Mistral-7B-v0.3-Chinese-Chat,3,0.3435621946945506,0.09605927100886698,0.3339835066382988,0.0,0.1262135922330097,0.1262135922330097,0.3417009211692762,0,2.262135922330097
-shenzhi-wang/Mistral-7B-v0.3-Chinese-Chat,5,0.34429078135481284,0.09638489591361771,0.33821501828549394,0.0,0.07590467784642542,0.07590467784642542,0.34316381414750663,2,3.3883495145631066
-shenzhi-wang/Mistral-7B-v0.3-Chinese-Chat,10,0.3408431061510028,0.09735670499814125,0.3331263575051163,0.0,0.10414827890556046,0.10414827890556046,0.339316280986861,11,6.558693733451015
-shenzhi-wang/Mistral-7B-v0.3-Chinese-Chat,50,0.36184201368489227,0.104864029030581,0.34553515243463534,0.0,0.5136804942630185,0.5136804942630185,0.35413782502473057,2,29.659311562224183
-gpt-4o-mini,0,0.3797696357415517,0.1208238389018596,0.37039368743666584,0.0,0.09532215357458076,0.09532215357458076,0.37821133607113916,0,1.5939982347749337
-gpt-4o-mini,1,0.37721414424357197,0.12013402254992751,0.36725632920414286,0.0,0.09179170344218888,0.09179170344218888,0.37572317024740703,0,1.5666372462488967
-gpt-4o-mini,3,0.3772985230936086,0.12400311006855895,0.36822411288225565,0.0,0.09179170344218888,0.09179170344218888,0.3758072155821894,0,1.2868490732568403
-gpt-4o-mini,5,0.35541821046691263,0.1202464326274801,0.3472080523493449,0.0,0.05030891438658429,0.05030891438658429,0.3546452926906339,0,1.203883495145631
-gpt-4o-mini,10,0.37335968903521094,0.1257600824824953,0.36566914304808296,0.0,0.0706090026478376,0.0706090026478376,0.37222227656264567,0,1.1879964695498677
-gpt-4o-mini,50,0.4044690970661121,0.13972883920222515,0.39117898438510224,0.0,0.08473080317740513,0.08473080317740513,0.4029924080114739,0,1.289496910856134
-gpt-4o,0,0.3797419877414444,0.12054600115274576,0.3704197875229349,0.0,0.09532215357458076,0.09532215357458076,0.37818380151840997,0,1.528684907325684
-gpt-4o,1,0.37588586538591867,0.12049862468096047,0.36572207722254313,0.0,0.09179170344218888,0.09179170344218888,0.3744001415355042,0,1.203883495145631
-gpt-4o,3,0.3768512103553621,0.12408746322526747,0.36677555510252513,0.0,0.09355692850838482,0.09355692850838482,0.3753332737090981,0,2.05207413945278
-gpt-4o,5,0.35772544915145654,0.12169683347842021,0.3480606673293163,0.0,0.0353045013239188,0.0353045013239188,0.3571787674657609,0,1.6840247131509267
-gpt-4o,10,0.3746444651189953,0.12498238983123719,0.36695107795536486,0.0,0.0706090026478376,0.0706090026478376,0.37350313867182305,0,1.7899382171226832
-gpt-4o,50,0.40413933252744955,0.13782450337569063,0.3905031632820436,0.0,0.07590467784642542,0.07590467784642542,0.402816463024093,0,2.025595763459841
+Qwen/Qwen2-72B-Instruct,0,0.4003638205699929,0.12223832517678616,0.38486660208459095,0.0,0.19593998234774934,0.19593998234774934,0.3970180421898014,1,8.894969108561341
+Qwen/Qwen2-72B-Instruct,1,0.4068727655718769,0.13151008586303575,0.3946283255752747,0.0,0.15798764342453664,0.15798764342453664,0.4041216347207881,1,8.983230361871138
+Qwen/Qwen2-72B-Instruct,3,0.4086244766794449,0.13771788946915253,0.3979712282250465,0.0,0.12709620476610767,0.12709620476610767,0.4063954239173824,0,11.657546337157987
+Qwen/Qwen2-72B-Instruct,5,0.4132330811975005,0.1439773872150899,0.40387035186513487,0.0,0.11915269196822595,0.11915269196822595,0.41111822769434864,0,17.167696381288614
+Qwen/Qwen2-72B-Instruct,10,0.41598174489789025,0.14493475334416772,0.4065933507975943,0.0,0.09620476610767872,0.09620476610767872,0.4142591929807702,0,29.728155339805824
+Qwen/Qwen2-72B-Instruct,50,0.4401536409204816,0.1538634893900684,0.41722880607716234,0.0,0.10150044130626655,0.10150044130626655,0.43823160654983345,0,112.50397175639894
+Qwen/Qwen2-7B-Instruct,0,0.377477070949433,0.11783492823424507,0.3678523300904837,0.0,0.07149161518093557,0.07149161518093557,0.3763128359886437,0,0.9805825242718447
+Qwen/Qwen2-7B-Instruct,1,0.38000752971097884,0.11731917392837622,0.371517786678723,0.0,0.07413945278022947,0.07413945278022947,0.37879237953430883,0,1.0529567519858782
+Qwen/Qwen2-7B-Instruct,3,0.38678180999660744,0.12368875746156333,0.3780278278830778,0.0,0.1412180052956752,0.1412180052956752,0.38444052153933106,0,1.6010591350397176
+Qwen/Qwen2-7B-Instruct,5,0.38784856371389564,0.1227725469820483,0.38246119910508375,0.0,0.09179170344218888,0.09179170344218888,0.38631555618548774,0,2.2894969108561343
+Qwen/Qwen2-7B-Instruct,10,0.38526484346757095,0.12535252418966952,0.38202725422893463,0.0,0.10326566637246248,0.10326566637246248,0.3835535147682633,0,4.006178287731686
+Qwen/Qwen2-7B-Instruct,50,0.3953455943001352,0.12949951844499932,0.3899754114871057,0.0,0.10061782877316858,0.10061782877316858,0.39363409715118836,0,17.46425419240953
+internlm/internlm2_5-7b-chat,0,0.36816799960793073,0.11360521358693174,0.3600058558701442,0.0,0.2144748455428067,0.2144748455428067,0.3648059323539847,0,1.2241835834068844
+internlm/internlm2_5-7b-chat,1,0.3719587471180722,0.1157707566176535,0.36379026028083117,0.0,0.14033539276257723,0.14033539276257723,0.36972107700643503,0,1.3124448367166814
+internlm/internlm2_5-7b-chat,3,0.3747105229822289,0.1154826016668525,0.36859373323449984,0.0,0.17740511915269197,0.17740511915269197,0.37187052462735126,0,1.8578993821712269
+internlm/internlm2_5-7b-chat,5,0.37285562384505977,0.11541534709366409,0.36845885184482197,0.0,0.14827890556045895,0.14827890556045895,0.37048732274065205,0,2.860547219770521
+internlm/internlm2_5-7b-chat,10,0.3750895095392996,0.11696492920010637,0.36774089220788087,0.0,0.13062665489849956,0.13062665489849956,0.37298723763770353,0,5.722859664607237
+internlm/internlm2_5-7b-chat,50,0.37213069871716603,0.11404688073207249,0.3627041392544321,0.0,0.16857899382171226,0.16857899382171226,0.3694484047441432,8,42.29214474845543
+shenzhi-wang/Llama3.1-70B-Chinese-Chat,0,0.3638704024273502,0.10874677881601094,0.35336472352140924,0.0,0.15445719329214475,0.15445719329214475,0.3614642386796342,0,7.8331862312444835
+shenzhi-wang/Llama3.1-70B-Chinese-Chat,1,0.37956764543783084,0.11805442002282653,0.36984338962652286,0.0,0.12533097969991175,0.12533097969991175,0.3775255236309064,0,8.307149161518094
+shenzhi-wang/Llama3.1-70B-Chinese-Chat,3,0.38622483411876246,0.12306660851355093,0.37461197525974343,0.0,0.14386584289496912,0.14386584289496912,0.38384366117983154,0,11.681376875551633
+shenzhi-wang/Llama3.1-70B-Chinese-Chat,5,0.3895488616778815,0.12582029733797498,0.37850976779334966,0.0,0.14386584289496912,0.14386584289496912,0.38714719527562863,0,17.23389232127096
+shenzhi-wang/Llama3.1-70B-Chinese-Chat,10,0.3897515010230098,0.12957008401715697,0.3804272354384455,0.0,0.1262135922330097,0.1262135922330097,0.3876399935245347,0,
+shenzhi-wang/Llama3.1-8B-Chinese-Chat,0,0.3476217085789916,0.09799438963103267,0.33493508618013,0.0,2.2162400706090026,2.2162400706090026,0.319817551404022,2,1.0750220653133276
+shenzhi-wang/Llama3.1-8B-Chinese-Chat,1,0.35430794534292803,0.10438367949419078,0.34360907692906495,0.0,0.19505736981465135,0.19505736981465135,0.3513601482457364,0,1.1571050308914386
+shenzhi-wang/Llama3.1-8B-Chinese-Chat,3,0.3600739839089376,0.10709900175348612,0.35151668174502293,0.0,0.14827890556045895,0.14827890556045895,0.3577868691137631,0,1.9814651368049427
+shenzhi-wang/Llama3.1-8B-Chinese-Chat,5,0.36231904915539526,0.11204735364530892,0.3555966636828387,0.0,0.15798764342453664,0.15798764342453664,0.3598691748988386,0,2.8146513680494265
+shenzhi-wang/Llama3.1-8B-Chinese-Chat,10,0.3615257973929306,0.10125226501021815,0.35395169595888565,0.0,0.8314210061782877,0.8314210061782877,0.34940648099916116,22,
+shenzhi-wang/Mistral-7B-v0.3-Chinese-Chat,0,0.3284997502705771,0.08313795089297474,0.31837381406868526,0.0,0.12797881729920565,0.12797881729920565,0.3266954815790356,0,1.204766107678729
+shenzhi-wang/Mistral-7B-v0.3-Chinese-Chat,1,0.3342115436248988,0.08857909016110346,0.32376944492764814,0.0,0.6690203000882613,0.6690203000882613,0.3250691235789747,1,1.4819064430714917
+shenzhi-wang/Mistral-7B-v0.3-Chinese-Chat,3,0.3435621946945506,0.09605927100886698,0.33411105509944494,0.0,0.1262135922330097,0.1262135922330097,0.3417009211692762,0,2.262135922330097
+shenzhi-wang/Mistral-7B-v0.3-Chinese-Chat,5,0.34429078135481284,0.09638489591361771,0.33818535378281456,0.0,0.07590467784642542,0.07590467784642542,0.34316381414750663,2,3.3883495145631066
+shenzhi-wang/Mistral-7B-v0.3-Chinese-Chat,10,0.3408431061510028,0.09735670499814125,0.3331892932821734,0.0,0.10414827890556046,0.10414827890556046,0.339316280986861,11,6.558693733451015
+shenzhi-wang/Mistral-7B-v0.3-Chinese-Chat,50,0.36184201368489227,0.104864029030581,0.3457669052738412,0.0,0.5136804942630185,0.5136804942630185,0.35413782502473057,2,29.659311562224183
+gpt-4o-mini,0,0.3797696357415517,0.1208238389018596,0.37055778050320864,0.0,0.09532215357458076,0.09532215357458076,0.37821133607113916,0,1.5939982347749337
+gpt-4o-mini,1,0.37721414424357197,0.12013402254992751,0.36752595751803463,0.0,0.09179170344218888,0.09179170344218888,0.37572317024740703,0,1.5666372462488967
+gpt-4o-mini,3,0.3772985230936086,0.12400311006855895,0.3682965259271725,0.0,0.09179170344218888,0.09179170344218888,0.3758072155821894,0,1.2868490732568403
+gpt-4o-mini,5,0.35541821046691263,0.1202464326274801,0.34743907979125577,0.0,0.05030891438658429,0.05030891438658429,0.3546452926906339,0,1.203883495145631
+gpt-4o-mini,10,0.37335968903521094,0.1257600824824953,0.3655455159774728,0.0,0.0706090026478376,0.0706090026478376,0.37222227656264567,0,1.1879964695498677
+gpt-4o-mini,50,0.4044690970661121,0.13972883920222515,0.3915950550621088,0.0,0.08473080317740513,0.08473080317740513,0.4029924080114739,0,1.289496910856134
+gpt-4o,0,0.3797419877414444,0.12054600115274576,0.37050277223396,0.0,0.09532215357458076,0.09532215357458076,0.37818380151840997,0,1.528684907325684
+gpt-4o,1,0.37588586538591867,0.12049862468096047,0.36605424160788713,0.0,0.09179170344218888,0.09179170344218888,0.3744001415355042,0,1.203883495145631
+gpt-4o,3,0.3768512103553621,0.12408746322526747,0.3667929041403734,0.0,0.09355692850838482,0.09355692850838482,0.3753332737090981,0,2.05207413945278
+gpt-4o,5,0.35772544915145654,0.12169683347842021,0.3484913675543446,0.0,0.0353045013239188,0.0353045013239188,0.3571787674657609,0,1.6840247131509267
+gpt-4o,10,0.3746444651189953,0.12498238983123719,0.3667923043349673,0.0,0.0706090026478376,0.0706090026478376,0.37350313867182305,0,1.7899382171226832
+gpt-4o,50,0.40413933252744955,0.13782450337569063,0.39078212423794856,0.0,0.07590467784642542,0.07590467784642542,0.402816463024093,0,2.025595763459841

scripts/tune-lf-4gpu.sh

@@ -0,0 +1,15 @@
+#!/bin/sh
+
+BASEDIR=$(dirname "$0")
+cd $BASEDIR/../llama-factory/
+echo Current Directory:
+pwd
+
+export ORG_NAME=$1
+export MODEL_NAME=$2
+export CHAT_TEMPLATE=$3
+export DATA_PATH=../datasets/mac/mac.tsv
+export YAML=config/mac_template_4gpu.yaml
+
+python ../llm_toolkit/setup_lf.py
+llamafactory-cli train config/models/$MODEL_NAME.yaml

scripts/tune-mac-4gpu.sh

@@ -14,6 +14,6 @@ grep MemTotal /proc/meminfo
 #pip install -r requirements.txt
 #cd ../LLaMA-Factory && pip install -e .[torch,metrics,vllm] && cd -
 
-./scripts/tune-lf.sh Qwen Qwen2-72B-Instruct qwen
+./scripts/tune-lf-4gpu.sh Qwen Qwen2-72B-Instruct qwen
 
-./scripts/tune-lf.sh shenzhi-wang Llama3.1-70B-Chinese-Chat llama3
+./scripts/tune-lf-4gpu.sh shenzhi-wang Llama3.1-70B-Chinese-Chat llama3