cmedina/agenx-lora-training
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: remove DeepSpeed, use torchrun without distributed training

Browse Source
This commit is contained in:
Christian Medina
2026-07-01 15:10:54 -04:00
parent 7de54746a2
commit 313b44381f
1 changed files with 14 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files

65
training/scripts/train.py
View File

@@ -59,69 +59,30 @@ def train(config_path):
except Exception as e:
print(f"✗ Failed: {e}")
# Strategy 1: 4-bit model variants
print("\n[1/6] Trying: 4-bit model AS-IS with DeepSpeed...")
# Strategy 1: Load 4-bit model AS-IS (no quantization, no DeepSpeed)
print("\n[1/2] Trying: 4-bit model AS-IS...")
try:
import deepspeed
model = AutoModelForCausalLM.from_pretrained(
config["base_model"],
torch_dtype=torch.float16,
device_map="auto",
trust_remote_code=True,
)
ds_config = {"zero_optimization": {"stage": 3, "offload_optimizer": {"device": "cpu"}, "offload_param": {"device": "cpu"}}, "fp16": {"enabled": True}}
optimizer = torch.optim.AdamW(model.parameters(), lr=float(config["train_params"]["learning_rate"]))
model, optimizer, _, _ = deepspeed.initialize(model=model, model_parameters=model.parameters(), optimizer=optimizer, config=ds_config)
print("✓ Success: 4-bit model AS-IS")
print("✓ Success: 4-bit model loaded")
except Exception as e:
print(f"✗ Failed: {e}")
print("\n[2/6] Trying: 4-bit model with bf16 compute...")
# Strategy 2: bf16 to CPU
print("\n[2/2] Trying: bf16 model to CPU...")
try:
from transformers import BitsAndBytesConfig
bnb_config = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True)
model = AutoModelForCausalLM.from_pretrained(config["base_model"], quantization_config=bnb_config, device_map="auto", trust_remote_code=True)
ds_config = {"zero_optimization": {"stage": 3, "offload_optimizer": {"device": "cpu"}, "offload_param": {"device": "cpu"}}, "bf16": {"enabled": True}}
optimizer = torch.optim.AdamW(model.parameters(), lr=float(config["train_params"]["learning_rate"]))
model, optimizer, _, _ = deepspeed.initialize(model=model, model_parameters=model.parameters(), optimizer=optimizer, config=ds_config)
print("✓ Success: 4-bit with bf16 compute")
except Exception as e:
print(f"✗ Failed: {e}")
print("\n[3/6] Trying: 4-bit model to CPU then DeepSpeed...")
try:
model = AutoModelForCausalLM.from_pretrained(config["base_model"], torch_dtype=torch.float16, device_map="cpu", trust_remote_code=True)
ds_config = {"zero_optimization": {"stage": 3, "offload_optimizer": {"device": "cpu"}, "offload_param": {"device": "cpu"}}, "fp16": {"enabled": True}}
optimizer = torch.optim.AdamW(model.parameters(), lr=float(config["train_params"]["learning_rate"]))
model, optimizer, _, _ = deepspeed.initialize(model=model, model_parameters=model.parameters(), optimizer=optimizer, config=ds_config)
print("✓ Success: 4-bit to CPU")
except Exception as e:
print(f"✗ Failed: {e}")
print("\n[4/6] Trying: 4-bit model fp32...")
try:
model = AutoModelForCausalLM.from_pretrained(config["base_model"], torch_dtype=torch.float32, device_map="auto", trust_remote_code=True)
ds_config = {"zero_optimization": {"stage": 3, "offload_optimizer": {"device": "cpu"}, "offload_param": {"device": "cpu"}}, "fp32": {"enabled": True}}
optimizer = torch.optim.AdamW(model.parameters(), lr=float(config["train_params"]["learning_rate"]))
model, optimizer, _, _ = deepspeed.initialize(model=model, model_parameters=model.parameters(), optimizer=optimizer, config=ds_config)
print("✓ Success: 4-bit fp32")
except Exception as e:
print(f"✗ Failed: {e}")
print("\n[5/6] Trying: bf16 model to CPU...")
try:
model = AutoModelForCausalLM.from_pretrained(config["base_model"], torch_dtype=torch.bfloat16, device_map="cpu", low_cpu_mem_usage=True, trust_remote_code=True)
ds_config = {"zero_optimization": {"stage": 3, "offload_optimizer": {"device": "cpu"}, "offload_param": {"device": "cpu"}}, "bf16": {"enabled": True}}
optimizer = torch.optim.AdamW(model.parameters(), lr=float(config["train_params"]["learning_rate"]))
model, optimizer, _, _ = deepspeed.initialize(model=model, model_parameters=model.parameters(), optimizer=optimizer, config=ds_config)
print("✓ Success: bf16 to CPU")
except Exception as e:
print(f"✗ Failed: {e}")
print("\n[6/6] Trying: bf16 model auto placement...")
try:
model = AutoModelForCausalLM.from_pretrained(config["base_model"], torch_dtype=torch.bfloat16, device_map="auto", low_cpu_mem_usage=True, trust_remote_code=True)
print("✓ Success: bf16 auto placement")
model = AutoModelForCausalLM.from_pretrained(
config["base_model"],
torch_dtype=torch.bfloat16,
device_map="cpu",
low_cpu_mem_usage=True,
trust_remote_code=True,
)
print("✓ Success: bf16 model loaded to CPU")
except Exception as e:
print(f"✗ Failed: {e}")
raise RuntimeError("All loading strategies failed!")