feat: add 4 loading strategies (no BnB for already-quantized model)

This commit is contained in:
Christian Medina
2026-07-01 15:57:42 -04:00
parent 3089f27901
commit 088521094e

View File

@@ -32,16 +32,63 @@ def train(config_path):
print(f"Loading model: {config['base_model']}")
# Load model AS-IS (already 4-bit quantized with CompressedTensors)
# Load model - try multiple strategies
print(f"\n[INFO] Loading {config['base_model']}...")
model = AutoModelForCausalLM.from_pretrained(
config["base_model"],
torch_dtype=torch.float16,
device_map="auto",
trust_remote_code=True,
)
print("✓ Success: Model loaded successfully")
# Strategy 1: 4-bit AS-IS (already quantized)
print("\n[1/4] Trying: 4-bit AS-IS...")
try:
model = AutoModelForCausalLM.from_pretrained(
config["base_model"],
torch_dtype=torch.float16,
device_map="auto",
trust_remote_code=True,
)
print("✓ Success: 4-bit AS-IS")
except Exception as e:
print(f"✗ Failed: {e}")
# Strategy 2: 4-bit to CPU
print("\n[2/4] Trying: 4-bit to CPU...")
try:
model = AutoModelForCausalLM.from_pretrained(
config["base_model"],
torch_dtype=torch.float16,
device_map="cpu",
trust_remote_code=True,
)
print("✓ Success: 4-bit to CPU")
except Exception as e:
print(f"✗ Failed: {e}")
# Strategy 3: bf16 to CPU
print("\n[3/4] Trying: bf16 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,
)
print("✓ Success: bf16 to CPU")
except Exception as e:
print(f"✗ Failed: {e}")
# Strategy 4: bf16 auto
print("\n[4/4] Trying: bf16 auto...")
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")
except Exception as e:
print(f"✗ Failed: {e}")
raise RuntimeError("All loading strategies failed!")
# Add LoRA
lora_config = LoraConfig(