tacit/nanochat
1
0
Fork 0
mirror of https://github.com/karpathy/nanochat.git synced 2025-12-07 12:52:16 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

fix: inference_mode, csv metadata, typo, DDP comment

Browse Source
This commit is contained in:
Dipesh Babu 2025-10-30 02:04:26 -04:00
parent 1ccbaf4416
commit a3e1352f6b
1 changed files with 51 additions and 21 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

72
scripts/base_eval.py
View File

@ -1,5 +1,5 @@
""" """
Evlauate the CORE metric for a given model. Evaluate the CORE metric for a given model.
Run on a single GPU: Run on a single GPU:
python base_eval.py python base_eval.py
@ -17,7 +17,7 @@ import random
import yaml import yaml
from contextlib import nullcontext from contextlib import nullcontext
import pandas as pd import csv
import torch import torch
from nanochat.common import compute_init, compute_cleanup, print0, get_base_dir, autodetect_device_type from nanochat.common import compute_init, compute_cleanup, print0, get_base_dir, autodetect_device_type
@ -28,6 +28,7 @@ from nanochat.core_eval import evaluate_task
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
# nanoChat specific function dealing with I/O etc. # nanoChat specific function dealing with I/O etc.
def evaluate_model(model, tokenizer, device, max_per_task=-1): def evaluate_model(model, tokenizer, device, max_per_task=-1):
""" """
Evaluate a base model on the CORE benchmark. Evaluate a base model on the CORE benchmark.
@ -43,7 +44,12 @@ def evaluate_model(model, tokenizer, device, max_per_task=-1):
with open(config_path, 'r') as f: with open(config_path, 'r') as f:
config = yaml.safe_load(f) config = yaml.safe_load(f)
tasks = config['icl_tasks'] tasks = config['icl_tasks']
eval_metadata = pd.read_csv(eval_meta_data)
eval_metadata = {}
with open(eval_meta_data, "r", newline="", encoding="utf-8") as f:
reader = csv.DictReader(f)
for row in reader:
eval_metadata[row["Eval Task"]] = row
# Evaluate each task # Evaluate each task
results = {} results = {}
@ -57,7 +63,8 @@ def evaluate_model(model, tokenizer, device, max_per_task=-1):
'num_fewshot': task['num_fewshot'][0], 'num_fewshot': task['num_fewshot'][0],
'continuation_delimiter': task.get('continuation_delimiter', ' ') 'continuation_delimiter': task.get('continuation_delimiter', ' ')
} }
print0(f"Evaluating: {label} ({task_meta['num_fewshot']}-shot, type: {task_meta['task_type']})... ", end='') print0(
f"Evaluating: {label} ({task_meta['num_fewshot']}-shot, type: {task_meta['task_type']})... ", end='')
# Load data for this task # Load data for this task
data_path = os.path.join(data_base_path, task_meta['dataset_uri']) data_path = os.path.join(data_base_path, task_meta['dataset_uri'])
@ -72,15 +79,24 @@ def evaluate_model(model, tokenizer, device, max_per_task=-1):
data = data[:max_per_task] data = data[:max_per_task]
# run the evaluation for this task # run the evaluation for this task
accuracy = evaluate_task(model, tokenizer, data, device, task_meta) # eval should be grad-free for stability/perf
with torch.inference_mode():
accuracy = evaluate_task(model, tokenizer, data, device, task_meta)
results[label] = accuracy results[label] = accuracy
row = eval_metadata[eval_metadata["Eval Task"] == label] # row = eval_metadata[eval_metadata["Eval Task"] == label]
random_baseline = row["Random baseline"].values[0] # random_baseline = row["Random baseline"].values[0]
centered_result = (accuracy - 0.01 * random_baseline) / (1.0 - 0.01 * random_baseline) row = eval_meta_data.get(label)
if row is None or "Random baseline" not in row:
raise KeyError(
f"Missing 'Random baseline' for task '{label}' in {eval_meta_data}")
random_baseline = float(row["Random baseline"])
centered_result = (accuracy - 0.01 * random_baseline) / \
(1.0 - 0.01 * random_baseline)
centered_results[label] = centered_result centered_results[label] = centered_result
end_time = time.time() end_time = time.time()
print0(f"accuracy: {accuracy:.4f} | centered: {centered_result:.4f} | time: {end_time - start_time:.2f}s") print0(
f"accuracy: {accuracy:.4f} | centered: {centered_result:.4f} | time: {end_time - start_time:.2f}s")
core_metric = sum(centered_results.values()) / len(centered_results) core_metric = sum(centered_results.values()) / len(centered_results)
out = { out = {
@ -93,8 +109,10 @@ def evaluate_model(model, tokenizer, device, max_per_task=-1):
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
# HuggingFace loading utilities and light wrappers for a model # HuggingFace loading utilities and light wrappers for a model
class ModelWrapper: class ModelWrapper:
"""Lightweight wrapper for a HuggingFace model""" """Lightweight wrapper for a HuggingFace model"""
def __init__(self, model, max_seq_len=None): def __init__(self, model, max_seq_len=None):
self.model = model self.model = model
self.max_seq_len = max_seq_len self.max_seq_len = max_seq_len
@ -104,6 +122,7 @@ class ModelWrapper:
logits = outputs.logits logits = outputs.logits
return logits return logits
def load_hf_model(hf_path: str, device): def load_hf_model(hf_path: str, device):
print0(f"Loading model from: {hf_path}") print0(f"Loading model from: {hf_path}")
# Load the model # Load the model
@ -118,17 +137,23 @@ def load_hf_model(hf_path: str, device):
return model, tokenizer return model, tokenizer
# ----------------------------------------------------------------------------- # -----------------------------------------------------------------------------
def main(): def main():
import argparse import argparse
parser = argparse.ArgumentParser() parser = argparse.ArgumentParser()
parser.add_argument('--hf-path', type=str, default=None, help='HuggingFace model path to evaluate') parser.add_argument('--hf-path', type=str, default=None,
parser.add_argument('--max-per-task', type=int, default=-1, help='Max examples per task to evaluate (-1 = disable)') help='HuggingFace model path to evaluate')
parser.add_argument('--max-per-task', type=int, default=-1,
help='Max examples per task to evaluate (-1 = disable)')
args = parser.parse_args() args = parser.parse_args()
# distributed / precision setup # distributed / precision setup
device_type = autodetect_device_type() device_type = autodetect_device_type()
ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init(device_type) ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init(
autocast_ctx = torch.amp.autocast(device_type=device_type, dtype=torch.bfloat16) if device_type == "cuda" else nullcontext() device_type)
autocast_ctx = torch.amp.autocast(
device_type=device_type, dtype=torch.bfloat16) if device_type == "cuda" else nullcontext()
# Load model and tokenizer from command line or from file system # Load model and tokenizer from command line or from file system
if args.hf_path is not None: if args.hf_path is not None:
@ -136,24 +161,27 @@ def main():
hf_path = args.hf_path hf_path = args.hf_path
print0(f"Loading huggingface model from: {hf_path}") print0(f"Loading huggingface model from: {hf_path}")
model, tokenizer = load_hf_model(hf_path, device) model, tokenizer = load_hf_model(hf_path, device)
model_name = hf_path # just for logging model_name = hf_path # just for logging
model_slug = hf_path.replace("/", "-") # for the output csv file model_slug = hf_path.replace("/", "-") # for the output csv file
else: else:
# load a local model from the file system # load a local model from the file system
model, tokenizer, meta = load_model("base", device, phase="eval") model, tokenizer, meta = load_model("base", device, phase="eval")
model_name = f"base_model (step {meta['step']})" # just for logging model_name = f"base_model (step {meta['step']})" # just for logging
model_slug = f"base_model_{meta['step']:06d}" # for the output csv file # for the output csv file
model_slug = f"base_model_{meta['step']:06d}"
# Evaluate the model # Evaluate the model
with autocast_ctx: with autocast_ctx:
out = evaluate_model(model, tokenizer, device, max_per_task=args.max_per_task) out = evaluate_model(model, tokenizer, device,
max_per_task=args.max_per_task)
# Write out the results to a csv file # Write out the results to a csv file
core_metric = None core_metric = None
centered_results = {} centered_results = {}
if ddp_rank == 0: if ddp_rank == 0:
base_dir = get_base_dir() base_dir = get_base_dir()
output_csv_path = os.path.join(base_dir, "base_eval", f"{model_slug}.csv") output_csv_path = os.path.join(
base_dir, "base_eval", f"{model_slug}.csv")
os.makedirs(os.path.dirname(output_csv_path), exist_ok=True) os.makedirs(os.path.dirname(output_csv_path), exist_ok=True)
results = out["results"] results = out["results"]
centered_results = out["centered_results"] centered_results = out["centered_results"]
@ -161,7 +189,8 @@ def main():
with open(output_csv_path, 'w') as f: with open(output_csv_path, 'w') as f:
f.write(f"{'Task':<35}, {'Accuracy':<10}, {'Centered':<10}\n") f.write(f"{'Task':<35}, {'Accuracy':<10}, {'Centered':<10}\n")
for label in results: for label in results:
f.write(f"{label:<35}, {results[label]:<10.6f}, {centered_results[label]:<10.6f}\n") f.write(
f"{label:<35}, {results[label]:<10.6f}, {centered_results[label]:<10.6f}\n")
f.write(f"{'CORE':<35}, {'':<10}, {core_metric:<10.6f}\n") f.write(f"{'CORE':<35}, {'':<10}, {core_metric:<10.6f}\n")
# Print the content of the csv file to console too # Print the content of the csv file to console too
print0("="*80) print0("="*80)
@ -177,10 +206,11 @@ def main():
"Model": model_name, "Model": model_name,
"CORE metric": core_metric, "CORE metric": core_metric,
}, },
centered_results, # the full table centered_results, # the full table
]) ])
compute_cleanup() compute_cleanup()
if __name__ == "__main__": if __name__ == "__main__":
main() main()