Merge f5d35391db into f66a780f68

README.md

@@ -184,6 +184,7 @@ python -m pytest tests/test_rustbpe.py -v -s
 │   ├── smoltalk.py                 # Conglomerate dataset of SmolTalk from HF
 │   └── spellingbee.py              # Task teaching model to spell/count letters
 ├── tests
+│   └── test_engine.py
 │   └── test_rustbpe.py
 └── uv.lock
 ```

nanochat/dataset.py

@@ -11,6 +11,7 @@ import os
 import argparse
 import time
 import requests
+import pyarrow.fs as fs
 import pyarrow.parquet as pq
 from multiprocessing import Pool
 
@@ -20,7 +21,7 @@ from nanochat.common import get_base_dir
 # The specifics of the current pretraining dataset
 
 # The URL on the internet where the data is hosted and downloaded from on demand
-BASE_URL = "https://huggingface.co/datasets/karpathy/fineweb-edu-100b-shuffle/resolve/main"
+BASE_URI = "hf://datasets/karpathy/fineweb-edu-100b-shuffle"
 MAX_SHARD = 1822 # the last datashard is shard_01822.parquet
 index_to_filename = lambda index: f"shard_{index:05d}.parquet" # format of the filenames
 base_dir = get_base_dir()
@@ -68,45 +69,17 @@ def download_single_file(index):
         return True
 
     # Construct the remote URL for this file
-    url = f"{BASE_URL}/{filename}"
+    uri = f"{BASE_URI}/{filename}"
     print(f"Downloading {filename}...")
-
-    # Download with retries
-    max_attempts = 5
-    for attempt in range(1, max_attempts + 1):
-        try:
-            response = requests.get(url, stream=True, timeout=30)
-            response.raise_for_status()
-            # Write to temporary file first
-            temp_path = filepath + f".tmp"
-            with open(temp_path, 'wb') as f:
-                for chunk in response.iter_content(chunk_size=1024 * 1024):  # 1MB chunks
-                    if chunk:
-                        f.write(chunk)
-            # Move temp file to final location
-            os.rename(temp_path, filepath)
-            print(f"Successfully downloaded {filename}")
-            return True
-
-        except (requests.RequestException, IOError) as e:
-            print(f"Attempt {attempt}/{max_attempts} failed for {filename}: {e}")
-            # Clean up any partial files
-            for path in [filepath + f".tmp", filepath]:
-                if os.path.exists(path):
-                    try:
-                        os.remove(path)
-                    except:
-                        pass
-            # Try a few times with exponential backoff: 2^attempt seconds
-            if attempt < max_attempts:
-                wait_time = 2 ** attempt
-                print(f"Waiting {wait_time} seconds before retry...")
-                time.sleep(wait_time)
-            else:
-                print(f"Failed to download {filename} after {max_attempts} attempts")
-                return False
-
-    return False
+    try:
+        # pyarrow.fs uses huggingface_hub with builtin exponential backoff
+        fs.copy_files(uri, filepath)
+    except (requests.RequestException, IOError) as e:
+        print(f"Failed to download {filename}: {e}")
+        return False
+    else:
+        print(f"Successfully downloaded {filename}")
+        return True
 
 
 if __name__ == "__main__":

nanochat/engine.py

@@ -17,8 +17,9 @@ import signal
 import warnings
 from contextlib import contextmanager
 from collections import deque
-from nanochat.common import compute_init
+from nanochat.common import compute_init, autodetect_device_type
 from nanochat.checkpoint_manager import load_model
+from contextlib import nullcontext 
 
 # -----------------------------------------------------------------------------
 # Calculator tool helpers
@@ -328,6 +329,9 @@ if __name__ == "__main__":
     import time
     # init compute
     ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init()
+    device_type = autodetect_device_type()
+    autocast_ctx = torch.amp.autocast(device_type=device_type, dtype=torch.bfloat16) if device_type == "cuda" else nullcontext()
+
     # load the model and tokenizer
     model, tokenizer, meta = load_model("base", device, phase="eval")
     bos_token_id = tokenizer.get_bos_token_id()
@@ -340,10 +344,11 @@ if __name__ == "__main__":
     torch.cuda.synchronize()
     t0 = time.time()
     stream = model.generate(prompt_tokens, **kwargs)
-    for token in stream:
-        generated_tokens.append(token)
-        chunk = tokenizer.decode([token])
-        print(chunk, end="", flush=True)
+    with autocast_ctx:
+        for token in stream:
+            generated_tokens.append(token)
+            chunk = tokenizer.decode([token])
+            print(chunk, end="", flush=True)
     print()
     torch.cuda.synchronize()
     t1 = time.time()
@@ -355,11 +360,12 @@ if __name__ == "__main__":
     stream = engine.generate(prompt_tokens, num_samples=1, **kwargs) # note: runs in fp32
     torch.cuda.synchronize()
     t0 = time.time()
-    for token_column, token_masks in stream:
-        token = token_column[0] # only print out the first row
-        generated_tokens.append(token)
-        chunk = tokenizer.decode([token])
-        print(chunk, end="", flush=True)
+    with autocast_ctx:
+        for token_column, token_masks in stream:
+            token = token_column[0] # only print out the first row
+            generated_tokens.append(token)
+            chunk = tokenizer.decode([token])
+            print(chunk, end="", flush=True)
     print()
     torch.cuda.synchronize()
     t1 = time.time()

nanochat/loss_eval.py

@@ -9,9 +9,9 @@ import torch.distributed as dist
 def evaluate_bpb(model, batches, steps, token_bytes):
     """
     Instead of the naive 'mean loss', this function returns the bits per byte (bpb),
-    which is a tokenization vocab size-indepedent metric, meaning you are still comparing
+    which is a tokenization vocab size-independent metric, meaning you are still comparing
     apples:apples if you change the vocab size. The way this works is that instead of just
-    calculating the average loss as usual, you calculate the sum loss, and indepependently
+    calculating the average loss as usual, you calculate the sum loss, and independently
     also the sum bytes (of all the target tokens), and divide. This normalizes the loss by
     the number of bytes that the target tokens represent.
 

pyproject.toml

@@ -9,6 +9,7 @@ dependencies = [
     "fastapi>=0.117.1",
     "files-to-prompt>=0.6",
     "psutil>=7.1.0",
+    "pyarrow>=21.0.0",
     "regex>=2025.9.1",
     "setuptools>=80.9.0",
     "tiktoken>=0.11.0",

scripts/chat_eval.py

@@ -1,6 +1,6 @@
 """
 Evaluate the Chat model.
-All the generic code lives here, and all the evlauation-specific
+All the generic code lives here, and all the evaluation-specific
 code lives in nanochat directory and is imported from here.
 
 Example runs:

scripts/chat_sft.py

@@ -192,7 +192,7 @@ for step in range(num_iterations):
         })
         model.train()
 
-    # evlauate accuracy of the multiple choice tasks (which are quick to run)
+    # evaluate accuracy of the multiple choice tasks (which are quick to run)
     if last_step or (step > 0 and step % eval_metrics_every == 0):
         model.eval()
         metrics = {}