Merge 2dc85662c3 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/common.py

@@ -113,12 +113,24 @@ def print_banner():
     """
     print0(banner)
 
-def is_ddp():
+def is_ddp_requested() -> bool:
-    # TODO is there a proper way
+    """
-    return int(os.environ.get('RANK', -1)) != -1
+    True if launched by torchrun (env present), even before init.
+    Used to decide whether we *should* initialize a PG.
+    """
+    return all(k in os.environ for k in ("RANK", "LOCAL_RANK", "WORLD_SIZE"))
+
+def is_ddp_initialized() -> bool:
+    """
+    True if torch.distributed is available and the process group is initialized.
+    Used at cleanup to avoid destroying a non-existent PG.
+    """
+    return dist.is_available() and dist.is_initialized()
 
 def get_dist_info():
-    if is_ddp():
+    if is_ddp_requested():
+        # We rely on torchrun's env to decide if we SHOULD init.
+        # (Initialization itself happens in compute init.)
         assert all(var in os.environ for var in ['RANK', 'LOCAL_RANK', 'WORLD_SIZE'])
         ddp_rank = int(os.environ['RANK'])
         ddp_local_rank = int(os.environ['LOCAL_RANK'])
@@ -161,8 +173,8 @@ def compute_init(device_type="cuda"): # cuda|cpu|mps
         torch.set_float32_matmul_precision("high") # uses tf32 instead of fp32 for matmuls
 
     # Distributed setup: Distributed Data Parallel (DDP), optional, and requires CUDA
-    ddp, ddp_rank, ddp_local_rank, ddp_world_size = get_dist_info()
+    is_ddp_requested, ddp_rank, ddp_local_rank, ddp_world_size = get_dist_info()
-    if ddp and device_type == "cuda":
+    if is_ddp_requested and device_type == "cuda":
         device = torch.device("cuda", ddp_local_rank)
         torch.cuda.set_device(device)  # make "cuda" default to this device
         dist.init_process_group(backend="nccl", device_id=device)
@@ -173,11 +185,11 @@ def compute_init(device_type="cuda"): # cuda|cpu|mps
     if ddp_rank == 0:
         logger.info(f"Distributed world size: {ddp_world_size}")
 
-    return ddp, ddp_rank, ddp_local_rank, ddp_world_size, device
+    return is_ddp_requested, ddp_rank, ddp_local_rank, ddp_world_size, device
 
 def compute_cleanup():
     """Companion function to compute_init, to clean things up before script exit"""
-    if is_ddp():
+    if is_ddp_initialized():
         dist.destroy_process_group()
 
 class DummyWandb:

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:
+    with autocast_ctx:
-        generated_tokens.append(token)
+        for token in stream:
-        chunk = tokenizer.decode([token])
+            generated_tokens.append(token)
-        print(chunk, end="", flush=True)
+            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:
+    with autocast_ctx:
-        token = token_column[0] # only print out the first row
+        for token_column, token_masks in stream:
-        generated_tokens.append(token)
+            token = token_column[0] # only print out the first row
-        chunk = tokenizer.decode([token])
+            generated_tokens.append(token)
-        print(chunk, end="", flush=True)
+            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.
 

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 = {}