Merge 42b05eea7e 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/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.
 

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 = {}
@@ -212,18 +212,30 @@ for step in range(num_iterations):
         break
 
     # evaluate the gradient
+    total_loss_sum = torch.tensor(0.0, device=device)  # sum of losses
     num_tokens = torch.tensor(0, device=device) # the number of "active" tokens of supervision seen
     for micro_step in range(grad_accum_steps):
         train_inputs, train_targets = next(train_iter)
         with autocast_ctx:
-            loss = model(train_inputs, train_targets)
-        train_loss = loss.detach() # for logging
-        loss = loss / grad_accum_steps # each .backward() is a grad sum => normalize loss here
+            loss = model(train_inputs, train_targets, loss_reduction='sum')
+        total_loss_sum += loss.detach() # for logging
         loss.backward() # accumulate the gradient
         num_tokens += (train_targets >= 0).sum()
     if ddp:
+        dist.all_reduce(total_loss_sum, op=dist.ReduceOp.SUM)
         dist.all_reduce(num_tokens, op=dist.ReduceOp.SUM) # sum over ranks
 
+    # scale gradients by total number of tokens
+    num_tokens_item = num_tokens.item()
+    if num_tokens_item == 0:
+        print0(f"Warning: the number of valid tokens in train targets is 0 at step {step}, skipping model update")
+        model.zero_grad(set_to_none=True)
+        continue
+
+    for param in model.parameters():
+        if param.grad is not None:
+            param.grad.div_(num_tokens_item)
+
     # learning rate scheduler
     lrm = get_lr_multiplier(step)
     for opt in optimizers:
@@ -236,8 +248,7 @@ for step in range(num_iterations):
     model.zero_grad(set_to_none=True)
 
     # logging
-    train_loss_item = train_loss.item()
-    num_tokens_item = num_tokens.item()
+    train_loss_item = total_loss_sum.item() / num_tokens_item
     print0(f"Step {step:05d}/{num_iterations:05d} | Training loss: {train_loss_item:.6f}| lrm: {lrm:.6f}| num_tokens: {num_tokens_item:,}")
     wandb_run.log({
         "step": step,