Merge 7bd999ba02 into f66a780f68

Before, when initiating a batch generation, the first sampled token was broadcasted to all sequences. This change now expands the output logits after a single efficient prefill, allowing for the sampling of a unique starting token for each sequence in the batch.

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
@@ -217,6 +218,8 @@ class Engine:
         ids = torch.tensor([tokens], dtype=torch.long, device=device)
         logits = self.model.forward(ids, kv_cache=kv_cache_prefill)
         logits = logits[:, -1, :]
+        # expand the logits to be of size (num_samples, vocab_size)
+        logits = logits.expand(num_samples, -1)
         next_ids = sample_next_token(logits, rng, temperature, top_k)  # (B, 1)
         sampled_tokens = next_ids[:, 0].tolist()
 
@@ -247,8 +250,6 @@ class Engine:
             # Get sampled tokens - either from prefill or from forward pass
             if first_iteration:
                 # Use the tokens we already sampled from prefill
-                sampled_tokens = [sampled_tokens[0]] * num_samples  # Broadcast first token to all rows
-                # TODO: we should sample a token for each row instead of broadcasting
                 first_iteration = False
             else:
                 # Forward the model and get the next token for each row
@@ -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 = {}