Merge 4715fdcf52 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.
 

rustbpe/Cargo.toml

@@ -13,3 +13,4 @@ pyo3-log = "0.12.4"
 ahash = "0.8.12"
 rayon = "1.11.0"
 compact_str = "0.9.0"
+num-format = "0.4"

rustbpe/src/lib.rs

@@ -6,6 +6,7 @@ use fancy_regex::Regex;
 use pyo3::prelude::*;
 
 use ahash::{AHashMap, AHashSet};
+use num_format::{Locale, ToFormattedString};
 use compact_str::CompactString;
 use rayon::prelude::*;
 
@@ -164,15 +165,15 @@ impl Tokenizer {
     fn train_core_incremental(&mut self, mut words: Vec<Word>, counts: Vec<i32>, vocab_size: u32) {
         assert!(vocab_size >= 256, "vocab_size must be at least 256");
         let num_merges = vocab_size - 256;
-        log::info!("Starting BPE training: {} merges to compute", num_merges);
+        log::info!("Starting BPE training: {} merges to compute", num_merges.to_formatted_string(&Locale::en));
         self.merges.clear();
 
         // ---- Initial pair_counts and where_to_update (parallel) ----
-        log::info!("Computing initial pair counts from {} unique sequences", words.len());
+        log::info!("Computing initial pair counts from {} unique sequences", words.len().to_formatted_string(&Locale::en));
         let (mut pair_counts, mut where_to_update) = count_pairs_parallel(&words, &counts);
 
         // ---- Build heap ----
-        log::info!("Building heap with {} unique pairs", pair_counts.len());
+        log::info!("Building heap with {} unique pairs", pair_counts.len().to_formatted_string(&Locale::en));
         let mut heap = OctonaryHeap::with_capacity(pair_counts.len());
         for (pair, pos) in where_to_update.drain() {
             let c = *pair_counts.get(&pair).unwrap_or(&0);
@@ -375,7 +376,7 @@ impl Tokenizer {
                 break;
             }
         }
-        log::info!("Processed {} sequences total, {} unique", total_sequences, counts.len());
+        log::info!("Processed {} sequences total, {} unique", total_sequences.to_formatted_string(&Locale::en), counts.len().to_formatted_string(&Locale::en));
 
         // Materialize words & counts
         let mut words = Vec::with_capacity(counts.len());

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 = {}

scripts/tok_eval.py

@@ -196,9 +196,9 @@ RESET = '\033[0m'
 
 # Print vocab sizes
 print(f"\nVocab sizes:")
-print(f"GPT-2: {vocab_sizes['gpt2']}")
-print(f"GPT-4: {vocab_sizes['gpt4']}")
-print(f"Ours: {vocab_sizes['ours']}")
+print(f"GPT-2: {vocab_sizes['gpt2']:,}")
+print(f"GPT-4: {vocab_sizes['gpt4']:,}")
+print(f"Ours: {vocab_sizes['ours']:,}")
 
 def print_comparison(baseline_name, baseline_results, ours_results, all_text):
     """Print comparison table between baseline tokenizer and ours."""