diff --git a/gpt_oss/generate.py b/gpt_oss/generate.py
index dfaaa6f1..87ee42d1 100644
--- a/gpt_oss/generate.py
+++ b/gpt_oss/generate.py
@@ -4,11 +4,18 @@
 # torchrun --nproc-per-node=4 -m gpt_oss.generate -p "why did the chicken cross the road?" model/
 
 import argparse
+import os
+from pathlib import Path
 
 from gpt_oss.tokenizer import get_tokenizer
 
 
-def main(args):
+def main(args: argparse.Namespace) -> None:
+    # Validate checkpoint path exists
+    checkpoint_path = Path(args.checkpoint)
+    if not checkpoint_path.exists():
+        raise FileNotFoundError(f"Checkpoint path does not exist: {args.checkpoint}")
+    
     match args.backend:
         case "torch":
             from gpt_oss.torch.utils import init_distributed
diff --git a/gpt_oss/tokenizer.py b/gpt_oss/tokenizer.py
index 866077f5..05c1080d 100644
--- a/gpt_oss/tokenizer.py
+++ b/gpt_oss/tokenizer.py
@@ -1,6 +1,7 @@
 import tiktoken
 
-def get_tokenizer():
+
+def get_tokenizer() -> tiktoken.Encoding:
     o200k_base = tiktoken.get_encoding("o200k_base")
     tokenizer = tiktoken.Encoding(
         name="o200k_harmony",
diff --git a/gpt_oss/torch/utils.py b/gpt_oss/torch/utils.py
index ce87a85d..47edf5c9 100644
--- a/gpt_oss/torch/utils.py
+++ b/gpt_oss/torch/utils.py
@@ -3,7 +3,7 @@
 import torch.distributed as dist
 
 
-def suppress_output(rank):
+def suppress_output(rank: int) -> None:
     """Suppress printing on the current device. Force printing with `force=True`."""
     import builtins as __builtin__
     builtin_print = __builtin__.print
@@ -20,21 +20,60 @@ def print(*args, **kwargs):
 
 def init_distributed() -> torch.device:
     """Initialize the model for distributed inference."""
+    # Check CUDA availability
+    if not torch.cuda.is_available():
+        raise RuntimeError(
+            "CUDA is not available. Please ensure CUDA is installed and accessible."
+        )
+    
     # Initialize distributed inference
     world_size = int(os.environ.get("WORLD_SIZE", 1))
     rank = int(os.environ.get("RANK", 0))
-    if world_size > 1:
-        dist.init_process_group(
-            backend="nccl", init_method="env://", world_size=world_size, rank=rank
+    
+    # Validate rank against available devices
+    if rank >= torch.cuda.device_count():
+        raise RuntimeError(
+            f"Rank {rank} exceeds available CUDA devices ({torch.cuda.device_count()}). "
+            f"Please set RANK to a value between 0 and {torch.cuda.device_count() - 1}."
         )
-    torch.cuda.set_device(rank)
-    device = torch.device(f"cuda:{rank}")
+    
+    try:
+        if world_size > 1:
+            dist.init_process_group(
+                backend="nccl", init_method="env://", world_size=world_size, rank=rank
+            )
+        
+        torch.cuda.set_device(rank)
+        device = torch.device(f"cuda:{rank}")
+        
+        # Test device accessibility
+        try:
+            torch.cuda.get_device_properties(device)
+        except RuntimeError as e:
+            raise RuntimeError(
+                f"Failed to access CUDA device {rank}: {e}. "
+                "Please check device availability and permissions."
+            ) from e
 
-    # Warm up NCCL to avoid first-time latency
-    if world_size > 1:
-        x = torch.ones(1, device=device)
-        dist.all_reduce(x)
-        torch.cuda.synchronize(device)
+        # Warm up NCCL to avoid first-time latency
+        if world_size > 1:
+            try:
+                x = torch.ones(1, device=device)
+                dist.all_reduce(x)
+                torch.cuda.synchronize(device)
+            except RuntimeError as e:
+                raise RuntimeError(
+                    f"Failed to initialize distributed communication on device {rank}: {e}"
+                ) from e
 
-    suppress_output(rank)
-    return device
+        suppress_output(rank)
+        return device
+        
+    except Exception as e:
+        # Clean up distributed process group if initialization failed
+        if world_size > 1 and dist.is_initialized():
+            try:
+                dist.destroy_process_group()
+            except Exception:
+                pass  # Ignore cleanup errors
+        raise
diff --git a/tests/test_tokenizer.py b/tests/test_tokenizer.py
new file mode 100644
index 00000000..645e13b0
--- /dev/null
+++ b/tests/test_tokenizer.py
@@ -0,0 +1,212 @@
+"""Unit tests for tokenizer encoding/decoding functionality."""
+
+import pytest
+from gpt_oss.tokenizer import get_tokenizer
+
+
+class TestTokenizerBasics:
+    """Test basic tokenizer functionality."""
+
+    def test_get_tokenizer_returns_encoding(self):
+        """Test that get_tokenizer returns a valid encoding."""
+        tokenizer = get_tokenizer()
+        assert tokenizer is not None
+        assert tokenizer.name == "o200k_harmony"
+
+    def test_tokenizer_has_harmony_special_tokens(self):
+        """Test that tokenizer includes Harmony special tokens."""
+        tokenizer = get_tokenizer()
+        special_tokens = tokenizer._special_tokens
+        
+        # Verify key Harmony tokens are present
+        assert "<|channel|>" in special_tokens
+        assert special_tokens["<|channel|>"] == 200005
+        assert "<|start|>" in special_tokens
+        assert special_tokens["<|start|>"] == 200006
+        assert "<|end|>" in special_tokens
+        assert special_tokens["<|end|>"] == 200007
+        assert "<|message|>" in special_tokens
+        assert special_tokens["<|message|>"] == 200008
+        assert "<|call|>" in special_tokens
+        assert special_tokens["<|call|>"] == 200012
+        assert "<|return|>" in special_tokens
+        assert special_tokens["<|return|>"] == 200002
+
+    def test_tokenizer_has_reserved_tokens(self):
+        """Test that tokenizer includes reserved token range."""
+        tokenizer = get_tokenizer()
+        special_tokens = tokenizer._special_tokens
+        
+        # Check reserved tokens exist in range
+        assert "<|reserved_200013|>" in special_tokens
+        assert special_tokens["<|reserved_200013|>"] == 200013
+        assert "<|reserved_201087|>" in special_tokens
+        assert special_tokens["<|reserved_201087|>"] == 201087
+
+
+class TestTokenizerEncoding:
+    """Test tokenizer encoding functionality."""
+
+    def test_encode_simple_text(self):
+        """Test encoding simple text."""
+        tokenizer = get_tokenizer()
+        text = "Hello, world!"
+        tokens = tokenizer.encode(text)
+        
+        assert isinstance(tokens, list)
+        assert len(tokens) > 0
+        assert all(isinstance(t, int) for t in tokens)
+
+    def test_encode_special_tokens(self):
+        """Test encoding text with special tokens."""
+        tokenizer = get_tokenizer()
+        text = "<|channel|>final<|message|>Hello<|return|>"
+        tokens = tokenizer.encode(text, allowed_special="all")
+        
+        assert 200005 in tokens  # <|channel|>
+        assert 200008 in tokens  # <|message|>
+        assert 200002 in tokens  # <|return|>
+
+    def test_encode_without_special_allowed_raises(self):
+        """Test that encoding special tokens without permission raises error."""
+        tokenizer = get_tokenizer()
+        text = "<|channel|>test"
+        
+        with pytest.raises(ValueError):
+            tokenizer.encode(text)
+
+    def test_encode_empty_string(self):
+        """Test encoding empty string."""
+        tokenizer = get_tokenizer()
+        tokens = tokenizer.encode("")
+        
+        assert isinstance(tokens, list)
+        assert len(tokens) == 0
+
+    def test_encode_unicode_text(self):
+        """Test encoding unicode text."""
+        tokenizer = get_tokenizer()
+        text = "Hello 世界 🌍"
+        tokens = tokenizer.encode(text)
+        
+        assert isinstance(tokens, list)
+        assert len(tokens) > 0
+
+
+class TestTokenizerDecoding:
+    """Test tokenizer decoding functionality."""
+
+    def test_decode_simple_tokens(self):
+        """Test decoding simple tokens."""
+        tokenizer = get_tokenizer()
+        text = "Hello, world!"
+        tokens = tokenizer.encode(text)
+        decoded = tokenizer.decode(tokens)
+        
+        assert decoded == text
+
+    def test_decode_with_special_tokens(self):
+        """Test decoding tokens including special tokens."""
+        tokenizer = get_tokenizer()
+        text = "<|channel|>final<|message|>Hello<|return|>"
+        tokens = tokenizer.encode(text, allowed_special="all")
+        decoded = tokenizer.decode(tokens)
+        
+        assert decoded == text
+
+    def test_decode_empty_list(self):
+        """Test decoding empty token list."""
+        tokenizer = get_tokenizer()
+        decoded = tokenizer.decode([])
+        
+        assert decoded == ""
+
+    def test_decode_single_token(self):
+        """Test decoding single token."""
+        tokenizer = get_tokenizer()
+        tokens = tokenizer.encode("a")
+        decoded = tokenizer.decode(tokens)
+        
+        assert decoded == "a"
+
+    def test_decode_unicode(self):
+        """Test decoding unicode tokens."""
+        tokenizer = get_tokenizer()
+        text = "Hello 世界 🌍"
+        tokens = tokenizer.encode(text)
+        decoded = tokenizer.decode(tokens)
+        
+        assert decoded == text
+
+
+class TestTokenizerRoundTrip:
+    """Test encode/decode round-trip consistency."""
+
+    @pytest.mark.parametrize("text", [
+        "Simple text",
+        "Text with numbers: 123456",
+        "Special chars: !@#$%^&*()",
+        "Unicode: 你好世界",
+        "Emoji: 🚀🌟💡",
+        "Mixed: Hello 世界 123 🎉",
+        "Newlines:\nand\ttabs",
+    ])
+    def test_roundtrip_consistency(self, text):
+        """Test that encode->decode returns original text."""
+        tokenizer = get_tokenizer()
+        tokens = tokenizer.encode(text)
+        decoded = tokenizer.decode(tokens)
+        
+        assert decoded == text
+
+    def test_roundtrip_with_harmony_format(self):
+        """Test round-trip with Harmony message format."""
+        tokenizer = get_tokenizer()
+        text = "<|channel|>analysis<|start|><|message|>Thinking...<|end|><|channel|>final<|message|>Answer<|return|>"
+        tokens = tokenizer.encode(text, allowed_special="all")
+        decoded = tokenizer.decode(tokens)
+        
+        assert decoded == text
+
+
+class TestTokenizerEdgeCases:
+    """Test edge cases and error handling."""
+
+    def test_encode_very_long_text(self):
+        """Test encoding very long text."""
+        tokenizer = get_tokenizer()
+        text = "a" * 10000
+        tokens = tokenizer.encode(text)
+        
+        assert isinstance(tokens, list)
+        assert len(tokens) > 0
+
+    def test_decode_invalid_token_ids(self):
+        """Test decoding with potentially invalid token IDs."""
+        tokenizer = get_tokenizer()
+        # Use valid token IDs from the special tokens
+        tokens = [200005, 200006, 200007]
+        decoded = tokenizer.decode(tokens)
+        
+        assert isinstance(decoded, str)
+
+    def test_multiple_tokenizer_instances_consistent(self):
+        """Test that multiple tokenizer instances behave consistently."""
+        tokenizer1 = get_tokenizer()
+        tokenizer2 = get_tokenizer()
+        
+        text = "Test consistency"
+        tokens1 = tokenizer1.encode(text)
+        tokens2 = tokenizer2.encode(text)
+        
+        assert tokens1 == tokens2
+
+    def test_special_token_ids_immutable(self):
+        """Test that special token IDs are consistent."""
+        tokenizer = get_tokenizer()
+        
+        # Get special tokens multiple times
+        channel_id_1 = tokenizer.encode("<|channel|>", allowed_special="all")[0]
+        channel_id_2 = tokenizer.encode("<|channel|>", allowed_special="all")[0]
+        
+        assert channel_id_1 == channel_id_2 == 200005