diff --git a/bert/convert.py b/bert/convert.py
index 63b448b4c..8806ba0d8 100644
--- a/bert/convert.py
+++ b/bert/convert.py
@@ -1,6 +1,6 @@
 import argparse
 
-import numpy
+import mlx.core as mx
 from transformers import AutoModel
 
 
@@ -23,9 +23,9 @@ def convert(bert_model: str, mlx_model: str) -> None:
     model = AutoModel.from_pretrained(bert_model)
     # save the tensors
     tensors = {
-        replace_key(key): tensor.numpy() for key, tensor in model.state_dict().items()
+        replace_key(key): mx.array(tensor) for key, tensor in model.state_dict().items()
     }
-    numpy.savez(mlx_model, **tensors)
+    mx.save_safetensors(mlx_model, tensors)
 
 
 if __name__ == "__main__":
@@ -39,7 +39,7 @@ def convert(bert_model: str, mlx_model: str) -> None:
     parser.add_argument(
         "--mlx-model",
         type=str,
-        default="weights/bert-base-uncased.npz",
+        default="bert-base-uncased.safetensors",
         help="The output path for the MLX BERT weights.",
     )
     args = parser.parse_args()
diff --git a/bert/model.py b/bert/model.py
index e7a19c99a..8fd234e8d 100644
--- a/bert/model.py
+++ b/bert/model.py
@@ -136,10 +136,7 @@ def load_model(
 
 def run(bert_model: str, mlx_model: str, batch: List[str]):
     model, tokenizer = load_model(bert_model, mlx_model)
-
-    tokens = tokenizer(batch, return_tensors="np", padding=True)
-    tokens = {key: mx.array(v) for key, v in tokens.items()}
-
+    tokens = tokenizer(batch, return_tensors="mlx", padding=True)
     return model(**tokens)
 
 
@@ -149,13 +146,13 @@ def run(bert_model: str, mlx_model: str, batch: List[str]):
         "--bert-model",
         type=str,
         default="bert-base-uncased",
-        help="The huggingface name of the BERT model to save.",
+        help="The huggingface name of the BERT model.",
     )
     parser.add_argument(
         "--mlx-model",
         type=str,
-        default="weights/bert-base-uncased.npz",
-        help="The path of the stored MLX BERT weights (npz file).",
+        default="bert-base-uncased.safetensors",
+        help="The path of the stored MLX BERT weights.",
     )
     parser.add_argument(
         "--text",
diff --git a/bert/requirements.txt b/bert/requirements.txt
index a6b564c5e..9bb67eecd 100644
--- a/bert/requirements.txt
+++ b/bert/requirements.txt
@@ -1,3 +1,2 @@
 mlx>=0.0.5
 transformers
-numpy
diff --git a/bert/test.py b/bert/test.py
index e5462ba3c..53e3c6b30 100644
--- a/bert/test.py
+++ b/bert/test.py
@@ -29,8 +29,8 @@ def run_torch(bert_model: str, batch: List[str]):
     parser.add_argument(
         "--mlx-model",
         type=str,
-        default="weights/bert-base-uncased.npz",
-        help="The path of the stored MLX BERT weights (npz file).",
+        default="bert-base-uncased.safetensors",
+        help="The path of the stored MLX BERT weights.",
     )
     parser.add_argument(
         "--text",
diff --git a/bert/weights/.gitignore b/bert/weights/.gitignore
deleted file mode 100644
index 44662642a..000000000
--- a/bert/weights/.gitignore
+++ /dev/null
@@ -1 +0,0 @@
-*.npz
\ No newline at end of file
diff --git a/llms/export/.gitignore b/llms/export/.gitignore
new file mode 100644
index 000000000..567609b12
--- /dev/null
+++ b/llms/export/.gitignore
@@ -0,0 +1 @@
+build/
diff --git a/llms/export/CMakeLists.txt b/llms/export/CMakeLists.txt
new file mode 100644
index 000000000..e2c18fc9f
--- /dev/null
+++ b/llms/export/CMakeLists.txt
@@ -0,0 +1,33 @@
+cmake_minimum_required(VERSION 3.27)
+
+project(mlxlm LANGUAGES CXX)
+
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+set(CMAKE_POSITION_INDEPENDENT_CODE ON)
+
+find_package(
+  Python 3.9
+  COMPONENTS Interpreter Development.Module
+  REQUIRED)
+execute_process(
+  COMMAND "${Python_EXECUTABLE}" -m mlx --cmake-dir
+  OUTPUT_STRIP_TRAILING_WHITESPACE
+  OUTPUT_VARIABLE MLX_ROOT)
+find_package(MLX CONFIG REQUIRED)
+
+add_library(mlxlm)
+target_link_libraries(mlxlm PUBLIC mlx)
+
+add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/third_party)
+
+target_sources(mlxlm
+  PRIVATE
+  mlxlm.cpp
+  tokenizer.cpp)
+
+add_executable(main main.cpp)
+target_link_libraries(main PRIVATE mlxlm)
+
+add_executable(test test.cpp)
+target_link_libraries(test PRIVATE mlxlm)
diff --git a/llms/export/README.md b/llms/export/README.md
new file mode 100644
index 000000000..3d6fac8c5
--- /dev/null
+++ b/llms/export/README.md
@@ -0,0 +1,34 @@
+# Export LLMs to C++
+
+Export language model inference from Python to run directly in C++.
+
+To run, first install the requirements:
+
+```bash
+pip install -U mlx-lm
+```
+
+Then generate text from Python with:
+
+```bash
+python export.py generate "How tall is K2?"
+```
+
+To export the generation function run:
+
+```bash
+python export.py export
+```
+
+Then build the C++ code (requires CMake):
+
+```bash
+cmake -B build -DCMAKE_BUILD_TYPE=Release
+cmake --build build
+```
+
+And run the generation from C++ with:
+
+```bash
+./build/main lama3.1-instruct-4bit "How tall is K2?"
+```
diff --git a/llms/export/export.py b/llms/export/export.py
new file mode 100644
index 000000000..36a2fe3a5
--- /dev/null
+++ b/llms/export/export.py
@@ -0,0 +1,171 @@
+import time
+from pathlib import Path
+
+import fire
+import mlx.core as mx
+from mlx_lm import load
+
+
+class ExportableCache:
+
+    def __init__(self, keys=None, values=None, offset=0):
+        self.offset = offset
+        self.keys = keys
+        self.values = values
+
+    def update_and_fetch(self, keys, values):
+        if self.keys is not None:
+            self.keys = mx.slice_update(self.keys, keys, self.offset, axes=(2,))
+            self.values = mx.slice_update(self.values, values, self.offset, axes=(2,))
+        else:
+            self.keys = keys
+            self.values = values
+        return self.keys, self.values
+
+    @property
+    def state(self):
+        return self.keys, self.values
+
+
+def expand(cache, mask=None, cache_step_size=256):
+    cache_size = cache[0].shape[-2]
+    new_size = cache_step_size * ((cache_size + cache_step_size) // cache_step_size)
+
+    def expand_kv(x):
+        B, n_heads, _, head_dim = x.shape
+        new_x = mx.zeros((B, n_heads, new_size, head_dim), x.dtype)
+        new_x[..., : x.shape[2], :] = x
+        return new_x
+
+    cache = [expand_kv(c) for c in cache]
+    if mask is None:
+        mask = mx.full(new_size, False)
+        mask[:cache_size] = True
+    else:
+        mask = mx.concatenate([mask, mx.full(cache_step_size, False)])
+    return cache, mask
+
+
+def causal_mask(N):
+    idx = mx.arange(N)
+    return idx[:, None] >= idx
+
+
+def step(model, y, *state):
+    mask = state[-1]
+    if len(state) > 1:
+        cache, offset = state[:-2], state[-2]
+        cache = [
+            ExportableCache(keys, values, offset)
+            for keys, values in zip(cache[::2], cache[1::2])
+        ]
+    else:
+        cache = [ExportableCache() for i in range(len(model.model.layers))]
+    logits = model(y, cache=cache, mask=mask)
+    cache = [y for x in cache for y in x.state]
+    return logits, *cache
+
+
+def generate_step(prompt, model, max_tokens):
+    mx.eval(model)
+
+    compiled_step = mx.compile(lambda *args: step(model, *args), shapeless=True)
+
+    def _step(*args):
+        logits, *cache = compiled_step(*args)
+        return mx.argmax(logits[:, -1], axis=-1), *cache
+
+    y, *cache = _step(prompt, causal_mask(prompt.size))
+    mx.async_eval(y)
+    offset = mx.array(prompt.size, mx.uint32)
+    cache, mask = expand(cache)
+    n = 0
+    while True:
+        if n < max_tokens - 1:
+            if mask.size <= (prompt.size + n):
+                cache, mask = expand(cache, mask)
+            mask[prompt.size + n] = True
+            next_y, *cache = _step(y[None], *cache, offset, mask)
+            mx.async_eval(next_y)
+            offset += 1
+        n += 1
+        yield y.item()
+        if n == max_tokens:
+            break
+        y = next_y
+
+
+def export(
+    model="mlx-community/Meta-Llama-3.1-8B-Instruct-4bit",
+    path="llama3.1-instruct-4bit",
+):
+    model, tokenizer = load(model)
+
+    mx.eval(model)
+
+    tokenizer.save_pretrained(path)
+
+    _step = lambda *args: step(model, *args)
+
+    # Make example inputs
+    y_prompt = mx.array([[0, 0]], mx.uint32)
+    y_gen = mx.array([[0]], mx.uint32)
+    offset = mx.array([0], mx.uint32)
+
+    mask = causal_mask(y_prompt.size)
+    _, *cache = _step(y_prompt, mask)
+
+    model_path = str(Path(path) / "model.mlxfn")
+    with mx.exporter(model_path, _step, shapeless=True) as exporter:
+        exporter(y_prompt, mask)
+        cache, mask = expand(cache)
+        exporter(y_gen, *cache, offset, mask)
+
+
+def generate(
+    prompt,
+    model="mlx-community/Meta-Llama-3.1-8B-Instruct-4bit",
+    max_tokens=128,
+):
+    print("[INFO] Loading model from disk.")
+    model, tokenizer = load(model)
+    prompt = tokenizer.apply_chat_template(
+        [{"role": "user", "content": prompt}],
+        add_generation_prompt=True,
+        return_tensors="mlx",
+    )
+
+    print("[INFO] Starting generation...")
+    tic = time.time()
+    tokens = []
+
+    detokenizer = tokenizer.detokenizer
+    detokenizer.reset()
+
+    for n, token in enumerate(generate_step(prompt, model, max_tokens)):
+        if n == 0:
+            prompt_tps = prompt.size / (time.time() - tic)
+            tic = time.time()
+
+        if token in tokenizer.eos_token_ids:
+            break
+        detokenizer.add_token(token)
+        print(detokenizer.last_segment, end="", flush=True)
+
+    detokenizer.finalize()
+    print(detokenizer.last_segment, flush=True)
+    gen_tps = (n + 1) / (time.time() - tic)
+    peak_memory = mx.metal.get_peak_memory() / 1e9
+    print("=" * 10)
+    print(f"Prompt: {prompt_tps:.3f} tokens-per-sec")
+    print(f"Generation: {gen_tps:.3f} tokens-per-sec")
+    print(f"Peak RAM: {peak_memory:.3f} GB")
+
+
+if __name__ == "__main__":
+    fire.Fire(
+        {
+            "generate": generate,
+            "export": export,
+        }
+    )
diff --git a/llms/export/main.cpp b/llms/export/main.cpp
new file mode 100644
index 000000000..262a3014d
--- /dev/null
+++ b/llms/export/main.cpp
@@ -0,0 +1,18 @@
+// Copyright © 2024 Apple Inc.
+
+#include <iostream>
+
+#include "mlxlm.h"
+
+int main(int argc, char *argv[]) {
+  if (argc < 3) {
+    std::cerr << "Must provide the model path and prompt." << std::endl;
+    return 1;
+  }
+  auto path = std::string(argv[1]);
+  auto prompt = std::string(argv[2]);
+
+  auto model = load_model(path + "/model.mlxfn");
+  auto tokenizer = load_tokenizer(path);
+  generate(model, tokenizer, prompt);
+}
diff --git a/llms/export/mlxlm.cpp b/llms/export/mlxlm.cpp
new file mode 100644
index 000000000..b5aa9a246
--- /dev/null
+++ b/llms/export/mlxlm.cpp
@@ -0,0 +1,120 @@
+// Copyright © 2024 Apple Inc.
+
+#include <chrono>
+#include <iomanip>
+#include <iostream>
+
+#include "mlxlm.h"
+
+namespace mx = mlx::core;
+
+#define seconds(x)                                                             \
+  (std::chrono::duration_cast<std::chrono::nanoseconds>(x).count() / 1e9)
+#define time_now() std::chrono::high_resolution_clock::now()
+
+// Maybe compile
+std::function<mx::Args(mx::Args)> load_model(const std::string &path) {
+  return mx::compile(mx::import_function(path), /* shapeless = */ true);
+}
+
+// Maybe make tokenizer virtual
+BPETokenizer load_tokenizer(const std::string &path) {
+  return BPETokenizer(path);
+}
+
+void generate(const std::function<mx::Args(mx::Args)> &model,
+              const BPETokenizer &tokenizer, const std::string &prompt,
+              int max_tokens /* = 256 */) {
+
+  auto prompt_tokens = tokenizer.encode(prompt);
+  int prompt_size = prompt_tokens.size();
+  auto y = mx::array(prompt_tokens.data(), {1, prompt_size}, mx::uint32);
+
+  auto create_causal_mask = [](int N) {
+    auto indices = mx::arange(N);
+    return mx::expand_dims(indices, 1) >= indices;
+  };
+
+  // Helper to expand the cache and mask
+  auto expand = [](auto &args, auto &mask) {
+    constexpr int cache_step_size = 256;
+    int cache_size = args[1].shape(-2);
+    int new_size =
+        cache_step_size * ((cache_size + cache_step_size) / cache_step_size);
+    for (auto it = args.begin() + 1; it != args.end(); ++it) {
+      auto &x = *it;
+      auto shape = x.shape();
+      shape[2] = new_size;
+      auto new_x = mx::zeros(shape, x.dtype());
+      shape[2] = cache_size;
+      *it =
+          mx::slice_update(new_x, x, mx::Shape(x.ndim(), 0), std::move(shape));
+    }
+    mask =
+        mx::slice_update(mx::full({new_size}, false), mask, {0}, {cache_size});
+  };
+
+  auto tic = time_now();
+  float prompt_time;
+  int n = 0;
+
+  mx::Args args;
+  {
+    args = model({y, create_causal_mask(y.size())});
+    auto logits = args[0];
+    logits = slice(logits, {0, -1, 0}, logits.shape());
+    y = argmax(logits, -1);
+    async_eval(y);
+  }
+
+  auto offset = mx::array(prompt_size, mx::uint32);
+  std::vector<int> tokens;
+
+  auto mask = mx::full({prompt_size}, true);
+  expand(args, mask);
+
+  for (; n < max_tokens; ++n) {
+    // Start next token decoding if needed
+    if (n < max_tokens - 1) {
+      args[0] = y;
+      auto m = prompt_size + n;
+      if (mask.size() <= m) {
+        expand(args, mask);
+      }
+      mask = mx::slice_update(mask, mx::array(true), {m}, {m + 1});
+      args.push_back(offset);
+      args.push_back(mask);
+      args = model(args);
+      args[0] = argmax(args[0], -1);
+      offset = offset + 1u;
+      async_eval(args[0]);
+    }
+
+    auto token = y.item<int>();
+    if (token == tokenizer.eos_token_id()) {
+      break;
+    }
+    tokens.push_back(token);
+    auto [result, complete] = tokenizer.try_decode(tokens);
+    if (complete) {
+      std::cout << result << std::flush;
+      tokens.clear();
+    }
+    if (n == 0) {
+      prompt_time = seconds(time_now() - tic);
+      tic = time_now();
+    }
+
+    if (n < max_tokens - 1) {
+      y = args[0];
+    }
+  }
+  auto result = tokenizer.decode(tokens);
+  std::cout << result << std::flush;
+
+  auto gen_time = seconds(time_now() - tic);
+  std::cout << std::endl;
+  std::cout << std::setprecision(5) << "Prompt toks/sec "
+            << prompt_size / prompt_time << "\nGeneration toks/sec "
+            << (n + 1) / gen_time << std::endl;
+}
diff --git a/llms/export/mlxlm.h b/llms/export/mlxlm.h
new file mode 100644
index 000000000..9dba3b0a1
--- /dev/null
+++ b/llms/export/mlxlm.h
@@ -0,0 +1,15 @@
+// Copyright © 2024 Apple Inc.
+
+#include <mlx/mlx.h>
+
+#include "tokenizer.h"
+
+namespace mx = mlx::core;
+
+std::function<mx::Args(mx::Args)> load_model(const std::string &path);
+
+BPETokenizer load_tokenizer(const std::string &path);
+
+void generate(const std::function<mx::Args(mx::Args)> &model,
+              const BPETokenizer &tokenizer, const std::string &prompt,
+              int max_tokens = 256);
diff --git a/llms/export/test.cpp b/llms/export/test.cpp
new file mode 100644
index 000000000..0c44ab85b
--- /dev/null
+++ b/llms/export/test.cpp
@@ -0,0 +1,23 @@
+// Copyright © 2024 Apple Inc.
+
+#include "tokenizer.h"
+#include <iostream>
+
+template <typename T, typename U = T> void check(const T &x, const U &y) {
+  if (x != y) {
+    std::cerr << "Mismatch" << std::endl;
+  }
+}
+
+void test_tokenizer(const std::string &path) {
+  BPETokenizer tokenizer(path);
+  check(tokenizer.encode("hello world!"), {128000, 15339, 1917, 0});
+  check(tokenizer.decode({15339}), "hello");
+  check(tokenizer.decode({0}), "!");
+  check(tokenizer.decode({1917}), " world");
+  check(tokenizer.encode("we'd  see   you say 世界你好真实好的很啊"),
+        {128000, 906, 4265, 220, 1518, 256, 499, 2019, 127365, 57668, 53901,
+         89151, 41073, 110085, 101600, 102856});
+}
+
+int main(int argc, char *argv[]) { test_tokenizer("."); }
diff --git a/llms/export/third_party/CMakeLists.txt b/llms/export/third_party/CMakeLists.txt
new file mode 100644
index 000000000..0496ea57e
--- /dev/null
+++ b/llms/export/third_party/CMakeLists.txt
@@ -0,0 +1,20 @@
+include(FetchContent)
+
+FetchContent_Declare(
+  json
+  URL https://github.com/nlohmann/json/releases/download/v3.11.3/json.tar.xz)
+FetchContent_MakeAvailable(json)
+target_include_directories(
+  mlxlm PRIVATE $<BUILD_INTERFACE:${json_SOURCE_DIR}/single_include/nlohmann>)
+
+execute_process(
+  COMMAND zsh "${CMAKE_CURRENT_SOURCE_DIR}/download_unicode.sh" "${CMAKE_CURRENT_BINARY_DIR}"
+  COMMAND_ERROR_IS_FATAL ANY
+)
+
+target_sources(mlxlm
+  PRIVATE
+  ${CMAKE_CURRENT_BINARY_DIR}/unicode.cpp 
+  ${CMAKE_CURRENT_BINARY_DIR}/unicode-data.cpp)
+
+target_include_directories(mlxlm PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/..)
diff --git a/llms/export/third_party/download_unicode.sh b/llms/export/third_party/download_unicode.sh
new file mode 100644
index 000000000..6e52fe387
--- /dev/null
+++ b/llms/export/third_party/download_unicode.sh
@@ -0,0 +1,9 @@
+#!/bin/bash
+
+commit=1204f9727005974587d6fc1dcd4d4f0ead87c856
+url=https://raw.githubusercontent.com/ggerganov/llama.cpp/${commit}/src/
+
+for file in 'unicode.cpp' 'unicode.h' 'unicode-data.cpp' 'unicode-data.h'
+do
+  curl -OL ${url}/${file} --output-dir $1 2>/dev/null
+done
diff --git a/llms/export/tokenizer.cpp b/llms/export/tokenizer.cpp
new file mode 100644
index 000000000..5e68a43bd
--- /dev/null
+++ b/llms/export/tokenizer.cpp
@@ -0,0 +1,201 @@
+
+#include <codecvt>
+#include <filesystem>
+#include <fstream>
+#include <json.hpp>
+#include <locale>
+
+#include "third_party/unicode.h"
+#include "tokenizer.h"
+
+using json = nlohmann::json;
+
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wdeprecated-declarations"
+std::pair<std::wstring, int> utf8_to_utf16(const std::string &s) {
+  static std::string replace_str = std::string(1, 0xFF);
+  static std::wstring replace_wstr = std::wstring(1, 0xFFFD);
+  std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> cvt(replace_str,
+                                                             replace_wstr);
+  auto out = cvt.from_bytes(s);
+  return {out, cvt.converted()};
+}
+#pragma GCC diagnostic pop
+
+auto make_byte_decoder() {
+  std::unordered_map<uint16_t, char> byte_decoder;
+  std::vector<uint16_t> limits = {0,        '!',  '~' + 1, L'¡',
+                                  L'¬' + 1, L'®', L'ÿ' + 1};
+  char n = 0;
+  for (int i = 0; i < limits.size() - 1; ++i) {
+    auto start = limits[i];
+    auto stop = limits[i + 1];
+    if (i % 2 == 0) {
+      for (int b = start; b < stop; ++b) {
+        byte_decoder[256 + n++] = b;
+      }
+    } else {
+      for (int b = start; b < stop; ++b) {
+        byte_decoder[b] = b;
+      }
+    }
+  }
+  return byte_decoder;
+}
+
+auto BPETokenizer::byte_decoder_ = make_byte_decoder();
+
+BPETokenizer::BPETokenizer(const std::string &path_) {
+  auto path = std::filesystem::path(path_);
+  std::ifstream ifs(path / "tokenizer.json");
+  auto tokenizer = json::parse(ifs);
+  auto model = tokenizer["model"];
+  token_to_id_ = model["vocab"];
+  id_to_token_.resize(token_to_id_.size());
+  for (auto &[s, id] : token_to_id_) {
+    if (id >= id_to_token_.size()) {
+      id_to_token_.resize(id + 1);
+    }
+    id_to_token_[id] = s;
+  }
+  std::string type = model["type"];
+  auto merges = model["merges"];
+  for (auto &s : merges) {
+    if (s.is_string()) {
+      merges_.emplace(s, merges_.size());
+    } else {
+      std::string s1 = s[0];
+      std::string s2 = s[1];
+      merges_.emplace(s1 + " " + s2, merges_.size());
+    }
+  }
+
+  auto added_tokens = tokenizer["added_tokens"];
+  for (auto &added_token : added_tokens) {
+    int id = added_token["id"];
+    if (id >= id_to_token_.size()) {
+      id_to_token_.resize(id + 1);
+    }
+    id_to_token_[id] = added_token["content"];
+    if (id_to_token_[id] == "<|begin_of_text|>") {
+      bos_id_ = id;
+    } else if (id_to_token_[id] == "<|eot_id|>") {
+      eos_id_ = id;
+    }
+  }
+
+  // Currently hardcoded to Llama3 BPE regex
+  pre_tokenizer_regex_ = {
+      "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}"
+      "\\p{N}]?\\p{L}+|\\p{N}{1,3}| "
+      "?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+"};
+}
+
+std::vector<int> BPETokenizer::encode(std::string text) const {
+
+  auto segments = unicode_regex_split(text, pre_tokenizer_regex_);
+
+  auto one_step_merge = [this](std::string segment, std::vector<int> &splits) {
+    int merge_idx;
+    int rank = INT32_MAX;
+    std::string candidate;
+    for (int i = 0; i < splits.size() - 2; ++i) {
+      auto start = splits[i];
+      auto mid = splits[i + 1];
+      auto end = splits[i + 2];
+      candidate.clear();
+      candidate.insert(candidate.end(), segment.begin() + start,
+                       segment.begin() + mid);
+      candidate += " ";
+      candidate.insert(candidate.end(), segment.begin() + mid,
+                       segment.begin() + end);
+      if (auto it = merges_.find(candidate); it != merges_.end()) {
+        if (it->second < rank) {
+          merge_idx = i;
+          rank = it->second;
+        }
+      }
+    }
+    if (rank == INT32_MAX) {
+      return false;
+    }
+    auto start = splits[merge_idx];
+    auto mid = splits[merge_idx + 1];
+    auto end = splits[merge_idx + 2];
+    std::string merge_l = segment.substr(start, mid - start);
+    std::string merge_r = segment.substr(mid, end - mid);
+    for (int i = splits.size() - 2; i >= 0; --i) {
+      auto start = splits[i];
+      auto mid = splits[i + 1];
+      auto end = splits[i + 2];
+      if (segment.substr(start, mid - start) == merge_l &&
+          segment.substr(mid, end - mid) == merge_r) {
+        splits.erase(splits.begin() + i + 1);
+        i -= 1;
+      }
+    }
+    return true;
+  };
+
+  std::vector<int> ids;
+  ids.push_back(bos_id_);
+
+  // Initialize merges to integer list
+  auto merge_segment = [&ids, &one_step_merge,
+                        this](const std::string &segment) {
+    std::vector<int> splits;
+    for (int i = 0; i < segment.size(); ++i) {
+      splits.push_back(i);
+      if (static_cast<unsigned char>(segment[i]) >= 128) {
+        i++;
+      }
+    }
+    splits.push_back(segment.size());
+
+    while (one_step_merge(segment, splits)) {
+    };
+    for (int i = 0; i < splits.size() - 1; ++i) {
+      auto start = splits[i];
+      auto end = splits[i + 1];
+      std::string s = segment.substr(start, end - start);
+      if (auto it = token_to_id_.find(s); it != token_to_id_.end()) {
+        ids.push_back(it->second);
+      } else {
+        throw std::runtime_error("UNK ENCOUNTERED");
+      }
+    }
+  };
+
+  for (auto &segment : segments) {
+    merge_segment(segment);
+  }
+  return ids;
+}
+
+std::string BPETokenizer::id_to_bytes(int id) const {
+  std::string token;
+  auto [wide_token, _] = utf8_to_utf16(id_to_token_[id]);
+  token.resize(wide_token.size());
+  for (int i = 0; i < wide_token.size(); ++i) {
+    token[i] = byte_decoder_[wide_token[i]];
+  }
+  return token;
+}
+
+std::pair<std::string, bool>
+BPETokenizer::try_decode(const std::vector<int> &ids) const {
+  std::string text;
+  for (auto id : ids) {
+    text += id_to_bytes(id);
+  }
+  auto [_, converted] = utf8_to_utf16(text);
+  bool complete = converted == text.size();
+  text.resize(converted);
+  return {text, complete};
+}
+
+std::string BPETokenizer::decode(const std::vector<int> &ids) const {
+  return try_decode(ids).first;
+}
+
+int BPETokenizer::eos_token_id() const { return eos_id_; }
diff --git a/llms/export/tokenizer.h b/llms/export/tokenizer.h
new file mode 100644
index 000000000..64dd2fe42
--- /dev/null
+++ b/llms/export/tokenizer.h
@@ -0,0 +1,37 @@
+// Copyright © 2024 Apple Inc.
+
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#pragma once
+
+/** BPE Tokenizer API */
+class BPETokenizer {
+public:
+  BPETokenizer(const std::string &path);
+
+  /** Encode a string of text to token integer ids. */
+  std::vector<int> encode(std::string text) const;
+
+  /** Try to decode the vector of ids to text. The text is truncated to
+   * include only the fully decodable tokens. */
+  std::string decode(const std::vector<int> &ids) const;
+
+  /** Try to decode the vector of ids to text. The second return value
+   * indicates if the decoding completed. The text is truncated to include
+   * only the fully decodable tokens. */
+  std::pair<std::string, bool> try_decode(const std::vector<int> &ids) const;
+
+  int eos_token_id() const;
+
+private:
+  std::unordered_map<std::string, int> token_to_id_;
+  std::vector<std::string> id_to_token_;
+  std::unordered_map<std::string, int> merges_;
+  int bos_id_;
+  int eos_id_;
+  static std::unordered_map<uint16_t, char> byte_decoder_;
+  std::string id_to_bytes(int id) const;
+  std::vector<std::string> pre_tokenizer_regex_;
+};
diff --git a/llms/mlx_lm/models/llama.py b/llms/mlx_lm/models/llama.py
index 7b452ea47..4a789dcf3 100644
--- a/llms/mlx_lm/models/llama.py
+++ b/llms/mlx_lm/models/llama.py
@@ -69,14 +69,12 @@ def __call__(
         mask: Optional[mx.array] = None,
         cache: Optional[Any] = None,
     ) -> mx.array:
-        B, L, D = x.shape
-
         queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
 
         # Prepare the queries, keys and values for the attention computation
-        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
-        keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
-        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        queries = mx.unflatten(queries, -1, (self.n_heads, -1)).transpose(0, 2, 1, 3)
+        keys = mx.unflatten(keys, -1, (self.n_kv_heads, -1)).transpose(0, 2, 1, 3)
+        values = mx.unflatten(values, -1, (self.n_kv_heads, -1)).transpose(0, 2, 1, 3)
 
         if cache is not None:
             queries = self.rope(queries, offset=cache.offset)
@@ -90,7 +88,7 @@ def __call__(
             queries, keys, values, cache=cache, scale=self.scale, mask=mask
         )
 
-        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        output = output.transpose(0, 2, 1, 3).flatten(-2, -1)
         return self.o_proj(output)