Remove Hazard tracking with Fences (#1509)

* remove hazard tracking

* with fence map

* no hazard tracking with fences

* nits

* fix fence retain

* cleanup

* fix quantized rebase

docs/src/usage/compile.rst

@@ -33,12 +33,12 @@ Let's start with a simple example:
   # Compile the function
   compiled_fun = mx.compile(fun)
 
-  # Prints: array(2.36788, dtype=float32) 
+  # Prints: array(2.36788, dtype=float32)
   print(compiled_fun(x, y))
 
 The output of both the regular function and the compiled function is the same
 up to numerical precision.
-   
+
 The first time you call a compiled function, MLX will build the compute
 graph, optimize it, and generate and compile code. This can be relatively
 slow. However, MLX will cache compiled functions, so calling a compiled
@@ -96,7 +96,7 @@ element-wise operations:
 
 .. code-block:: python
 
-  def gelu(x):  
+  def gelu(x):
       return x * (1 + mx.erf(x / math.sqrt(2))) / 2
 
 If you use this function with small arrays, it will be overhead bound. If you
@@ -136,13 +136,6 @@ Now make an array, and benchmark both functions:
 On an M1 Max the times are 15.5 and 3.1 milliseconds. The compiled ``gelu`` is
 five times faster.
 
-.. note::
-
-  As of the latest MLX, CPU functions are not fully compiled. Compiling CPU
-  functions can still be helpful, but won't typically result in as large a
-  speedup as compiling operations that run on the GPU.
-
-
 Debugging
 ---------
 
@@ -287,7 +280,7 @@ to the function. In some cases this can be pretty inconvenient. Hence,
   print(fun(mx.array(1.0)))
 
 
-Compiling Training Graphs 
+Compiling Training Graphs
 -------------------------
 
 This section will step through how to use :func:`compile` with a simple example
@@ -297,7 +290,7 @@ full forward, backward, and update with :func:`compile`.
 
 To start, here is the simple example without any compilation:
 
-.. code-block:: python 
+.. code-block:: python
 
   import mlx.core as mx
   import mlx.nn as nn
@@ -330,7 +323,7 @@ To start, here is the simple example without any compilation:
 To compile the update we can put it all in a function and compile it with the
 appropriate input and output captures. Here's the same example but compiled:
 
-.. code-block:: python 
+.. code-block:: python
 
   import mlx.core as mx
   import mlx.nn as nn
@@ -355,7 +348,7 @@ appropriate input and output captures. Here's the same example but compiled:
 
   # The state that will be captured as input and output
   state = [model.state, optimizer.state]
-      
+
   @partial(mx.compile, inputs=state, outputs=state)
   def step(x, y):
       loss_and_grad_fn = nn.value_and_grad(model, loss_fn)
@@ -410,7 +403,7 @@ Compiling transformed functions works just as expected:
 
    In order to compile as much as possible, a transformation of a compiled
    function will not by default be compiled. To compile the transformed
-   function simply pass it through :func:`compile`. 
+   function simply pass it through :func:`compile`.
 
 You can also compile functions which themselves call compiled functions. A
 good practice is to compile the outer most function to give :func:`compile`

docs/src/usage/function_transforms.rst

@@ -25,7 +25,7 @@ Here is a simple example:
 
 The output of :func:`grad` on :func:`sin` is simply another function. In this
 case it is the gradient of the sine function which is exactly the cosine
-function. To get the second derivative you can do: 
+function. To get the second derivative you can do:
 
 .. code-block:: shell
 
@@ -50,7 +50,7 @@ Automatic Differentiation
 .. _auto diff:
 
 Automatic differentiation in MLX works on functions rather than on implicit
-graphs. 
+graphs.
 
 .. note::
 
@@ -114,7 +114,7 @@ way to do that is the following:
 
    def loss_fn(params, x, y):
       w, b = params["weight"], params["bias"]
-      h = w * x + b 
+      h = w * x + b
       return mx.mean(mx.square(h - y))
 
    params = {"weight": mx.array(1.0), "bias": mx.array(0.0)}
@@ -132,7 +132,7 @@ way to do that is the following:
 
 Notice the tree structure of the parameters is preserved in the gradients.
 
-In some cases you may want to stop gradients from propagating through a 
+In some cases you may want to stop gradients from propagating through a
 part of the function. You can use the :func:`stop_gradient` for that.
 
 
@@ -166,14 +166,14 @@ A naive way to add the elements from two sets of vectors is with a loop:
 Instead you can use :func:`vmap` to automatically vectorize the addition:
 
 .. code-block:: python
-   
+
    # Vectorize over the second dimension of x and the
    # first dimension of y
    vmap_add = mx.vmap(lambda x, y: x + y, in_axes=(1, 0))
 
 The ``in_axes`` parameter can be used to specify which dimensions of the
 corresponding input to vectorize over. Similarly, use ``out_axes`` to specify
-where the vectorized axes should be in the outputs. 
+where the vectorized axes should be in the outputs.
 
 Let's time these two different versions:
 

docs/src/usage/indexing.rst

@@ -51,7 +51,7 @@ You can also use an :obj:`array` to index another :obj:`array`:
 .. code-block:: shell
 
   >>> arr = mx.arange(10)
-  >>> idx = mx.array([5, 7]) 
+  >>> idx = mx.array([5, 7])
   >>> arr[idx]
   array([5, 7], dtype=int32)
 
@@ -82,7 +82,7 @@ general, MLX has limited support for operations for which outputs
 operations which MLX does not yet support include :func:`numpy.nonzero` and the
 single input version of :func:`numpy.where`.
 
-In Place Updates 
+In Place Updates
 ----------------
 
 In place updates to indexed arrays are possible in MLX. For example:

docs/src/usage/lazy_evaluation.rst

@@ -13,7 +13,7 @@ compute graph is recorded. The actual computation only happens if an
 :func:`eval` is performed.
 
 MLX uses lazy evaluation because it has some nice features, some of which we
-describe below. 
+describe below.
 
 Transforming Compute Graphs
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
@@ -116,7 +116,7 @@ saving functions) will also evaluate the array.
 
 Calling :func:`array.item` on a scalar array will also evaluate it. In the
 example above, printing the loss (``print(loss)``) or adding the loss scalar to
-a list (``losses.append(loss.item())``) would cause a graph evaluation. If 
+a list (``losses.append(loss.item())``) would cause a graph evaluation. If
 these lines are before ``mx.eval(loss, model.parameters())`` then this
 will be a partial evaluation, computing only the forward pass.
 

docs/src/usage/numpy.rst

@@ -3,10 +3,10 @@
 Conversion to NumPy and Other Frameworks
 ========================================
 
-MLX array supports conversion between other frameworks with either:  
+MLX array supports conversion between other frameworks with either:
 
-* The `Python Buffer Protocol <https://docs.python.org/3/c-api/buffer.html>`_. 
-* `DLPack <https://dmlc.github.io/dlpack/latest/>`_.  
+* The `Python Buffer Protocol <https://docs.python.org/3/c-api/buffer.html>`_.
+* `DLPack <https://dmlc.github.io/dlpack/latest/>`_.
 
 Let's convert an array to NumPy and back.
 
@@ -66,7 +66,7 @@ even though no in-place operations on MLX memory are executed.
 PyTorch
 -------
 
-.. warning:: 
+.. warning::
 
    PyTorch Support for :obj:`memoryview` is experimental and can break for
    multi-dimensional arrays. Casting to NumPy first is advised for now.

docs/src/usage/quick_start.rst

@@ -64,4 +64,4 @@ Other gradient transformations include :func:`vjp` for vector-Jacobian products
 and :func:`jvp` for Jacobian-vector products.
 
 Use :func:`value_and_grad` to efficiently compute both a function's output and
-gradient with respect to the function's input. 
+gradient with respect to the function's input.

docs/src/usage/saving_and_loading.rst

@@ -8,33 +8,33 @@ Saving and Loading Arrays
 MLX supports multiple array serialization formats.
 
 .. list-table:: Serialization Formats
-   :widths: 20 8 25 25 
+   :widths: 20 8 25 25
    :header-rows: 1
 
-   * - Format 
-     - Extension 
+   * - Format
+     - Extension
      - Function
-     - Notes 
-   * - NumPy 
-     - ``.npy`` 
+     - Notes
+   * - NumPy
+     - ``.npy``
      - :func:`save`
      - Single arrays only
-   * - NumPy archive 
-     - ``.npz`` 
+   * - NumPy archive
+     - ``.npz``
      - :func:`savez` and :func:`savez_compressed`
-     - Multiple arrays 
+     - Multiple arrays
    * - Safetensors
-     - ``.safetensors`` 
+     - ``.safetensors``
      - :func:`save_safetensors`
-     - Multiple arrays 
-   * - GGUF 
-     - ``.gguf`` 
+     - Multiple arrays
+   * - GGUF
+     - ``.gguf``
      - :func:`save_gguf`
      - Multiple arrays
 
 The :func:`load` function will load any of the supported serialization
 formats. It determines the format from the extensions. The output of
-:func:`load` depends on the format. 
+:func:`load` depends on the format.
 
 Here's an example of saving a single array to a file:
 

docs/src/usage/unified_memory.rst

@@ -20,7 +20,7 @@ Both ``a`` and ``b`` live in unified memory.
 
 In MLX, rather than moving arrays to devices, you specify the device when you
 run the operation. Any device can perform any operation on ``a`` and ``b``
-without needing to move them from one memory location to another. For example: 
+without needing to move them from one memory location to another. For example:
 
 .. code-block:: python