Parallel setup / Auto Sharding mode for general tensor operation?

[DiagRisker]

Hello,

This is an attempt to get a simple working guide on parallel operations:
I searched for a lot of examples, but none did highlight the Auto sharding mode (jax.sharding.AxisType.Auto) in a way I can work with.
I have a lot of applications in mind where stating how my mesh dispatch an Array is not advantageous (I rather have JAX decide mesh configuration arbitrarely before the computation)

cf https://docs.jax.dev/en/latest/sharded-computation.html
Here are 3 applications I'd like to tackle:

import multiprocessing, os 
PUcount = multiprocessing.cpu_count() #not trustworthy on servors
os.environ["XLA_FLAGS"] = f'--xla_force_host_platform_device_count={PUcount}' # Use N CPU core as devices
import jax, jax.numpy as jnp, jax.lax as lax ; from jax.tree_util import Partial
Pmode = 1
if Pmode:
    from jax.sharding import Mesh, PartitionSpec, NamedSharding, Sharding
    # meshing how devices behave given tensor axis to be treated in parallel
    axsn = ('cathair',) 
    mesh = Mesh( jax.devices(), 
            axis_names = axsn, 
            axis_types =(jax.sharding.AxisType.Auto,),
            ) 
    # Partition
    spec = PartitionSpec(*axsn) # jax.P is the alias of jax.sharding.PartitionSpec
    # Sharding object
    sharding = jax.sharding.NamedSharding(mesh, spec)
    tileme= Partial(jax.device_put , src = sharding, may_alias = True) # dispatch for concurrent compute
else:
    tileme= jax.jit(lambda arg: arg)

Y = jnp.ones( (32,64,7) )
print(tileme(Y).sharding) # -> SingleDeviceSharding(device=CpuDevice(id=0), memory_kind=device) This is unexpected..

#case 1 - using psum and force multi device compute
O = lax.psum( tileme(Y), tuple(range(1,2)))
print(O.sharding) # same suprise : SingleDeviceSharding..

#case 2 - having a dynamic tensor contraction function, how to map the mesh in an automatic way?
precision = jax.lax.Precision('high')
@jax.jit
def f(x: jax.Array):  #, **args):
    A = jnp.ones((x.shape[-1],3), dtype = jnp.float32)
    return lax.dot( tileme(x), tileme(A), dimension_numbers = (((x.ndim-1,),(0,)),((),())), 
            precision = precision) )
O = f(tileme(Y)) ; print(O.shape, tileme(O).sharding) # same SingleDeviceSharding outcome

#case 3 - sharding a pytree like structure
# say that I want to init a pytree whose size is known
mytree= [jnp.ones((4, 3, 2)), jnp.ones((7, 5, 1))]
ade = jax.jit( lambda x: x) #can be replaced by a wide set of functions
result = jax.tree_util.tree_map(ade, mytree)

how to parallelize case 3? given : #11394
This question also remains in a case where I want to allocate a contiguous chunk (or chunks) of memory to "fill in parallel" with case 3 solution (if we can get one). Case 3 is also useful if I want to do mix sized mini batches (images of different sizes)
Is the SingleDeviceSharding state due to jax.device_put working in eager mode ? (dispatching on the mesh only for computation)

Thanks in advance!

NB : setting : Pmode =1 vs Pmode =0
have a slight negative impact on performance for jacobian reverse applications (measuring with timeit.timeit and .block_until_ready() method). I don't see substantial speedup like in a truly parallelized case..

[DiagRisker]

Update:
Changing PUcount = 16 and tileme= Partial(jax.device_put , src = sharding, may_alias = True) to tileme= Partial(jax.device_put , device = sharding, may_alias = True), give an execution error for case 2:
ValueError: One of device_put args was given the sharding of NamedSharding(mesh=Mesh('cathair': 16, axis_types=(Auto,)), spec=PartitionSpec('cathair',), memory_kind=device), which implies that the global size of its dimension 0 should be divisible by 16, but it is equal to 7 (full shape: (7,))

Why JAX does not automatically let some devices idle if needed? (quite rigid out of ML training scope)
is there an alternative using jax.sharding.Sharding (documentation is not clear on the arguments) for sharding tensors?

I tried coding a jax.device_put macro for automating device count with tensor dimension:

import jax, jax.numpy as jnp, jax.lax as lax, numpy as np; from jax.tree_util import Partial
from typing import Sequence

uint = lambda el, b = 1: el // b + b * bool(el % b)
lmax =  lambda *l : int(.5*(sum(l) + abs( l[0]-l[1])) )

def tileme( info : Sequence[int] | int = (8,), devices = np.array(jax.devices()) ):
    """ automatic dispatch with convenient parameters
    devices = None or list of devices
    """
    axsn = tuple( str(el) for el in range(len(info)) )#('cathair',)
    # if not devices:
        # devices = jax.devices()
        # maxDev = (len(jax.devices()),)
    # else:
    devices = np.asarray(devices)
    maxDev = np.array(devices).shape
    mesh = Mesh( devices[tuple( slice( lmax( el, el//uint(el/al) ) ) for el,al in zip(info,maxDev) )], 
        axis_names = axsn, 
        axis_types =(jax.sharding.AxisType.Auto,),
        )
    spec = jax.P(*axsn)
    sharding = jax.sharding.NamedSharding(mesh, spec)
    return Partial(jax.device_put , device = sharding, may_alias = True)

precision = jax.lax.Precision('high')
@jax.jit
def f(x: jax.Array):
    A = jnp.ones((x.shape[-1],3), dtype = jnp.float32)
    return jax.lax.dot( tileme(info=x.shape[-1:])(x), tileme(info=x.shape[-1:])(A),  dimension_numbers = (((x.ndim-1,),(0,)),((),())), 
            precision = precision) 

This works on x=jnp.arange(6,dtype = jnp.float32) but not with:

z = jnp.ones((4,3,2,7))
f(z)

ValueError: One of device_put args was given the sharding of NamedSharding(mesh=Mesh('0': 7, axis_types=(Auto,)), spec=PartitionSpec('0',), memory_kind=device), which implies that the global size of its dimension 0 should be divisible by 7, but it is equal to 4 (full shape: (4, 3, 2, 7))

It seems I must explicitly state every bit of the tensor shape as axis info of the jax.device_put call, however it should not be needed (since jax.Array same as numpy.array, tensor indexing is virtually calling a very lengthy contiguous vector with regular stepping)

Same exercise for case 1:

Y = jnp.ones( (32,64,7) )
O = lax.psum( tileme(info = Y.shape[1:2])(Y), tuple(range(1,2)))
O.sharding # NamedSharding(mesh=Mesh('0': 16, axis_types=(Auto,)), spec=PartitionSpec('0',), memory_kind=device) 

Now the sharding works , case 1 solved !
but case 2 and 3 need a flexible solution

---

Just a gentle nudge for @jakevdp, @emilyfertig , @superbobry, @hawkinsp, @cgarciae on the topic.
If you have a moment to share your thoughts on just one of the 3 cases I gave, it would help :)