From 91d3292923c8b7cd4cb93b678a3fb4a63fb3ec32 Mon Sep 17 00:00:00 2001 From: Wassim Kabalan Date: Fri, 28 Feb 2025 13:46:41 +0100 Subject: [PATCH 1/6] add correct annotations for weights in painting and warning for cic_paint in distributed pm --- jaxpm/painting.py | 33 ++++++++++++++++++++------------- 1 file changed, 20 insertions(+), 13 deletions(-) diff --git a/jaxpm/painting.py b/jaxpm/painting.py index 3083f08..f8797f2 100644 --- a/jaxpm/painting.py +++ b/jaxpm/painting.py @@ -12,7 +12,7 @@ from jaxpm.kernels import cic_compensation, fftk from jaxpm.painting_utils import gather, scatter -def _cic_paint_impl(grid_mesh, positions, weight=None): +def _cic_paint_impl(grid_mesh, positions, weight=1.): """ Paints positions onto mesh mesh: [nx, ny, nz] displacement field: [nx, ny, nz, 3] @@ -27,12 +27,11 @@ def _cic_paint_impl(grid_mesh, positions, weight=None): neighboor_coords = floor + connection kernel = 1. - jnp.abs(positions - neighboor_coords) kernel = kernel[..., 0] * kernel[..., 1] * kernel[..., 2] - if weight is not None: - if jnp.isscalar(weight): - kernel = jnp.multiply(jnp.expand_dims(weight, axis=-1), kernel) - else: - kernel = jnp.multiply(weight.reshape(*positions.shape[:-1]), - kernel) + if jnp.isscalar(weight): + kernel = jnp.multiply(jnp.expand_dims(weight, axis=-1), kernel) + else: + kernel = jnp.multiply(weight.reshape(*positions.shape[:-1]), + kernel) neighboor_coords = jnp.mod( neighboor_coords.reshape([-1, 8, 3]).astype('int32'), @@ -48,7 +47,13 @@ def _cic_paint_impl(grid_mesh, positions, weight=None): @partial(jax.jit, static_argnums=(3, 4)) -def cic_paint(grid_mesh, positions, weight=None, halo_size=0, sharding=None): +def cic_paint(grid_mesh, positions, weight=1., halo_size=0, sharding=None): + + if sharding is not None: + print(""" + WARNING : absolute painting is not recommended in multi-device mode. + Please use relative painting instead. + """) positions = positions.reshape((*grid_mesh.shape, 3)) @@ -57,9 +62,11 @@ def cic_paint(grid_mesh, positions, weight=None, halo_size=0, sharding=None): gpu_mesh = sharding.mesh if isinstance(sharding, NamedSharding) else None spec = sharding.spec if isinstance(sharding, NamedSharding) else P() + weight_spec = P() if jnp.isscalar(weight) else spec + grid_mesh = autoshmap(_cic_paint_impl, gpu_mesh=gpu_mesh, - in_specs=(spec, spec, P()), + in_specs=(spec, spec, weight_spec), out_specs=spec)(grid_mesh, positions, weight) grid_mesh = halo_exchange(grid_mesh, halo_extents=halo_extents, @@ -151,7 +158,7 @@ def cic_paint_2d(mesh, positions, weight): return mesh -def _cic_paint_dx_impl(displacements, halo_size, weight=1., chunk_size=2**24): +def _cic_paint_dx_impl(displacements, weight=1. , halo_size=0 , chunk_size=2**24): halo_x, _ = halo_size[0] halo_y, _ = halo_size[1] @@ -190,13 +197,13 @@ def cic_paint_dx(displacements, gpu_mesh = sharding.mesh if isinstance(sharding, NamedSharding) else None spec = sharding.spec if isinstance(sharding, NamedSharding) else P() + weight_spec = P() if jnp.isscalar(weight) else spec grid_mesh = autoshmap(partial(_cic_paint_dx_impl, halo_size=halo_size, - weight=weight, chunk_size=chunk_size), gpu_mesh=gpu_mesh, - in_specs=spec, - out_specs=spec)(displacements) + in_specs=(spec, weight_spec), + out_specs=spec)(displacements , weight) grid_mesh = halo_exchange(grid_mesh, halo_extents=halo_extents, From eb0778aa245a8c61b13066a2d3ac7bc8f78ddca9 Mon Sep 17 00:00:00 2001 From: Wassim Kabalan Date: Fri, 28 Feb 2025 13:47:00 +0100 Subject: [PATCH 2/6] update test_against_fpm --- tests/test_against_fpm.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tests/test_against_fpm.py b/tests/test_against_fpm.py index 6d17939..e02eff4 100644 --- a/tests/test_against_fpm.py +++ b/tests/test_against_fpm.py @@ -76,7 +76,7 @@ def test_nbody_absolute(simulation_config, initial_conditions, a=lpt_scale_factor, order=order) - ode_fn = ODETerm(make_diffrax_ode(cosmo, mesh_shape)) + ode_fn = ODETerm(make_diffrax_ode(mesh_shape)) solver = Dopri5() controller = PIDController(rtol=1e-8, @@ -122,7 +122,7 @@ def test_nbody_relative(simulation_config, initial_conditions, dx, p, _ = lpt(cosmo, initial_conditions, a=lpt_scale_factor, order=order) ode_fn = ODETerm( - make_diffrax_ode(cosmo, mesh_shape, paint_absolute_pos=False)) + make_diffrax_ode(mesh_shape, paint_absolute_pos=False)) solver = Dopri5() controller = PIDController(rtol=1e-9, From 580387ce1c91d25ed96d47d748e118963363bf2b Mon Sep 17 00:00:00 2001 From: Wassim Kabalan Date: Fri, 28 Feb 2025 13:47:22 +0100 Subject: [PATCH 3/6] update distributed tests and add jacfwd jacrev and vmap tests --- tests/test_distributed_pm.py | 280 ++++++++++++++++++++++++++++++++++- 1 file changed, 273 insertions(+), 7 deletions(-) diff --git a/tests/test_distributed_pm.py b/tests/test_distributed_pm.py index eb44456..aa68d54 100644 --- a/tests/test_distributed_pm.py +++ b/tests/test_distributed_pm.py @@ -15,17 +15,29 @@ from jax.sharding import PartitionSpec as P # noqa : E402 from jaxpm.distributed import uniform_particles # noqa : E402 from jaxpm.painting import cic_paint, cic_paint_dx # noqa : E402 -from jaxpm.pm import lpt, make_diffrax_ode # noqa : E402 - -_TOLERANCE = 3.0 # 🙃🙃 +from jaxpm.pm import lpt, make_diffrax_ode , pm_forces # noqa : E402 +from functools import partial # noqa : E402 +import jax_cosmo as jc # noqa : E402 +from jaxpm.distributed import fft3d , ifft3d +from jaxdecomp import get_fft_output_sharding +_TOLERANCE = 1e-1 # 🙃🙃 +pdims = [(1, 8) , (8 , 1) , (4 , 2), (2 , 4)] @pytest.mark.distributed @pytest.mark.parametrize("order", [1, 2]) +@pytest.mark.parametrize("pdims", pdims) @pytest.mark.parametrize("absolute_painting", [True, False]) -def test_distrubted_pm(simulation_config, initial_conditions, cosmo, order, +def test_distrubted_pm(simulation_config, initial_conditions, cosmo, order,pdims, absolute_painting): + if absolute_painting: + pytest.skip("Absolute painting is not recommended in distributed mode") + + painting_str = "absolute" if absolute_painting else "relative" + print("="*50) + print(f"Running with {painting_str} painting and pdims {pdims} ...") + mesh_shape, box_shape = simulation_config # SINGLE DEVICE RUN cosmo._workspace = {} @@ -60,6 +72,7 @@ def test_distrubted_pm(simulation_config, initial_conditions, cosmo, order, t1=1.0, dt0=None, y0=y0, + args=cosmo, stepsize_controller=controller, saveat=saveat) @@ -72,7 +85,7 @@ def test_distrubted_pm(simulation_config, initial_conditions, cosmo, order, print("Done with single device run") # MULTI DEVICE RUN - mesh = jax.make_mesh((1, 8), ('x', 'y')) + mesh = jax.make_mesh(pdims, ('x', 'y')) sharding = NamedSharding(mesh, P('x', 'y')) halo_size = mesh_shape[0] // 2 @@ -128,16 +141,23 @@ def test_distrubted_pm(simulation_config, initial_conditions, cosmo, order, t1=1.0, dt0=None, y0=y0, + args=cosmo, stepsize_controller=controller, saveat=saveat) + final_field = solutions.ys[-1, 0] + print(f"Final field sharding is {final_field.sharding}") + + assert final_field.sharding.is_equivalent_to(sharding , ndim=3) \ + , f"Final field sharding is not correct .. should be {sharding} it is instead {final_field.sharding}" + if absolute_painting: multi_device_final_field = cic_paint(jnp.zeros(shape=mesh_shape), - solutions.ys[-1, 0], + final_field, halo_size=halo_size, sharding=sharding) else: - multi_device_final_field = cic_paint_dx(solutions.ys[-1, 0], + multi_device_final_field = cic_paint_dx(final_field, halo_size=halo_size, sharding=sharding) @@ -148,3 +168,249 @@ def test_distrubted_pm(simulation_config, initial_conditions, cosmo, order, print(f"MSE is {mse}") assert mse < _TOLERANCE + + + +@pytest.mark.distributed +@pytest.mark.parametrize("order", [1, 2]) +@pytest.mark.parametrize("pdims", pdims) +def test_distrubted_gradients(simulation_config, initial_conditions, cosmo, + order, nbody_from_lpt1, nbody_from_lpt2 , pdims): + + + mesh_shape, box_shape = simulation_config + # SINGLE DEVICE RUN + cosmo._workspace = {} + + + mesh = jax.make_mesh(pdims, ('x', 'y')) + sharding = NamedSharding(mesh, P('x', 'y')) + halo_size = mesh_shape[0] // 2 + + + initial_conditions = lax.with_sharding_constraint(initial_conditions, + sharding) + + print(f"sharded initial conditions {initial_conditions.sharding}") + cosmo._workspace = {} + + + @jax.jit + def forward_model(initial_conditions, cosmo): + + + dx, p, _ = lpt(cosmo, + initial_conditions, + a=0.1, + order=order, + halo_size=halo_size, + sharding=sharding) + ode_fn = ODETerm( + make_diffrax_ode(mesh_shape, + paint_absolute_pos=False, + halo_size=halo_size, + sharding=sharding)) + y0 = jax.tree.map(lambda dx, p: jnp.stack([dx, p]), dx, p) + + solver = Dopri5() + controller = PIDController(rtol=1e-8, + atol=1e-8, + pcoeff=0.4, + icoeff=1, + dcoeff=0) + + + saveat = SaveAt(t1=True) + solutions = diffeqsolve(ode_fn, + solver, + t0=0.1, + t1=1.0, + dt0=None, + y0=y0, + args=cosmo, + stepsize_controller=controller, + saveat=saveat) + + + + multi_device_final_field = cic_paint_dx(solutions.ys[-1, 0], + halo_size=halo_size, + sharding=sharding) + + return multi_device_final_field + + + @jax.jit + def model(initial_conditions, cosmo): + final_field = forward_model(initial_conditions, cosmo) + return MSE(final_field, + nbody_from_lpt1 if order == 1 else nbody_from_lpt2) + + obs_val = model(initial_conditions, cosmo) + + shifted_initial_conditions = initial_conditions + jax.random.normal( + jax.random.key(42), initial_conditions.shape) * 5 + + + good_grads = jax.grad(model)(initial_conditions, cosmo) + off_grads = jax.grad(model)(shifted_initial_conditions, cosmo) + + + assert good_grads.sharding.is_equivalent_to(initial_conditions.sharding , ndim=3) + assert off_grads.sharding.is_equivalent_to(initial_conditions.sharding , ndim=3) + + + +@pytest.mark.distributed +@pytest.mark.parametrize("pdims", pdims) +def test_fwd_rev_gradients(cosmo,pdims): + + + mesh_shape, box_shape = (8, 8, 8), (20.0, 20.0, 20.0) + # SINGLE DEVICE RUN + cosmo._workspace = {} + + + mesh = jax.make_mesh(pdims, ('x', 'y')) + sharding = NamedSharding(mesh, P('x', 'y')) + halo_size = mesh_shape[0] // 2 + + + initial_conditions = jax.random.normal(jax.random.PRNGKey(42), mesh_shape) + initial_conditions = lax.with_sharding_constraint(initial_conditions, + sharding) + print(f"sharded initial conditions {initial_conditions.sharding}") + cosmo._workspace = {} + + + @partial(jax.jit, static_argnums=(2, 3, 4)) + def compute_forces(initial_conditions, + cosmo, + a=0.5, + halo_size=0, + sharding=None): + + + paint_absolute_pos = False + particles = jnp.zeros_like(initial_conditions, + shape=(*initial_conditions.shape, 3)) + + + a = jnp.atleast_1d(a) + E = jnp.sqrt(jc.background.Esqr(cosmo, a)) + + initial_conditions = jax.lax.with_sharding_constraint(initial_conditions,sharding) + delta_k = fft3d(initial_conditions) + out_sharding = get_fft_output_sharding(sharding) + delta_k = jax.lax.with_sharding_constraint(delta_k, out_sharding) + + initial_force = pm_forces(particles, + delta=delta_k, + paint_absolute_pos=paint_absolute_pos, + halo_size=halo_size, + sharding=sharding) + + + return initial_force[..., 0] + + + forces = compute_forces(initial_conditions, + cosmo, + halo_size=halo_size, + sharding=sharding) + back_gradient = jax.jacrev(compute_forces)(initial_conditions, + cosmo, + halo_size=halo_size, + sharding=sharding) + fwd_gradient = jax.jacfwd(compute_forces)(initial_conditions, + cosmo, + halo_size=halo_size, + sharding=sharding) + + + print(f"Forces sharding is {forces.sharding}") + print(f"Backward gradient sharding is {back_gradient.sharding}") + print(f"Forward gradient sharding is {fwd_gradient.sharding}") + assert forces.sharding.is_equivalent_to(initial_conditions.sharding , ndim=3) + assert back_gradient[0, 0, 0, ...].sharding.is_equivalent_to(initial_conditions.sharding , ndim=3) + assert fwd_gradient.sharding.is_equivalent_to(initial_conditions.sharding , ndim=3) + + + +@pytest.mark.distributed +@pytest.mark.parametrize("pdims", pdims) +def test_vmap(cosmo,pdims): + + mesh_shape, box_shape = (8, 8, 8), (20.0, 20.0, 20.0) + # SINGLE DEVICE RUN + cosmo._workspace = {} + + + mesh = jax.make_mesh(pdims, ('x', 'y')) + sharding = NamedSharding(mesh, P('x', 'y')) + halo_size = mesh_shape[0] // 2 + + + single_dev_initial_conditions = jax.random.normal(jax.random.PRNGKey(42), mesh_shape) + initial_conditions = lax.with_sharding_constraint(single_dev_initial_conditions, + sharding) + + single_ics = jnp.stack([single_dev_initial_conditions, single_dev_initial_conditions , single_dev_initial_conditions]) + sharded_ics = jnp.stack([initial_conditions, initial_conditions , initial_conditions]) + print(f"unsharded initial conditions batch {single_ics.sharding}") + print(f"sharded initial conditions batch {sharded_ics.sharding}") + cosmo._workspace = {} + + + @partial(jax.jit, static_argnums=(2, 3, 4)) + def compute_forces(initial_conditions, + cosmo, + a=0.5, + halo_size=0, + sharding=None): + + + paint_absolute_pos = False + particles = jnp.zeros_like(initial_conditions, + shape=(*initial_conditions.shape, 3)) + + + a = jnp.atleast_1d(a) + E = jnp.sqrt(jc.background.Esqr(cosmo, a)) + + initial_conditions = jax.lax.with_sharding_constraint(initial_conditions,sharding) + delta_k = fft3d(initial_conditions) + out_sharding = get_fft_output_sharding(sharding) + delta_k = jax.lax.with_sharding_constraint(delta_k, out_sharding) + + initial_force = pm_forces(particles, + delta=delta_k, + paint_absolute_pos=paint_absolute_pos, + halo_size=halo_size, + sharding=sharding) + + + return initial_force[..., 0] + + def fn(ic): + return compute_forces(ic, + cosmo, + halo_size=halo_size, + sharding=sharding) + + v_compute_forces = jax.vmap(fn) + + print(f"single_ics shape {single_ics.shape}") + print(f"sharded_ics shape {sharded_ics.shape}") + + single_dev_forces = v_compute_forces(single_ics) + sharded_forces = v_compute_forces(sharded_ics) + + assert single_dev_forces.ndim == 4 + assert sharded_forces.ndim == 4 + + + print(f"Sharded forces {sharded_forces.sharding}") + + assert sharded_forces[0].sharding.is_equivalent_to(initial_conditions.sharding , ndim=3) + assert sharded_forces.sharding.spec[0] == None \ No newline at end of file From 9f494da317bbe03cfbbb579ad2b8ec28ed000312 Mon Sep 17 00:00:00 2001 From: Wassim Kabalan Date: Fri, 28 Feb 2025 13:47:43 +0100 Subject: [PATCH 4/6] format --- jaxpm/painting.py | 10 +-- tests/test_against_fpm.py | 3 +- tests/test_distributed_pm.py | 124 ++++++++++++++++------------------- 3 files changed, 62 insertions(+), 75 deletions(-) diff --git a/jaxpm/painting.py b/jaxpm/painting.py index f8797f2..78d63ef 100644 --- a/jaxpm/painting.py +++ b/jaxpm/painting.py @@ -30,8 +30,7 @@ def _cic_paint_impl(grid_mesh, positions, weight=1.): if jnp.isscalar(weight): kernel = jnp.multiply(jnp.expand_dims(weight, axis=-1), kernel) else: - kernel = jnp.multiply(weight.reshape(*positions.shape[:-1]), - kernel) + kernel = jnp.multiply(weight.reshape(*positions.shape[:-1]), kernel) neighboor_coords = jnp.mod( neighboor_coords.reshape([-1, 8, 3]).astype('int32'), @@ -158,7 +157,10 @@ def cic_paint_2d(mesh, positions, weight): return mesh -def _cic_paint_dx_impl(displacements, weight=1. , halo_size=0 , chunk_size=2**24): +def _cic_paint_dx_impl(displacements, + weight=1., + halo_size=0, + chunk_size=2**24): halo_x, _ = halo_size[0] halo_y, _ = halo_size[1] @@ -203,7 +205,7 @@ def cic_paint_dx(displacements, chunk_size=chunk_size), gpu_mesh=gpu_mesh, in_specs=(spec, weight_spec), - out_specs=spec)(displacements , weight) + out_specs=spec)(displacements, weight) grid_mesh = halo_exchange(grid_mesh, halo_extents=halo_extents, diff --git a/tests/test_against_fpm.py b/tests/test_against_fpm.py index e02eff4..5ef5211 100644 --- a/tests/test_against_fpm.py +++ b/tests/test_against_fpm.py @@ -121,8 +121,7 @@ def test_nbody_relative(simulation_config, initial_conditions, # Initial displacement dx, p, _ = lpt(cosmo, initial_conditions, a=lpt_scale_factor, order=order) - ode_fn = ODETerm( - make_diffrax_ode(mesh_shape, paint_absolute_pos=False)) + ode_fn = ODETerm(make_diffrax_ode(mesh_shape, paint_absolute_pos=False)) solver = Dopri5() controller = PIDController(rtol=1e-9, diff --git a/tests/test_distributed_pm.py b/tests/test_distributed_pm.py index aa68d54..6e94f56 100644 --- a/tests/test_distributed_pm.py +++ b/tests/test_distributed_pm.py @@ -2,8 +2,11 @@ from conftest import initialize_distributed initialize_distributed() # ignore : E402 +from functools import partial # noqa : E402 + import jax # noqa : E402 import jax.numpy as jnp # noqa : E402 +import jax_cosmo as jc # noqa : E402 import pytest # noqa : E402 from diffrax import SaveAt # noqa : E402 from diffrax import Dopri5, ODETerm, PIDController, diffeqsolve @@ -12,30 +15,30 @@ from jax import lax # noqa : E402 from jax.experimental.multihost_utils import process_allgather # noqa : E402 from jax.sharding import NamedSharding from jax.sharding import PartitionSpec as P # noqa : E402 +from jaxdecomp import get_fft_output_sharding from jaxpm.distributed import uniform_particles # noqa : E402 +from jaxpm.distributed import fft3d, ifft3d from jaxpm.painting import cic_paint, cic_paint_dx # noqa : E402 -from jaxpm.pm import lpt, make_diffrax_ode , pm_forces # noqa : E402 -from functools import partial # noqa : E402 -import jax_cosmo as jc # noqa : E402 -from jaxpm.distributed import fft3d , ifft3d -from jaxdecomp import get_fft_output_sharding +from jaxpm.pm import lpt, make_diffrax_ode, pm_forces # noqa : E402 + _TOLERANCE = 1e-1 # 🙃🙃 -pdims = [(1, 8) , (8 , 1) , (4 , 2), (2 , 4)] +pdims = [(1, 8), (8, 1), (4, 2), (2, 4)] + @pytest.mark.distributed @pytest.mark.parametrize("order", [1, 2]) @pytest.mark.parametrize("pdims", pdims) @pytest.mark.parametrize("absolute_painting", [True, False]) -def test_distrubted_pm(simulation_config, initial_conditions, cosmo, order,pdims, - absolute_painting): +def test_distrubted_pm(simulation_config, initial_conditions, cosmo, order, + pdims, absolute_painting): if absolute_painting: pytest.skip("Absolute painting is not recommended in distributed mode") painting_str = "absolute" if absolute_painting else "relative" - print("="*50) + print("=" * 50) print(f"Running with {painting_str} painting and pdims {pdims} ...") mesh_shape, box_shape = simulation_config @@ -170,46 +173,40 @@ def test_distrubted_pm(simulation_config, initial_conditions, cosmo, order,pdims assert mse < _TOLERANCE - @pytest.mark.distributed @pytest.mark.parametrize("order", [1, 2]) @pytest.mark.parametrize("pdims", pdims) def test_distrubted_gradients(simulation_config, initial_conditions, cosmo, - order, nbody_from_lpt1, nbody_from_lpt2 , pdims): - + order, nbody_from_lpt1, nbody_from_lpt2, pdims): mesh_shape, box_shape = simulation_config # SINGLE DEVICE RUN cosmo._workspace = {} - mesh = jax.make_mesh(pdims, ('x', 'y')) sharding = NamedSharding(mesh, P('x', 'y')) halo_size = mesh_shape[0] // 2 - initial_conditions = lax.with_sharding_constraint(initial_conditions, sharding) print(f"sharded initial conditions {initial_conditions.sharding}") cosmo._workspace = {} - @jax.jit def forward_model(initial_conditions, cosmo): - dx, p, _ = lpt(cosmo, - initial_conditions, - a=0.1, - order=order, - halo_size=halo_size, - sharding=sharding) + initial_conditions, + a=0.1, + order=order, + halo_size=halo_size, + sharding=sharding) ode_fn = ODETerm( make_diffrax_ode(mesh_shape, - paint_absolute_pos=False, - halo_size=halo_size, - sharding=sharding)) + paint_absolute_pos=False, + halo_size=halo_size, + sharding=sharding)) y0 = jax.tree.map(lambda dx, p: jnp.stack([dx, p]), dx, p) solver = Dopri5() @@ -219,7 +216,6 @@ def test_distrubted_gradients(simulation_config, initial_conditions, cosmo, icoeff=1, dcoeff=0) - saveat = SaveAt(t1=True) solutions = diffeqsolve(ode_fn, solver, @@ -231,15 +227,12 @@ def test_distrubted_gradients(simulation_config, initial_conditions, cosmo, stepsize_controller=controller, saveat=saveat) - - multi_device_final_field = cic_paint_dx(solutions.ys[-1, 0], halo_size=halo_size, sharding=sharding) return multi_device_final_field - @jax.jit def model(initial_conditions, cosmo): final_field = forward_model(initial_conditions, cosmo) @@ -251,38 +244,33 @@ def test_distrubted_gradients(simulation_config, initial_conditions, cosmo, shifted_initial_conditions = initial_conditions + jax.random.normal( jax.random.key(42), initial_conditions.shape) * 5 - good_grads = jax.grad(model)(initial_conditions, cosmo) off_grads = jax.grad(model)(shifted_initial_conditions, cosmo) - - assert good_grads.sharding.is_equivalent_to(initial_conditions.sharding , ndim=3) - assert off_grads.sharding.is_equivalent_to(initial_conditions.sharding , ndim=3) - + assert good_grads.sharding.is_equivalent_to(initial_conditions.sharding, + ndim=3) + assert off_grads.sharding.is_equivalent_to(initial_conditions.sharding, + ndim=3) @pytest.mark.distributed @pytest.mark.parametrize("pdims", pdims) -def test_fwd_rev_gradients(cosmo,pdims): - +def test_fwd_rev_gradients(cosmo, pdims): mesh_shape, box_shape = (8, 8, 8), (20.0, 20.0, 20.0) # SINGLE DEVICE RUN cosmo._workspace = {} - mesh = jax.make_mesh(pdims, ('x', 'y')) sharding = NamedSharding(mesh, P('x', 'y')) halo_size = mesh_shape[0] // 2 - initial_conditions = jax.random.normal(jax.random.PRNGKey(42), mesh_shape) initial_conditions = lax.with_sharding_constraint(initial_conditions, sharding) print(f"sharded initial conditions {initial_conditions.sharding}") cosmo._workspace = {} - @partial(jax.jit, static_argnums=(2, 3, 4)) def compute_forces(initial_conditions, cosmo, @@ -290,16 +278,15 @@ def test_fwd_rev_gradients(cosmo,pdims): halo_size=0, sharding=None): - paint_absolute_pos = False particles = jnp.zeros_like(initial_conditions, shape=(*initial_conditions.shape, 3)) - a = jnp.atleast_1d(a) E = jnp.sqrt(jc.background.Esqr(cosmo, a)) - - initial_conditions = jax.lax.with_sharding_constraint(initial_conditions,sharding) + + initial_conditions = jax.lax.with_sharding_constraint( + initial_conditions, sharding) delta_k = fft3d(initial_conditions) out_sharding = get_fft_output_sharding(sharding) delta_k = jax.lax.with_sharding_constraint(delta_k, out_sharding) @@ -310,10 +297,8 @@ def test_fwd_rev_gradients(cosmo,pdims): halo_size=halo_size, sharding=sharding) - return initial_force[..., 0] - forces = compute_forces(initial_conditions, cosmo, halo_size=halo_size, @@ -327,41 +312,44 @@ def test_fwd_rev_gradients(cosmo,pdims): halo_size=halo_size, sharding=sharding) - print(f"Forces sharding is {forces.sharding}") print(f"Backward gradient sharding is {back_gradient.sharding}") print(f"Forward gradient sharding is {fwd_gradient.sharding}") - assert forces.sharding.is_equivalent_to(initial_conditions.sharding , ndim=3) - assert back_gradient[0, 0, 0, ...].sharding.is_equivalent_to(initial_conditions.sharding , ndim=3) - assert fwd_gradient.sharding.is_equivalent_to(initial_conditions.sharding , ndim=3) - + assert forces.sharding.is_equivalent_to(initial_conditions.sharding, + ndim=3) + assert back_gradient[0, 0, 0, ...].sharding.is_equivalent_to( + initial_conditions.sharding, ndim=3) + assert fwd_gradient.sharding.is_equivalent_to(initial_conditions.sharding, + ndim=3) @pytest.mark.distributed @pytest.mark.parametrize("pdims", pdims) -def test_vmap(cosmo,pdims): +def test_vmap(cosmo, pdims): mesh_shape, box_shape = (8, 8, 8), (20.0, 20.0, 20.0) # SINGLE DEVICE RUN cosmo._workspace = {} - mesh = jax.make_mesh(pdims, ('x', 'y')) sharding = NamedSharding(mesh, P('x', 'y')) halo_size = mesh_shape[0] // 2 + single_dev_initial_conditions = jax.random.normal(jax.random.PRNGKey(42), + mesh_shape) + initial_conditions = lax.with_sharding_constraint( + single_dev_initial_conditions, sharding) - single_dev_initial_conditions = jax.random.normal(jax.random.PRNGKey(42), mesh_shape) - initial_conditions = lax.with_sharding_constraint(single_dev_initial_conditions, - sharding) - - single_ics = jnp.stack([single_dev_initial_conditions, single_dev_initial_conditions , single_dev_initial_conditions]) - sharded_ics = jnp.stack([initial_conditions, initial_conditions , initial_conditions]) + single_ics = jnp.stack([ + single_dev_initial_conditions, single_dev_initial_conditions, + single_dev_initial_conditions + ]) + sharded_ics = jnp.stack( + [initial_conditions, initial_conditions, initial_conditions]) print(f"unsharded initial conditions batch {single_ics.sharding}") print(f"sharded initial conditions batch {sharded_ics.sharding}") cosmo._workspace = {} - @partial(jax.jit, static_argnums=(2, 3, 4)) def compute_forces(initial_conditions, cosmo, @@ -369,16 +357,15 @@ def test_vmap(cosmo,pdims): halo_size=0, sharding=None): - paint_absolute_pos = False particles = jnp.zeros_like(initial_conditions, shape=(*initial_conditions.shape, 3)) - a = jnp.atleast_1d(a) E = jnp.sqrt(jc.background.Esqr(cosmo, a)) - - initial_conditions = jax.lax.with_sharding_constraint(initial_conditions,sharding) + + initial_conditions = jax.lax.with_sharding_constraint( + initial_conditions, sharding) delta_k = fft3d(initial_conditions) out_sharding = get_fft_output_sharding(sharding) delta_k = jax.lax.with_sharding_constraint(delta_k, out_sharding) @@ -389,14 +376,13 @@ def test_vmap(cosmo,pdims): halo_size=halo_size, sharding=sharding) - return initial_force[..., 0] def fn(ic): return compute_forces(ic, - cosmo, - halo_size=halo_size, - sharding=sharding) + cosmo, + halo_size=halo_size, + sharding=sharding) v_compute_forces = jax.vmap(fn) @@ -409,8 +395,8 @@ def test_vmap(cosmo,pdims): assert single_dev_forces.ndim == 4 assert sharded_forces.ndim == 4 - print(f"Sharded forces {sharded_forces.sharding}") - assert sharded_forces[0].sharding.is_equivalent_to(initial_conditions.sharding , ndim=3) - assert sharded_forces.sharding.spec[0] == None \ No newline at end of file + assert sharded_forces[0].sharding.is_equivalent_to( + initial_conditions.sharding, ndim=3) + assert sharded_forces.sharding.spec[0] == None From 4e4d3745f0b4ada88c9f2ac0b7e5bf0d7b6f1953 Mon Sep 17 00:00:00 2001 From: Wassim Kabalan Date: Fri, 28 Feb 2025 13:48:08 +0100 Subject: [PATCH 5/6] add Caveats to notebook readme --- notebooks/README.md | 47 +++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 47 insertions(+) diff --git a/notebooks/README.md b/notebooks/README.md index 43d9d0b..872fdd4 100644 --- a/notebooks/README.md +++ b/notebooks/README.md @@ -37,3 +37,50 @@ Each notebook includes installation instructions and guidelines for configuring - **SLURM** for job scheduling on clusters (if running multi-host setups) > **Note**: These notebooks are tested on the **Jean Zay** supercomputer and may require configuration changes for different HPC clusters. + +## Caveats + +### Cloud-in-Cell (CIC) Painting (Single Device) + +There is two ways to perform the CIC painting in JAXPM. The first one is to use the `cic_paint` which paints absolute particle positions to the mesh. The second one is to use the `cic_paint_dx` which paints relative particle positions to the mesh (using uniform particles). The absolute version is faster at the cost of more memory usage. + +inorder to use relative painting you need to : + + - Set the `particles` argument in `lpt` function from `jaxpm.pm` to `None` + - Set `paint_absolute_pos` to `False` in `make_ode_fn` or `make_diffrax_ode` function from `jaxpm.pm` (it is True by default) + +Otherwise you set `particles` to the starting particles of your choice and leave `paint_absolute_pos` to `True` (default value). + +### Cloud-in-Cell (CIC) Painting (Multi Device) + +Both `cic_paint` and `cic_paint_dx` functions are available in multi-device mode. + +You need to set the arguments `sharding` and `halo_size` which is explained in the notebook [03-MultiGPU_PM_Halo.ipynb](03-MultiGPU_PM_Halo.ipynb). + +One thing to note that `cic_paint` is not as accurate as `cic_paint_dx` in multi-device mode and therefor is not recommended. + +Using relative painting in multi-device mode is just like in single device mode.\ +You need to set the `particles` argument in `lpt` function from `jaxpm.pm` to `None` and set `paint_absolute_pos` to `False` + +### Distributed PM + +To run a distributed PM follow the examples in notebooks [03](03-MultiGPU_PM_Halo.ipynb) and [05](05-MultiHost_PM.ipynb) for multi-host. + +In short you need to set the arguments `sharding` and `halo_size` in `lpt` , `linear_field` the `make_ode` functions and `pm_forces` if you use it. + +Missmatching the shardings will give you errors and unexpected results. + +You can also use `normal_field` and `uniform_particles` from `jaxpm.pm.distributed` to create the fields and particles with a sharding. + +### Choosing the right pdims + +pdims are processor dimensions.\ +Explained more in the jaxdecomp paper [here](https://github.com/DifferentiableUniverseInitiative/jaxDecomp). + +For 8 devices there are three decompositions that are possible: +- (1 , 8) +- (2 , 4) , (4 , 2) +- (8 , 1) + +(1 , X) should be the fastest (2 , X) or (X , 2) is more accurate but slightly slower.\ +and (X , 1) is giving the least accurate results for some reason so it is not recommended. From e1daa8cba4ece2c9449d2fdbe5b195a58fe19caf Mon Sep 17 00:00:00 2001 From: Wassim Kabalan Date: Fri, 28 Feb 2025 14:03:33 +0100 Subject: [PATCH 6/6] final touches --- tests/test_against_fpm.py | 2 ++ tests/test_distributed_pm.py | 2 -- tests/test_gradients.py | 5 +++-- 3 files changed, 5 insertions(+), 4 deletions(-) diff --git a/tests/test_against_fpm.py b/tests/test_against_fpm.py index 5ef5211..5ebcbc2 100644 --- a/tests/test_against_fpm.py +++ b/tests/test_against_fpm.py @@ -95,6 +95,7 @@ def test_nbody_absolute(simulation_config, initial_conditions, t1=1.0, dt0=None, y0=y0, + args=cosmo, stepsize_controller=controller, saveat=saveat) @@ -140,6 +141,7 @@ def test_nbody_relative(simulation_config, initial_conditions, t1=1.0, dt0=None, y0=y0, + args=cosmo, stepsize_controller=controller, saveat=saveat) diff --git a/tests/test_distributed_pm.py b/tests/test_distributed_pm.py index 6e94f56..69c37ed 100644 --- a/tests/test_distributed_pm.py +++ b/tests/test_distributed_pm.py @@ -258,7 +258,6 @@ def test_distrubted_gradients(simulation_config, initial_conditions, cosmo, def test_fwd_rev_gradients(cosmo, pdims): mesh_shape, box_shape = (8, 8, 8), (20.0, 20.0, 20.0) - # SINGLE DEVICE RUN cosmo._workspace = {} mesh = jax.make_mesh(pdims, ('x', 'y')) @@ -328,7 +327,6 @@ def test_fwd_rev_gradients(cosmo, pdims): def test_vmap(cosmo, pdims): mesh_shape, box_shape = (8, 8, 8), (20.0, 20.0, 20.0) - # SINGLE DEVICE RUN cosmo._workspace = {} mesh = jax.make_mesh(pdims, ('x', 'y')) diff --git a/tests/test_gradients.py b/tests/test_gradients.py index bb48920..1f611aa 100644 --- a/tests/test_gradients.py +++ b/tests/test_gradients.py @@ -39,7 +39,7 @@ def test_nbody_grad(simulation_config, initial_conditions, lpt_scale_factor, particles, a=lpt_scale_factor, order=order) - ode_fn = ODETerm(make_diffrax_ode(cosmo, mesh_shape)) + ode_fn = ODETerm(make_diffrax_ode(mesh_shape)) y0 = jnp.stack([particles + dx, p]) else: @@ -48,7 +48,7 @@ def test_nbody_grad(simulation_config, initial_conditions, lpt_scale_factor, a=lpt_scale_factor, order=order) ode_fn = ODETerm( - make_diffrax_ode(cosmo, mesh_shape, paint_absolute_pos=False)) + make_diffrax_ode(mesh_shape, paint_absolute_pos=False)) y0 = jnp.stack([dx, p]) solver = Dopri5() @@ -66,6 +66,7 @@ def test_nbody_grad(simulation_config, initial_conditions, lpt_scale_factor, t1=1.0, dt0=None, y0=y0, + args=cosmo, adjoint=adjoint, stepsize_controller=controller, saveat=saveat)