Skip to content

Use shard_map instead of pmap #39

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
teddykoker wants to merge 3 commits into
base: main
Choose a base branch
from
Open

Use shard_map instead of pmap #39

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
349e78e
use shard_map instead of pmap
teddykoker Mar 18, 2026
6bfe609
comment on step.copy()
teddykoker Mar 18, 2026
8fe3e71
rm break
teddykoker Mar 18, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
58 changes: 40 additions & 18 deletions nequix/train.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,4 @@
import argparse
import functools
import os
import time
from collections import defaultdict
Expand All @@ -9,7 +8,9 @@
import equinox as eqx
import jax
import jax.numpy as jnp
import jax.sharding as jshard
import jraph
from jax.experimental.shard_map import shard_map
import optax
import yaml
from wandb_osh.hooks import TriggerWandbSyncHook
Expand Down Expand Up @@ -305,8 +306,13 @@ def train(config_path: str):

opt_state = optim.init(eqx.filter(model, eqx.is_array))

model = jax.device_put_replicated(model, list(jax.devices()))
opt_state = jax.device_put_replicated(opt_state, list(jax.devices()))
mesh = jax.sharding.Mesh(jax.devices(), ("device",))
P = jshard.PartitionSpec
data_sharding = jshard.NamedSharding(mesh, P("device"))
model_sharding = jshard.NamedSharding(mesh, P())

model = eqx.filter_shard(model, model_sharding)
opt_state = eqx.filter_shard(opt_state, model_sharding)
ema_model = jax.tree.map(lambda x: x.copy(), model) # copy model
step = jnp.array(0)
start_epoch = 0
Expand All @@ -324,6 +330,9 @@ def train(config_path: str):
best_val_loss,
wandb_run_id,
) = load_training_state(config["resume_from"])
model = eqx.filter_shard(model, model_sharding)
ema_model = eqx.filter_shard(ema_model, model_sharding)
opt_state = eqx.filter_shard(opt_state, model_sharding)

wandb_init_kwargs = {"project": "nequix", "config": config}
if wandb_run_id:
Expand All @@ -333,10 +342,9 @@ def train(config_path: str):
wandb.run.summary["param_count"] = param_count
wandb_run_id = getattr(wandb.run, "id", None)

# @eqx.filter_jit
@functools.partial(eqx.filter_pmap, in_axes=(0, 0, None, 0, 0), axis_name="device")
def train_step(model, ema_model, step, opt_state, batch):
# training step
# remove leading device-shard dim (shard_map preserves rank; model expects no device axis)
batch = jax.tree.map(lambda x: x.squeeze(0), batch)
(total_loss, metrics), grads = eqx.filter_value_and_grad(loss, has_aux=True)(
model,
batch,
Expand All @@ -345,7 +353,7 @@ def train_step(model, ema_model, step, opt_state, batch):
config["stress_weight"],
config["loss_type"],
)
grads = jax.lax.pmean(grads, axis_name="device")
grads = jax.lax.pmean(grads, "device")
metrics["grad_norm"] = optax.global_norm(grads)
updates, opt_state = optim.update(grads, opt_state, eqx.filter(model, eqx.is_array))
model = eqx.apply_updates(model, updates)
Expand All @@ -360,22 +368,37 @@ def train_step(model, ema_model, step, opt_state, batch):
)
ema_model = eqx.combine(ema_static, new_ema_params)

return (
model,
ema_model,
opt_state,
total_loss,
metrics,
)
# add axis so shard_map can concat on per-device axis
metrics = jax.tree.map(lambda x: x[None], metrics)
total_loss = total_loss[None]

return model, ema_model, opt_state, total_loss, metrics

metrics_spec = {
"energy_mae_per_atom": P("device"),
"force_mae": P("device"),
"stress_mae_per_atom": P("device"),
"grad_norm": P("device"),
}
train_step_sharded = shard_map(
train_step,
mesh,
in_specs=(P(), P(), P(), P(), P("device")),
out_specs=(P(), P(), P(), P("device"), metrics_spec),
check_rep=False, # NOTE: needed for kernels
)
train_step = eqx.filter_jit(donate="all")(train_step_sharded)

for epoch in range(start_epoch, config["n_epochs"]):
start_time = time.time()
train_loader.loader.set_epoch(epoch)
for batch in prefetch(train_loader):
batch_time = time.time() - start_time
start_time = time.time()
batch_sharded = eqx.filter_shard(batch, data_sharding)
# NB: step is copied because it is donated in train_step
(model, ema_model, opt_state, total_loss, metrics) = train_step(
model, ema_model, step, opt_state, batch
model, ema_model, step.copy(), opt_state, batch_sharded
)
train_time = time.time() - start_time
step = step + 1
Expand All @@ -395,9 +418,8 @@ def train_step(model, ema_model, step, opt_state, batch):
wandb_sync()
start_time = time.time()

ema_model_single = jax.tree.map(lambda x: x[0], ema_model)
val_metrics = evaluate(
ema_model_single,
ema_model,
val_loader,
config["energy_weight"],
config["force_weight"],
Expand All @@ -407,7 +429,7 @@ def train_step(model, ema_model, step, opt_state, batch):

if val_metrics["loss"] < best_val_loss:
best_val_loss = val_metrics["loss"]
save_model(Path(wandb.run.dir) / "checkpoint.nqx", ema_model_single, config)
save_model(Path(wandb.run.dir) / "checkpoint.nqx", ema_model, config)

save_training_state(
Path(wandb.run.dir) / "state.pkl",
Expand Down
Loading