Files
lerobot/tests/optim/test_optimizers.py
T
Haoming Song ef88d4e52b feat(train): parallel training framework — FSDP2, HSDP, gradient accumulation, and DCP checkpoints (#4010)
* feat(train): parallel training engine with FSDP2, HSDP, and DCP checkpoints

Replace the FSDP1 training path with a config-owned parallel-training
engine:

- Topology and runtime configs (--parallelism.*, --accelerator.*):
  dp_replicate x dp_shard degrees select single-process, DDP (unchanged
  default), FSDP2, or HSDP; mixed precision, first-class gradient
  accumulation, and FSDP/DDP tuning knobs are mirrored as plain
  dataclasses that build the accelerate objects at runtime, so every
  run is reproducible from its train_config.json alone. Accelerate env
  vars are guarded against configuring the engine behind the config
  system's back.
- Declarative policy surface: policies declare FSDP2 wrap units
  (_fsdp_wrap_modules) and non-forward entry points
  (_fsdp_forward_methods); a shared engine resolves them around
  accelerator.prepare(). Context-parallel fields are reserved and
  validated to 1.
- Checkpoints: selectable --checkpoint_format (safetensors | dcp |
  safetensors_dcp); the sharded optimizer channel is always DCP;
  two-phase resume (step+RNG before prepare, DCP model/optimizer after)
  reshards across GPU-topology changes; lerobot-convert-dcp merges DCP
  shards into a distributable model.safetensors offline.
- Publishing: PreTrainedPolicy.push_model_to_hub is replaced by the
  free publish_trained_model (model + processors + card + train config,
  all-ranks gather with main-rank writes);
  PreTrainedPolicy._save_pretrained gathers state dicts internally,
  removing the state_dict= threading from save_pretrained.
- lerobot_train is restructured around the engine: optimizer built
  before the single prepare() call, deferred weight load on DCP
  resumes, collective save_checkpoint with no call-site rank branches,
  dp-world-size-based sample accounting.

Breaking changes: FSDP checkpoints from lerobot <= 0.6.x are not
resumable (weights stay loadable via from_pretrained; pin
lerobot==0.6.x to finish old runs); the `accelerate launch
--config_file` yaml flow is superseded by the config flags; training
autocast is owned exclusively by --accelerator.mixed_precision
(policy.dtype only casts parameters).

Also fixes: reward-model hub publishing crash (TypeError on extra
kwargs).

Verified by ~200 new CPU tests (config round-trips, checkpoint
round-trips per format, two-phase resume, publisher contracts,
converter equivalence, accelerate canaries), a 5-test 4-GPU suite
(FSDP2 save/resume bit-exactness, HSDP/DDP loss parity,
changed-topology resume, all-ranks save_pretrained, grad-accum
equivalence), and end-to-end ACT (1/4/8 GPUs) + FastWAM 6B
(FSDP2 + HSDP) training runs.
2026-08-06 19:16:41 +08:00

243 lines
8.6 KiB
Python

# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import pytest
import torch
from lerobot.optim.optimizers import (
AdamConfig,
AdamWConfig,
MultiAdamConfig,
SGDConfig,
load_optimizer_state,
save_optimizer_state,
)
from lerobot.utils.constants import (
OPTIMIZER_PARAM_GROUPS,
OPTIMIZER_STATE,
)
@pytest.mark.parametrize(
"config_cls, expected_class",
[
(AdamConfig, torch.optim.Adam),
(AdamWConfig, torch.optim.AdamW),
(SGDConfig, torch.optim.SGD),
(MultiAdamConfig, dict),
],
)
def test_optimizer_build(config_cls, expected_class, model_params):
config = config_cls()
if config_cls == MultiAdamConfig:
params_dict = {"default": model_params}
optimizer = config.build(params_dict)
assert isinstance(optimizer, expected_class)
assert isinstance(optimizer["default"], torch.optim.Adam)
assert optimizer["default"].defaults["lr"] == config.lr
else:
optimizer = config.build(model_params)
assert isinstance(optimizer, expected_class)
assert optimizer.defaults["lr"] == config.lr
def test_save_optimizer_state(optimizer, tmp_path):
save_optimizer_state(optimizer, tmp_path)
assert (tmp_path / OPTIMIZER_STATE).is_file()
assert (tmp_path / OPTIMIZER_PARAM_GROUPS).is_file()
def test_save_and_load_optimizer_state(model_params, optimizer, tmp_path):
save_optimizer_state(optimizer, tmp_path)
loaded_optimizer = AdamConfig().build(model_params)
loaded_optimizer = load_optimizer_state(loaded_optimizer, tmp_path)
torch.testing.assert_close(optimizer.state_dict(), loaded_optimizer.state_dict())
@pytest.fixture
def base_params_dict():
return {
"actor": [torch.nn.Parameter(torch.randn(10, 10))],
"critic": [torch.nn.Parameter(torch.randn(5, 5))],
"temperature": [torch.nn.Parameter(torch.randn(3, 3))],
}
@pytest.mark.parametrize(
"config_params, expected_values",
[
# Test 1: Basic configuration with different learning rates
(
{
"lr": 1e-3,
"weight_decay": 1e-4,
"optimizer_groups": {
"actor": {"lr": 1e-4},
"critic": {"lr": 5e-4},
"temperature": {"lr": 2e-3},
},
},
{
"actor": {"lr": 1e-4, "weight_decay": 1e-4, "betas": (0.9, 0.999)},
"critic": {"lr": 5e-4, "weight_decay": 1e-4, "betas": (0.9, 0.999)},
"temperature": {"lr": 2e-3, "weight_decay": 1e-4, "betas": (0.9, 0.999)},
},
),
# Test 2: Different weight decays and beta values
(
{
"lr": 1e-3,
"weight_decay": 1e-4,
"optimizer_groups": {
"actor": {"lr": 1e-4, "weight_decay": 1e-5},
"critic": {"lr": 5e-4, "weight_decay": 1e-6},
"temperature": {"lr": 2e-3, "betas": (0.95, 0.999)},
},
},
{
"actor": {"lr": 1e-4, "weight_decay": 1e-5, "betas": (0.9, 0.999)},
"critic": {"lr": 5e-4, "weight_decay": 1e-6, "betas": (0.9, 0.999)},
"temperature": {"lr": 2e-3, "weight_decay": 1e-4, "betas": (0.95, 0.999)},
},
),
# Test 3: Epsilon parameter customization
(
{
"lr": 1e-3,
"weight_decay": 1e-4,
"optimizer_groups": {
"actor": {"lr": 1e-4, "eps": 1e-6},
"critic": {"lr": 5e-4, "eps": 1e-7},
"temperature": {"lr": 2e-3, "eps": 1e-8},
},
},
{
"actor": {"lr": 1e-4, "weight_decay": 1e-4, "betas": (0.9, 0.999), "eps": 1e-6},
"critic": {"lr": 5e-4, "weight_decay": 1e-4, "betas": (0.9, 0.999), "eps": 1e-7},
"temperature": {"lr": 2e-3, "weight_decay": 1e-4, "betas": (0.9, 0.999), "eps": 1e-8},
},
),
],
)
def test_multi_adam_configuration(base_params_dict, config_params, expected_values):
# Create config with the given parameters
config = MultiAdamConfig(**config_params)
optimizers = config.build(base_params_dict)
# Verify optimizer count and keys
assert len(optimizers) == len(expected_values)
assert set(optimizers.keys()) == set(expected_values.keys())
# Check that all optimizers are Adam instances
for opt in optimizers.values():
assert isinstance(opt, torch.optim.Adam)
# Verify hyperparameters for each optimizer
for name, expected in expected_values.items():
optimizer = optimizers[name]
for param, value in expected.items():
assert optimizer.defaults[param] == value
@pytest.fixture
def multi_optimizers(base_params_dict):
config = MultiAdamConfig(
lr=1e-3,
optimizer_groups={
"actor": {"lr": 1e-4},
"critic": {"lr": 5e-4},
"temperature": {"lr": 2e-3},
},
)
return config.build(base_params_dict)
def test_save_multi_optimizer_state(multi_optimizers, tmp_path):
# Save optimizer states
save_optimizer_state(multi_optimizers, tmp_path)
# Verify that directories were created for each optimizer
for name in multi_optimizers:
assert (tmp_path / name).is_dir()
assert (tmp_path / name / OPTIMIZER_STATE).is_file()
assert (tmp_path / name / OPTIMIZER_PARAM_GROUPS).is_file()
def test_save_and_load_multi_optimizer_state(base_params_dict, multi_optimizers, tmp_path):
# Option 1: Add a minimal backward pass to populate optimizer states
for name, params in base_params_dict.items():
if name in multi_optimizers:
# Create a dummy loss and do backward
dummy_loss = params[0].sum()
dummy_loss.backward()
# Perform an optimization step
multi_optimizers[name].step()
# Zero gradients for next steps
multi_optimizers[name].zero_grad()
# Save optimizer states
save_optimizer_state(multi_optimizers, tmp_path)
# Create new optimizers with the same config
config = MultiAdamConfig(
lr=1e-3,
optimizer_groups={
"actor": {"lr": 1e-4},
"critic": {"lr": 5e-4},
"temperature": {"lr": 2e-3},
},
)
new_optimizers = config.build(base_params_dict)
# Load optimizer states
loaded_optimizers = load_optimizer_state(new_optimizers, tmp_path)
# Verify state dictionaries match
for name in multi_optimizers:
torch.testing.assert_close(multi_optimizers[name].state_dict(), loaded_optimizers[name].state_dict())
def test_save_and_load_empty_multi_optimizer_state(base_params_dict, tmp_path):
"""Test saving and loading optimizer states even when the state is empty (no backward pass)."""
# Create config and build optimizers
config = MultiAdamConfig(
lr=1e-3,
optimizer_groups={
"actor": {"lr": 1e-4},
"critic": {"lr": 5e-4},
"temperature": {"lr": 2e-3},
},
)
optimizers = config.build(base_params_dict)
# Save optimizer states without any backward pass (empty state)
save_optimizer_state(optimizers, tmp_path)
# Create new optimizers with the same config
new_optimizers = config.build(base_params_dict)
# Load optimizer states
loaded_optimizers = load_optimizer_state(new_optimizers, tmp_path)
# Verify hyperparameters match even with empty state
for name, optimizer in optimizers.items():
assert optimizer.defaults["lr"] == loaded_optimizers[name].defaults["lr"]
assert optimizer.defaults["weight_decay"] == loaded_optimizers[name].defaults["weight_decay"]
assert optimizer.defaults["betas"] == loaded_optimizers[name].defaults["betas"]
# Verify state dictionaries match (they will be empty)
torch.testing.assert_close(
optimizer.state_dict()["param_groups"], loaded_optimizers[name].state_dict()["param_groups"]
)