|
|
@ -112,13 +112,6 @@ def is_model_parallel_parameter(p):
|
|
|
|
return hasattr(p, IS_TENSOR_PARALLEL) and getattr(p, IS_TENSOR_PARALLEL)
|
|
|
|
return hasattr(p, IS_TENSOR_PARALLEL) and getattr(p, IS_TENSOR_PARALLEL)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def is_moe_parallel_parameter(p):
|
|
|
|
|
|
|
|
# FIXME(HHC): clip_grad need to changed to adapted for MoE
|
|
|
|
|
|
|
|
# This return value must set to False, otherwise it will raise
|
|
|
|
|
|
|
|
# an error in training
|
|
|
|
|
|
|
|
return False
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def _calc_l2_norm(grads):
|
|
|
|
def _calc_l2_norm(grads):
|
|
|
|
norm = 0.0
|
|
|
|
norm = 0.0
|
|
|
|
if len(grads) > 0:
|
|
|
|
if len(grads) > 0:
|
|
|
@ -214,14 +207,11 @@ def clip_grad_norm_fp32(parameters, max_norm, norm_type=2):
|
|
|
|
else:
|
|
|
|
else:
|
|
|
|
tensor_parallel_grads = []
|
|
|
|
tensor_parallel_grads = []
|
|
|
|
no_tensor_parallel_grads = []
|
|
|
|
no_tensor_parallel_grads = []
|
|
|
|
moe_parallel_grads = [] # used to collect moe tensor parallel gradients
|
|
|
|
|
|
|
|
zero_sharded_grads = []
|
|
|
|
zero_sharded_grads = []
|
|
|
|
for p in params:
|
|
|
|
for p in params:
|
|
|
|
if is_model_parallel_parameter(p):
|
|
|
|
if is_model_parallel_parameter(p):
|
|
|
|
reductor = (gpc.get_world_size(ParallelMode.TENSOR) / getattr(p, NUM_PARTITIONS))**(1 / norm_type)
|
|
|
|
reductor = (gpc.get_world_size(ParallelMode.TENSOR) / getattr(p, NUM_PARTITIONS))**(1 / norm_type)
|
|
|
|
tensor_parallel_grads.append(p.grad.data / reductor)
|
|
|
|
tensor_parallel_grads.append(p.grad.data / reductor)
|
|
|
|
elif is_moe_parallel_parameter(p):
|
|
|
|
|
|
|
|
moe_parallel_grads.append(p.grad.data)
|
|
|
|
|
|
|
|
elif hasattr(p, 'zero_is_sharded'):
|
|
|
|
elif hasattr(p, 'zero_is_sharded'):
|
|
|
|
zero_sharded_grads.append(p.grad.data)
|
|
|
|
zero_sharded_grads.append(p.grad.data)
|
|
|
|
else:
|
|
|
|
else:
|
|
|
@ -230,28 +220,21 @@ def clip_grad_norm_fp32(parameters, max_norm, norm_type=2):
|
|
|
|
if norm_type == 2.0 and enable_cuda_kernels:
|
|
|
|
if norm_type == 2.0 and enable_cuda_kernels:
|
|
|
|
tensor_parallel_norm = _calc_l2_norm(tensor_parallel_grads)**norm_type
|
|
|
|
tensor_parallel_norm = _calc_l2_norm(tensor_parallel_grads)**norm_type
|
|
|
|
no_tensor_parallel_norm = _calc_l2_norm(no_tensor_parallel_grads)**norm_type
|
|
|
|
no_tensor_parallel_norm = _calc_l2_norm(no_tensor_parallel_grads)**norm_type
|
|
|
|
moe_parallel_norm = _calc_l2_norm(moe_parallel_grads)**norm_type
|
|
|
|
|
|
|
|
zero_sharded_norm = _calc_l2_norm(zero_sharded_grads)**norm_type
|
|
|
|
zero_sharded_norm = _calc_l2_norm(zero_sharded_grads)**norm_type
|
|
|
|
else:
|
|
|
|
else:
|
|
|
|
tensor_parallel_norm = _calc_lp(tensor_parallel_grads, norm_type)
|
|
|
|
tensor_parallel_norm = _calc_lp(tensor_parallel_grads, norm_type)
|
|
|
|
no_tensor_parallel_norm = _calc_lp(no_tensor_parallel_grads, norm_type)
|
|
|
|
no_tensor_parallel_norm = _calc_lp(no_tensor_parallel_grads, norm_type)
|
|
|
|
moe_parallel_norm = _calc_lp(moe_parallel_grads, norm_type)
|
|
|
|
|
|
|
|
zero_sharded_norm = _calc_lp(zero_sharded_grads, norm_type)
|
|
|
|
zero_sharded_norm = _calc_lp(zero_sharded_grads, norm_type)
|
|
|
|
|
|
|
|
|
|
|
|
# If grads are on CPU, the norms is also on CPU. Cast them to CUDA tensors
|
|
|
|
# If grads are on CPU, the norms is also on CPU. Cast them to CUDA tensors
|
|
|
|
if not enable_cuda_kernels:
|
|
|
|
if not enable_cuda_kernels:
|
|
|
|
tensor_parallel_norm = _move_norm_to_cuda(tensor_parallel_norm)
|
|
|
|
tensor_parallel_norm = _move_norm_to_cuda(tensor_parallel_norm)
|
|
|
|
no_tensor_parallel_norm = _move_norm_to_cuda(no_tensor_parallel_norm)
|
|
|
|
no_tensor_parallel_norm = _move_norm_to_cuda(no_tensor_parallel_norm)
|
|
|
|
moe_parallel_norm = _move_norm_to_cuda(moe_parallel_norm)
|
|
|
|
|
|
|
|
zero_sharded_norm = _move_norm_to_cuda(zero_sharded_norm)
|
|
|
|
zero_sharded_norm = _move_norm_to_cuda(zero_sharded_norm)
|
|
|
|
|
|
|
|
|
|
|
|
# Sum across all model-parallel GPUs.
|
|
|
|
# Sum across all model-parallel GPUs.
|
|
|
|
if gpc.is_initialized(ParallelMode.TENSOR) and len(tensor_parallel_grads) > 0:
|
|
|
|
if gpc.is_initialized(ParallelMode.TENSOR) and len(tensor_parallel_grads) > 0:
|
|
|
|
dist.all_reduce(tensor_parallel_norm, op=dist.ReduceOp.SUM, group=gpc.get_group(ParallelMode.TENSOR))
|
|
|
|
dist.all_reduce(tensor_parallel_norm, op=dist.ReduceOp.SUM, group=gpc.get_group(ParallelMode.TENSOR))
|
|
|
|
# Sum across all moe-tensor-parallel GPUs
|
|
|
|
|
|
|
|
if len(moe_parallel_grads) > 0:
|
|
|
|
|
|
|
|
dist.all_reduce(moe_parallel_norm, group=gpc.get_group(ParallelMode.MOE_MODEL))
|
|
|
|
|
|
|
|
no_tensor_parallel_norm += moe_parallel_norm
|
|
|
|
|
|
|
|
# Sum across all zero sharded GPUs
|
|
|
|
# Sum across all zero sharded GPUs
|
|
|
|
if len(zero_sharded_grads) > 0:
|
|
|
|
if len(zero_sharded_grads) > 0:
|
|
|
|
dist.all_reduce(zero_sharded_norm, group=gpc.get_group(ParallelMode.DATA))
|
|
|
|
dist.all_reduce(zero_sharded_norm, group=gpc.get_group(ParallelMode.DATA))
|
|
|
|