423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2026-05-05 07:38:55 +08:00
Code Issues Packages Projects Releases Wiki Activity

[v1] fix device_mesh and sp for fsdp2 (#10429)

Browse Source
This commit is contained in:
sunyi0505
2026-04-28 11:20:11 +08:00
committed by GitHub
parent 2f0bef207a
commit 50945ef850
3 changed files with 23 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
src/llamafactory/v1/core/base_trainer.py
View File

@@ -269,26 +269,13 @@ class BaseTrainer:
# deepspeed: engine.step() already ran inside backward at the sync boundary
grad_norm = self._deepspeed_engine.get_grad_norm()
else:
grad_norm = torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.args.max_grad_norm).item()
if self.args.dist_config and self.args.dist_config.get("cp_size", 1) > 1:
from torch.nn.utils.clip_grad import _clip_grads_with_norm_, _get_total_norm
grad_norm = grad_norm**2
grad_norm = DistributedInterface().all_reduce(grad_norm, op=ReduceOp.SUM, dim=Dim.CP)
grad_norm = grad_norm**0.5
parameters = self.model.parameters()
if isinstance(parameters, torch.Tensor):
parameters = [parameters]
else:
parameters = list(parameters)
grads = [p.grad for p in parameters if p.grad is not None]
grad_norm = _get_total_norm(grads)
grad_norm = grad_norm.to(self.device)
_clip_grads_with_norm_(parameters, self.args.max_grad_norm, grad_norm)
if isinstance(grad_norm, torch.distributed._tensor.DTensor):
grad_norm = grad_norm.full_tensor().item()
else:
grad_norm = torch.nn.utils.clip_grad_norm_(
self.model.parameters(), self.args.max_grad_norm
).item()
# isfinite(): argument 'input' (position 1) must be Tensor, not float
if not torch.isfinite(torch.tensor(grad_norm)): # type: ignore # pyright: ignore [reportUnknownReturnType]
logger.warning_rank0(f"Gradient norm is not finite: {grad_norm}")
else:

7
src/llamafactory/v1/plugins/model_plugins/parallelization/sequence_parallel.py
View File

@@ -175,9 +175,9 @@ def sequence_parallel_loss(model, model_inputs):
global_labels = [torch.empty_like(labels) for _ in range(cp_world_size)]
dist.all_gather(global_labels, labels, group=cp_group)
labels = torch.cat(global_labels, dim=1).contiguous()
shift_labels = labels[..., 1:].view(-1).contiguous()
shift_labels = labels[..., 1:].contiguous()
shift_labels = F.pad(shift_labels, (0, 1), value=-100)
shift_labels = torch.chunk(shift_labels, chunks=cp_world_size, dim=-1)[cp_rank].contiguous()
shift_labels = torch.chunk(shift_labels, chunks=cp_world_size, dim=1)[cp_rank].contiguous()
# use all_gather to collect loss_weights from all sequence parallel processes
loss_weights = model_inputs["loss_weights"]
@@ -186,7 +186,8 @@ def sequence_parallel_loss(model, model_inputs):
shift_loss_weights = torch.cat(global_loss_weights, dim=1).contiguous()
shift_loss_weights = shift_loss_weights[..., 1:].contiguous()
shift_logits = logits.view(shift_labels.size(0), -1).contiguous()
shift_logits = logits.view(-1, logits.size(-1)).contiguous()
shift_labels = shift_labels.view(-1).contiguous()
# use all_gather to collect log_probs from all sequence parallel processes
log_probs = -F.cross_entropy(shift_logits, shift_labels, reduction="none").view(batch_size, -1)