[Fix] fix SP for InternS1 VL RL

[tina-wen]

Root Cause

Under sequence parallelism, entropy statistics and GRPO batch loss calibration were computed from sharded tensors, which could produce incorrect token counts and inconsistent metrics across SP ranks. In addition, GRPO loss batching needed to accept the forwarded SP context from the worker.

Fix

Gather shifted labels and logprobs before entropy aggregation under SP, pass sp_mesh into RL loss batching from the training worker, and make GRPO batch construction use the SP-aware token counting path.

[CyCle1024]

each sp rank should calculate its own rank loss, this change makes all ranks in a sp_group caculate the same loss, and it would be all_reduced sum later, thus the sum_entropy would times sp_size

[CyCle1024]

All sp ranks in one sp group get the same rank_grad_tokens. Same problem above.

[CyCle1024]

@tina-wen reference code block:

xtuner/xtuner/v1/loss/ce_loss.py

Lines 141 to 161 in fad9de6

	if sp_mesh is not None:
	# gather shifted_labels from different sp ranks to compute the correct loss weight
	shifted_labels = sp_gather(shifted_labels, sp_mesh=sp_mesh, dim=1)
	mask = (shifted_labels != loss_cfg.ignore_idx).int()
	num_grad_tokens = torch.zeros_like(boundaries, dtype=torch.int32)
	prev_idx = 0
	for j, boundary in enumerate(boundaries):
	num_grad_tokens[j] = mask[0, prev_idx:boundary].sum()
	prev_idx = boundary
	if loss_cfg.loss_reduction == "sample":
	loss_weight = 1.0 / num_grad_tokens
	elif loss_cfg.loss_reduction == "square":
	loss_weight = 1.0 / torch.sqrt(num_grad_tokens.float())
	else:
	raise NotImplementedError(loss_cfg.loss_reduction)
	loss_weight = loss_weight.repeat_interleave(num_tokens).unsqueeze(0)
	if sp_mesh is not None:
	loss_weight = sp_split(loss_weight, sp_mesh=sp_mesh, split_dim=1, padding_value=0.0)
	shifted_labels = sp_split(shifted_labels, sp_mesh=sp_mesh, split_dim=1, padding_value=-100)

not the same compute logic, but the sp processing.