[docs] update megatron docs

docs/source/Megatron-SWIFT/Command-line-parameters.md

@@ -192,7 +192,8 @@
 - moe_enable_deepep: 启用 DeepEP 以实现 MoE 模型中的高效 token 调度和合并。仅在通过设置 `--moe_token_dispatcher_type flex` 使用弹性 token 调度器时有效。
 - 🔥moe_grouped_gemm: 当每个rank包含多个专家时，通过在多个流中启动多个本地 GEMM 内核，利用 TransformerEngine中的GroupedLinear提高利用率和性能。默认为True。
 - 🔥moe_permute_fusion: 在令牌分发过程中融合令牌重排操作。默认为False。
-- 🔥moe_aux_loss_coeff: 默认为0，不使用aux_loss。**通常情况下，该值设置的越大，训练效果越差，但MoE负载越均衡**，请根据实验效果，选择合适的值。
+- 🔥moe_aux_loss_coeff: 默认为0，不使用aux_loss。通常情况下，该值设置的越大，训练效果越差，但MoE负载越均衡，请根据实验效果，选择合适的值。
+  - 注意：moe_aux_loss在`padding_free`为 False 的情况下，"megatron-core<0.16"存在对padding token计算routing loss的情况，参考[这个PR](https://github.com/NVIDIA/Megatron-LM/pull/2142)。此外请使用"mcore-bridge>=1.4.0.dev"，参考[这个PR](https://github.com/modelscope/mcore-bridge/pull/79)。
 - moe_z_loss_coeff: z-loss 的缩放系数。默认为None。
 - 🔥moe_shared_expert_overlap: 启用共享专家计算与调度器通信之间的重叠。如果不启用此选项，共享专家将在路由专家之后执行。仅在设置了`moe_shared_expert_intermediate_size`时有效。默认为False。
 - 🔥moe_expert_capacity_factor: 每个专家的容量因子，None表示不会丢弃任何token。默认为None。通过设置 `--moe_expert_capacity_factor`，超出专家容量的 token 会基于其被选中的概率被丢弃。可以**令训练负载均匀，提升训练速度**（例如设置为1或2）。

docs/source_en/Megatron-SWIFT/Command-line-parameters.md

@@ -203,7 +203,8 @@ For guidance on selecting parallelization strategies, please refer to the [Train
 - moe_enable_deepep: Enable DeepEP for efficient token dispatching and combine in MoE models. Only works with flex token dispatcher by setting `--moe_token_dispatcher_type flex`.
 - 🔥moe_grouped_gemm: When each rank contains multiple experts, multiple local GEMM kernels can be launched in parallel streams to improve utilization and performance by using GroupedLinear from TransformerEngine. Default is True.
 - 🔥moe_permute_fusion: Fuses token permutation operations during token dispatch. Default is False.
-- 🔥moe_aux_loss_coeff: Defaults to 0, meaning the auxiliary loss is not used. **Generally, a higher value leads to worse training performance but more balanced MoE expert utilization.** Please choose an appropriate value based on experimental results.
+- 🔥moe_aux_loss_coeff: Defaults to 0, meaning the auxiliary loss is not used. Generally, a higher value leads to worse training performance but more balanced MoE expert utilization. Please choose an appropriate value based on experimental results.
+  - Note: For `moe_aux_loss`, when `padding_free` is set to False, versions of `megatron-core < 0.16` have an issue where routing loss is computed on padding tokens. Please refer to [this PR](https://github.com/NVIDIA/Megatron-LM/pull/2142). Additionally, please use `mcore-bridge >= 1.4.0.dev`. Please refer to [this PR](https://github.com/modelscope/mcore-bridge/pull/79).
 - moe_z_loss_coeff: Scaling coefficient for z-loss. Default is None.
 - 🔥moe_shared_expert_overlap: Enables overlap between shared expert computation and the dispatcher. If not enabled, shared expert computation will be performed after routing experts. Only effective when `moe_shared_expert_intermediate_size` is set. Default is False.
 - 🔥moe_expert_capacity_factor: Capacity factor for each expert. `None` means no tokens will be dropped. Default is `None`. When `--moe_expert_capacity_factor` is set, tokens exceeding an expert’s capacity will be dropped based on their selection probability. This can **balance the training load and improve training speed** (for example, set it to 1. or 2.).

swift/megatron/arguments/megatron_args.py

@@ -623,7 +623,9 @@ def _set_default(self):
                     'mcore-bridge>=1.3.0.dev',
                     'Please install mcore-bridge>=1.3.0.dev to use mlp_padding_free with sequence parallel.')
             if self.context_parallel_size > 1:
-                raise ValueError('mlp_padding_free is not compatible with context parallel.')
+                require_version(
+                    'mcore-bridge>=1.4.0.dev',
+                    'Please install mcore-bridge>=1.4.0.dev to use mlp_padding_free with context parallel.')
         if self.local_rank is None:
             self.local_rank = get_dist_setting()[1]
         if self.lr is None: