frameworks/slime/docker/npu_patch/mindspeed.patch

diff --git a/mindspeed/core/fusions/fused_rope.py b/mindspeed/core/fusions/fused_rope.py
index a6f02e07..70f7cb08 100644
--- a/mindspeed/core/fusions/fused_rope.py
+++ b/mindspeed/core/fusions/fused_rope.py
@@ -126,5 +126,6 @@ def apply_rotary_pos_emb(
         freqs,
         rotary_interleaved=config.rotary_interleaved,
         multi_latent_attention=config.multi_latent_attention,
-        mscale=mscale
+        mscale=mscale,
+        cp_group=cp_group
     )
diff --git a/mindspeed/megatron_adaptor.py b/mindspeed/megatron_adaptor.py
index f14b231d..4615590e 100644
--- a/mindspeed/megatron_adaptor.py
+++ b/mindspeed/megatron_adaptor.py
@@ -57,6 +57,7 @@ def delete_lock_file():
 def repatch(args):
     MindSpeedFeaturesManager.remove_patches()
     full_args = get_full_args()
+    args = vars(args)
     for k, v in args.items():
         setattr(full_args, k, v)
     MindSpeedFeaturesManager.apply_features_pre_patches(full_args)
diff --git a/mindspeed/te/pytorch/attention/dot_product_attention/dot_product_attention.py b/mindspeed/te/pytorch/attention/dot_product_attention/dot_product_attention.py
index ac4eabe5..78c5866d 100644
--- a/mindspeed/te/pytorch/attention/dot_product_attention/dot_product_attention.py
+++ b/mindspeed/te/pytorch/attention/dot_product_attention/dot_product_attention.py
@@ -330,6 +330,8 @@ class DotProductAttention(torch.nn.Module):
         inference_params: Any = None,
         pad_between_seqs: Optional[bool] = None,
         fp8_output: Optional[bool] = False,
+        local_cp_size=None,
+        cp_group=None,
     ) -> torch.Tensor:
         """
         Dot Product Attention Layer.
@@ -659,7 +661,9 @@ class MindSpeedTEDotProductAttention(DotProductAttention):
         ) and not getattr(self.config, 'is_llava', False):
             self.config.sparse_mode = 2
             attention_mask = get_attention_mask(self.config)
-
+        attention_mask = torch.triu(
+            torch.ones((2048, 2048),
+                       device=query.device, dtype=torch.bool), diagonal=1)
         packed_seq_kwargs = (
             {key: getattr(packed_seq_params, key) for key in self.kept_packed_seq_params}
             if packed_seq_params is not None