diff --git a/python/sglang/srt/layers/attention/flashattention_backend.py b/python/sglang/srt/layers/attention/flashattention_backend.py
index 336ffc6ef..526bb5d1c 100644
--- a/python/sglang/srt/layers/attention/flashattention_backend.py
+++ b/python/sglang/srt/layers/attention/flashattention_backend.py
@@ -995,17 +995,24 @@ class FlashAttentionBackend(AttentionBackend):
                     else forward_batch.encoder_out_cache_loc
                 )
                 if self.use_mla:
-                    # MLA: under CP, k and k_rope arrive full-sequence
-                    # (rebuild_cp_kv_cache ran upstream in
-                    # forward_absorb_prepare); rank-local otherwise.
-                    # out_cache_loc is never zigzag-split, so the write
-                    # lands in the right slots on every rank in either case.
-                    self.token_to_kv_pool.set_mla_kv_buffer(
-                        layer,
-                        cache_loc,
-                        k,
-                        k_rope,
-                    )
+                    if is_cp_v2_active(forward_batch):
+                        # CP-v2: k/k_rope are rank-local; the strategy gathers
+                        # the latent to full sequence and writes it.
+                        cp_strategy = get_cp_strategy()
+                        assert cp_strategy is not None
+                        cp_strategy.materialize_full_mla_kv(
+                            forward_batch, layer, k, k_rope
+                        )
+                    else:
+                        # CP-v1: k/k_rope arrive full-sequence (rebuild_cp_kv_cache
+                        # ran upstream); rank-local when CP is off. out_cache_loc is
+                        # never zigzag-split, so the write lands in the right slots.
+                        self.token_to_kv_pool.set_mla_kv_buffer(
+                            layer,
+                            cache_loc,
+                            k,
+                            k_rope,
+                        )
                 elif is_cp_mode:
                     # Dense-MHA CP: k, v are still rank-local; backend
                     # all-gathers and writes to the per-rank pool.
@@ -1438,9 +1445,20 @@ class FlashAttentionBackend(AttentionBackend):
                             ver=self.fa_impl_ver,
                         )
 
-                    o = cp_attn_forward_extend(
-                        forward_batch, q_fused, self.device, _mla_cp_attn
-                    )
+                    if is_cp_v2_active(forward_batch):
+                        cp_strategy = get_cp_strategy()
+                        assert cp_strategy is not None
+                        o = cp_strategy.run_attention(
+                            q_fused,
+                            forward_batch,
+                            self.device,
+                            _mla_cp_attn,
+                            attention_backend=CPAttentionBackendKind.FLASH_ATTENTION,
+                        )
+                    else:
+                        o = cp_attn_forward_extend(
+                            forward_batch, q_fused, self.device, _mla_cp_attn
+                        )
                 else:
                     result = flash_attn_with_kvcache(
                         q=q_rope,
diff --git a/python/sglang/srt/layers/cp/utils.py b/python/sglang/srt/layers/cp/utils.py
index fbadc0c74..1ba7545ca 100644
--- a/python/sglang/srt/layers/cp/utils.py
+++ b/python/sglang/srt/layers/cp/utils.py
@@ -40,6 +40,7 @@ if TYPE_CHECKING:
 CP_V2_DEFAULT_MODEL_CLASSES = frozenset(
     {
         "Qwen3MoeForCausalLM",
+        "DeepseekV3ForCausalLM",
     }
 )
 
diff --git a/python/sglang/srt/layers/cp/zigzag.py b/python/sglang/srt/layers/cp/zigzag.py
index b53236994..d353f0d79 100644
--- a/python/sglang/srt/layers/cp/zigzag.py
+++ b/python/sglang/srt/layers/cp/zigzag.py
@@ -354,6 +354,21 @@ class ZigzagCPStrategy(ContextParallelStrategy):
             layer.v_scale,
         )
 
+    def materialize_full_mla_kv(
+        self, forward_batch, layer: Any, k_nope: Any, k_rope: Any
+    ) -> None:
+        kv_lora_rank = k_nope.shape[-1]
+        latent = torch.cat([k_nope, k_rope], dim=-1).contiguous()
+        latent_full = self.gather_kv_cache(
+            latent, forward_batch, torch.cuda.current_stream()
+        )
+        get_token_to_kv_pool().set_mla_kv_buffer(
+            layer,
+            forward_batch.out_cache_loc,
+            latent_full[..., :kv_lora_rank],
+            latent_full[..., kv_lora_rank:],
+        )
+
     def _all_gather_reorganized(self, x: torch.Tensor, forward_batch, stream):
         meta = forward_batch.attn_cp_metadata
         max_len = meta.max_rank_len[0]
diff --git a/python/sglang/srt/model_executor/runner/eager_runner.py b/python/sglang/srt/model_executor/runner/eager_runner.py
index f6f4afd26..a7a76c2ed 100644
--- a/python/sglang/srt/model_executor/runner/eager_runner.py
+++ b/python/sglang/srt/model_executor/runner/eager_runner.py
@@ -285,21 +285,7 @@ class EagerRunner(BaseRunner):
             model_runner.attn_backend.init_forward_metadata(forward_batch)
 
         cp_v2_active = is_cp_v2_active(forward_batch)
-        forward_positions = forward_batch.positions
-        if cp_v2_active:
-            prepare_cp_forward(forward_batch)
-            complete_hidden_states = kwargs.get("input_embeds")
-            if complete_hidden_states is None:
-                embed_layer = model_runner.model.get_input_embeddings()
-                complete_hidden_states = embed_layer(forward_batch.input_ids)
-            sharded_hidden_states, sharded_positions = cp_split_before_forward(
-                complete_hidden_states,
-                forward_batch.positions,
-                forward_batch,
-            )
-            kwargs["input_embeds"] = sharded_hidden_states
-            forward_positions = sharded_positions
-        else:
+        if not cp_v2_active:
             forward_batch.attn_cp_metadata = None
 
         category = (
@@ -335,51 +321,67 @@ class EagerRunner(BaseRunner):
                 ):
                     ret = model_runner.model.forward(
                         forward_batch.input_ids,
-                        forward_positions,
+                        forward_batch.positions,
                         forward_batch,
                         **kwargs,
                     )
             elif cp_v2_active:
-                # CP-V2: drive .model directly to gather across CP ranks before logits.
-                hidden_states = model_runner.model.model(
-                    forward_batch.input_ids,
-                    forward_positions,
-                    forward_batch,
-                    input_embeds=kwargs.get("input_embeds"),
-                    pp_proxy_tensors=kwargs.get("pp_proxy_tensors"),
-                )
-                aux_hidden_states = None
-                capture_aux_hidden_states = getattr(
-                    model_runner.model, "capture_aux_hidden_states", False
-                )
-                if capture_aux_hidden_states:
-                    hidden_states, aux_hidden_states = hidden_states
-                if model_runner.model.pp_group.is_last_rank:
-                    hidden_states = cp_gather_after_forward(
-                        hidden_states,
-                        forward_batch,
-                        torch.cuda.current_stream(),
-                    )
-                    ret = model_runner.model.logits_processor(
-                        forward_batch.input_ids,
-                        hidden_states,
-                        model_runner.model.lm_head,
-                        forward_batch,
-                        aux_hidden_states,
-                    )
-                elif capture_aux_hidden_states:
-                    ret = hidden_states, aux_hidden_states
-                else:
-                    ret = hidden_states
+                ret = self._execute_extend_cp_v2(forward_batch, kwargs)
             else:
                 ret = model_runner.model.forward(
                     forward_batch.input_ids,
-                    forward_positions,
+                    forward_batch.positions,
                     forward_batch,
                     **kwargs,
                 )
         return ret
 
+    def _execute_extend_cp_v2(
+        self, forward_batch: ForwardBatch, kwargs: dict
+    ) -> Union[LogitsProcessorOutput, PPProxyTensors]:
+        """CP-v2 extend: shard inputs at the model boundary, run the body on the
+        rank-local slice, then gather hidden states before the logits step.
+        """
+        model = self.model_runner.model
+
+        prepare_cp_forward(forward_batch)
+        input_embeds = kwargs.get("input_embeds")
+        if input_embeds is None:
+            input_embeds = model.get_input_embeddings()(forward_batch.input_ids)
+        input_embeds, positions = cp_split_before_forward(
+            input_embeds, forward_batch.positions, forward_batch
+        )
+
+        hidden_states = model.model(
+            forward_batch.input_ids,
+            positions,
+            forward_batch,
+            input_embeds=input_embeds,
+            pp_proxy_tensors=kwargs.get("pp_proxy_tensors"),
+        )
+        capture_aux_hidden_states = getattr(model, "capture_aux_hidden_states", False)
+        aux_hidden_states = None
+        if capture_aux_hidden_states:
+            hidden_states, aux_hidden_states = hidden_states
+
+        if not model.pp_group.is_last_rank:
+            return (
+                (hidden_states, aux_hidden_states)
+                if capture_aux_hidden_states
+                else hidden_states
+            )
+
+        hidden_states = cp_gather_after_forward(
+            hidden_states, forward_batch, torch.cuda.current_stream()
+        )
+        return model.logits_processor(
+            forward_batch.input_ids,
+            hidden_states,
+            model.lm_head,
+            forward_batch,
+            aux_hidden_states,
+        )
+
     def _execute_idle(
         self, forward_batch: ForwardBatch, pp_proxy_tensors=None
     ) -> Union[LogitsProcessorOutput, PPProxyTensors]:
diff --git a/python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mla.py b/python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mla.py
index c3ed46742..c399cf39f 100644
--- a/python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mla.py
+++ b/python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mla.py
@@ -19,6 +19,7 @@ from sglang.srt.layers.attention.dsa.utils import (
     is_graph_dsa_split_op_surface,
 )
 from sglang.srt.layers.communicator import get_attn_tp_context
+from sglang.srt.layers.cp.utils import is_cp_v2_active
 from sglang.srt.layers.dcp import (
     all_gather_kv_cache_for_mla_extend,
     all_gather_q_for_mla_decode,
@@ -565,8 +566,13 @@ class DeepseekMLAForwardMixin:
             dsa_prefill_cp=dsa_prefill_cp,
             fuse_rope_for_trtllm_mla=fuse_rope_for_trtllm_mla,
         )
-        if (dsa_prefill_cp or mla_prefill_cp) and not defer_kv_gather_until_after_rope:
-            # support allgather+rerrange
+        if (
+            (dsa_prefill_cp or mla_prefill_cp)
+            and not defer_kv_gather_until_after_rope
+            and not is_cp_v2_active(forward_batch)
+        ):
+            # CP-v1 gathers the latent here; CP-v2 gathers it in the attention
+            # backend via the strategy (materialize_full_mla_kv).
             k_nope, k_pe = self.rebuild_cp_kv_cache(
                 latent_cache, forward_batch, k_nope, k_pe
             )
diff --git a/python/sglang/srt/models/deepseek_nextn.py b/python/sglang/srt/models/deepseek_nextn.py
index 35fcedeea..222effb03 100644
--- a/python/sglang/srt/models/deepseek_nextn.py
+++ b/python/sglang/srt/models/deepseek_nextn.py
@@ -35,6 +35,7 @@ from sglang.srt.layers.attention.dsa.utils import (
     is_dsa_enable_prefill_cp,
     is_dsa_prefill_cp_round_robin_split,
 )
+from sglang.srt.layers.cp.utils import is_cp_v2_active
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.linear import ReplicatedLinear
 from sglang.srt.layers.logits_processor import LogitsProcessor
@@ -251,10 +252,12 @@ class DeepseekModelNextN(nn.Module):
                 else:
                     hidden_states = self.eh_proj(eh_input)
 
-            use_cp = dsa_use_prefill_cp(
-                forward_batch, self.dsa_enable_prefill_cp
-            ) or mla_use_prefill_cp(forward_batch, self.mla_enable_prefill_cp)
-            if use_cp:
+            # CP-v2 shards/gathers at the eager-runner boundary instead.
+            use_cp_v1 = (
+                dsa_use_prefill_cp(forward_batch, self.dsa_enable_prefill_cp)
+                or mla_use_prefill_cp(forward_batch, self.mla_enable_prefill_cp)
+            ) and not is_cp_v2_active(forward_batch)
+            if use_cp_v1:
                 hidden_states = cp_split_and_rebuild_data(forward_batch, hidden_states)
                 positions = cp_split_and_rebuild_position(forward_batch, positions)
             residual = None
@@ -285,7 +288,7 @@ class DeepseekModelNextN(nn.Module):
                 else:
                     hidden_states = self.shared_head.norm(hidden_states)
 
-                if use_cp:
+                if use_cp_v1:
                     local_num_tokens = hidden_states.shape[0]
                     hidden_states = cp_all_gather_rerange_output(
                         hidden_states,
diff --git a/python/sglang/srt/models/deepseek_v2.py b/python/sglang/srt/models/deepseek_v2.py
index a086013be..a406c0823 100644
--- a/python/sglang/srt/models/deepseek_v2.py
+++ b/python/sglang/srt/models/deepseek_v2.py
@@ -77,6 +77,7 @@ from sglang.srt.layers.communicator_dsa_cp import (
     DSACPLayerCommunicator,
     maybe_prefetch_next_full_attention_kv,
 )
+from sglang.srt.layers.cp.utils import is_cp_v2_active
 from sglang.srt.layers.dcp.planner import (
     prepare_decode_context_parallel_metadata,
 )
@@ -2554,9 +2555,13 @@ class DeepseekV2Model(nn.Module):
             else None
         )
 
-        if dsa_use_prefill_cp(
-            forward_batch, self.dsa_enable_prefill_cp
-        ) or mla_use_prefill_cp(forward_batch, self.mla_enable_prefill_cp):
+        # CP-v2 shards/gathers at the eager-runner boundary instead.
+        use_cp_v1 = (
+            dsa_use_prefill_cp(forward_batch, self.dsa_enable_prefill_cp)
+            or mla_use_prefill_cp(forward_batch, self.mla_enable_prefill_cp)
+        ) and not is_cp_v2_active(forward_batch)
+
+        if use_cp_v1:
             if self.pp_group.is_first_rank:
                 hidden_states = cp_split_and_rebuild_data(forward_batch, hidden_states)
             positions = cp_split_and_rebuild_position(forward_batch, positions)
@@ -2659,10 +2664,7 @@ class DeepseekV2Model(nn.Module):
                 else:
                     hidden_states, _ = self.norm(hidden_states, residual)
 
-        if self.pp_group.is_last_rank and (
-            dsa_use_prefill_cp(forward_batch, self.dsa_enable_prefill_cp)
-            or mla_use_prefill_cp(forward_batch, self.mla_enable_prefill_cp)
-        ):
+        if self.pp_group.is_last_rank and use_cp_v1:
             # allgather + rerrange
             hidden_states = cp_all_gather_rerange_output(
                 hidden_states,