[Feat] Shared expert dp for deepseek and deepseek_mtp (#3495)

### What this PR does / why we need it?
shared expert dp for deepseek and deepseek_mtp, could be combined with
sp to improve performance.

### How was this patch tested?

- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

---------

Signed-off-by: zhaozx-cn <zhaozx2116@163.com>
Co-authored-by: realliujiaxu <realliujiaxu@163.com>

vllm_ascend/attention/mla_v1.py

@@ -1245,7 +1245,8 @@ class AscendMLAImpl(MLAAttentionImpl):
             current_ms_metadata = get_multistream_comm_context()
             if current_ms_metadata is not None:
                 with torch.npu.stream(current_ms_metadata.comm_stream):
-                    o_proj_input[num_decode_tokens:] = output_prefill
+                    o_proj_input[
+                        num_decode_tokens:num_actual_tokens] = output_prefill
                     current_ms_metadata.after_comm_event.record()
             else:
                 o_proj_input[

vllm_ascend/models/deepseek_mtp.py

@@ -88,6 +88,8 @@ class CustomDeepSeekMultiTokenPredictorLayer(DeepSeekMultiTokenPredictorLayer):
         spec_step_index: int = 0,
     ) -> torch.Tensor:
         assert inputs_embeds is not None
+        inputs_embeds = torch.ops.vllm.maybe_all_gather_and_maybe_unpad(
+            inputs_embeds, True)
         # masking inputs at position 0, as not needed by MTP
         inputs_embeds = torch.where((positions == 0).unsqueeze(-1),
                                     torch.zeros_like(inputs_embeds),
@@ -200,4 +202,6 @@ class CustomDeepSeekMTP(DeepSeekMTP):
         hidden_states = self.model(input_ids, positions, kv_caches,
                                    attn_metadata, previous_hidden_states,
                                    inputs_embeds, spec_step_idx)
+        hidden_states = torch.ops.vllm.maybe_all_gather_and_maybe_unpad(
+            hidden_states, True)
         return hidden_states

vllm_ascend/models/layers/mla.py

@@ -122,8 +122,17 @@ class AscendMultiHeadLatentAttention(MultiHeadLatentAttention):
             hidden_states: torch.Tensor,
             kv_cache: Optional[torch.Tensor] = None,
             attn_metadata: Optional[AttentionMetadata] = None) -> torch.Tensor:
-        need_gather_q_kv = get_forward_context().sp_enabled
-        output_shape = hidden_states.shape
+        forward_context = get_forward_context()
+        sp_enabled = forward_context.sp_enabled
+        need_gather_q_kv = False
+        if sp_enabled and self.debug_layer_idx < self.layers:
+            need_gather_q_kv = True
+        if not sp_enabled or self.debug_layer_idx < self.layers:
+            output_shape = hidden_states.shape
+        else:
+            # used in deepseek mtp layer
+            output_shape = torch.chunk(hidden_states, self.tp_size,
+                                       dim=0)[0].shape
         # FIXME: This does not seem right, should make sure the buffer is fixed
         output = torch.empty(output_shape,
                              dtype=hidden_states.dtype,

vllm_ascend/ops/layernorm.py

@@ -99,6 +99,7 @@ class AscendRMSNorm(RMSNorm):
         import torch_npu
 
         if residual is not None:
+            residual = torch.ops.vllm.maybe_chunk_residual(x, residual)
             assert x.size(0) == residual.size(0)
             x, residual = _addrmsnorm_forward_oot(
                 self, x, residual, self.next_need_quant_fusion_linear,

vllm_ascend/ops/register_custom_ops.py

@@ -2,6 +2,7 @@ import torch
 import torch.nn.functional as F
 import torch_npu
 from vllm.distributed import (get_dp_group, get_ep_group,
+                              get_tensor_model_parallel_rank,
                               get_tensor_model_parallel_world_size,
                               tensor_model_parallel_all_gather,
                               tensor_model_parallel_all_reduce,
@@ -15,6 +16,27 @@ from vllm_ascend.ops.weight_prefetch import maybe_npu_prefetch
 from vllm_ascend.utils import npu_stream_switch, prefetch_stream
 
 
+def _maybe_chunk_residual_impl(x: torch.Tensor,
+                               residual: torch.Tensor) -> torch.Tensor:
+    try:
+        forward_context = get_forward_context()
+    except AssertionError:
+        return residual
+
+    if x.size(0) != residual.size(0):
+        sp_enabled = forward_context.sp_enabled
+        assert sp_enabled is True, ("Currently, this situation only occurs "
+                                    "when sp is enabled")
+        pad_size = forward_context.pad_size
+        if pad_size > 0:
+            residual = F.pad(residual, (0, 0, 0, pad_size))
+        tp_size = get_tensor_model_parallel_world_size()
+        tp_rank = get_tensor_model_parallel_rank()
+        residual = torch.chunk(residual, tp_size, dim=0)[tp_rank]
+
+    return residual
+
+
 def _maybe_all_gather_and_maybe_unpad_impl(
         x: torch.Tensor,
         label: bool,
@@ -235,6 +257,11 @@ def _maybe_all_reduce_tensor_model_parallel_impl(
         return tensor_model_parallel_all_reduce(final_hidden_states)
 
 
+direct_register_custom_op(op_name="maybe_chunk_residual",
+                          op_func=_maybe_chunk_residual_impl,
+                          fake_impl=lambda x, residual: x,
+                          mutates_args=[],
+                          dispatch_key="PrivateUse1")
 direct_register_custom_op(op_name="maybe_all_gather_and_maybe_unpad",
                           op_func=_maybe_all_gather_and_maybe_unpad_impl,
                           fake_impl=_maybe_all_gather_and_maybe_unpad_fake,

vllm_ascend/platform.py

@@ -273,7 +273,7 @@ class NPUPlatform(Platform):
         if parallel_config and parallel_config.worker_cls == "auto":
             # TODO: this is a tricky way to disable `use_sequence_parallel_moe` in vllm.
             os.environ["VLLM_ALL2ALL_BACKEND"] = "flashinfer_all2allv"
-            if ascend_config.torchair_graph_config.enabled or ascend_config.enable_shared_expert_dp:
+            if ascend_config.torchair_graph_config.enabled:
                 parallel_config.worker_cls = "vllm_ascend.torchair.torchair_worker.NPUTorchairWorker"
             else:
                 parallel_config.worker_cls = "vllm_ascend.worker.worker_v1.NPUWorker"
@@ -320,8 +320,6 @@ class NPUPlatform(Platform):
         ascend_config = get_ascend_config()
 
         if use_mla and ascend_config.enable_shared_expert_dp:
-            if use_mla and not use_sparse:
-                return "vllm_ascend.torchair.torchair_mla.AscendMLATorchairBackend"
             if use_mla and use_sparse:
                 return "vllm_ascend.torchair.torchair_sfa.AscendSFATorchairBackend"
 

vllm_ascend/spec_decode/mtp_proposer.py

@@ -82,9 +82,7 @@ class MtpProposer(Proposer):
         with set_default_torch_dtype(
                 draft_model_config.dtype), set_current_vllm_config(
                     self.vllm_config):
-            if self.torchair_graph_enabled or (
-                    self.enable_shared_expert_dp
-                    and self.vllm_config.model_config.use_mla):
+            if self.torchair_graph_enabled:
                 self.model = TorchairDeepSeekMTP(
                     vllm_config=self.vllm_config).to(target_device)
             else: