feat: support data parallel for deepseek (vllm-project#1012)

### What this PR does / why we need it?
feat: support data parallel for deepseek

### Does this PR introduce _any_ user-facing change?
Yes, support dp for deepseek

### How was this patch tested?

```
export VLLM_ENABLE_MC2=0
export VLLM_USE_V1=1
export TASK_QUEUE_ENABLE=1

source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh

nohup python -m vllm.entrypoints.openai.api_server
--model=/path/to/DeepSeek-R1-W8A8 \
    --quantization ascend \
    --served-model-name auto \
    --trust-remote-code \
    --distributed-executor-backend=mp \
    --port 8006 \
    -tp=8 \
    -dp=2 \
    --max-num-seqs 24 \
    --max-model-len 4096 \
    --max-num-batched-tokens 4096 \
    --block-size 128 \
    -O 0 \
    --no-enable-prefix-caching \
--additional-config
'{"torchair_graph_batch_sizes":[24],"expert_tensor_parallel_size":16,"ascend_scheduler_config":{},"enable_graph_mode":true}'
\
    --gpu-memory-utilization 0.95 &> run.log &
disown
```

Signed-off-by: boying <897013703@qq.com>

Author: NeverRaR

vllm_ascend/attention/mla_v1.py

@@ -117,6 +117,8 @@ class AscendMLAMetadata:
     # For logging.
     num_input_tokens: int = 0  # Number of tokens including padding.
 
+    with_prefill_across_dp: bool = False
+
     # The dimension of the attention heads
     head_dim: Optional[int] = None
     attn_mask: torch.Tensor = None
@@ -260,6 +262,10 @@ def build_dummy(self, num_reqs: int,
                                   PAD_SLOT_ID,
                                   dtype=torch.int32,
                                   device=device)
+        query_start_loc = torch.full((num_reqs, ),
+                                     -1,
+                                     dtype=torch.int32,
+                                     device=device)
         decode_metadata = AscendMLADecodeMetadata(
             input_positions=input_positions,
             block_table=block_table,
@@ -278,15 +284,21 @@ def build_dummy(self, num_reqs: int,
             attn_state=AscendAttentionState.DecodeOnly,
             prefill=None,
             decode=decode_metadata,
+            query_start_loc=query_start_loc,
+            seq_lens=seq_lens,
+            block_tables=block_table,
         )
 
-    def build(self,
-              num_reqs: int,
-              num_actual_tokens: int,
-              max_query_len: int,
-              common_attn_metadata: CommonAttentionMetadata,
-              common_prefix_len: Optional[int] = None,
-              graph_pad_size: int = -1) -> AscendMLAMetadata:
+    def build(
+        self,
+        num_reqs: int,
+        num_actual_tokens: int,
+        max_query_len: int,
+        common_attn_metadata: CommonAttentionMetadata,
+        common_prefix_len: Optional[int] = None,
+        graph_pad_size: int = -1,
+        with_prefill_across_dp: bool = False,
+    ) -> AscendMLAMetadata:
         assert self._num_decodes + self._num_prefills == num_reqs
 
         # Note(simon): be careful about the CPU <> GPU memory movement in this
@@ -388,6 +400,7 @@ def build(self,
             query_start_loc=query_start_loc,
             block_tables=block_table,
             seq_lens=seq_lens,
+            with_prefill_across_dp=with_prefill_across_dp,
         )
 
 
@@ -621,7 +634,7 @@ def exec_kv(
         kv = self.kv_a_proj_with_mqa(hidden_states)[0]
         # npu_kv_rmsnorm_rope_cache needs [B, N, S, D]
         kv = kv.view(B, N, S, self.kv_lora_rank + self.qk_rope_head_dim)
-        k_pe, k_nope, _, _ = torch.ops.npu_inference.npu_kv_rmsnorm_rope_cache(
+        k_pe, k_nope, _, _ = torch_npu.npu_kv_rmsnorm_rope_cache(
             kv,
             self.kv_a_layernorm.weight,
             cos,
@@ -643,7 +656,7 @@ def rope_single(
         B, N, D = x.shape
         S = 1
         x = x.view(B, N, S, D)
-        x = torch.ops.npu_inference.npu_interleave_rope(x, cos, sin)
+        x = torch_npu.npu_interleave_rope(x, cos, sin)
         return x.view(B, N, D)
 
     def _forward_decode(
@@ -766,6 +779,7 @@ def forward(
                 sin = sin[attn_metadata.decode.input_positions]
                 cos = cos[:, None, None, :]
                 sin = sin[:, None, None, :]
+
                 decode_q_pe = self.rope_single(decode_q_pe, cos, sin)
                 decode_k_pe, decode_k_nope = self.exec_kv(
                     hidden_states_or_kv_c_normed, cos, sin, kv_cache,

vllm_ascend/models/deepseek_v2.py

@@ -212,6 +212,14 @@ def __init__(
         self.tp_group = get_tp_group().device_group
         self.tp_rank = get_tp_group().rank_in_group
 
+        self.params_dtype = torch.get_default_dtype()
+
+        self.enable_graph_mode = False
+        additional_config = get_current_vllm_config().additional_config
+        if additional_config:
+            self.enable_graph_mode = additional_config.get(
+                "enable_graph_mode", False)
+
     def forward(
             self,
             hidden_states: torch.Tensor,
@@ -228,52 +236,65 @@ def forward(
         else:
             is_prefill = attn_metadata.num_prefills > 0
             enable_force_load_balance = False
-        num_tokens, hidden_dim = hidden_states.shape
+            if hasattr(attn_metadata, 'with_prefill_across_dp'):
+                is_prefill = is_prefill or attn_metadata.with_prefill_across_dp
+
+        num_tokens, hidden_size = hidden_states.shape
 
         if self.n_shared_experts is not None:
             shared_output = self.shared_experts(hidden_states)
 
         if self.tp_size > 1:
-            # pass
-            num_tokens, hidden_size = hidden_states.shape
-            if num_tokens < self.tp_size:
-                target_size = self.tp_size
-                new_hidden_states = torch.empty([target_size, hidden_size],
-                                                dtype=hidden_states.dtype,
-                                                device=hidden_states.device)
-                new_hidden_states[:num_tokens] = hidden_states
-                hidden_states = new_hidden_states
-            chunk_hidden_states = torch.tensor_split(hidden_states,
-                                                     self.tp_size,
-                                                     dim=0)
-            local_hidden_states = chunk_hidden_states[self.tp_rank]
-        else:
-            local_hidden_states = hidden_states
+            if envs_ascend.VLLM_ENABLE_MC2 and not is_prefill:
+                chunks = torch.chunk(hidden_states, self.tp_size, dim=0)
+                hidden_states = chunks[self.tp_rank]
+            elif not self.enable_graph_mode:
+                num_padding_tokens = (self.tp_size -
+                                      num_tokens % self.tp_size) % self.tp_size
+                # Pad hidden_states to make it divisible by tp_size to avoid cross-ring AllGatherV on 910B2C
+                if num_padding_tokens > 0:
+                    hidden_states = nn.functional.pad(
+                        hidden_states, (0, 0, 0, num_padding_tokens))
+                chunk_hidden_states = torch.tensor_split(hidden_states,
+                                                         self.tp_size,
+                                                         dim=0)
+                hidden_states = chunk_hidden_states[self.tp_rank]
 
         # router_logits: (num_tokens, n_experts)
-        router_logits, _ = self.gate(local_hidden_states)
+        router_logits, _ = self.gate(hidden_states)
 
-        router_hidden_states = self.experts(
-            hidden_states=local_hidden_states,
+        hidden_states = self.experts(
+            hidden_states=hidden_states,
             router_logits=router_logits,
             is_prefill=is_prefill,
             top_k=CustomDeepseekV2MoE.top_k,
             enable_force_load_balance=enable_force_load_balance,
         ) * self.routed_scaling_factor
 
         if self.tp_size > 1:
-            dist.all_gather(list(chunk_hidden_states), router_hidden_states,
-                            self.tp_group)
-            final_hidden_states = torch.cat(chunk_hidden_states, dim=0)
-            if num_tokens < self.tp_size:
-                final_hidden_states = final_hidden_states[:num_tokens]
-        else:
-            final_hidden_states = router_hidden_states
+            if self.enable_graph_mode:
+                if envs_ascend.VLLM_ENABLE_MC2 and not is_prefill:
+                    final_hidden_states = torch.zeros(
+                        [num_tokens, hidden_size],
+                        dtype=self.params_dtype,
+                        device="npu")
+                    dist.all_gather_into_tensor(final_hidden_states,
+                                                hidden_states, self.tp_group)
+                    hidden_states = final_hidden_states
+                else:
+                    hidden_states = tensor_model_parallel_all_reduce(
+                        hidden_states)
+            else:
+                dist.all_gather(list(chunk_hidden_states), hidden_states,
+                                self.tp_group)
+                hidden_states = torch.cat(chunk_hidden_states, dim=0)
+                if num_padding_tokens > 0:
+                    hidden_states = hidden_states[:-num_padding_tokens]
 
         if shared_output is not None:
-            final_hidden_states = final_hidden_states + shared_output
+            hidden_states = hidden_states + shared_output
 
-        return final_hidden_states.view(num_tokens, hidden_dim)
+        return hidden_states.view(num_tokens, hidden_size)
 
 
 class CustomDeepseekV2MLAAttention(DeepseekV2MLAAttention):

vllm_ascend/ops/fused_moe.py

@@ -587,6 +587,12 @@ def __init__(self, moe: MoEConfig = None):
         self.global_batch_size = vllm_config.scheduler_config.max_num_seqs
         self.local_batch_size = self.global_batch_size // self.ep_size
 
+        self.enable_graph_mode = False
+        additional_config = get_current_vllm_config().additional_config
+        if additional_config:
+            self.enable_graph_mode = additional_config.get(
+                "enable_graph_mode", False)
+
         try:
             device_group = ep_group.device_group
             # TODO: Try local_rank = ep_group.rank_in_group
@@ -664,7 +670,7 @@ def apply(
                 top_k=top_k,
                 expert_map=expert_map,
                 moe_all_to_all_group_name=self.moe_all_to_all_group_name)
-        elif get_ep_group().world_size == 1:
+        elif self.enable_graph_mode or get_ep_group().world_size == 1:
             return fused_experts(hidden_states=x,
                                  w1=layer.w13_weight,
                                  w2=layer.w2_weight,
@@ -750,26 +756,20 @@ def __init__(
         self.expert_map = None
         self.activation = activation
 
-        if self.ep_size > 1:
-            # Create a tensor of size num_experts filled with -1
-            self.local_num_experts, self.expert_map = determine_expert_map(
-                self.ep_size,
-                get_ep_group().rank_in_group, self.global_num_experts)
-
-            self.moe_parallel_config.tp_rank = get_etp_group().rank_in_group
-            self.moe_parallel_config.ep_rank = get_ep_group().rank_in_group
+        # Create a tensor of size num_experts filled with -1
+        self.local_num_experts, self.expert_map = determine_expert_map(
+            self.ep_size,
+            get_ep_group().rank_in_group, self.global_num_experts)
 
-        else:
-            # Adjust TP size for DP attention
-            # haven't test its functionality yet, may remove in the future
+        self.moe_parallel_config.tp_rank = get_etp_group().rank_in_group
+        self.moe_parallel_config.ep_rank = get_ep_group().rank_in_group
 
-            self.moe_parallel_config.tp_rank = self.tp_size * self.dp_rank
-            self.moe_parallel_config.ep_rank = 0
-            self.moe_parallel_config.tp_size = self.tp_size * self.dp_size
-            self.moe_parallel_config.ep_size = 1
+        self.enable_graph_mode = False
+        additional_config = get_current_vllm_config().additional_config
+        if additional_config:
+            self.enable_graph_mode = additional_config.get(
+                "enable_graph_mode", False)
 
-            self.local_num_experts, self.expert_map = (self.global_num_experts,
-                                                       None)
         if self.scoring_func != "softmax" and not self.use_grouped_topk:
             raise ValueError("Only softmax scoring function is supported for "
                              "non-grouped topk.")
@@ -807,8 +807,15 @@ def __init__(
                 in ("GPTQMarlinMoEMethod", "CompressedTensorsWNA16MoEMethod")):
             moe_quant_params["intermediate_size_full"] = intermediate_size
 
+        self.ep_group = get_ep_group()
         self.quant_method.create_weights(layer=self, **moe_quant_params)
 
+        self.enable_graph_mode = False
+        additional_config = get_current_vllm_config().additional_config
+        if additional_config:
+            self.enable_graph_mode = additional_config.get(
+                "enable_graph_mode", False)
+
     def forward(self,
                 hidden_states: torch.Tensor,
                 router_logits: torch.Tensor,
@@ -822,11 +829,28 @@ def forward(self,
         else:
             real_top_k = self.top_k
 
-        if VLLM_ENABLE_MC2 and not is_prefill:
-            ...
+        #                MC2   ag/rs  broadcast/all_reduce
+        #  prefill_req     x      x            √
+        #  decode_req     √      x            √
+        #  graph_mode     √      √            x
+        if self.dp_size > 1:
+            if VLLM_ENABLE_MC2 and not is_prefill:
+                ...
+            elif self.enable_graph_mode:
+                if USING_LCCL_COM:  # type: ignore
+                    hidden_states = get_dp_group().all_gather(
+                        hidden_states, 0, False)
+                    router_logits = get_dp_group().all_gather(
+                        router_logits, 0, False)
+                elif self.enable_graph_mode and not is_prefill:
+                    hidden_states = get_dp_group().all_gather(hidden_states, 0)
+                    router_logits = get_dp_group().all_gather(router_logits, 0)
+                else:
+                    hidden_states, router_logits = get_ep_group().dispatch(
+                        hidden_states, router_logits)
 
         # Matrix multiply.
-        final_hidden_states = self.quant_method.apply(
+        hidden_states = self.quant_method.apply(
             layer=self,
             x=hidden_states,
             router_logits=router_logits,
@@ -843,11 +867,26 @@ def forward(self,
             is_prefill=is_prefill,
             enable_force_load_balance=enable_force_load_balance)
 
-        if VLLM_ENABLE_MC2 and not is_prefill:
-            ...
+        if self.dp_size > 1:
+            if VLLM_ENABLE_MC2 and not is_prefill:
+                ...
+            elif self.enable_graph_mode:
+                if USING_LCCL_COM:  # type: ignore
+                    hidden_states = dist._functional_collectives.reduce_scatter_tensor(
+                        hidden_states,
+                        "sum",
+                        scatter_dim=0,
+                        group=get_dp_group().device_group)
+                elif self.enable_graph_mode and not is_prefill:
+                    hidden_states = dist._functional_collectives.reduce_scatter_tensor(
+                        hidden_states,
+                        "sum",
+                        scatter_dim=0,
+                        group=get_dp_group().device_group)
+                else:
+                    hidden_states = get_ep_group().combine(hidden_states)
 
         if self.reduce_results and (self.tp_size > 1 or self.ep_size > 1):
-            final_hidden_states = tensor_model_parallel_all_reduce(
-                final_hidden_states)
+            hidden_states = tensor_model_parallel_all_reduce(hidden_states)
 
-        return final_hidden_states
+        return hidden_states

vllm_ascend/platform.py

@@ -138,7 +138,7 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
 
             # Calculate expert parallel size based on world size
             parallel_config.expert_parallel_size = (
-                parallel_config.world_size //
+                parallel_config.world_size_across_dp //
                 parallel_config.expert_tensor_parallel_size)
 
         if model_config is None:
@@ -167,6 +167,8 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
                         raise NotImplementedError(
                             "enable_graph_mode only works with deepseek model."
                         )
+                # Set compilation level to NO_COMPILATION to disable ACL Graph
+                compilation_config.level = CompilationLevel.NO_COMPILATION
 
         elif envs.VLLM_USE_V1 and model_config is not None and not enforce_eager:
             model_type = model_config.hf_config.model_type

vllm_ascend/quantization/w8a8_dynamic.py

@@ -20,6 +20,7 @@
 import torch
 import torch.distributed as dist
 import torch_npu
+from vllm.config import get_current_vllm_config
 from vllm.distributed import GroupCoordinator
 
 import vllm_ascend.envs as envs_ascend
@@ -508,6 +509,12 @@ def __init__(self):
 
         self.ep_group = get_ep_group()
 
+        self.enable_graph_mode = False
+        additional_config = get_current_vllm_config().additional_config
+        if additional_config:
+            self.enable_graph_mode = additional_config.get(
+                "enable_graph_mode", False)
+
         try:
             device_group = self.ep_group.device_group
             # TODO: Try local_rank = ep_group.rank_in_group
@@ -629,7 +636,7 @@ def apply(
                 top_k=top_k,
                 expert_map=expert_map,
                 moe_all_to_all_group_name=self.moe_all_to_all_group_name)
-        elif self.ep_group.world_size == 1:
+        elif self.enable_graph_mode or self.ep_group.world_size == 1:
             return fused_experts(hidden_states=x,
                                  w1=layer.w13_weight,
                                  w1_scale=layer.w13_weight_scale,