replace all_reduce for kv_consumer and support different num_tokens among all ranks

Signed-off-by: linfeng-yuan <1102311262@qq.com>

Author: linfeng-yuan

vllm_ascend/ops/fused_moe/fused_moe.py

@@ -126,7 +126,8 @@ def apply(self,
         # to avoid accumulating too much tokens on a single rank.
         # currently it is only activated when doing profile runs.
         if enable_force_load_balance:
-            topk_ids = torch.randint_like(topk_ids, 0, global_num_experts)
+            random_matrix = torch.rand(topk_ids.size(0), global_num_experts, device=topk_ids.device)
+            topk_ids = torch.argsort(random_matrix, dim=1)[:, :topk_ids.size(1)].to(topk_ids.dtype)
 
         moe_comm_method = get_forward_context().moe_comm_method
         return moe_comm_method.fused_experts(

vllm_ascend/ops/fused_moe/token_dispatcher.py

@@ -25,6 +25,7 @@
 
 import torch
 import torch_npu
+from vllm.config import get_current_vllm_config
 from vllm.distributed.parallel_state import get_ep_group
 
 from vllm_ascend.distributed.parallel_state import get_mc2_group
@@ -100,15 +101,34 @@ def __init__(self, **kwargs):
         self.need_extra_args = (
             get_ascend_device_type() == AscendDeviceType._910_93)
 
-        # NOTE: Currently, when in A3, we need to pass in some extra param into dispatch & combine
-        self.a3_need_extra_args = \
-            get_ascend_device_type() == AscendDeviceType._910_93
         # NOTE: When in A2, setting the environment variables HCCL_INTRA_PCIE_ENABLE=1 and
         # HCCL_INTRA_ROCE_ENABLE=0 can reduce cross-machine communication traffic and significantly
         # improve communication performance.
         self.need_expert_scale = is_hierarchical_communication_enabled()
         self.with_quant = False
 
+        # Here we need to calculate the global_bs = max_bs_per_rank * ep_world_size to execute
+        # dispatch & combine operators with different input num_tokens per rank.
+        vllm_config = get_current_vllm_config()
+        scheduler_config = vllm_config.scheduler_config
+        compilation_config = vllm_config.compilation_config
+        speculative_config = vllm_config.speculative_config
+        tp_size = vllm_config.parallel_config.tensor_parallel_size
+        uniform_decode_query_len = 1 if not speculative_config else \
+            1 + speculative_config.num_speculative_tokens
+        decode_max_num_seqs = getattr(scheduler_config,
+                                      'decode_max_num_seqs', 0)
+        max_num_reqs = max(scheduler_config.max_num_seqs,
+                                decode_max_num_seqs)
+        if compilation_config.cudagraph_capture_sizes:
+            max_num_tokens = compilation_config.max_cudagraph_capture_size
+        else:
+            max_num_tokens = min(
+                max_num_reqs * uniform_decode_query_len, 512)
+        num_tokens_per_tp_rank = (max_num_tokens + tp_size - 1) // tp_size
+        mc2_tokens_capacity = num_tokens_per_tp_rank * tp_size
+        self.global_bs = mc2_tokens_capacity * self.ep_world_size
+
     def get_dispatch_mc2_kwargs(
         self,
         hidden_states: torch.Tensor,
@@ -130,7 +150,7 @@ def get_dispatch_mc2_kwargs(
             "expert_shard_type": 0,
             "shared_expert_rank_num": 0,
             "moe_expert_num": moe_expert_num,
-            "global_bs": 0,
+            "global_bs": self.global_bs,
             "expert_token_nums_type": 0,
         }
 
@@ -147,10 +167,6 @@ def get_dispatch_mc2_kwargs(
                 "tp_world_size": 1,
                 "tp_rank_id": 0,
             })
-        if self.a3_need_extra_args and self.enable_dispatch_v2:
-            stage1_kwargs.update({
-                "x_active_mask": mc2_mask,
-            })
         if self.need_expert_scale:
             stage1_kwargs.update({
                 "expert_scales":
@@ -256,7 +272,7 @@ def get_combine_mc_kwargs(self, hidden_states: torch.Tensor,
             "expert_shard_type": 0,
             "shared_expert_rank_num": 0,
             "moe_expert_num": moe_expert_num,
-            "global_bs": 0,
+            "global_bs": self.global_bs,
         }
 
         if self.with_quant:
@@ -285,9 +301,6 @@ def get_combine_mc_kwargs(self, hidden_states: torch.Tensor,
                 "tp_rank_id": 0,
             })
 
-        if self.a3_need_extra_args and self.enable_dispatch_v2:
-            stage3_kwargs["x_active_mask"] = mc2_mask
-
         kwargs_mc2.update(stage3_kwargs)
         return kwargs_mc2
 

vllm_ascend/quantization/w4a8_dynamic.py

@@ -371,8 +371,8 @@ def apply(
         # to avoid accumulating too much tokens on a single rank.
         # currently it is only activated when doing profile runs.
         if enable_force_load_balance:
-            topk_ids = torch.randint_like(
-                topk_ids, 0, global_num_experts - global_redundant_expert_num)
+            random_matrix = torch.rand(topk_ids.size(0), global_num_experts, device=topk_ids.device)
+            topk_ids = torch.argsort(random_matrix, dim=1)[:, :topk_ids.size(1)].to(topk_ids.dtype)
 
         topk_weights = topk_weights.to(x.dtype)
 

vllm_ascend/quantization/w8a8_dynamic.py

@@ -216,8 +216,8 @@ def apply(
         # to avoid accumulating too much tokens on a single rank.
         # currently it is only activated when doing profile runs.
         if enable_force_load_balance:
-            topk_ids = torch.randint_like(
-                topk_ids, 0, global_num_experts - global_redundant_expert_num)
+            random_matrix = torch.rand(topk_ids.size(0), global_num_experts, device=topk_ids.device)
+            topk_ids = torch.argsort(random_matrix, dim=1)[:, :topk_ids.size(1)].to(topk_ids.dtype)
 
         topk_weights = topk_weights.to(self.in_dtype)
 

vllm_ascend/worker/model_runner_v1.py

@@ -910,6 +910,17 @@ def _sync_metadata_across_dp(
         if self.dp_size == 1:
             return num_tokens, None, with_prefill
 
+        # NOTE: Here we can skip the all_reduce operation and avoid paading tokens
+        # to max_tokens_acrodd_dp in D nodes. In MoE models, we must ensure that
+        # num_tokens DOES NOT exceed mc2_tokens_capacity which means that moe_comm_method
+        # of each rank is MC2. It is recommended to enable recompute scheduler for D Noes.
+        if self.is_kv_consumer and not self.in_profile_run:
+            num_tokens_after_padding = torch.tensor([num_tokens] *
+                                        self.dp_size,
+                                        device="cpu",
+                                        dtype=torch.int32)
+            return num_tokens, num_tokens_after_padding, with_prefill
+
         # Sync num_tokens, with_prefill across dp ranks
         num_tokens_tensor = torch.tensor([
             num_tokens if i == self.dp_rank else 0 for i in range(self.dp_size)