[Improve]: Remove dlblas from lmdeploy (#4682)

* remove dlblas

* resolve comments

* resolve comment

Author: RunningLeon

docker/install.sh

@@ -53,7 +53,7 @@ else
 fi
 
 pip install /wheels/*.whl
-pip install dlblas==0.0.7 dlslime==0.0.2.post1
+pip install dlslime==0.0.2.post1
 
 pip install ninja einops packaging
 

lmdeploy/pytorch/backends/cuda/graph_runner.py

@@ -308,7 +308,7 @@ def prepare_inputs_for_generation(
         """Prepare inputs."""
 
         if get_deepep_state().enabled():
-            from dlblas.layers.moe.token_dispatcher import DeepEPBuffer, DeepEPMode
+            from lmdeploy.pytorch.backends.cuda.token_dispatcher import DeepEPBuffer, DeepEPMode
             deepep_mode = DeepEPMode.LOW_LATENCY if context.global_is_decoding() else DeepEPMode.NORMAL
             DeepEPBuffer.set_deepep_mode(deepep_mode)
 
@@ -322,7 +322,7 @@ def reset(self):
         """Remove all graphs to prevent hanging on exit."""
         self._runner_map.clear()
         if get_deepep_state().enabled():
-            from dlblas.layers.moe.token_dispatcher import DeepEPBuffer
+            from lmdeploy.pytorch.backends.cuda.token_dispatcher import DeepEPBuffer
 
             if hasattr(DeepEPBuffer, 'destroy'):
                 from torch import distributed as dist

lmdeploy/pytorch/backends/cuda/moe/blocked_fp8.py

@@ -81,8 +81,9 @@ def __init__(
         layer_index: int = 0,
         top_k: int = 8,
         out_dtype: torch.dtype = torch.bfloat16,
+        num_max_dispatch_tokens_per_rank: int = 128,
     ):
-        from dlblas.layers.moe.token_dispatcher import DeepEPTokenDispatcherNormal
+        from lmdeploy.pytorch.backends.cuda.token_dispatcher import DeepEPTokenDispatcherNormal
         self.layer_index = layer_index
         self.top_k = top_k
         self.num_experts = num_experts
@@ -94,6 +95,7 @@ def __init__(
             num_local_experts=self.num_local_experts,
             hidden_size=hidden_dim,
             params_dtype=out_dtype,
+            num_max_dispatch_tokens_per_rank=num_max_dispatch_tokens_per_rank,
         )
 
     def forward(
@@ -148,7 +150,7 @@ def fusedmoe_forward(self, state, up_weight, down_weight):
 
 
 def _disposible_tensor(tensor):
-    from dlblas.utils.utils import DisposibleTensor
+    from lmdeploy.pytorch.backends.cuda.token_dispatcher import DisposibleTensor
     if isinstance(tensor, torch.Tensor):
         tensor = DisposibleTensor(tensor)
     else:
@@ -237,8 +239,9 @@ def __init__(
         hidden_dim: int,
         layer_index: int,
         out_dtype: torch.dtype = torch.bfloat16,
+        num_max_dispatch_tokens_per_rank: int = 128,
     ):
-        from dlblas.layers.moe.token_dispatcher import DeepEPTokenDispatcherLowLatency
+        from lmdeploy.pytorch.backends.cuda.token_dispatcher import DeepEPTokenDispatcherLowLatency
         self.num_experts = num_experts
         self.layer_index = layer_index
         self.out_dtype = out_dtype
@@ -248,6 +251,7 @@ def __init__(
             num_local_experts=num_experts // ep_size,
             hidden_size=hidden_dim,
             params_dtype=out_dtype,
+            num_max_dispatch_tokens_per_rank=num_max_dispatch_tokens_per_rank,
         )
 
     def experts(
@@ -258,8 +262,7 @@ def experts(
         masked_m: torch.Tensor,
         expected_m: int,
     ):
-        from dlblas.utils.utils import DisposibleTensor
-
+        from lmdeploy.pytorch.backends.cuda.token_dispatcher import DisposibleTensor
         from lmdeploy.pytorch.kernels.cuda.activation import silu_and_mul_moe_ep
         from lmdeploy.pytorch.third_party.deep_gemm import m_grouped_bf16_gemm_nt_masked
         num_groups, m, _ = hidden_states.shape
@@ -339,22 +342,25 @@ def build_deepep_moe(
     top_k: int,
     layer_idx: int = 0,
     out_dtype: torch.dtype = torch.bfloat16,
+    num_max_dispatch_tokens_per_rank: int = 128,
 ):
     if low_latency_mode:
         return FusedMoELowLatency(ep_size=ep_size,
                                   ep_group=ep_group,
                                   num_experts=num_experts,
                                   hidden_dim=hidden_dim,
                                   layer_index=layer_idx,
-                                  out_dtype=out_dtype)
+                                  out_dtype=out_dtype,
+                                  num_max_dispatch_tokens_per_rank=num_max_dispatch_tokens_per_rank)
     else:
         return FusedMoENormal(ep_size=ep_size,
                               ep_group=ep_group,
                               num_experts=num_experts,
                               hidden_dim=hidden_dim,
                               layer_index=layer_idx,
                               top_k=top_k,
-                              out_dtype=out_dtype)
+                              out_dtype=out_dtype,
+                              num_max_dispatch_tokens_per_rank=num_max_dispatch_tokens_per_rank)
 
 
 class FusedMoEEPImpl(TritonFusedMoEImpl):
@@ -370,6 +376,7 @@ def __init__(
         renormalize: bool = False,
         layer_idx: int = 0,
         out_dtype: torch.dtype = torch.bfloat16,
+        num_max_dispatch_tokens_per_rank: int = 128,
     ):
         super().__init__(top_k, num_experts, renormalize)
         self.num_experts = num_experts
@@ -378,19 +385,21 @@ def __init__(
         self.hidden_dim = hidden_dim
         self.layer_idx = layer_idx
         self.out_dtype = out_dtype
+        self.num_max_dispatch_tokens_per_rank = num_max_dispatch_tokens_per_rank
 
         try:
             import deep_gemm  # noqa: F401
         except ImportError:
             logger.exception('DeepGEMM is required for DeepEP MoE implementation.')
+            raise
 
-        try:
-            from dlblas.layers.moe.token_dispatcher import DeepEPBuffer, DeepEPMode, use_deepep  # noqa: F401
-            get_deepep_state().enable()
-            if hasattr(DeepEPBuffer, 'set_explicitly_destroy'):
-                DeepEPBuffer.set_explicitly_destroy()
-        except ImportError:
-            logger.warning('For higher performance, please install DeepEP https://github.com/deepseek-ai/DeepEP')
+        from lmdeploy.pytorch.backends.cuda.token_dispatcher import DeepEPBuffer, use_deepep
+        if not use_deepep:
+            raise ImportError('DeepEP is required for DeepEP MoE implementation. Please install '
+                              'https://github.com/deepseek-ai/DeepEP.')
+        get_deepep_state().enable()
+        if hasattr(DeepEPBuffer, 'set_explicitly_destroy'):
+            DeepEPBuffer.set_explicitly_destroy()
 
         # pre-allocate buffer
         self.fusedmoe_build(True)
@@ -440,7 +449,8 @@ def fusedmoe_build(self, low_latency_mode: bool = False):
                                       self.hidden_dim,
                                       self.top_k,
                                       layer_idx=self.layer_idx,
-                                      out_dtype=self.out_dtype)
+                                      out_dtype=self.out_dtype,
+                                      num_max_dispatch_tokens_per_rank=self.num_max_dispatch_tokens_per_rank)
         return deepep_moe
 
 
@@ -457,6 +467,7 @@ def build(
         ep_group: dist.ProcessGroup = None,
         layer_idx: int = 0,
         out_dtype: torch.dtype = torch.bfloat16,
+        num_max_dispatch_tokens_per_rank: int = 128,
     ):
         """Build from mlp."""
         if ep_size > 1:
@@ -467,5 +478,6 @@ def build(
                                   hidden_dim=hidden_dim,
                                   renormalize=renormalize,
                                   layer_idx=layer_idx,
-                                  out_dtype=out_dtype)
+                                  out_dtype=out_dtype,
+                                  num_max_dispatch_tokens_per_rank=num_max_dispatch_tokens_per_rank)
         return TritonFusedMoEImpl(top_k=top_k, num_experts=num_experts, renormalize=renormalize)