Fused MOE for Mixtral (vllm-project#2542)

pcmoritz · scv119 · alexm-redhat · commit a6c40bf1cede · 2024-02-13T11:24:42.000-05:00
Co-authored-by: chen shen &lt;scv119@gmail.com&gt;
diff --git a/csrc/moe_align_block_size_kernels.cu b/csrc/moe_align_block_size_kernels.cu
@@ -95,7 +95,7 @@ void moe_align_block_size(
     const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
     assert(num_experts <= NUM_MAX_EXPERTS);
     VLLM_DISPATCH_INTEGRAL_TYPES(
-        topk_ids.scalar_type(), "moe_alig_block_size_kernel", [&] {
+        topk_ids.scalar_type(), "moe_align_block_size_kernel", [&] {
         vllm::moe_align_block_size_kernel<scalar_t><<<1, num_experts, 0, stream>>>(
             topk_ids.data_ptr<scalar_t>(), 
             sorted_token_ids.data_ptr<int32_t>(), 
diff --git a/csrc/ops.h b/csrc/ops.h
@@ -82,6 +82,13 @@ void gptq_shuffle(
   torch::Tensor q_weight,
   torch::Tensor q_perm);
 
+void moe_align_block_size(
+  torch::Tensor topk_ids,
+  int num_experts,
+  int block_size,
+  torch::Tensor sorted_token_ids,
+  torch::Tensor experts_ids,
+  torch::Tensor num_tokens_post_pad);
 
 #ifndef USE_ROCM
 using fptr_t = uint64_t;
@@ -103,12 +110,3 @@ std::pair<std::vector<uint8_t>, std::vector<int64_t>> get_graph_buffer_ipc_meta(
 void register_graph_buffers(fptr_t _fa, const std::vector<std::string> &handles,
                             const std::vector<std::vector<int64_t>> &offsets);
 #endif
-
-void moe_align_block_size(
-  torch::Tensor topk_ids,
-  int num_experts,
-  int block_size,
-  torch::Tensor sorted_token_ids,
-  torch::Tensor experts_ids,
-  torch::Tensor num_tokens_post_pad
-  );
diff --git a/csrc/pybind.cpp b/csrc/pybind.cpp
@@ -58,9 +58,9 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
   ops.def("gptq_shuffle", &gptq_shuffle, "Post processing for GPTQ");
   ops.def("squeezellm_gemm", &squeezellm_gemm, "Quantized GEMM for SqueezeLLM");
   ops.def(
-      "moe_align_block_size",
-      &moe_align_block_size,
-      "Aligning the number of tokens to be processed by each expert such that it is divisible by the block size.");
+    "moe_align_block_size",
+    &moe_align_block_size,
+    "Aligning the number of tokens to be processed by each expert such that it is divisible by the block size.");
 
   // Cache ops
   pybind11::module cache_ops = m.def_submodule("cache_ops", "vLLM cache ops");
diff --git a/vllm/model_executor/models/mixtral.py b/vllm/model_executor/models/mixtral.py
@@ -23,8 +23,6 @@
 """Inference-only Mixtral model."""
 from typing import List, Optional, Tuple
 
-import numpy as np
-
 import torch
 import torch.nn.functional as F
 
@@ -33,10 +31,11 @@
 
 from vllm.model_executor.input_metadata import InputMetadata
 from vllm.model_executor.layers.attention import PagedAttention
+from vllm.model_executor.layers.fused_moe import fused_moe
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (LinearMethodBase,
-                                               ReplicatedLinear,
                                                QKVParallelLinear,
+                                               ReplicatedLinear,
                                                RowParallelLinear)
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.sampler import Sampler
@@ -47,92 +46,85 @@
 from vllm.model_executor.parallel_utils.parallel_state import (
     get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size)
 from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.model_executor.utils import set_weight_attrs
 from vllm.model_executor.weight_utils import (default_weight_loader,
                                               hf_model_weights_iterator)
 from vllm.sequence import SamplerOutput
 
 KVCache = Tuple[torch.Tensor, torch.Tensor]
 
 
-class MixtralMLP(nn.Module):
+class MixtralMoE(nn.Module):
+    """A tensor-parallel MoE implementation for Mixtral that shards each expert
+    across all ranks.
+
+    Each expert's weights are sharded across all ranks and a fused MoE
+    kernel is used for the forward pass, and finally we reduce the outputs
+    across ranks.
+    """
 
     def __init__(
         self,
         num_experts: int,
+        top_k: int,
         hidden_size: int,
         intermediate_size: int,
-        linear_method: Optional[LinearMethodBase] = None,
-    ) -> None:
+        params_dtype: Optional[torch.dtype] = None,
+    ):
         super().__init__()
-        self.num_experts = num_experts
-        self.ffn_dim = intermediate_size
-        self.hidden_dim = hidden_size
-
-        self.w1 = ReplicatedLinear(self.hidden_dim,
-                                   self.ffn_dim,
-                                   bias=False,
-                                   linear_method=linear_method)
-        self.w2 = ReplicatedLinear(self.ffn_dim,
-                                   self.hidden_dim,
-                                   bias=False,
-                                   linear_method=linear_method)
-        self.w3 = ReplicatedLinear(self.hidden_dim,
-                                   self.ffn_dim,
-                                   bias=False,
-                                   linear_method=linear_method)
-
-        # TODO: Use vllm's SiluAndMul
-        self.act_fn = nn.SiLU()
-
-    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
-        w1_out, _ = self.w1(hidden_states)
-        w1_out = self.act_fn(w1_out)
-        w3_out, _ = self.w3(hidden_states)
-        current_hidden_states = w1_out * w3_out
-        current_hidden_states, _ = self.w2(current_hidden_states)
-        return current_hidden_states
-
+        tp_size = get_tensor_model_parallel_world_size()
+        self.num_total_experts = num_experts
+        self.top_k = top_k
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size // tp_size
 
-class MixtralMoE(nn.Module):
+        if params_dtype is None:
+            params_dtype = torch.get_default_dtype()
+        self.params_dtype = params_dtype
 
-    def __init__(
-        self,
-        config: MixtralConfig,
-        linear_method: Optional[LinearMethodBase] = None,
-    ):
-        super().__init__()
-        self.config = config
-        self.rank = get_tensor_model_parallel_rank()
-        self.tp_size = get_tensor_model_parallel_world_size()
-        self.num_total_experts = config.num_local_experts
-        self.top_k = config.num_experts_per_tok
-        if self.tp_size > self.num_total_experts:
-            raise ValueError(
-                f"Tensor parallel size {self.tp_size} is greater than "
-                f"the number of experts {self.num_total_experts}.")
-        # Split experts equally between ranks
-        self.expert_indicies = np.array_split(range(
-            self.num_total_experts), self.tp_size)[self.rank].tolist()
-        if not self.expert_indicies:
-            raise ValueError(
-                f"Rank {self.rank} has no experts assigned to it.")
-
-        self.experts = nn.ModuleList([
-            MixtralMLP(self.num_total_experts,
-                       config.hidden_size,
-                       config.intermediate_size,
-                       linear_method=linear_method)
-            if idx in self.expert_indicies else None
-            for idx in range(self.num_total_experts)
-        ])
-        self.gate = ReplicatedLinear(config.hidden_size,
+        self.gate = ReplicatedLinear(self.hidden_size,
                                      self.num_total_experts,
                                      bias=False,
+                                     params_dtype=self.params_dtype,
                                      linear_method=None)
 
+        self.ws = nn.Parameter(
+            torch.empty(self.num_total_experts,
+                        2 * self.intermediate_size,
+                        self.hidden_size,
+                        device="cuda",
+                        dtype=self.params_dtype))
+        self.w2s = nn.Parameter(
+            torch.empty(self.num_total_experts,
+                        self.hidden_size,
+                        self.intermediate_size,
+                        device="cuda",
+                        dtype=self.params_dtype))
+
+        set_weight_attrs(self.ws, {
+            "weight_loader": self.weight_loader,
+        })
+        set_weight_attrs(self.w2s, {
+            "weight_loader": self.weight_loader,
+        })
+
+    def weight_loader(self, param: nn.Parameter, loaded_weight: torch.Tensor,
+                      weight_name: str, expert_id: int):
+        tp_rank = get_tensor_model_parallel_rank()
+        param_data = param.data
+        shard_size = self.intermediate_size
+        shard = slice(tp_rank * shard_size, (tp_rank + 1) * shard_size)
+        if weight_name.endswith("w1.weight"):
+            param_data[expert_id, 0:shard_size, :] = loaded_weight[shard, :]
+        if weight_name.endswith("w3.weight"):
+            param_data[expert_id,
+                       shard_size:2 * shard_size, :] = loaded_weight[shard, :]
+        if weight_name.endswith("w2.weight"):
+            param_data[expert_id, :, :] = loaded_weight[:, shard]
+
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
-        batch_size, sequence_length, hidden_dim = hidden_states.shape
-        hidden_states = hidden_states.view(-1, hidden_dim)
+        batch_size, sequence_length, hidden_size = hidden_states.shape
+        hidden_states = hidden_states.view(-1, self.hidden_size)
         # router_logits: (batch * sequence_length, n_experts)
         router_logits, _ = self.gate(hidden_states)
 
@@ -142,22 +134,18 @@ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
                                                        dim=-1)
         routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
 
-        final_hidden_states = None
-        for expert_idx in self.expert_indicies:
-            expert_layer = self.experts[expert_idx]
-            expert_mask = (selected_experts == expert_idx)
-            expert_weights = (routing_weights * expert_mask).sum(dim=-1,
-                                                                 keepdim=True)
-
-            current_hidden_states = expert_layer(hidden_states).mul_(
-                expert_weights)
-            if final_hidden_states is None:
-                final_hidden_states = current_hidden_states
-            else:
-                final_hidden_states.add_(current_hidden_states)
+        final_hidden_states = fused_moe(hidden_states,
+                                        self.ws,
+                                        self.w2s,
+                                        routing_weights,
+                                        selected_experts,
+                                        inplace=True)
+
+        final_hidden_states = tensor_model_parallel_all_reduce(
+            final_hidden_states)
 
-        return tensor_model_parallel_all_reduce(final_hidden_states).view(
-            batch_size, sequence_length, hidden_dim)
+        return final_hidden_states.view(batch_size, sequence_length,
+                                        hidden_size)
 
 
 class MixtralAttention(nn.Module):
@@ -257,8 +245,11 @@ def __init__(
             rope_theta=rope_theta,
             sliding_window=config.sliding_window,
             linear_method=linear_method)
-        self.block_sparse_moe = MixtralMoE(config=config,
-                                           linear_method=linear_method)
+        self.block_sparse_moe = MixtralMoE(
+            num_experts=config.num_local_experts,
+            top_k=config.num_experts_per_tok,
+            hidden_size=config.hidden_size,
+            intermediate_size=config.intermediate_size)
         self.input_layernorm = RMSNorm(config.hidden_size,
                                        eps=config.rms_norm_eps)
         self.post_attention_layernorm = RMSNorm(config.hidden_size,
@@ -378,6 +369,14 @@ def load_weights(self,
             ("qkv_proj", "v_proj", "v"),
         ]
 
+        expert_params_mapping = [
+            # (param_name, weight_name, expert_id)
+            ("ws" if weight_name in ["w1", "w3"] else "w2s",
+             f"experts.{expert_id}.{weight_name}.weight", expert_id)
+            for expert_id in range(self.config.num_local_experts)
+            for weight_name in ["w1", "w2", "w3"]
+        ]
+
         params_dict = dict(self.named_parameters())
         for name, loaded_weight in hf_model_weights_iterator(
                 model_name_or_path,
@@ -387,6 +386,7 @@ def load_weights(self,
                 fall_back_to_pt=False):
             if "rotary_emb.inv_freq" in name:
                 continue
+
             for (param_name, weight_name, shard_id) in stacked_params_mapping:
                 if weight_name not in name:
                     continue
@@ -399,14 +399,22 @@ def load_weights(self,
                 weight_loader(param, loaded_weight, shard_id)
                 break
             else:
-                # Skip loading extra bias for GPTQ models.
-                if name.endswith(".bias") and name not in params_dict:
-                    continue
-                # Skip experts that are not assigned to this worker.
-                if ("block_sparse_moe.experts." in name
-                        and name not in params_dict):
-                    continue
-                param = params_dict[name]
-                weight_loader = getattr(param, "weight_loader",
-                                        default_weight_loader)
-                weight_loader(param, loaded_weight)
+                for param_name, weight_name, expert_id in expert_params_mapping:
+                    if weight_name not in name:
+                        continue
+                    name = name.replace(weight_name, param_name)
+                    param = params_dict[name]
+                    weight_loader = param.weight_loader
+                    weight_loader(param,
+                                  loaded_weight,
+                                  weight_name,
+                                  expert_id=expert_id)
+                    break
+                else:
+                    # Skip loading extra bias for GPTQ models.
+                    if name.endswith(".bias") and name not in params_dict:
+                        continue
+                    param = params_dict[name]
+                    weight_loader = getattr(param, "weight_loader",
+                                            default_weight_loader)
+                    weight_loader(param, loaded_weight)
