MoE for mixtral 8x7b (#2535)

* MoE for mixtral 8x7b
* removing bnb_sparse for now

Author: vince62s

onmt/decoders/transformer.py

@@ -9,6 +9,7 @@
 from onmt.modules import MultiHeadedAttention, AverageAttention
 from onmt.modules.position_ffn import PositionwiseFeedForward
 from onmt.modules.position_ffn import ActivationFunction
+from onmt.modules.moe import MoE
 from onmt.utils.misc import sequence_mask
 
 try:
@@ -43,6 +44,8 @@ def __init__(
         parallel_gpu=1,
         sliding_window=0,
         rotary_interleave=True,
+        num_experts=0,
+        num_experts_per_tok=2,
     ):
         """
         Args:
@@ -109,18 +112,34 @@ def __init__(
                 d_model, dropout=attention_dropout, aan_useffn=aan_useffn
             )
 
-        self.feed_forward = PositionwiseFeedForward(
-            d_model,
-            d_ff,
-            dropout,
-            pos_ffn_activation_fn,
-            add_ffnbias,
-            parallel_residual,
-            layer_norm,
-            norm_eps,
-            use_ckpting=use_ckpting,
-            parallel_gpu=parallel_gpu,
-        )
+        if num_experts > 0:
+            self.feed_forward = MoE(
+                num_experts,
+                num_experts_per_tok,
+                d_model,
+                d_ff,
+                dropout,
+                pos_ffn_activation_fn,
+                add_ffnbias,
+                parallel_residual,
+                layer_norm,
+                norm_eps,
+                use_ckpting=use_ckpting,
+                parallel_gpu=parallel_gpu,
+            )
+        else:
+            self.feed_forward = PositionwiseFeedForward(
+                d_model,
+                d_ff,
+                dropout,
+                pos_ffn_activation_fn,
+                add_ffnbias,
+                parallel_residual,
+                layer_norm,
+                norm_eps,
+                use_ckpting=use_ckpting,
+                parallel_gpu=parallel_gpu,
+            )
         self.parallel_residual = parallel_residual
         self.shared_layer_norm = shared_layer_norm
         if layer_norm == "standard":
@@ -261,6 +280,8 @@ def __init__(
         parallel_gpu=1,
         sliding_window=0,
         rotary_interleave=True,
+        num_experts=0,
+        num_experts_per_tok=2,
     ):
         """
         Args:
@@ -290,6 +311,8 @@ def __init__(
             parallel_gpu=parallel_gpu,
             sliding_window=sliding_window,
             rotary_interleave=rotary_interleave,
+            num_experts=num_experts,
+            num_experts_per_tok=num_experts_per_tok,
         )
         self.context_attn = MultiHeadedAttention(
             heads,
@@ -450,6 +473,8 @@ def from_opt(cls, opt, embeddings):
             else 1,
             sliding_window=opt.sliding_window,
             rotary_interleave=opt.rotary_interleave,
+            num_experts=opt.num_experts,
+            num_experts_per_tok=opt.num_experts_per_tok,
         )
 
     def init_state(self, src, enc_out, enc_final_hs):
@@ -569,6 +594,8 @@ def __init__(
         parallel_gpu=1,
         sliding_window=0,
         rotary_interleave=True,
+        num_experts=0,
+        num_experts_per_tok=2,
     ):
         super(TransformerDecoder, self).__init__(
             d_model, copy_attn, embeddings, alignment_layer, layer_norm, norm_eps
@@ -600,6 +627,8 @@ def __init__(
                     parallel_gpu=parallel_gpu,
                     sliding_window=sliding_window,
                     rotary_interleave=rotary_interleave,
+                    num_experts=num_experts,
+                    num_experts_per_tok=num_experts_per_tok,
                 )
                 for i in range(num_layers)
             ]
@@ -836,6 +865,8 @@ def __init__(
         parallel_gpu=1,
         sliding_window=0,
         rotary_interleave=True,
+        num_experts=0,
+        num_experts_per_tok=2,
     ):
         super(TransformerLMDecoder, self).__init__(
             d_model, copy_attn, embeddings, alignment_layer, layer_norm, norm_eps
@@ -866,6 +897,8 @@ def __init__(
                     parallel_gpu=parallel_gpu,
                     sliding_window=sliding_window,
                     rotary_interleave=rotary_interleave,
+                    num_experts=num_experts,
+                    num_experts_per_tok=num_experts_per_tok,
                 )
                 for i in range(num_layers)
             ]

onmt/modules/bnb_linear.py

@@ -7,7 +7,12 @@
 try:
     os.environ["BITSANDBYTES_NOWELCOME"] = "1"
     from bitsandbytes import MatmulLtState
-    from bitsandbytes.nn import Linear4bit, Linear8bitLt, Params4bit, Int8Params
+    from bitsandbytes.nn import (
+        Linear4bit,
+        Linear8bitLt,
+        Params4bit,
+        Int8Params,
+    )
 except ImportError:
     raise ImportError("Install bitsandbytes to use 4/8bit compression")
 

onmt/modules/moe.py

@@ -0,0 +1,63 @@
+"""MoE mixture of experts"."""
+import torch
+import torch.nn as nn
+from onmt.modules.position_ffn import PositionwiseFeedForward
+
+
+class MoE(nn.Module):
+    def __init__(
+        self,
+        num_experts,
+        num_experts_per_tok,
+        d_model,
+        d_ff,
+        dropout,
+        pos_ffn_activation_fn,
+        add_ffnbias,
+        parallel_residual,
+        layer_norm,
+        norm_eps,
+        use_ckpting=[],
+        parallel_gpu=1,
+    ):
+        super().__init__()
+        self.experts = nn.ModuleList(
+            [
+                PositionwiseFeedForward(
+                    d_model,
+                    d_ff,
+                    dropout,
+                    pos_ffn_activation_fn,
+                    add_ffnbias,
+                    parallel_residual,
+                    layer_norm,
+                    norm_eps,
+                    use_ckpting=use_ckpting,
+                    parallel_gpu=parallel_gpu,
+                )
+                for i in range(num_experts)
+            ]
+        )
+        self.gate = nn.Linear(d_model, num_experts, bias=False)
+        self.num_experts_per_tok = num_experts_per_tok
+
+    def forward(self, x):
+        orig_shape = x.shape
+        x = x.view(-1, x.shape[-1])
+
+        scores = self.gate(x)
+        expert_weights, expert_indices = torch.topk(
+            scores, self.num_experts_per_tok, dim=-1
+        )
+        expert_weights = expert_weights.softmax(dim=-1)
+        flat_expert_indices = expert_indices.view(-1)
+
+        x = x.repeat_interleave(self.num_experts_per_tok, dim=0)
+        y = torch.empty_like(x)
+        for i, expert in enumerate(self.experts):
+            if torch.any(flat_expert_indices == i):
+                y[flat_expert_indices == i] = expert(x[flat_expert_indices == i])
+        y = (y.view(*expert_weights.shape, -1) * expert_weights.unsqueeze(-1)).sum(
+            dim=1
+        )
+        return y.view(*orig_shape)

onmt/opts.py

@@ -901,6 +901,20 @@ def model_opts(parser):
         default=2048,
         help="Size of hidden transformer feed-forward",
     )
+    group.add(
+        "--num_experts",
+        "-num_experts",
+        type=int,
+        default=0,
+        help="Number of experts",
+    )
+    group.add(
+        "--num_experts_per_tok",
+        "-num_experts_per_tok",
+        type=int,
+        default=2,
+        help="Number of experts per token",
+    )
     group.add(
         "--aan_useffn",
         "-aan_useffn",