add built-in llama_ffn; add helloworld_custom_expert_sharded;

tutel/examples/helloworld.py

@@ -37,6 +37,7 @@
 parser.add_argument('--capacity_factor', type=float, default=1.0)  # 0.0 for dMoE (dropless-MoE), negative for no-padded capacity.
 parser.add_argument('--megablocks_size', type=int, default=0)
 parser.add_argument('--use_tensorcore', default=False, action='store_true')
+parser.add_argument('--expert_type', type=str, default='ffn')
 
 args = parser.parse_args()
 
@@ -74,7 +75,7 @@ def __init__(self):
 
         self._moe_layer = tutel_moe.moe_layer(
             gate_type = {'type': 'top', 'k': top_value, 'fp32_gate': args.fp32_gate, 'capacity_factor': args.capacity_factor},
-            experts = {'type': 'ffn', 'count_per_node': num_local_experts, 'hidden_size_per_expert': hidden_size, 'activation_fn': lambda x: F.relu(x)},
+            experts = {'type': args.expert_type, 'count_per_node': num_local_experts, 'hidden_size_per_expert': hidden_size, 'activation_fn': lambda x: F.relu(x)},
             model_dim = model_dim,
             scan_expert_func = lambda name, param: setattr(param, 'skip_allreduce', True),
             seeds = (1, dist_rank + 1, 1),

tutel/examples/helloworld_custom_expert.py

@@ -68,7 +68,8 @@ def __init__(self, model_dim, local_experts, sharded_count, my_config):
         self.reset_parameters()
 
     def reset_parameters(self):
-        pass
+        with torch.no_grad():
+          self.W.normal_(0, 0.001)
 
     def forward(self, x, ctx):
         if ctx.sharded_count > 1:

tutel/examples/helloworld_custom_expert_sharded.py

@@ -0,0 +1,176 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+import os
+import torch
+import torch.optim as optim
+import torch.nn.functional as F
+from torch import nn
+import argparse
+
+from tutel import system
+from tutel import moe as tutel_moe
+from tutel import net
+
+parser = argparse.ArgumentParser()
+
+parser.add_argument('--local_rank', type=int, default=-1)
+parser.add_argument('--batch_size', type=int, default=16)
+parser.add_argument('--num_tokens', type=int, default=512)
+parser.add_argument('--model_dim', type=int, default=2048)
+parser.add_argument('--num_local_experts', type=int, default=2)
+parser.add_argument('--dtype', type=str, default='float32')
+parser.add_argument('--fp32_gate', default=False, action='store_true')
+parser.add_argument('--top', type=int, default=2)
+parser.add_argument('--l_aux_wt', type=float, default=0.0)
+parser.add_argument('--a2a_ffn_overlap_degree', type=int, default=1)
+parser.add_argument('--allreduce_degree', type=int, default=1)
+parser.add_argument('--num_steps', type=int, default=100)
+parser.add_argument('--checkpoint_path', type=str, default='')
+parser.add_argument('--device', type=str, default='cuda' if torch.cuda.is_available() else 'cpu')
+parser.add_argument('--use_2dh', default=False, action='store_true')
+parser.add_argument('--eval', default=False, action='store_true')
+parser.add_argument('--capacity_factor', type=float, default=1.0)  # 0.0 for dMoE (dropless-MoE), negative for no-padded capacity.
+
+args = parser.parse_args()
+
+parallel_env = system.init_data_model_parallel(backend='nccl' if args.device == 'cuda' else 'gloo')
+dist_rank, dist_world_size, dist_print = parallel_env.global_rank, parallel_env.global_size, parallel_env.dist_print
+args.local_rank = parallel_env.local_device.index
+
+batch_size = args.batch_size
+num_tokens = args.num_tokens
+model_dim = args.model_dim
+num_local_experts = args.num_local_experts
+top_value = args.top
+a2a_ffn_overlap_degree = args.a2a_ffn_overlap_degree
+device = parallel_env.local_device
+
+if args.dtype == 'float32':
+    torch.set_default_dtype(torch.float32)
+elif args.dtype == 'float64':
+    torch.set_default_dtype(torch.float64)
+elif args.dtype == 'float16':
+    torch.set_default_dtype(torch.float16)
+elif args.dtype == 'bfloat16':
+    torch.set_default_dtype(torch.bfloat16)
+else:
+    raise Exception('Unrecognized data type specified: %s' % args.dtype)
+
+
+class CustomExpertDemo(torch.nn.Module):
+
+    def _create_sharded_param(self, *full_shape, **kwargs):
+        full_shape = torch.Size(full_shape)
+        sharded_shape = (full_shape.numel() + self.sharded_count - 1) // self.sharded_count
+        return torch.nn.Parameter(torch.empty(sharded_shape, **kwargs)), full_shape
+
+    def _get_gathered_param(self, param, full_shape):
+        sharded_group = net.create_groups_from_world(group_count=-self.sharded_count).model_group
+        return net.zero_gather(param, group=sharded_group).view(-1).narrow(0, 0, full_shape.numel()).view(full_shape)
+
+    def __init__(self, model_dim, local_experts, sharded_count, my_config):
+        super().__init__()
+        self.sharded_count = sharded_count
+        self.W, self.W_full_shape = self._create_sharded_param(local_experts, model_dim, model_dim)
+        self.my_activation = torch.nn.functional.relu if my_config == 'relu' else None
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        with torch.no_grad():
+          self.W.normal_(0, 0.001)
+
+    def forward(self, x, ctx):
+        W_full = self._get_gathered_param(self.W, self.W_full_shape)
+        y = torch.matmul(x, W_full)
+        if self.my_activation is not None:
+          y = self.my_activation(y)
+        return y
+
+
+class ExampleModel(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+        self._moe_layer = tutel_moe.moe_layer(
+            gate_type = {'type': 'top', 'k': top_value, 'fp32_gate': args.fp32_gate, 'capacity_factor': args.capacity_factor},
+            experts = {'type': 'custom', 'module': CustomExpertDemo, 'count_per_node': num_local_experts, 'my_config': None},
+            model_dim = model_dim,
+            scan_expert_func = lambda name, param: setattr(param, 'skip_allreduce', True),
+            seeds = (1, dist_rank + 1, 1),
+            a2a_ffn_overlap_degree = a2a_ffn_overlap_degree,
+            use_2dh=args.use_2dh,
+        )
+
+        # Summary of different parameter types: gate, local_experts
+        local_count = sum([torch.numel(param) for name, param in self._moe_layer.get_parameter_iterator(param_type='local_experts')])
+        shared_count = sum([torch.numel(param) for name, param in self._moe_layer.get_parameter_iterator(param_type='gate')])
+        dist_print('[Statistics] param count for MoE local_experts = %s, param count for MoE gate = %s.\n' % (local_count, shared_count))
+
+    def forward(self, input):
+        result = self._moe_layer(input)
+        result = F.log_softmax(torch.sum(result, dim=2), dim=1)
+        return result
+
+model = ExampleModel().to(device)
+dist_print(model)
+
+if args.checkpoint_path:
+    checkpoint_path = system.apply_rank_size_from_pattern(args.checkpoint_path, rank=parallel_env.global_rank, size=parallel_env.global_size)
+    if os.path.exists(checkpoint_path):
+        model.load_state_dict(torch.load(checkpoint_path))
+    else:
+        print('Checkpoint not loaded: file `%s` is not found. Will train the model from start.' % checkpoint_path)
+
+optimizer = torch.optim.SGD(model.parameters(), lr=1e-5)
+
+torch.manual_seed(0)
+x = torch.tensor(torch.randn([batch_size, num_tokens, model_dim], dtype=torch.float32, device='cpu').detach().numpy(), dtype=torch.get_default_dtype(), requires_grad=False, device=device)
+y = torch.LongTensor(batch_size).random_(1).to(device)
+
+tuples = (dist_world_size, args.dtype, model_dim, batch_size * num_tokens, num_local_experts, top_value, a2a_ffn_overlap_degree, device)
+dist_print('[Benchmark] world_size = %s, dtype = %s, model_dim = %s, samples = %s, num_local_experts = %s, topK = %s, a2a_ffn_overlap_degree = %s, device = `%s`' % tuples)
+
+average_time, num_steps = 0, args.num_steps
+
+if args.allreduce_degree == -1:
+    params_for_all_reduce = []
+else:
+    params_for_all_reduce = [p for p in model.parameters() if not hasattr(p, 'skip_allreduce') and getattr(p, 'requires_grad', False)]
+
+for i in range(num_steps):
+    t_start = system.record_time()
+
+    if not args.eval:
+        optimizer.zero_grad()
+        output = model(x)
+        loss = F.nll_loss(output, y)
+        if args.l_aux_wt:
+            loss += args.l_aux_wt * model._moe_layer.l_aux
+        loss.backward()
+        if dist_world_size > 1:
+            for p in params_for_all_reduce:
+                p.grad /= dist_world_size
+                p.grad = net.simple_all_reduce(p.grad)
+        optimizer.step()
+    else:
+        with torch.no_grad():
+            output = model(x)
+            loss = F.nll_loss(output, y)
+
+    t_stop = system.record_time()
+
+    num_global_experts = tutel_moe.moe_layer.global_expert_count(num_local_experts, group=system.get_local_session().model_group)
+    mm_ceof, cap_ceof = 1 if args.eval else 3, min(args.top, num_global_experts)
+    dist_print('STEP-%s: loss = %.5f, step_time = %.6f sec.' % (i, float(loss.data), t_stop - t_start))
+
+    if i + 10 >= num_steps:
+        average_time += t_stop - t_start
+
+average_time /= 10
+dist_print('\n[Summary] Average synchronized step_time = %s sec.' % average_time)
+
+if args.checkpoint_path:
+    torch.save(model.state_dict(), checkpoint_path)

tutel/experts/llama_ffn.py

@@ -0,0 +1,45 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+import torch
+from .. import net
+
+class LlamaFFNNetwork(torch.nn.Module):
+
+    def _create_sharded_param(self, *full_shape, **kwargs):
+        full_shape = torch.Size(full_shape)
+        sharded_shape = (full_shape.numel() + self.sharded_count - 1) // self.sharded_count
+        return torch.nn.Parameter(torch.empty(sharded_shape, **kwargs)), full_shape
+
+    def _get_gathered_param(self, param, full_shape):
+        sharded_group = net.create_groups_from_world(group_count=-self.sharded_count).model_group
+        return net.zero_gather(param, group=sharded_group).view(-1).narrow(0, 0, full_shape.numel()).view(full_shape)
+
+    def __init__(self, model_dim, hidden_size_per_expert, local_experts, sharded_count, activation_fn=torch.nn.functional.silu):
+        super().__init__()
+        self.sharded_count = sharded_count
+        self.W_fc1, self.W_fc1_full_shape = self._create_sharded_param(local_experts, model_dim, hidden_size_per_expert)
+        self.W_fc2, self.W_fc2_full_shape = self._create_sharded_param(local_experts, model_dim, hidden_size_per_expert)
+        self.W_fc3, self.W_fc3_full_shape = self._create_sharded_param(local_experts, hidden_size_per_expert, model_dim)
+        self.activation_fn = activation_fn
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        with torch.no_grad():
+          self.W_fc1.normal_(0, 0.01)
+          self.W_fc2.normal_(0, 0.01)
+          self.W_fc3.normal_(0, 0.01)
+
+    def forward(self, x, ctx):
+        W_fc1_full = self._get_gathered_param(self.W_fc1, self.W_fc1_full_shape)
+        W_fc2_full = self._get_gathered_param(self.W_fc2, self.W_fc2_full_shape)
+        W_fc3_full = self._get_gathered_param(self.W_fc3, self.W_fc3_full_shape)
+
+        y1 = torch.matmul(x, W_fc1_full)
+        y2 = torch.matmul(x, W_fc2_full)
+        y = self.activation_fn(y1 * y2)
+        y = torch.matmul(y, W_fc3_full)
+        return y
+
+
+ExpertModule = LlamaFFNNetwork