Merge pull request #174 from laekov/fix-test

Fix tests

cuda/fastermoe/smart_schedule.h

@@ -338,7 +338,7 @@ void fmoe_cuda_fused_backward_impl(
             collect_fn(si, i / num_expert, 0);
             if (i / num_expert == rank) {
                 cudaEventCreate(evt_reduce + i % num_expert);
-                cudaEventRecord(evt_reduce[i % num_expert], smgr->stream(num_expert));
+                cudaEventRecord(evt_reduce[i % num_expert], smgr->stream(0));
             }
             ++si;
         }
@@ -367,7 +367,6 @@ void fmoe_cuda_fused_backward_impl(
     for (long i = 0, si = 0; i < world_size * num_expert; ++i) {
         if (stored_models[i]) {
             if (i / num_expert == rank) {
-                FMOE_SWE(smgr->stream(0), evt_reduce[i % num_expert]);
                 FMOE_SWE(smgr->torchStream(), evt_reduce[i % num_expert]);
                 set_grad_fn(si, i % num_expert);
             }

setup.py

@@ -41,7 +41,7 @@
 if __name__ == '__main__':
     setuptools.setup(
         name='fastmoe',
-        version='1.0.2',
+        version='1.1.0',
         description='An efficient Mixture-of-Experts system for PyTorch',
         author=', '.join(authors),
         author_email='[email protected]',

tests/moe.py

@@ -78,7 +78,7 @@ def forward(self, inp, gate_idx, gate_score):
 class NaiveExpert(nn.Module):
     def __init__(self, d_model):
         super(NaiveExpert, self).__init__()
-        self.linear = nn.Linear(d_model, d_model).cuda()
+        self.linear = nn.Linear(d_model, d_model)
 
     def forward(self, x, fec=None):
         return self.linear(x)
@@ -89,7 +89,7 @@ def __init__(self, d_model):
         super(LinearExpert, self).__init__()
         self.model = nn.Sequential(
             nn.Linear(d_model, d_model * 2), nn.ReLU(), nn.Linear(d_model * 2, d_model),
-        ).cuda()
+        )
 
     def forward(self, x, fec=None):
         return self.model(x)

tests/test.sh

@@ -30,4 +30,4 @@ fi
 
 export CUDA_VISIBLE_DEVICES=$localrank
 
-exec $@
+exec $@ 2>&1 | tee $RANK.log

tests/test_ddp.py

@@ -4,6 +4,7 @@
 import sys
 from typing import Dict
 import random
+import socket as sock
 
 import pytest
 import torch
@@ -24,6 +25,8 @@ def _ensure_initialized():
         dist.init_process_group(backend="nccl")
 
 
+port_count = 0
+
 def _run_distributed(func, world_size, args: Dict, script=__file__, env=dict()):
     device_count = torch.cuda.device_count()
     if device_count < world_size:
@@ -33,7 +36,9 @@ def _run_distributed(func, world_size, args: Dict, script=__file__, env=dict()):
 
     ps = []
     env["MASTER_ADDR"] = "localhost"
-    env["MASTER_PORT"] = str(random.randint(50000, 60000))
+    global port_count
+    env["MASTER_PORT"] = str(9010 + port_count)
+    port_count += 1
     env["OMPI_COMM_WORLD_SIZE"] = str(world_size)
     env["LD_LIBRARY_PATH"] = os.environ.get("LD_LIBRARY_PATH")
 
@@ -58,7 +63,7 @@ def _run_distributed(func, world_size, args: Dict, script=__file__, env=dict()):
 @pytest.mark.parametrize("d_model", [16])
 @pytest.mark.parametrize("d_hidden", [32])
 @pytest.mark.parametrize("mp_size", [1, 2])
-@pytest.mark.parametrize("data_type", ['torch.FloatTensor', 'torch.DoubleTensor', 'torch.HalfTensor'])
+@pytest.mark.parametrize("data_type", ['torch.float32', 'torch.bfloat16', 'torch.float16'])
 def test_fmoe_linear_distributed(
     num_expert, top_k, batch_size, d_model, d_hidden, mp_size, data_type
 ):
@@ -83,7 +88,8 @@ def test_fmoe_linear_distributed(
 @pytest.mark.parametrize("d_model", [16])
 @pytest.mark.parametrize("expert", ["NaiveExpert", "LinearExpert"])
 @pytest.mark.parametrize("mp_size", [1, 2])
-def test_fmoe_distributed(num_expert, top_k, batch_size, d_model, expert, mp_size):
+@pytest.mark.parametrize("data_type", ['torch.float32', 'torch.bfloat16', 'torch.float16'])
+def test_fmoe_distributed(num_expert, top_k, batch_size, d_model, expert, mp_size, data_type):
     _run_distributed(
         "_test_fmoe",
         mp_size * 2,
@@ -94,6 +100,7 @@ def test_fmoe_distributed(num_expert, top_k, batch_size, d_model, expert, mp_siz
             "d_model": d_model,
             "expert": expert,
             "mp_size": mp_size,
+            "data_type": data_type,
         },
     )
 
@@ -137,8 +144,29 @@ def test_fmoe_local_ddp(mp_size):
         del args["mp_size"]
         locals()[sys.argv[1]](**args)
     else:
-        test_fmoe_local_ddp(mp_size=2)
-        test_fmoe_linear_distributed(
-            num_expert=4, top_k=2, batch_size=4, d_model=8, d_hidden=8, mp_size=2,
-            data_type="torch.HalfTensor"
+        torch.distributed.init_process_group(backend="nccl")
+        args = dict(mp_size=1, data_type='torch.float16')
+        args["rank"] = torch.distributed.get_rank()
+        args["world_size"] = torch.distributed.get_world_size()
+        args["mp_group"] = [
+            torch.distributed.new_group(
+                ranks=[j * args["mp_size"] + i for i in range(args["mp_size"])],
+                backend="nccl",
+            )
+            for j in range(args["world_size"] // args["mp_size"])
+        ][args["rank"] // args["mp_size"]]
+        args["dp_group"] = [
+            torch.distributed.new_group(
+                ranks=[
+                    i * args["mp_size"] + j
+                    for i in range(args["world_size"] // args["mp_size"])
+                ],
+                backend="nccl",
+            )
+            for j in range(args["mp_size"])
+        ][args["rank"] % args["mp_size"]]
+        args["world_group"] = torch.distributed.new_group(
+            ranks=list(range(args["world_size"])), backend="nccl",
         )
+        del args["mp_size"]
+        _test_fmoe(4, 2, 16, 2, 'NaiveExpert', **args)

tests/test_faster_schedule.py

@@ -18,7 +18,7 @@
 @pytest.mark.parametrize("n_process", [8])
 @pytest.mark.parametrize("d_model", [1024])
 @pytest.mark.parametrize("batch_size", [16])
-@pytest.mark.parametrize("n_expert", [1])
+@pytest.mark.parametrize("n_expert", [1, 4])
 @pytest.mark.parametrize("group_sz", [1, 2, 4])
 def test_faster_schedule(n_process, d_model, batch_size, n_expert, group_sz):
     _run_distributed('_test_faster_schedule',
@@ -45,28 +45,39 @@ def _test_faster_schedule(d_model, batch_size, n_expert):
     x2 = x1.data.clone()
     x2.requires_grad = True
     topk_idx = torch.randint(0, world_size * n_expert, (batch_size, 2)).cuda()
-    m1 = torch.nn.Linear(d_model, d_model).cuda()
-    m2 = torch.nn.Linear(d_model, d_model).cuda()
+    m1s = [torch.nn.Linear(d_model, d_model).cuda() for _ in range(n_expert)]
+    m2s = [torch.nn.Linear(d_model, d_model).cuda() for _ in range(n_expert)]
     with torch.no_grad():
-        m2.weight.copy_(m1.weight)
-        m2.bias.copy_(m1.bias)
+        for m1, m2 in zip(m1s, m2s):
+            m2.weight.copy_(m1.weight)
+            m2.bias.copy_(m1.bias)
+
+    def ef1(x, fec, eidx):
+        return m1s[eidx](x)
 
-    def ef1(x, fec):
-        y = m1(x)
-        return y
     def ef2(x, fec):
-        y = m2(x)
+        o = 0
+        ys = []
+        for m, i in zip(m2s, fec):
+            if i > 0:
+                ys.append(m(x[o:o + i]))
+                o += i
+        y = torch.cat(ys)
         return y
 
     ensure_comm(x1, None)
-    y1 = smart_fwd(x1, topk_idx, ef1, n_expert, world_size)
+    y1 = smart_fwd(x1, topk_idx, ef1, n_expert, world_size, experts=m1s)
     y1.sum().backward()
 
-    y2 = naive_fwd(x2, topk_idx, ef2, n_expert, world_size)
+    y2 = naive_fwd(x2, topk_idx, ef2, n_expert, world_size, experts=m2s)
     y2.sum().backward()
-    _assert_numerical(['out', 'grad_in', 'grad_bias', 'grad_weight'],
-            [y1, x1.grad, m1.bias.grad, m1.weight.grad],
-            [y2, x2.grad, m2.bias.grad, m2.weight.grad], rank)
+    _assert_numerical(['out', 'grad_in'],
+            [y1, x1.grad],
+            [y2, x2.grad], rank)
+    for i in range(n_expert):
+        _assert_numerical([f'grad_bias_{i}', f'grad_weight_{i}'],
+                          [m1s[i].bias.grad, m1s[i].weight.grad],
+                          [m2s[i].bias.grad, m2s[i].weight.grad], rank)
 
 
 if __name__ == '__main__':
@@ -75,4 +86,4 @@ def ef2(x, fec):
         locals()[sys.argv[1]](**args)
     else:
         # test_faster_schedule(8, 16, 16, 1, 2)
-        _test_faster_schedule(4, 2, 1)
+        _test_faster_schedule(4, 2, 4)

tests/test_faster_shadow.py

@@ -20,7 +20,7 @@
 @pytest.mark.parametrize("batch_size", [16, 512])
 @pytest.mark.parametrize("n_expert", [1])
 @pytest.mark.parametrize("group_sz", [1, 2, 4])
-@pytest.mark.parametrize("pass_stored", [False, True])
+@pytest.mark.parametrize("pass_stored", [True, False])
 def test_faster_shadow(n_process, d_model, batch_size, n_expert, group_sz, pass_stored):
     _run_distributed('_test_faster_shadow',
             n_process,
@@ -54,30 +54,31 @@ def _test_faster_shadow(d_model, batch_size, n_expert, pass_stored):
         m2.weight.copy_(m1.weight)
         m2.bias.copy_(m1.bias)
 
-    def ef1(x, fec):
+    def ef1(x, fec, eidx):
         y = m1(x)
         return y
     def ef2(x, fec):
         y = m2(x)
         return y
 
     if pass_stored:
-        stored_models = torch.randint(0, 2, (world_size,)).bool().cuda()
+        stored_models = torch.randint(0, 2, (world_size * n_expert,)).bool().cuda()
+        while stored_models.sum().item() == 0:
+            stored_models = torch.randint(0, 2, (world_size * n_expert,)).bool().cuda()
+        stored_models[-1] = True
         dist.broadcast(stored_models, 0)
         stored_models = stored_models.cpu()
-
-    # if rank == 0:
-         # print('stored models {}'.format(stored_models))
+        print(stored_models)
 
     ensure_comm(x1, None)
     if pass_stored:
-        y1 = smart_fwd(x1, topk_idx, ef1, n_expert, world_size, experts=m1, 
+        y1 = smart_fwd(x1, topk_idx, ef1, n_expert, world_size, experts=[m1],
                 stored_models=stored_models)
     else:
-        y1 = smart_fwd(x1, topk_idx, ef1, n_expert, world_size, experts=m1)
+        y1 = smart_fwd(x1, topk_idx, ef1, n_expert, world_size, experts=[m1])
     y1.sum().backward()
 
-    y2 = naive_fwd(x2, topk_idx, ef2, n_expert, world_size, experts=m2)
+    y2 = naive_fwd(x2, topk_idx, ef2, n_expert, world_size, experts=[m2])
     y2.sum().backward()
     _assert_numerical(['out', 'grad_in', 'grad_bias', 'grad_weight'],
             [y1, x1.grad, m1.bias.grad, m1.weight.grad],

tests/test_numerical.py

@@ -50,8 +50,12 @@ def _perform_forward(
 
 def _assert_numerical(names, moe_out_list, raw_out_list, rank, precision=1e-3):
     for name, mo, ro in zip(names, moe_out_list, raw_out_list):
+        if mo is None and ro is None:
+            continue
+        if mo is None or ro is None:
+            assert False
         err = (mo - ro).abs().max()
-        if err.dtype == torch.bfloat16:
+        if err.dtype == torch.bfloat16 or err.dtype == torch.float16:
             precision *= 100
         print("Rank {} {} abs err {}".format(rank, name, err))
         if err > precision:
@@ -93,7 +97,7 @@ def __init__(
 @pytest.mark.parametrize("mp_group", [None])
 @pytest.mark.parametrize("dp_group", [None])
 @pytest.mark.parametrize("world_group", [None])
-@pytest.mark.parametrize("data_type", ['torch.FloatTensor', 'torch.DoubleTensor', 'torch.HalfTensor'])
+@pytest.mark.parametrize("data_type", ['torch.float32', 'torch.bfloat16', 'torch.float16'])
 def test_fmoe_linear(
     num_expert,
     top_k,
@@ -111,6 +115,9 @@ def test_fmoe_linear(
     torch.manual_seed(42 + rank)
     torch.cuda.manual_seed(42 + rank)
 
+    if isinstance(data_type, str):
+        data_type = eval(data_type)
+
     moe = MyMoE(
         num_expert, d_model, d_hidden, world_size, mp_group, top_k, activation
     ).type(data_type).cuda()
@@ -238,6 +245,9 @@ def test_fmoe(
 
     if isinstance(expert, str):
         expert = globals()[expert]
+        assert(expert is not None)
+    if isinstance(data_type, str):
+        data_type = eval(data_type)
 
     moe = FMoE(
         num_expert=num_expert,
@@ -247,15 +257,15 @@ def test_fmoe(
         mp_group=mp_group,
         expert=expert,
         top_k=top_k,
-    ).cuda().to(data_type)
+    ).cuda().type(data_type)
 
     moe_raw = BruteForceMoE(
         expert=expert,
         num_expert=num_expert,
         d_model=d_model,
         world_size=world_size,
         top_k=top_k,
-    ).cuda().to(data_type)
+    ).cuda().type(data_type)
 
     if world_size == 1:
         for expert_moe, expert_raw in zip(moe.experts, moe_raw.experts):
@@ -266,9 +276,11 @@ def test_fmoe(
     else:
         assert len(moe.experts) >= 1
         for idx, para in enumerate(moe.experts[0].parameters()):
+            assert(para.device.type == 'cuda')
             para_tensor = torch.cat(
                 [list(expert.parameters())[idx].unsqueeze(0) for expert in moe.experts]
             )
+            assert(para_tensor.device.type == 'cuda')
             para_array = [torch.empty_like(para_tensor) for _ in range(world_size)]
             torch.distributed.all_gather(para_array, para_tensor)
             para_tensor_gathered = torch.cat(para_array, dim=0)
@@ -419,6 +431,8 @@ def test_fmoe_experts(
 
     if isinstance(expert, str):
         expert = globals()[expert]
+    if isinstance(data_type, str):
+        data_type = eval(data_type)
 
     moe = FMoE(
         num_expert=num_expert,
@@ -428,7 +442,7 @@ def test_fmoe_experts(
         mp_group=mp_group,
         expert=expert,
         top_k=top_k,
-    ).cuda().to(data_type)
+    ).cuda().type(data_type)
 
     moe_raw = BruteForceMoE(
         expert=expert,
@@ -447,9 +461,12 @@ def test_fmoe_experts(
     else:
         assert len(moe.experts) >= 1
         for idx, para in enumerate(moe.experts[0].parameters()):
+            for ep in expert.parameters():
+                assert(ep.device.type == 'cuda')
             para_tensor = torch.cat(
                 [list(expert.parameters())[idx].unsqueeze(0) for expert in moe.experts]
             )
+            assert(para_tensor.device.type == 'cuda')
             para_array = [torch.empty_like(para_tensor) for _ in range(world_size)]
             torch.distributed.all_gather(para_array, para_tensor)
             para_tensor_gathered = torch.cat(para_array, dim=0)