Merge pull request #21 from bias_improvement

Bias improvement #15

cuda/moe.cpp

@@ -32,37 +32,44 @@ std::vector<torch::Tensor> moe_local_gather(
     return moe_cuda_local_gather(output_buf, pos);
 }
 
-
 std::vector<torch::Tensor> moe_forward(
         torch::Tensor input_buf,     		// [batch_size x in_feat]
+        torch::Tensor expert_count,  		// [num_expert]
         torch::Tensor weight,        		// [num_expert x out_feat x in_feat]
-        torch::Tensor expert_count   // [batch_size]
+        at::optional<torch::Tensor> bias_o  // [num_expert x out_feat] or None
         ) {
+
     CHECK_INPUT(input_buf);
     CHECK_INPUT(weight);
-    /*
-        The bias term should have been merged into weight. Note the following fact that 
-        Wx+b = [W b] [x]
-                     [1]  
-    */
-    return moe_cuda_forward(input_buf, weight, expert_count);
+    
+    // check if bias is valid in case it exists
+    if (bias_o.has_value()) {
+        auto bias = bias_o.value();
+        CHECK_INPUT(bias);
+    }
+    
+    return moe_cuda_forward(input_buf, expert_count, weight, bias_o);
 }
 
 std::vector<torch::Tensor> moe_backward(
         torch::Tensor grad_output_buf, 		// [batch_size x out_feat]
-        torch::Tensor input_buf,       // [batch_size x out_feat]
+        torch::Tensor input_buf,       		// [batch_size x in_feat]
+        torch::Tensor expert_count,         // [num_expert]
         torch::Tensor weight,           	// [num_expert x out_feat x in_feat]
-        torch::Tensor expert_count
+        at::optional<torch::Tensor> bias_o  // [num_expert x out_feat] or None
         ) {
+    
     CHECK_INPUT(grad_output_buf);
     CHECK_INPUT(input_buf);
     CHECK_INPUT(weight);
-    /*
-        The bias term should have been merged into weight. Note the following fact that 
-        Wx+b = [W b] [x]
-                     [1]  
-    */
-    return moe_cuda_backward(grad_output_buf, input_buf, weight, expert_count);
+
+    // check if bias is valid in case it exists
+    if (bias_o.has_value()) {
+        auto bias = bias_o.value();
+        CHECK_INPUT(bias);
+    }
+    
+    return moe_cuda_backward(grad_output_buf, input_buf, expert_count, weight, bias_o);
 }
 
 #ifdef MOE_USE_NCCL

cuda/moe_compute_kernel.cu

@@ -37,6 +37,58 @@ void batch_scatter_kernel(size_t wid, const long* pos,
     }
 }
 
+
+/*
+    This function is to be called with one block per each column
+*/
+template <typename scalar_t>
+__global__ 
+void column_reduce(const scalar_t * matrix, scalar_t * result, 
+    int m /* lines */, int n /* columns*/) {
+    
+    // https://stackoverflow.com/questions/27570552/templated-cuda-kernel-with-dynamic-shared-memory
+    extern __shared__ unsigned char my_smem[];
+    scalar_t *sdata = reinterpret_cast<scalar_t *>(my_smem);
+
+    // normal tid
+    int tid = threadIdx.x + threadIdx.y * blockDim.x;
+    
+    // transposed tid for shared memory
+    int new_tid = threadIdx.y + threadIdx.x * blockDim.y;
+
+    // true x value in the matrix
+    int real_x = threadIdx.x + blockDim.x * blockIdx.x;
+    
+    int i = real_x + n * threadIdx.y;
+    const int it = n*blockDim.y;
+    int offset = it;
+    float accumulator = 0;
+
+    if (threadIdx.y < m && real_x < n) {
+        // store all the values from this column in a warped way
+        accumulator = matrix[i];
+        while (i + offset < n*m) {
+            accumulator += matrix[i + offset];
+            offset += it;
+        }
+    }
+
+    // save column reduction data in a transposed way
+    sdata[new_tid] = accumulator;
+    __syncthreads();
+
+    for (size_t t= 16; t > 0; t>>=1) {
+        if (tid < 32 * 32 - 16)
+            sdata[tid] += sdata[tid + t];
+        __syncthreads();
+    }
+    
+    if (threadIdx.y == 0 && real_x < n) 
+        result[real_x] = sdata[new_tid];
+    
+}
+
+
 void moe_cuda_expert_count_impl(
         const int* d_gate,
         int* expert_count,
@@ -118,11 +170,12 @@ void moe_cuda_forward_impl(
         const scalar_t* weight,
         const long* expert_count,
         scalar_t* output_buf,
+        const bool has_bias,
         const size_t in_feat,
         const size_t out_feat,
         const size_t num_expert,
         CudaStreamManager* smgr) {
-	scalar_t alpha = 1, beta = 0; 
+    scalar_t alpha = 1, beta = has_bias ? 1 : 0; 
 
     for (int i = 0, ptr = 0; i < num_expert; ++i) {
         if (expert_count[i] == 0) {
@@ -154,6 +207,8 @@ void moe_cuda_backward_impl(
         const long* expert_count,
         scalar_t* grad_input_buf,
         scalar_t* grad_weight,
+        scalar_t* grad_bias,
+        const bool has_bias,
         const size_t batch_size,
         const size_t in_feat,
         const size_t out_feat,
@@ -161,10 +216,16 @@ void moe_cuda_backward_impl(
         CudaStreamManager* smgr) {
     scalar_t alpha = 1, beta = 0;
 
+    // bias
+    dim3 block_threads(32, 32);
+    dim3 grid_threads(out_feat / 32 + (out_feat % 32 ? 1 : 0), 1);
+    
+
     for (int i = 0, ptr = 0; i < num_expert; ++i) {
         if (expert_count[i] == 0) {
             cudaMemset(grad_weight + i * in_feat * out_feat, 0, 
                     sizeof(scalar_t) * in_feat * out_feat);
+            cudaMemset(grad_bias + i * out_feat, 0, sizeof(scalar_t) * out_feat);
             continue;
         }
         // Use T(B) x T(A) = T(C) to produce row-major C
@@ -195,6 +256,17 @@ void moe_cuda_backward_impl(
                 grad_weight + i * in_feat * out_feat, in_feat
                 ));
         
+        if (has_bias) {
+            column_reduce
+            <<<grid_threads, block_threads, sizeof(scalar_t)*1024, smgr->stream(0)>>>
+            (
+                grad_output_buf + ptr * out_feat,
+                grad_bias + i * out_feat,
+                expert_count[i],
+                out_feat
+            );
+        }
+
         ptr += expert_count[i];
     }
     smgr->sync(num_expert);
@@ -275,8 +347,9 @@ std::vector<torch::Tensor> moe_cuda_local_gather(
 
 std::vector<torch::Tensor> moe_cuda_forward(
         torch::Tensor input_buf,
+        torch::Tensor expert_count,
         torch::Tensor weight,
-		torch::Tensor expert_count
+        at::optional<torch::Tensor> bias
         ) {
     auto smgr = getCudaStreamManager(input_buf.device().index());
     const auto batch_size = input_buf.size(0);
@@ -288,10 +361,17 @@ std::vector<torch::Tensor> moe_cuda_forward(
     printf("[forward] expert=%ld, in_feat (d_model)=%ld, out_feat (d_ffn)=%ld\n", 
             num_expert, in_feat, out_feat);
 #endif
+
+    torch::Tensor output;
+    
+    if (bias.has_value()) {
+        output = bias.value().repeat_interleave(expert_count.to(bias.value().device()), 0);
+    } else{
         auto out_options = torch::TensorOptions()
             .device(input_buf.device())
             .dtype(input_buf.dtype());
-    auto output = torch::empty({batch_size, out_feat}, out_options);
+        output = torch::empty({batch_size, out_feat}, out_options);
+    }
         
     AT_DISPATCH_FLOATING_TYPES_AND_HALF(input_buf.scalar_type(), "moe_forward_cuda", 
             ([&] {
@@ -300,6 +380,7 @@ std::vector<torch::Tensor> moe_cuda_forward(
             weight.data_ptr<scalar_t>(),
             expert_count.data_ptr<long>(),
             output.data_ptr<scalar_t>(),
+            bias.has_value(),
             in_feat,
             out_feat,
             num_expert,
@@ -313,8 +394,9 @@ std::vector<torch::Tensor> moe_cuda_forward(
 std::vector<torch::Tensor> moe_cuda_backward(
     torch::Tensor grad_output_buf, 	// [batch_size x out_feat]
     torch::Tensor input_buf, 		// [batch_size x out_feat]
+    torch::Tensor expert_count,
     torch::Tensor weight, 			// [num_expert x out_feat x in_feat]
-	torch::Tensor expert_count
+    at::optional<torch::Tensor> bias
 ) {
     auto smgr = getCudaStreamManager(input_buf.device().index());
     const auto batch_size = input_buf.size(0);
@@ -330,6 +412,7 @@ std::vector<torch::Tensor> moe_cuda_backward(
 
     auto grad_input_buf = grad_output_buf.new_empty({batch_size, in_feat}); 
     auto grad_weight = grad_output_buf.new_empty({num_expert, out_feat, in_feat});
+    auto grad_bias = grad_output_buf.new_empty({num_expert, out_feat});
 
     AT_DISPATCH_FLOATING_TYPES_AND_HALF(input_buf.scalar_type(), "moe_cuda_backward", ([&] {
         moe_cuda_backward_impl<scalar_t>(
@@ -339,6 +422,8 @@ std::vector<torch::Tensor> moe_cuda_backward(
             expert_count.data_ptr<long>(),
             grad_input_buf.data_ptr<scalar_t>(),
             grad_weight.data_ptr<scalar_t>(),
+            grad_bias.data_ptr<scalar_t>(),
+            bias.has_value(),
             batch_size,
             in_feat,
             out_feat,
@@ -347,5 +432,5 @@ std::vector<torch::Tensor> moe_cuda_backward(
         );
     }));
 
-    return {grad_input_buf, grad_weight};
+    return {grad_input_buf, grad_weight, grad_bias};
 }

cuda/moe_cuda_kernel.h

@@ -19,14 +19,16 @@ std::vector<torch::Tensor> moe_cuda_local_gather(
 
 std::vector<torch::Tensor> moe_cuda_forward(
     torch::Tensor input_buf,
+	torch::Tensor expert_count,
     torch::Tensor weight,
-	torch::Tensor expert_count);
+	at::optional<torch::Tensor> bias);
 
 std::vector<torch::Tensor> moe_cuda_backward(
     torch::Tensor grad_output_buf,
     torch::Tensor input_buf,
+	torch::Tensor expert_count,
     torch::Tensor weight,
-	torch::Tensor expert_count);
+	at::optional<torch::Tensor> bias);
 
 #ifdef MOE_USE_NCCL
 

fmoe/functions.py

@@ -110,21 +110,25 @@ class MOELinear(Function):
     """
 
     @staticmethod
-    def forward(ctx, global_input_buf, weight, fwd_expert_count):
+    def forward(ctx, global_input_buf, fwd_expert_count, weight, bias=None):
         (global_output_buf,) = fmoe_cuda.forward(
-            global_input_buf, weight, fwd_expert_count
+            global_input_buf, fwd_expert_count, weight, bias
         )
-        variables = (global_input_buf, weight, fwd_expert_count)
+        variables = (global_input_buf, fwd_expert_count, weight, bias)
         ctx.save_for_backward(*variables)
         return global_output_buf
 
     @staticmethod
     def backward(ctx, grad_out):
-        (input_buf, weight, fwd_expert_count) = ctx.saved_tensors
-        grad_inp_buf, grad_weight = fmoe_cuda.backward(
-            grad_out, input_buf, weight, fwd_expert_count
+        (input_buf, fwd_expert_count, weight, bias) = ctx.saved_tensors
+        grad_inp_buf, grad_weight, grad_bias = fmoe_cuda.backward(
+            grad_out, input_buf, fwd_expert_count, weight, bias
         )
-        return grad_inp_buf, grad_weight, None
+
+        if not torch.is_tensor(bias):
+            grad_bias = None
+
+        return grad_inp_buf, None, grad_weight, grad_bias
 
 
 class MOEGather(Function):

fmoe/layers.py

@@ -41,37 +41,7 @@ class FMoELinear(nn.Module):
         r"""
         Call MOE function
         """
-        x = MOELinear.apply(inp, self.weight, fwd_expert_count)
-        if self.bias is not None:
-            # TODO: torch.repeat_interleave seems have numerical
-            # instability in backward, leading to incorrect
-            # gradient computation for solution 1 and 2.
-            # Solution 3 uses a for-loop to expand the bias,
-            # but is 50% slower.
-            # This part should finally goes to MOELinear.apply,
-            # like MOELinear.apply(x, weight, bias, count)
-
-            # Solution 1
-            bias = torch.repeat_interleave(
-                self.bias, fwd_expert_count.to(self.bias.device), dim=0
-            )
-
-            # Solution 2
-            # bias_idx = torch.arange(self.num_expert)\
-            #     .repeat_interleave(fwd_expert_count)
-            # bias = self.bias[bias_idx]
-
-            # Solution 3
-            # bias = []
-            # for i in range(self.num_expert):
-            #    if fwd_expert_count[i] > 0:
-            #        bias.append(
-            #            self.bias[i].unsqueeze(0).expand(
-            #                fwd_expert_count[i], -1
-            #            )
-            #        )
-            # bias = torch.cat(bias, dim=0)
-            x = x + bias
+        x = MOELinear.apply(inp, fwd_expert_count, self.weight, self.bias)
         return x
 
     def extra_repr(self) -> str:

tests/test_ddp.py

@@ -41,8 +41,9 @@ def _run_distributed(func, world_size, args: 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'])
 def test_fmoe_linear_distributed(
-    num_expert, top_k, batch_size, d_model, d_hidden, mp_size
+    num_expert, top_k, batch_size, d_model, d_hidden, mp_size, data_type
 ):
     _run_distributed(
         "_test_fmoe_linear",
@@ -54,6 +55,7 @@ def test_fmoe_linear_distributed(
             "d_model": d_model,
             "d_hidden": d_hidden,
             "mp_size": mp_size,
+            "data_type": data_type
         },
     )
 
@@ -120,5 +122,6 @@ if __name__ == "__main__":
     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
+            num_expert=4, top_k=2, batch_size=4, d_model=8, d_hidden=8, mp_size=2,
+            data_type="torch.HalfTensor"
         )

tests/test_numerical.py

@@ -17,15 +17,15 @@ from moe import BruteForceMoELinear, BruteForceMoE, NaiveExpert, LinearExpert
 
 
 def _perform_forward(
-    moe: nn.Module, moe_raw: nn.Module, batch_size, d_model, top_k, rank, mp_group
+    moe: nn.Module, moe_raw: nn.Module, batch_size, d_model, top_k, rank, mp_group, data_type='torch.FloatTensor'
 ):
     moe.zero_grad()
     moe_raw.zero_grad()
-    if not mp_group:
-        inp = torch.rand(batch_size, d_model).cuda()
-    else:
+
+    inp = torch.rand(batch_size, d_model).type(data_type).cuda()
+        
+    if mp_group:
         group_sender = rank // mp_group.size() * mp_group.size()
-        inp = torch.rand(batch_size, d_model).cuda()
         torch.distributed.broadcast(inp, group_sender, group=mp_group)
         torch.distributed.broadcast(
             moe.gate.gate.weight.data, group_sender, group=mp_group
@@ -49,15 +49,17 @@ def _perform_forward(
     return moe_out, raw_out, inp.grad, inp_raw.grad
 
 
-def _assert_numercial(names, moe_out_list, raw_out_list, rank):
+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):
         err = (mo - ro).abs().sum()
         print("Rank {} {} abs err {}".format(rank, name, err))
-        if err > 1e-3:
+        if err > precision:
             sys.stderr.write(f"=========== {name} moe out ==============\n")
             sys.stderr.write("{}\n".format(mo))
             sys.stderr.write(f"=========== {name} raw out ==============\n")
             sys.stderr.write("{}\n".format(ro))
+            sys.stderr.write(f"=========== {name} diff ==============\n")
+            sys.stderr.write("{}\n{}\n".format((mo - ro).abs(), err))
             assert False
 
 
@@ -90,6 +92,7 @@ class MyMoE(FMoE):
 @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'])
 def test_fmoe_linear(
     num_expert,
     top_k,
@@ -101,6 +104,7 @@ def test_fmoe_linear(
     mp_group,
     dp_group,
     world_group,
+    data_type,
     activation=torch.nn.functional.gelu,
 ):
     torch.manual_seed(42 + rank)
@@ -108,7 +112,7 @@ def test_fmoe_linear(
 
     moe = MyMoE(
         num_expert, d_model, d_hidden, world_size, mp_group, top_k, activation
-    ).cuda()
+    ).type(data_type).cuda()
 
     moe_raw = BruteForceMoELinear(
         activation=activation,
@@ -117,7 +121,7 @@ def test_fmoe_linear(
         d_hidden=d_hidden,
         world_size=world_size,
         top_k=top_k,
-    ).cuda()
+    ).type(data_type).cuda()
 
     if world_size == 1:
         moe_raw.weight_htoh4.data = moe.experts.htoh4.weight.data.clone()
@@ -148,7 +152,7 @@ def test_fmoe_linear(
         moe_raw.bias_h4toh.data = torch.cat(bias_h4toh_array, dim=0)
 
     moe_out, raw_out, moe_grad_in, raw_grad_in = _perform_forward(
-        moe, moe_raw, batch_size, d_model, top_k, rank, mp_group
+        moe, moe_raw, batch_size, d_model, top_k, rank, mp_group, data_type=data_type
     )
 
     moe_out_list = (
@@ -198,7 +202,10 @@ def test_fmoe_linear(
         "h4toh bias grad",
     ]
 
-    _assert_numercial(names, moe_out_list, raw_out_list, rank)
+
+    precision = 5e-1 if data_type == 'torch.HalfTensor' else 1e-3
+        
+    _assert_numerical(names, moe_out_list, raw_out_list, rank, precision=precision)
 
 
 @pytest.mark.parametrize("batch_size", [4])
@@ -299,7 +306,7 @@ def test_fmoe(
     raw_out_list = [raw_out, raw_grad, raw_grad_in]
     names = ["forward", "backward", "grad_in"]
 
-    _assert_numercial(names, moe_out_list, raw_out_list, rank)
+    _assert_numerical(names, moe_out_list, raw_out_list, rank)
 
 
 class MyModule(nn.Module):
@@ -375,7 +382,7 @@ def _test_fmoe_local_ddp(rank, world_size, mp_group, dp_group, world_group):
 
     names = ["mp grad", "dp grad", "wp grad"]
 
-    _assert_numercial(names, ddp_out_list, raw_out_list, rank)
+    _assert_numerical(names, ddp_out_list, raw_out_list, rank)
 
 
 if __name__ == "__main__":