[Performance] Performance improvements in non-blockwise fp8 CUTLASS MoE (vllm-project#20762)

ElizaWszola · web-flow · commit 9fb2d22032ce · 2025-07-17T09:56:44.000-04:00
Signed-off-by: ElizaWszola &lt;ewszola@redhat.com&gt;
diff --git a/benchmarks/kernels/benchmark_grouped_gemm_cutlass.py b/benchmarks/kernels/benchmark_grouped_gemm_cutlass.py
@@ -80,6 +80,11 @@ def bench_run(
         a, score, topk, renormalize=False
     )
 
+    ab_strides1 = torch.full((num_experts,), k, device="cuda", dtype=torch.int64)
+    ab_strides2 = torch.full((num_experts,), n, device="cuda", dtype=torch.int64)
+    c_strides1 = torch.full((num_experts,), 2 * n, device="cuda", dtype=torch.int64)
+    c_strides2 = torch.full((num_experts,), k, device="cuda", dtype=torch.int64)
+
     def run_triton_moe(
         a: torch.Tensor,
         w1: torch.Tensor,
@@ -111,6 +116,10 @@ def run_cutlass_moe(
         w2: torch.Tensor,
         w1_scale: torch.Tensor,
         w2_scale: torch.Tensor,
+        ab_strides1: torch.Tensor,
+        ab_strides2: torch.Tensor,
+        c_strides1: torch.Tensor,
+        c_strides2: torch.Tensor,
         topk_weights: torch.Tensor,
         topk_ids: torch.Tensor,
         per_act_token: bool,
@@ -125,6 +134,10 @@ def run_cutlass_moe(
                 topk_ids,
                 w1_scale,
                 w2_scale,
+                ab_strides1,
+                ab_strides2,
+                c_strides1,
+                c_strides2,
                 per_act_token,
                 a1_scale=None,
             )
@@ -136,6 +149,10 @@ def run_cutlass_from_graph(
         w2_q: torch.Tensor,
         w1_scale: torch.Tensor,
         w2_scale: torch.Tensor,
+        ab_strides1: torch.Tensor,
+        ab_strides2: torch.Tensor,
+        c_strides1: torch.Tensor,
+        c_strides2: torch.Tensor,
         topk_weights: torch.Tensor,
         topk_ids: torch.Tensor,
     ):
@@ -150,6 +167,10 @@ def run_cutlass_from_graph(
                 topk_ids,
                 w1_scale,
                 w2_scale,
+                ab_strides1,
+                ab_strides2,
+                c_strides1,
+                c_strides2,
                 per_act_token,
                 a1_scale=None,
             )
@@ -194,6 +215,10 @@ def replay_graph(graph, num_repeats):
             w2_q,
             w1_scale,
             w2_scale,
+            ab_strides1,
+            ab_strides2,
+            c_strides1,
+            c_strides2,
             topk_weights,
             topk_ids,
         )
@@ -231,6 +256,10 @@ def replay_graph(graph, num_repeats):
         "w1_scale": w1_scale,
         "w2_scale": w2_scale,
         "per_act_token": per_act_token,
+        "ab_strides1": ab_strides1,
+        "ab_strides2": ab_strides2,
+        "c_strides1": c_strides1,
+        "c_strides2": c_strides2,
         # cuda graph params
         "cutlass_graph": cutlass_graph,
         "triton_graph": triton_graph,
@@ -289,6 +318,10 @@ def replay_graph(graph, num_repeats):
         w2_q,
         w1_scale,
         w2_scale,
+        ab_strides1,
+        ab_strides2,
+        c_strides1,
+        c_strides2,
         topk_weights,
         topk_ids,
         per_act_token,
@@ -297,7 +330,7 @@ def replay_graph(graph, num_repeats):
 
     results.append(
         benchmark.Timer(
-            stmt="run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, topk_weights, topk_ids, per_act_token, num_runs)",  # noqa: E501
+            stmt="run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, ab_strides1, ab_strides2, c_strides1, c_strides2, topk_weights, topk_ids, per_act_token, num_runs)",  # noqa: E501
             globals=globals,
             label=label,
             sub_label=sub_label,
diff --git a/csrc/moe/moe_permute_unpermute_op.cu b/csrc/moe/moe_permute_unpermute_op.cu
@@ -160,6 +160,30 @@ __global__ void shuffleInputRowsKernel(const T* input,
   }
 }
 
+template <typename T>
+__global__ void shuffleInputRowsKernelSlow(const T* input,
+                                           const int32_t* dst2src_map,
+                                           T* output, int64_t num_src_rows,
+                                           int64_t num_dst_rows,
+                                           int64_t num_cols) {
+  int64_t dest_row_idx = blockIdx.x;
+  int64_t const source_row_idx = dst2src_map[dest_row_idx];
+
+  if (blockIdx.x < num_dst_rows) {
+    // Duplicate and permute rows
+    auto const* source_row_ptr = input + source_row_idx * num_cols;
+    auto* dest_row_ptr = output + dest_row_idx * num_cols;
+
+    int64_t const start_offset = threadIdx.x;
+    int64_t const stride = blockDim.x;
+
+    for (int elem_index = start_offset; elem_index < num_cols;
+         elem_index += stride) {
+      dest_row_ptr[elem_index] = source_row_ptr[elem_index];
+    }
+  }
+}
+
 void shuffle_rows(const torch::Tensor& input_tensor,
                   const torch::Tensor& dst2src_map,
                   torch::Tensor& output_tensor) {
@@ -173,17 +197,24 @@ void shuffle_rows(const torch::Tensor& input_tensor,
   int64_t const num_src_rows = input_tensor.size(0);
   int64_t const num_cols = input_tensor.size(1);
 
-  TORCH_CHECK(!(num_cols % (128 / sizeof(input_tensor.scalar_type()) / 8)),
-              "num_cols must be divisible by 128 / "
-              "sizeof(input_tensor.scalar_type()) / 8");
-
-  MOE_DISPATCH(input_tensor.scalar_type(), [&] {
-    shuffleInputRowsKernel<scalar_t><<<blocks, threads, 0, stream>>>(
-        reinterpret_cast<scalar_t*>(input_tensor.data_ptr()),
-        dst2src_map.data_ptr<int32_t>(),
-        reinterpret_cast<scalar_t*>(output_tensor.data_ptr()), num_src_rows,
-        num_dest_rows, num_cols);
-  });
+  if (num_cols % (128 / sizeof(input_tensor.scalar_type()) / 8)) {
+    // use slow kernel if num_cols can't be aligned to 128 bits
+    MOE_DISPATCH(input_tensor.scalar_type(), [&] {
+      shuffleInputRowsKernelSlow<scalar_t><<<blocks, threads, 0, stream>>>(
+          reinterpret_cast<scalar_t*>(input_tensor.data_ptr()),
+          dst2src_map.data_ptr<int32_t>(),
+          reinterpret_cast<scalar_t*>(output_tensor.data_ptr()), num_src_rows,
+          num_dest_rows, num_cols);
+    });
+  } else {
+    MOE_DISPATCH(input_tensor.scalar_type(), [&] {
+      shuffleInputRowsKernel<scalar_t><<<blocks, threads, 0, stream>>>(
+          reinterpret_cast<scalar_t*>(input_tensor.data_ptr()),
+          dst2src_map.data_ptr<int32_t>(),
+          reinterpret_cast<scalar_t*>(output_tensor.data_ptr()), num_src_rows,
+          num_dest_rows, num_cols);
+    });
+  }
 }
 
 #else
diff --git a/tests/kernels/moe/test_cutlass_moe.py b/tests/kernels/moe/test_cutlass_moe.py
@@ -206,6 +206,10 @@ def run_8_bit(moe_tensors: MOETensors8Bit,
         'topk_ids': topk_ids,
         'w1_scale': moe_tensors.w1_scale,
         'w2_scale': moe_tensors.w2_scale,
+        'ab_strides1': moe_tensors.ab_strides1,
+        'ab_strides2': moe_tensors.ab_strides2,
+        'c_strides1': moe_tensors.c_strides1,
+        'c_strides2': moe_tensors.c_strides2,
         'per_act_token': per_act_token,
         'a1_scale': None  #moe_tensors.a_scale
     }
@@ -439,6 +443,11 @@ def test_run_cutlass_moe_fp8(
         expert_map[start:end] = list(range(num_local_experts))
         expert_map = torch.tensor(expert_map, dtype=torch.int32, device="cuda")
 
+        ab_strides1 = torch.full((e, ), k, device="cuda", dtype=torch.int64)
+        ab_strides2 = torch.full((e, ), n, device="cuda", dtype=torch.int64)
+        c_strides1 = torch.full((e, ), 2 * n, device="cuda", dtype=torch.int64)
+        c_strides2 = torch.full((e, ), k, device="cuda", dtype=torch.int64)
+
         activation = lambda o, i: torch.ops._C.silu_and_mul(o, i)
         a1q, a1q_scale = moe_kernel_quantize_input(mt.a, mt.a_scale,
                                                    torch.float8_e4m3fn,
@@ -447,8 +456,9 @@ def test_run_cutlass_moe_fp8(
         func = lambda output: run_cutlass_moe_fp8(
             output, a1q, mt.w1_q, mt.w2_q, topk_ids, activation,
             global_num_experts, expert_map, mt.w1_scale, mt.w2_scale,
-            a1q_scale, None, workspace13, workspace2, None, mt.a.dtype,
-            per_act_token, per_out_channel, False)
+            a1q_scale, None, ab_strides1, ab_strides2, c_strides1, c_strides2,
+            workspace13, workspace2, None, mt.a.dtype, per_act_token,
+            per_out_channel, False)
 
         workspace13.random_()
         output_random_workspace = torch.empty(output_shape,
diff --git a/tests/kernels/moe/test_pplx_cutlass_moe.py b/tests/kernels/moe/test_pplx_cutlass_moe.py
@@ -75,6 +75,7 @@ def pplx_cutlass_moe(
     assert torch.cuda.current_device() == pgi.local_rank
 
     num_tokens, hidden_dim = a.shape
+    intermediate_dim = w2.shape[2]
     num_experts = w1.shape[0]
     block_size = hidden_dim  # TODO support more cases
     device = pgi.device
@@ -123,10 +124,31 @@ def pplx_cutlass_moe(
         num_local_experts=num_local_experts,
         num_dispatchers=num_dispatchers)
 
+    ab_strides1 = torch.full((num_local_experts, ),
+                             hidden_dim,
+                             device="cuda",
+                             dtype=torch.int64)
+    ab_strides2 = torch.full((num_local_experts, ),
+                             intermediate_dim,
+                             device="cuda",
+                             dtype=torch.int64)
+    c_strides1 = torch.full((num_local_experts, ),
+                            2 * intermediate_dim,
+                            device="cuda",
+                            dtype=torch.int64)
+    c_strides2 = torch.full((num_local_experts, ),
+                            hidden_dim,
+                            device="cuda",
+                            dtype=torch.int64)
+
     experts = CutlassExpertsFp8(num_local_experts,
                                 out_dtype,
                                 per_act_token,
                                 per_out_ch,
+                                ab_strides1,
+                                ab_strides2,
+                                c_strides1,
+                                c_strides2,
                                 num_dispatchers=num_dispatchers,
                                 use_batched_format=True)
 
diff --git a/vllm/model_executor/layers/fused_moe/cutlass_moe.py b/vllm/model_executor/layers/fused_moe/cutlass_moe.py
@@ -13,8 +13,7 @@
     MoEPrepareAndFinalizeNoEP)
 from vllm.model_executor.layers.fused_moe.topk_weight_and_reduce import (
     TopKWeightAndReduceDelegate)
-from vllm.model_executor.layers.fused_moe.utils import (_fp8_perm,
-                                                        _fp8_quantize,
+from vllm.model_executor.layers.fused_moe.utils import (_fp8_quantize,
                                                         _resize_cache)
 from vllm.scalar_type import scalar_types
 
@@ -34,6 +33,10 @@ def run_cutlass_moe_fp8(
     w2_scale: Optional[torch.Tensor],
     a1q_scale: Optional[torch.Tensor],
     a2_scale: Optional[torch.Tensor],
+    ab_strides1: torch.Tensor,
+    ab_strides2: torch.Tensor,
+    c_strides1: torch.Tensor,
+    c_strides2: torch.Tensor,
     workspace13: torch.Tensor,
     workspace2: torch.Tensor,
     expert_num_tokens: Optional[torch.Tensor],
@@ -152,27 +155,11 @@ def run_cutlass_moe_fp8(
                                     problem_sizes1, problem_sizes2, a_map,
                                     c_map, global_num_experts, N, K)
 
-        a1q = _fp8_perm(a1q, a_map)
-        a1q_scale = a1q_scale[a_map] if per_act_token else a1q_scale
+        a1q = ops.shuffle_rows(a1q, a_map)
+        a1q_scale = (ops.shuffle_rows(a1q_scale, a_map)
+                     if per_act_token else a1q_scale)
         expert_offsets = expert_offsets[:-1]
 
-    ab_strides1 = torch.full((w1.size(0), ),
-                             K,
-                             device=device,
-                             dtype=torch.int64)
-    c_strides1 = torch.full((w1.size(0), ),
-                            2 * N,
-                            device=device,
-                            dtype=torch.int64)
-    ab_strides2 = torch.full((w1.size(0), ),
-                             N,
-                             device=device,
-                             dtype=torch.int64)
-    c_strides2 = torch.full((w1.size(0), ),
-                            K,
-                            device=device,
-                            dtype=torch.int64)
-
     if use_batched_format:
         c1 = _resize_cache(workspace13, (local_E * padded_M, N * 2))
         c2 = _resize_cache(workspace2, (local_E * padded_M, N))
@@ -209,7 +196,8 @@ def run_cutlass_moe_fp8(
     else:
         # We can't do this inplace because output may point to the same tensor
         # as c3.
-        output.copy_(c3[c_map].view(M * topk, K), non_blocking=True)
+        output.copy_(ops.shuffle_rows(c3, c_map).view(M * topk, K),
+                     non_blocking=True)
 
 
 # TODO (bnell): split class batched vs. non-batched?
@@ -222,6 +210,10 @@ def __init__(
         out_dtype: Optional[torch.dtype],
         per_act_token_quant: bool,
         per_out_ch_quant: bool,
+        ab_strides1: torch.Tensor,
+        ab_strides2: torch.Tensor,
+        c_strides1: torch.Tensor,
+        c_strides2: torch.Tensor,
         block_shape: Optional[list[int]] = None,
         num_dispatchers: Optional[int] = None,
         use_batched_format: bool = False,
@@ -238,6 +230,10 @@ def __init__(
         self.max_experts_per_worker = max_experts_per_worker
         self.num_dispatchers = num_dispatchers
         self.out_dtype = out_dtype
+        self.ab_strides1 = ab_strides1
+        self.ab_strides2 = ab_strides2
+        self.c_strides1 = c_strides1
+        self.c_strides2 = c_strides2
         self.use_batched_format = use_batched_format
 
     @property
@@ -316,7 +312,8 @@ def apply(self, output: torch.Tensor, hidden_states: torch.Tensor,
         run_cutlass_moe_fp8(
             output, hidden_states, w1, w2, topk_ids, activation_callable,
             global_num_experts, expert_map, w1_scale, w2_scale, a1q_scale,
-            a2_scale, workspace13, workspace2, expert_num_tokens,
+            a2_scale, self.ab_strides1, self.ab_strides2, self.c_strides1,
+            self.c_strides2, workspace13, workspace2, expert_num_tokens,
             self.out_dtype if self.out_dtype is not None else in_dtype,
             self.per_act_token_quant, self.per_out_ch_quant,
             self.use_batched_format)
@@ -330,6 +327,10 @@ def cutlass_moe_fp8(
     topk_ids: torch.Tensor,
     w1_scale: torch.Tensor,
     w2_scale: torch.Tensor,
+    ab_strides1: torch.Tensor,
+    ab_strides2: torch.Tensor,
+    c_strides1: torch.Tensor,
+    c_strides2: torch.Tensor,
     per_act_token: Optional[bool] = None,
     activation: str = "silu",
     a1_scale: Optional[torch.Tensor] = None,
@@ -357,6 +358,17 @@ def cutlass_moe_fp8(
         Shape: [num_experts] or [num_experts, 2N]
     - w2_scale (torch.Tensor): The fp32 scale to dequantize w2_q.
         Shape: [num_experts] or [num_experts, K]
+    - ab_strides1 (torch.Tensor): The input/weight strides for the first gemm.
+        Shape: [num_experts]
+    - ab_strides2 (torch.Tensor): The input/weight strides for the second gemm.
+        Shape: [num_experts]
+    - c_strides1 (torch.Tensor): The output strides for the first gemm.
+        Shape: [num_experts]
+    - c_strides2 (torch.Tensor): The output strides for the second gemm.
+        Shape: [num_experts]
+    - per_act_token (Optional[bool]): Whether the scale is per-token or
+                                      per-tensor.
+    - activation (str): The activation function to use.
     - a1_scale (Optional[torch.Tensor]): The optional fp32 scale to quantize a.
         Shape: scalar or [M]
     - a2_scale (Optional[torch.Tensor]): The optional fp32 scale to
@@ -389,6 +401,10 @@ def cutlass_moe_fp8(
             out_dtype=a.dtype,
             per_act_token_quant=per_act_token,
             per_out_ch_quant=per_out_ch,
+            ab_strides1=ab_strides1,
+            ab_strides2=ab_strides2,
+            c_strides1=c_strides1,
+            c_strides2=c_strides2,
             use_batched_format=False,
         ),
     )
diff --git a/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py b/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
