Fixes and cleanup

csrc/activation_kernels.cu

@@ -70,6 +70,9 @@ __device__ __forceinline__ T gelu_tanh_kernel(const T& x) {
   int64_t num_tokens = input.numel() / input.size(-1);                   \
   dim3 grid(num_tokens);                                                 \
   dim3 block(std::min(d, 1024));                                         \
+  if (num_tokens == 0) {                                                 \
+    return;                                                              \
+  }                                                                      \
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input));      \
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();          \
   VLLM_DISPATCH_FLOATING_TYPES(                                          \

csrc/dispatch_utils.h

@@ -65,5 +65,19 @@
   AT_DISPATCH_CASE(at::ScalarType::Int, __VA_ARGS__)   \
   AT_DISPATCH_CASE(at::ScalarType::Long, __VA_ARGS__)
 
+#define VLLM_DISPATCH_CASE_INTEGRAL_AND_UNSIGNED_TYPES(...) \
+  AT_DISPATCH_CASE(at::ScalarType::Byte, __VA_ARGS__)       \
+  AT_DISPATCH_CASE(at::ScalarType::Char, __VA_ARGS__)       \
+  AT_DISPATCH_CASE(at::ScalarType::Short, __VA_ARGS__)      \
+  AT_DISPATCH_CASE(at::ScalarType::Int, __VA_ARGS__)        \
+  AT_DISPATCH_CASE(at::ScalarType::Long, __VA_ARGS__)       \
+  AT_DISPATCH_CASE(at::ScalarType::UInt16, __VA_ARGS__)     \
+  AT_DISPATCH_CASE(at::ScalarType::UInt32, __VA_ARGS__)     \
+  AT_DISPATCH_CASE(at::ScalarType::UInt64, __VA_ARGS__)
+
 #define VLLM_DISPATCH_INTEGRAL_TYPES(TYPE, NAME, ...) \
   AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_INTEGRAL_TYPES(__VA_ARGS__))
+
+#define VLLM_DISPATCH_INTEGRAL_AND_UNSIGNED_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                              \
+      TYPE, NAME, VLLM_DISPATCH_CASE_INTEGRAL_AND_UNSIGNED_TYPES(__VA_ARGS__))

csrc/moe/moe_align_sum_kernels.cu

@@ -326,7 +326,7 @@ void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
   }
 
   if (use_global_memory) {
-    VLLM_DISPATCH_INTEGRAL_TYPES(
+    VLLM_DISPATCH_INTEGRAL_AND_UNSIGNED_TYPES(
         topk_ids.scalar_type(), "moe_align_block_size_global_mem_kernel", [&] {
           // calc needed amount of shared mem for `tokens_cnts` and `cumsum`
           // tensors
@@ -351,7 +351,7 @@ void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
               cumsum_buffer.data_ptr<int32_t>());
         });
   } else if (use_i16) {
-    VLLM_DISPATCH_INTEGRAL_TYPES(
+    VLLM_DISPATCH_INTEGRAL_AND_UNSIGNED_TYPES(
         topk_ids.scalar_type(), "moe_align_block_size_kernel", [&] {
           // set dynamic shared mem
           auto kernel =
@@ -366,7 +366,7 @@ void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
               topk_ids.numel());
         });
   } else {
-    VLLM_DISPATCH_INTEGRAL_TYPES(
+    VLLM_DISPATCH_INTEGRAL_AND_UNSIGNED_TYPES(
         topk_ids.scalar_type(), "moe_align_block_size_kernel", [&] {
           auto kernel =
               vllm::moe::moe_align_block_size_kernel<scalar_t, int32_t>;
@@ -391,7 +391,7 @@ void sgl_moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
   TORCH_CHECK(num_experts == 256,
               "sgl_moe_align_block_size kernel only supports deepseek v3.");
 
-  VLLM_DISPATCH_INTEGRAL_TYPES(
+  VLLM_DISPATCH_INTEGRAL_AND_UNSIGNED_TYPES(
       topk_ids.scalar_type(), "sgl_moe_align_block_size_kernel", [&] {
         // calc needed amount of shared mem for `cumsum` tensors
         auto options_int =

csrc/moe/topk_softmax_kernels.cu

@@ -108,9 +108,17 @@ __launch_bounds__(TPB) __global__
     }
 }
 
-template <int TPB>
-__launch_bounds__(TPB) __global__ void moeTopK(const float* inputs_after_softmax, const bool* finished, float* output,
-    int* indices, int* source_rows, const int num_experts, const int k, const int start_expert, const int end_expert)
+template <int TPB, typename IndType>
+__launch_bounds__(TPB) __global__ void moeTopK(
+    const float* inputs_after_softmax,
+    const bool* finished,
+    float* output,
+    IndType* indices,
+    int* source_rows,
+    const int num_experts,
+    const int k,
+    const int start_expert,
+    const int end_expert)
 {
 
     using cub_kvp = cub::KeyValuePair<int, float>;
@@ -182,9 +190,9 @@ __launch_bounds__(TPB) __global__ void moeTopK(const float* inputs_after_softmax
   2) This implementation assumes k is small, but will work for any k.
 */
 
-template <int VPT, int NUM_EXPERTS, int WARPS_PER_CTA, int BYTES_PER_LDG>
+template <int VPT, int NUM_EXPERTS, int WARPS_PER_CTA, int BYTES_PER_LDG, typename IndType>
 __launch_bounds__(WARPS_PER_CTA* WARP_SIZE) __global__
-    void topkGatingSoftmax(const float* input, const bool* finished, float* output, const int num_rows, int* indices,
+    void topkGatingSoftmax(const float* input, const bool* finished, float* output, const int num_rows, IndType* indices,
         int* source_rows, const int k, const int start_expert, const int end_expert)
 {
     // We begin by enforcing compile time assertions and setting up compile time constants.
@@ -397,8 +405,8 @@ struct TopkConstants
 };
 } // namespace detail
 
-template <int EXPERTS, int WARPS_PER_TB>
-void topkGatingSoftmaxLauncherHelper(const float* input, const bool* finished, float* output, int* indices,
+template <int EXPERTS, int WARPS_PER_TB, typename IndType>
+void topkGatingSoftmaxLauncherHelper(const float* input, const bool* finished, float* output, IndType* indices,
     int* source_row, const int num_rows, const int k, const int start_expert, const int end_expert, cudaStream_t stream)
 {
     static constexpr std::size_t MAX_BYTES_PER_LDG = 16;
@@ -421,10 +429,11 @@ void topkGatingSoftmaxLauncherHelper(const float* input, const bool* finished, f
         token_expert_indices, num_tokens, topk, 0, num_experts,         \
         stream);
 
+template <typename IndType>
 void topkGatingSoftmaxKernelLauncher(
     const float* gating_output,
     float* topk_weights,
-    int* topk_indicies,
+    IndType* topk_indicies,
     int* token_expert_indices,
     float* softmax_workspace,
     const int num_tokens,
@@ -493,14 +502,32 @@ void topk_softmax(
     const at::cuda::OptionalCUDAGuard device_guard(device_of(gating_output));
     const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
     torch::Tensor softmax_workspace = torch::empty({workspace_size}, gating_output.options());
-    vllm::moe::topkGatingSoftmaxKernelLauncher(
-        gating_output.data_ptr<float>(),
-        topk_weights.data_ptr<float>(),
-        topk_indices.data_ptr<int>(),
-        token_expert_indices.data_ptr<int>(),
-        softmax_workspace.data_ptr<float>(),
-        num_tokens,
-        num_experts,
-        topk,
-        stream);
+
+    if(topk_indices.scalar_type() == at::ScalarType::Int)
+    {
+        vllm::moe::topkGatingSoftmaxKernelLauncher(
+            gating_output.data_ptr<float>(),
+            topk_weights.data_ptr<float>(),
+            topk_indices.data_ptr<int>(),
+            token_expert_indices.data_ptr<int>(),
+            softmax_workspace.data_ptr<float>(),
+            num_tokens,
+            num_experts,
+            topk,
+            stream);
+    }
+    else
+    {
+        assert(topk_indices.scalar_type() == at::ScalarType::UInt32);
+        vllm::moe::topkGatingSoftmaxKernelLauncher(
+            gating_output.data_ptr<float>(),
+            topk_weights.data_ptr<float>(),
+            topk_indices.data_ptr<uint32_t>(),
+            token_expert_indices.data_ptr<int>(),
+            softmax_workspace.data_ptr<float>(),
+            num_tokens,
+            num_experts,
+            topk,
+            stream);
+    }
 }

examples/offline_inference/data_parallel.py

@@ -65,11 +65,17 @@ def parse_args():
                         type=int,
                         default=0,
                         help="Master node port")
+    parser.add_argument("--enforce-eager",
+                        action='store_true',
+                        help="Enforce eager mode execution.")
+    parser.add_argument("--trust-remote-code",
+                        action='store_true',
+                        help="Trust remote code.")
     return parser.parse_args()
 
 
 def main(model, dp_size, local_dp_rank, global_dp_rank, dp_master_ip,
-         dp_master_port, GPUs_per_dp_rank):
+         dp_master_port, GPUs_per_dp_rank, enforce_eager, trust_remote_code):
     os.environ["VLLM_DP_RANK"] = str(global_dp_rank)
     os.environ["VLLM_DP_RANK_LOCAL"] = str(local_dp_rank)
     os.environ["VLLM_DP_SIZE"] = str(dp_size)
@@ -109,10 +115,13 @@ def main(model, dp_size, local_dp_rank, global_dp_rank, dp_master_ip,
                                      max_tokens=[16, 20][global_dp_rank % 2])
 
     # Create an LLM.
-    llm = LLM(model=model,
-              tensor_parallel_size=GPUs_per_dp_rank,
-              enforce_eager=True,
-              enable_expert_parallel=True)
+    llm = LLM(
+        model=model,
+        tensor_parallel_size=GPUs_per_dp_rank,
+        enforce_eager=enforce_eager,
+        enable_expert_parallel=True,
+        trust_remote_code=trust_remote_code,
+    )
     outputs = llm.generate(prompts, sampling_params)
     # Print the outputs.
     for i, output in enumerate(outputs):
@@ -155,7 +164,8 @@ def main(model, dp_size, local_dp_rank, global_dp_rank, dp_master_ip,
         proc = Process(target=main,
                        args=(args.model, dp_size, local_dp_rank,
                              global_dp_rank, dp_master_ip, dp_master_port,
-                             tp_size))
+                             tp_size, args.enforce_eager,
+                             args.trust_remote_code))
         proc.start()
         procs.append(proc)
     exit_code = 0

tests/kernels/moe/test_batched_moe.py

@@ -0,0 +1,107 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from dataclasses import dataclass
+
+import pytest
+import torch
+import triton.language as tl
+
+from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
+    invoke_moe_batched_triton_kernel)
+
+
+@dataclass
+class BatchedMMConfig:
+    dtype: torch.dtype
+    num_experts: int
+    max_tokens_per_expert: int
+    K: int
+    N: int
+
+
+@dataclass
+class BatchedMMTensors:
+    A: torch.Tensor  # [E, max_tokens, K]
+    B: torch.Tensor  # [E, K, N] - column major
+    C: torch.Tensor  # [E, max_tokens, N]
+    num_expert_tokens: torch.Tensor  # [E]
+
+    @staticmethod
+    def make_tensors(config: BatchedMMConfig):
+        A = torch.randn(
+            (config.num_experts, config.max_tokens_per_expert, config.K),
+            device="cuda",
+            dtype=config.dtype) / 50.0
+        B = torch.randn((config.num_experts, config.N, config.K),
+                        device="cuda",
+                        dtype=config.dtype) / 50.0
+        C = torch.zeros(
+            (config.num_experts, config.max_tokens_per_expert, config.N),
+            device="cuda",
+            dtype=config.dtype)
+        num_expert_tokens = torch.randint(low=0,
+                                          high=config.max_tokens_per_expert,
+                                          size=(config.num_experts, ),
+                                          device="cuda",
+                                          dtype=torch.int32)
+        return BatchedMMTensors(A, B, C, num_expert_tokens)
+
+
+def ref_impl(A: torch.Tensor, B: torch.Tensor, C: torch.Tensor,
+             num_expert_tokens: torch.Tensor) -> torch.Tensor:
+
+    num_expert_tokens_cpu = num_expert_tokens.clone()
+    num_expert_tokens_cpu = num_expert_tokens_cpu.to(device="cpu")
+    num_experts = num_expert_tokens.size(0)
+
+    for e in range(num_experts):
+        num_tokens = num_expert_tokens_cpu[e]
+        C[e, :num_tokens, :] = A[e, :num_tokens, :] @ B[e].transpose(0, 1)
+
+    return C
+
+
+@pytest.mark.parametrize("num_experts", [16, 32])
+@pytest.mark.parametrize("max_tokens_per_expert",
+                         [32, 64, 128, 192, 224, 256, 512])
+@pytest.mark.parametrize("K", [128, 256, 1024])
+@pytest.mark.parametrize("N", [128, 256, 512, 1024])
+@pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16])
+def test_batched_mm(num_experts: int, max_tokens_per_expert: int, K: int,
+                    N: int, dtype: torch.dtype):
+
+    config = BatchedMMConfig(dtype, num_experts, max_tokens_per_expert, K, N)
+    tensors = BatchedMMTensors.make_tensors(config)
+
+    test_output = tensors.C
+    ref_output = test_output.clone()
+
+    compute_tl_dtype = {
+        torch.float16: tl.float16,
+        torch.bfloat16: tl.bfloat16,
+        torch.float32: tl.float32
+    }[test_output.dtype]
+    invoke_moe_batched_triton_kernel(
+        tensors.A,
+        tensors.B,
+        test_output,
+        tensors.num_expert_tokens,
+        compute_tl_dtype,
+        # Quantization data
+        None,
+        None,
+        None,
+        # Quantization schemes
+        False,
+        False,
+        False,
+        config={
+            "BLOCK_SIZE_M": 16,
+            "BLOCK_SIZE_N": 16,
+            "BLOCK_SIZE_K": 16
+        })
+
+    ref_output = ref_impl(tensors.A, tensors.B, ref_output,
+                          tensors.num_expert_tokens)
+
+    torch.testing.assert_close(test_output, ref_output, atol=1e-3, rtol=1e-3)