Add topk logits torch op for DS3.2. (#25945)

Signed-off-by: Daniel Campora <961215+dcampora@users.noreply.github.com>
Signed-off-by: Daniel Cámpora <961215+dcampora@users.noreply.github.com>
Co-authored-by: youkaichao <youkaichao@gmail.com>

csrc/ops.h

@@ -100,6 +100,11 @@ void apply_repetition_penalties_(torch::Tensor& logits,
                                  const torch::Tensor& output_mask,
                                  const torch::Tensor& repetition_penalties);
 
+void top_k_per_row(const torch::Tensor& logits, const torch::Tensor& rowStarts,
+                   const torch::Tensor& rowEnds, torch::Tensor& indices,
+                   torch::Tensor& values, int64_t numRows, int64_t stride0,
+                   int64_t stride1);
+
 void rms_norm_static_fp8_quant(torch::Tensor& out, torch::Tensor& input,
                                torch::Tensor& weight, torch::Tensor& scale,
                                double epsilon);

csrc/sampler.cu

@@ -44,6 +44,245 @@ __global__ void apply_repetition_penalties_kernel(
   }
 }
 
+static inline __device__ uint16_t extractBinIdx(float x) {
+  union {
+    __half h;
+    uint16_t u16;
+  } tmp;
+  tmp.h = __float2half_rn(x);
+  tmp.u16 = (x < 0.f) ? (~tmp.u16 & 0xffff) : (tmp.u16 | 0x8000);
+  return 511 - (tmp.u16 >> 7);
+}
+
+template <int kNumThreadsPerBlock = 512>
+static __global__ void topKPerRow(const float* logits, const int* rowStarts,
+                                  const int* rowEnds, int* outIndices,
+                                  float* outLogits, int stride0, int stride1) {
+  // The number of bins in the histogram.
+  static constexpr int kNumBins = 512;
+
+  // The top-k width.
+  static constexpr int kTopK = 2048;
+  // The number of elements per thread for the final top-k sort.
+  static constexpr int kNumTopKItemsPerThread = kTopK / kNumThreadsPerBlock;
+  // The class to sort the elements during the final top-k sort.
+  using TopKSort = cub::BlockRadixSort<float, kNumThreadsPerBlock,
+                                       kNumTopKItemsPerThread, int>;
+
+  // The number of slots for the final pass.
+  static constexpr int kNumFinalItems = 3072;
+  // The number of elements per thread for the final sort.
+  static constexpr int kNumFinalItemsPerThread =
+      kNumFinalItems / kNumThreadsPerBlock;
+  // The class to sort the elements during the final pass.
+  using FinalSort = cub::BlockRadixSort<float, kNumThreadsPerBlock,
+                                        kNumFinalItemsPerThread, int>;
+
+  // The class to compute the inclusive prefix-sum over the histogram.
+  using Scan = cub::BlockScan<int, kNumThreadsPerBlock>;
+
+  // Shared memory to compute the block scan.
+  __shared__ typename Scan::TempStorage smemScan;
+
+  // The structure to store the final items (for the final pass).
+  struct FinalItems {
+    // Shared memory to store the indices for the final pass.
+    int indices[kNumFinalItems];
+    // Shared memory to store the logits for the final pass.
+    float logits[kNumFinalItems];
+  };
+
+  // Shared memory to compute the block sort.
+  __shared__ union {
+    FinalItems items;
+    typename FinalSort::TempStorage finalSort;
+    typename TopKSort::TempStorage topKSort;
+  } smemFinal;
+
+  // Shared memory to store the histogram.
+  __shared__ int smemHistogram[kNumBins];
+  // Shared memory to store the selected indices.
+  __shared__ int smemIndices[kTopK];
+  // Shared memory to store the selected logits.
+  __shared__ float smemLogits[kTopK];
+  // Shared memory to store the threshold bin.
+  __shared__ int smemThresholdBinIdx[1];
+  // Shared memory counter to register the candidates for the final phase.
+  __shared__ int smemFinalDstIdx[1];
+
+  // The row computed by this block.
+  int rowIdx = blockIdx.x;
+  // The range of logits within the row.
+  int rowStart = rowStarts[rowIdx], rowEnd = rowEnds[rowIdx];
+  // The length of the row.
+  int rowLen = rowEnd - rowStart;
+
+  // Shortcut if the length of the row is smaller than Top-K. Indices are not
+  // sorted by their corresponding logit.
+  if (rowLen <= kTopK) {
+    for (int rowIt = threadIdx.x; rowIt < rowLen;
+         rowIt += kNumThreadsPerBlock) {
+      int idx = rowStart + rowIt;
+      outIndices[rowIdx * kTopK + rowIt] = idx - rowStart;
+      outLogits[rowIdx * kTopK + rowIt] =
+          logits[rowIdx * stride0 + idx * stride1];
+    }
+    for (int rowIt = rowLen + threadIdx.x; rowIt < kTopK;
+         rowIt += kNumThreadsPerBlock) {
+      outIndices[rowIdx * kTopK + rowIt] = -1;
+      outLogits[rowIdx * kTopK + rowIt] = -FLT_MAX;
+    }
+    return;
+  }
+
+  // Clear the histogram.
+  if (threadIdx.x < kNumBins) {
+    smemHistogram[threadIdx.x] = 0;
+  }
+
+  // Make sure the histogram is ready.
+  __syncthreads();
+
+  // Fetch elements one-by-one.
+  for (int rowIt = rowStart + threadIdx.x; rowIt < rowEnd;
+       rowIt += kNumThreadsPerBlock) {
+    uint16_t idx = extractBinIdx(logits[rowIdx * stride0 + rowIt * stride1]);
+    atomicAdd(&smemHistogram[idx], 1);
+  }
+
+  // Make sure the histogram is ready.
+  __syncthreads();
+
+  // Read the values from SMEM.
+  int binCount{0};
+  if (threadIdx.x < kNumBins) {
+    binCount = smemHistogram[threadIdx.x];
+  }
+
+  // Make sure each thread has read its value.
+  __syncthreads();
+
+  // Compute the prefix sum.
+  int prefixSum{0}, totalSum{0};
+  Scan(smemScan).ExclusiveSum(binCount, prefixSum, totalSum);
+
+  // Update the histogram with the prefix sums.
+  if (threadIdx.x < kNumBins) {
+    smemHistogram[threadIdx.x] = prefixSum;
+  }
+
+  // Make sure the data is in shared memory.
+  __syncthreads();
+
+  // Find the last valid bin.
+  if (threadIdx.x < kNumBins) {
+    int nextPrefixSum =
+        threadIdx.x == kNumBins - 1 ? totalSum : smemHistogram[threadIdx.x + 1];
+    if (prefixSum < kTopK && nextPrefixSum >= kTopK) {
+      smemThresholdBinIdx[0] = threadIdx.x;
+    }
+  }
+
+  // Clear the counter to store the items for the final phase.
+  if (threadIdx.x == 0) {
+    smemFinalDstIdx[0] = 0;
+  }
+
+  // Make sure the data is in shared memory.
+  __syncthreads();
+
+  // The threshold bin.
+  int thresholdBinIdx = smemThresholdBinIdx[0];
+
+  // Fetch elements one-by-one and populate the shared memory buffers.
+  for (int rowIt = rowStart + threadIdx.x; rowIt < rowEnd;
+       rowIt += kNumThreadsPerBlock) {
+    float logit = logits[rowIdx * stride0 + rowIt * stride1];
+    uint16_t idx = extractBinIdx(logit);
+    if (idx < thresholdBinIdx) {
+      int dstIdx = atomicAdd(&smemHistogram[idx], 1);
+      smemLogits[dstIdx] = logit;
+      smemIndices[dstIdx] = rowIt;
+    } else if (idx == thresholdBinIdx) {
+      int dstIdx = atomicAdd(&smemFinalDstIdx[0], 1);
+      if (dstIdx < kNumFinalItems) {
+        smemFinal.items.logits[dstIdx] = logit;
+        smemFinal.items.indices[dstIdx] = rowIt;
+      }
+    }
+  }
+
+  // Make sure the elements are in shared memory.
+  __syncthreads();
+
+  // The logits of the elements to be sorted in the final pass.
+  float finalLogits[kNumFinalItemsPerThread];
+  // The indices of the elements to be sorted in the final pass.
+  int finalIndices[kNumFinalItemsPerThread];
+
+// Init.
+#pragma unroll
+  for (int ii = 0; ii < kNumFinalItemsPerThread; ++ii) {
+    finalLogits[ii] = -FLT_MAX;
+  }
+
+// Read the elements from SMEM.
+#pragma unroll
+  for (int ii = 0; ii < kNumFinalItemsPerThread; ++ii) {
+    int srcIdx = ii * kNumThreadsPerBlock + threadIdx.x;
+    if (srcIdx < smemFinalDstIdx[0]) {
+      finalLogits[ii] = smemFinal.items.logits[srcIdx];
+      finalIndices[ii] = smemFinal.items.indices[srcIdx];
+    }
+  }
+
+  // Make sure the shared memory has been read.
+  __syncthreads();
+
+  // Sort the elements.
+  FinalSort(smemFinal.finalSort)
+      .SortDescendingBlockedToStriped(finalLogits, finalIndices);
+
+  // Copy the data back to the shared memory storage.
+  int baseIdx = thresholdBinIdx > 0 ? smemHistogram[thresholdBinIdx - 1] : 0;
+#pragma unroll
+  for (int ii = 0; ii < kNumFinalItemsPerThread; ++ii) {
+    int srcIdx = ii * kNumThreadsPerBlock + threadIdx.x;
+    int dstIdx = baseIdx + srcIdx;
+    if (dstIdx < kTopK) {
+      smemLogits[dstIdx] = finalLogits[ii];
+      smemIndices[dstIdx] = finalIndices[ii];
+    }
+  }
+
+  // Make sure the data is in shared memory.
+  __syncthreads();
+
+  // The topK logits.
+  float topKLogits[kNumTopKItemsPerThread];
+  // The topK indices.
+  int topKIndices[kNumTopKItemsPerThread];
+
+// Load from shared memory.
+#pragma unroll
+  for (int ii = 0; ii < kNumTopKItemsPerThread; ++ii) {
+    topKLogits[ii] = smemLogits[ii * kNumThreadsPerBlock + threadIdx.x];
+    topKIndices[ii] = smemIndices[ii * kNumThreadsPerBlock + threadIdx.x];
+  }
+
+  // Sort the elements.
+  TopKSort(smemFinal.topKSort)
+      .SortDescendingBlockedToStriped(topKLogits, topKIndices);
+
+// Store to global memory.
+#pragma unroll
+  for (int ii = 0; ii < kNumTopKItemsPerThread; ++ii) {
+    int offset = rowIdx * kTopK + ii * kNumThreadsPerBlock + threadIdx.x;
+    outIndices[offset] = topKIndices[ii] - rowStart;
+    outLogits[offset] = topKLogits[ii];
+  }
+}
+
 }  // namespace vllm
 
 void apply_repetition_penalties_(
@@ -85,4 +324,20 @@ void apply_repetition_penalties_(
                 repetition_penalties.data_ptr<scalar_t>(), num_seqs, vocab_size,
                 tile_size);
       });
-}
+}
+
+void top_k_per_row(const torch::Tensor& logits, const torch::Tensor& rowStarts,
+                   const torch::Tensor& rowEnds, torch::Tensor& indices,
+                   torch::Tensor& values, int64_t numRows, int64_t stride0,
+                   int64_t stride1) {
+  // Compute the results on the device.
+  constexpr int kNumThreadsPerBlock = 512;
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+
+  vllm::topKPerRow<kNumThreadsPerBlock>
+      <<<numRows, kNumThreadsPerBlock, 0, stream>>>(
+          logits.data_ptr<float>(), rowStarts.data_ptr<int>(),
+          rowEnds.data_ptr<int>(), indices.data_ptr<int>(),
+          values.data_ptr<float>(), static_cast<int>(stride0),
+          static_cast<int>(stride1));
+}

csrc/torch_bindings.cpp

@@ -188,6 +188,13 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.impl("apply_repetition_penalties_", torch::kCUDA,
            &apply_repetition_penalties_);
 
+  // Optimized top-k per row operation
+  ops.def(
+      "top_k_per_row(Tensor logits, Tensor rowStarts, Tensor rowEnds, "
+      "Tensor! indices, Tensor! values, int numRows, int stride0, "
+      "int stride1) -> ()");
+  ops.impl("top_k_per_row", torch::kCUDA, &top_k_per_row);
+
   // Layernorm-quant
   // Apply Root Mean Square (RMS) Normalization to the input tensor.
   ops.def(

tests/kernels/test_top_k_per_row.py

@@ -0,0 +1,143 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import numpy as np
+import pytest
+import torch
+
+from vllm.platforms import current_platform
+
+# Test parameters
+NUM_ROWS = [1, 32, 2050]
+TOP_K_VALUES = [2048]
+
+
+def create_random_logits(
+    row_starts: torch.Tensor,
+    row_ends: torch.Tensor,
+    vocab_size: int,
+    dtype: torch.dtype,
+    seed: int,
+) -> torch.Tensor:
+    """Create random logits tensor for testing."""
+    torch.manual_seed(seed)
+    np.random.seed(seed)
+    # Generate logits with some structure to make testing more meaningful
+    logits = torch.randn(row_starts.shape[0], max(row_ends), dtype=dtype, device="cuda")
+    for i, end in enumerate(row_ends):
+        logits[i, end:] = float("-inf")
+    return logits
+
+
+def create_row_boundaries(
+    seq_len: int, vocab_size: int
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """Create row start and end indices for testing."""
+    row_starts = torch.zeros(seq_len, dtype=torch.int32, device="cuda")
+    row_ends = torch.arange(1, seq_len + 1, device="cuda", dtype=torch.int32)
+    return row_starts, row_ends
+
+
+def compare_top_k_results(
+    cuda_indices: torch.Tensor,
+    cuda_values: torch.Tensor,
+    torch_indices: torch.Tensor,
+    torch_values: torch.Tensor,
+    row_starts: torch.Tensor,
+    row_ends: torch.Tensor,
+    top_k: int,
+    tolerance: float = 1e-5,
+) -> bool:
+    """
+    Compare results from CUDA top_k_per_row with torch.topk.
+    Both results should be sorted and contain the same top-k elements.
+    """
+    num_rows = cuda_indices.shape[0]
+
+    for row_idx in range(num_rows):
+        # Get valid elements using row boundaries
+        row_start = row_starts[row_idx].item()
+        row_end = row_ends[row_idx].item()
+        row_length = row_end - row_start
+        num_valid = min(top_k, row_length)
+        cuda_row_indices = cuda_indices[row_idx][:num_valid].cpu()
+        torch_row_indices = torch_indices[row_idx][:num_valid].cpu()
+
+        # Compare the sets of indices first
+        cuda_set = set(cuda_row_indices.tolist())
+        torch_set = set(torch_row_indices.tolist())
+        if cuda_set == torch_set:
+            continue
+
+        # Any difference in elements, compare the values
+        cuda_row_values = cuda_values[row_idx][:num_valid].cpu()
+        torch_row_values = torch_values[row_idx][:num_valid].cpu()
+
+        cuda_only_values, torch_only_values = [], []
+        for idx in cuda_set - torch_set:
+            cuda_pos = (cuda_row_indices == idx).nonzero(as_tuple=True)[0]
+            cuda_only_values.append(cuda_row_values[cuda_pos[0]])
+
+        for idx in torch_set - cuda_set:
+            torch_pos = (torch_row_indices == idx).nonzero(as_tuple=True)[0]
+            torch_only_values.append(torch_row_values[torch_pos[0]])
+
+        if len(cuda_only_values) != len(torch_only_values):
+            return False
+        if not torch.allclose(
+            torch.tensor(cuda_only_values),
+            torch.tensor(torch_only_values),
+            rtol=tolerance,
+            atol=tolerance,
+        ):
+            return False
+
+    return True
+
+
+@pytest.mark.parametrize("num_rows", NUM_ROWS)
+@pytest.mark.parametrize("top_k", TOP_K_VALUES)
+@pytest.mark.skipif(not current_platform.is_cuda(), reason="This test requires CUDA")
+@torch.inference_mode()
+def test_top_k_per_row(
+    num_rows: int,
+    top_k: int,
+) -> None:
+    """
+    Test top_k_per_row.
+    """
+    torch.set_default_device("cuda:0")
+
+    # Create test data
+    vocab_size = 20000
+    row_starts, row_ends = create_row_boundaries(num_rows, vocab_size)
+    logits = create_random_logits(row_starts, row_ends, vocab_size, torch.float32, 42)
+
+    # Create output tensors
+    indices = torch.empty((num_rows, 2048), dtype=torch.int32, device="cuda")
+    values = torch.empty((num_rows, 2048), dtype=torch.float32, device="cuda")
+
+    # Run CUDA implementation
+    torch.ops._C.top_k_per_row(
+        logits,
+        row_starts,
+        row_ends,
+        indices,
+        values,
+        num_rows,
+        top_k,
+        logits.stride(0),
+        logits.stride(1),
+    )
+
+    # Run reference implementation
+    torch_values, torch_indices = logits.topk(min(top_k, max(row_ends)), dim=-1)
+    mask_lo = torch_indices >= 0
+    mask_hi = (torch_indices - (row_ends - row_starts)[:, None]) < 0
+    mask = mask_lo & mask_hi
+    torch_indices = torch_indices.masked_fill(~mask, -1)
+
+    # Compare results
+    assert compare_top_k_results(
+        indices, values, torch_indices, torch_values, row_starts, row_ends, top_k
+    ), "CUDA top_k_per_row results don't match torch.topk"

vllm/model_executor/models/deepseek_v2.py

@@ -643,17 +643,24 @@ def sparse_attn_indexer(
                 chunk.cu_seqlen_ks,
                 chunk.cu_seqlen_ke,
             )
-            topk_indices = logits.topk(min(topk_tokens, logits.shape[-1]), dim=-1)[1]
+            num_rows = logits.shape[0]
-            topk_indices -= chunk.cu_seqlen_ks[:, None]
+            assert topk_tokens == 2048, "top_k_per_row assumes size 2048"
-            mask_lo = topk_indices >= 0
+            topk_indices = torch.empty(
-            mask_hi = (
+                num_rows, topk_tokens, dtype=torch.int32, device=logits.device
-                topk_indices - (chunk.cu_seqlen_ke - chunk.cu_seqlen_ks)[:, None] < 0
             )
-            mask = torch.full_like(
+            topk_values = torch.empty(
-                topk_indices, False, dtype=torch.bool, device=topk_indices.device
+                num_rows, topk_tokens, dtype=logits.dtype, device=logits.device
+            )
+            torch.ops._C.top_k_per_row(
+                logits,
+                chunk.cu_seqlen_ks,
+                chunk.cu_seqlen_ke,
+                topk_indices,
+                topk_values,
+                num_rows,
+                logits.stride(0),
+                logits.stride(1),
             )
-            mask = mask_lo & mask_hi
-            topk_indices = topk_indices.masked_fill(~mask, -1)
             topk_indices_buffer[
                 chunk.token_start : chunk.token_end, : topk_indices.shape[-1]
             ] = topk_indices.to(dtype=torch.int32)
@@ -693,28 +700,32 @@ def sparse_attn_indexer(
         # padded query len
         current_device = padded_q_fp8_decode_tokens.device
         padded_num_tokens = batch_size * next_n
-        positions = (
-            torch.arange(max_model_len, device=current_device)
-            .unsqueeze(0)
-            .expand(batch_size * next_n, -1)
-        )
         row_indices = torch.arange(padded_num_tokens, device=current_device) // next_n
         next_n_offset = (
             torch.arange(padded_num_tokens, device=padded_q_fp8_decode_tokens.device)
             % next_n
         )
         index_end_pos = (
-            decode_metadata.seq_lens[row_indices] - next_n + next_n_offset
+            decode_metadata.seq_lens[row_indices] - next_n + next_n_offset + 1
         ).unsqueeze(1)
-        # index_end_pos: [B * N, 1]
+        num_rows = logits.shape[0]
-        mask = positions <= index_end_pos
+        assert topk_tokens == 2048, "top_k_per_row assumes size 2048"
-        # mask: [B * N, L]
+        topk_indices = torch.empty(
-        logits = logits.masked_fill(~mask, float("-inf"))
+            num_rows, topk_tokens, dtype=torch.int32, device=logits.device
-        topk_indices = logits.topk(topk_tokens, dim=-1)[1].to(torch.int32)  # [B * N, K]
+        )
-        # ensure we don't set indices for the top k
+        topk_values = torch.empty(
-        # that is out of range(masked already)
+            num_rows, topk_tokens, dtype=logits.dtype, device=logits.device
-        # this will happen if context length is shorter than K
+        )
-        topk_indices[topk_indices > index_end_pos] = -1
+        torch.ops._C.top_k_per_row(
+            logits,
+            torch.zeros(num_rows, dtype=torch.int32, device=logits.device),
+            index_end_pos.to(dtype=torch.int32, device=logits.device),
+            topk_indices,
+            topk_values,
+            num_rows,
+            logits.stride(0),
+            logits.stride(1),
+        )
         if decode_metadata.requires_padding:
             # if padded, we need to unpack
             # the topk indices removing padded tokens