use flashinfer vec_dtypes in sgl_kernel (#3083)

sgl-kernel/src/sgl-kernel/csrc/sampling_scaling_penalties.cu

@@ -1,11 +1,12 @@
 #include <ATen/ATen.h>
 #include <ATen/cuda/CUDAContext.h>
 #include <c10/cuda/CUDAGuard.h>
+#include <pytorch_extension_utils.h>
 
 #include <THC/THCAtomics.cuh>
+#include <flashinfer/vec_dtypes.cuh>
 
 #include "utils.h"
-#include "vectorization.cuh"
 
 template <typename scalar_t>
 __global__ void sampling_scaling_penalties_kernel(const scalar_t* logits, const scalar_t* scaling_penalties,
@@ -13,31 +14,31 @@ __global__ void sampling_scaling_penalties_kernel(const scalar_t* logits, const
   const int32_t tid = blockIdx.x * blockDim.x + threadIdx.x;
   const int32_t stride = blockDim.x * gridDim.x;
 
-  auto const* vectorized_logits = reinterpret_cast<vec4_t<scalar_t> const*>(logits);
-  auto const* vectorized_penalties = reinterpret_cast<vec4_t<scalar_t> const*>(scaling_penalties);
-  auto* vectorized_output = reinterpret_cast<vec4_t<scalar_t>*>(output);
+  constexpr uint32_t vec_size = 16 / sizeof(scalar_t);
+  using vec_t = flashinfer::vec_t<scalar_t, vec_size>;
 
-  const int32_t num_vec_elems = numel >> 2;
+  const int32_t num_vec_elems = numel / vec_size;
 
-#pragma unroll 4
+#pragma unroll 1
   for (int32_t i = tid; i < num_vec_elems; i += stride) {
-    vec4_t<scalar_t> logits_vec = vectorized_logits[i];
-    vec4_t<scalar_t> penalties_vec = vectorized_penalties[i];
-    vec4_t<scalar_t> out_vec;
+    vec_t logits_vec, penalties_vec, out_vec;
+    logits_vec.cast_load(logits + i * vec_size);
+    penalties_vec.cast_load(scaling_penalties + i * vec_size);
 
-    out_vec.x = logits_vec.x > 0 ? logits_vec.x / penalties_vec.x : logits_vec.x * penalties_vec.x;
-    out_vec.y = logits_vec.y > 0 ? logits_vec.y / penalties_vec.y : logits_vec.y * penalties_vec.y;
-    out_vec.z = logits_vec.z > 0 ? logits_vec.z / penalties_vec.z : logits_vec.z * penalties_vec.z;
-    out_vec.w = logits_vec.w > 0 ? logits_vec.w / penalties_vec.w : logits_vec.w * penalties_vec.w;
+#pragma unroll
+    for (uint32_t j = 0; j < vec_size; ++j) {
+      out_vec[j] = logits_vec[j] > scalar_t(0.0f) ? logits_vec[j] / penalties_vec[j] : logits_vec[j] * penalties_vec[j];
+    }
 
-    vectorized_output[i] = out_vec;
+    out_vec.cast_store(output + i * vec_size);
   }
 
-  const int32_t start_idx = num_vec_elems * 4;
+  // process the remaining elements
+  const int32_t start_idx = num_vec_elems * vec_size;
   for (int32_t i = start_idx + tid; i < numel; i += stride) {
     scalar_t logit = logits[i];
     scalar_t penalty = scaling_penalties[i];
-    output[i] = logit > 0 ? logit / penalty : logit * penalty;
+    output[i] = logit > scalar_t(0.0f) ? logit / penalty : logit * penalty;
   }
 }
 
@@ -48,12 +49,14 @@ torch::Tensor sampling_scaling_penalties(const torch::Tensor& logits, const torc
 
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
-  AT_DISPATCH_FLOATING_TYPES_AND2(
-      at::ScalarType::Half, at::ScalarType::BFloat16, logits.scalar_type(), "sampling_scaling_penalties_kernel", ([&] {
-        const int blocks = (numel + threads * 4 - 1) / (threads * 4);
-        sampling_scaling_penalties_kernel<scalar_t><<<blocks, threads, 0, stream>>>(
-            logits.data_ptr<scalar_t>(), scaling_penalties.data_ptr<scalar_t>(), output.data_ptr<scalar_t>(), numel);
-      }));
+  DISPATCH_PYTORCH_DTYPE_TO_CTYPE_FP16(logits.scalar_type(), scalar_t, [&] {
+    uint32_t vec_size = 16 / sizeof(scalar_t);
+    const int blocks = (numel + threads * vec_size - 1) / (threads * vec_size);
+    sampling_scaling_penalties_kernel<scalar_t><<<blocks, threads, 0, stream>>>(
+        static_cast<scalar_t*>(logits.data_ptr()), static_cast<scalar_t*>(scaling_penalties.data_ptr()),
+        static_cast<scalar_t*>(output.data_ptr()), numel);
+    return true;
+  });
 
   return output;
 }

sgl-kernel/src/sgl-kernel/csrc/vectorization.cuh

@@ -1,29 +0,0 @@
-// Adapted from https://github.com/vllm-project/vllm/blob/main/csrc/quantization/vectorization.cuh
-#pragma once
-/**
- * __device__ datatypes vectorized by 4
- */
-
-// Include both AMD and NVIDIA fp8 types to avoid circular import
-// TODO(luka/varun) use FP8_TYPE instead after refactoring
-#include <c10/util/Float8_e4m3fn.h>
-#include <c10/util/Float8_e4m3fnuz.h>
-
-// Vectorization containers
-template <typename scalar_t>
-struct __align__(8) vec4_t {
-  scalar_t x;
-  scalar_t y;
-  scalar_t z;
-  scalar_t w;
-};
-
-template <typename quant_type_t>
-struct __align__(4) q8x4_t {
-  static_assert(std::is_same_v<quant_type_t, int8_t> || std::is_same_v<quant_type_t, c10::Float8_e4m3fn> ||
-                std::is_same_v<quant_type_t, c10::Float8_e4m3fnuz>);
-  quant_type_t x;
-  quant_type_t y;
-  quant_type_t z;
-  quant_type_t w;
-};

sgl-kernel/tests/test_sampling_scaling_penalties.py

@@ -1,37 +1,34 @@
+import pytest
 import torch
 from sgl_kernel import sampling_scaling_penalties
 
 
-def test_sampling_scaling_penalties():
-    batch_sizes = [1, 2, 4, 8, 16, 32, 64, 65]
-    vocab_sizes = [2048, 4096, 8192, 16384, 32768, 32767]
-    dtypes = [torch.float32, torch.half, torch.bfloat16]
+@pytest.mark.parametrize("batch_size", [1, 2, 4, 8, 16, 32, 64, 65])
+@pytest.mark.parametrize("vocab_size", [2048, 4096, 8192, 16384, 32768, 32767])
+@pytest.mark.parametrize("dtype", [torch.half, torch.bfloat16])
+def test_sampling_scaling_penalties(batch_size, vocab_size, dtype):
     device = torch.device("cuda")
+    rtol = 1e-3
+    atol = 1e-3
 
-    for dtype in dtypes:
-        rtol = 1e-3
-        atol = 1e-3
+    logits = torch.randn(batch_size, vocab_size, device=device, dtype=dtype)
+    scaling_penalties = (
+        torch.rand(batch_size, vocab_size, device=device, dtype=dtype) + 0.5
+    )
 
-        for bs in batch_sizes:
-            for vocab_size in vocab_sizes:
-                logits = torch.randn(bs, vocab_size, device=device, dtype=dtype)
-                scaling_penalties = (
-                    torch.rand(bs, vocab_size, device=device, dtype=dtype) + 0.5
-                )
+    ref_output = torch.where(
+        logits > 0, logits / scaling_penalties, logits * scaling_penalties
+    )
 
-                ref_output = torch.where(
-                    logits > 0, logits / scaling_penalties, logits * scaling_penalties
-                )
+    kernel_output = sampling_scaling_penalties(logits, scaling_penalties)
 
-                kernel_output = sampling_scaling_penalties(logits, scaling_penalties)
-
-                torch.testing.assert_close(
-                    kernel_output,
-                    ref_output,
-                    rtol=rtol,
-                    atol=atol,
-                    msg=f"Failed for batch_size={bs}, vocab_size={vocab_size}, dtype={dtype}",
-                )
+    torch.testing.assert_close(
+        kernel_output,
+        ref_output,
+        rtol=rtol,
+        atol=atol,
+        msg=f"Failed for batch_size={batch_size}, vocab_size={vocab_size}, dtype={dtype}",
+    )
 
 
 if __name__ == "__main__":