[1/3] fix dsv3 awq issue (#4556)

Co-authored-by: leoneo <1320612015@qq.com>

sgl-kernel/csrc/gemm/awq_kernel.cu

@@ -3,6 +3,16 @@
 #include <c10/cuda/CUDAGuard.h>
 #include <cuda_fp16.h>
 #include <torch/all.h>
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
+#include <cuda_bf16.h>
+#endif
+
+template <int lut>
+__device__ inline int lop3(int a, int b, int c) {
+  int res;
+  asm volatile("lop3.b32 %0, %1, %2, %3, %4;\n" : "=r"(res) : "r"(a), "r"(b), "r"(c), "n"(lut));
+  return res;
+}
 
 __device__ uint4 dequantize_s4_to_fp16x2(uint32_t const& source) {
 #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 750
@@ -68,32 +78,102 @@ __device__ uint4 dequantize_s4_to_fp16x2(uint32_t const& source) {
 #endif
 }
 
+__device__ uint4 dequantize_s4_to_bf16x2(uint32_t const& source) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
+  uint4 result;
+  uint32_t* h = reinterpret_cast<uint32_t*>(&result);
+  uint32_t const i4s = source;
+
+  // Define masks and constants
+  static constexpr uint32_t MASK = 0x000f000f;
+  static constexpr uint32_t EX = 0x43004300;
+  static constexpr uint32_t MUL = 0x3F803F80;
+  static constexpr uint32_t ADD = 0xC300C300;
+
+  int lo0 = lop3<(0xf0 & 0xcc) | 0xaa>(i4s, MASK, EX);
+  int hi0 = lop3<(0xf0 & 0xcc) | 0xaa>(i4s >> 4, MASK, EX);
+  int lo1 = lop3<(0xf0 & 0xcc) | 0xaa>(i4s >> 8, MASK, EX);
+  int hi1 = lop3<(0xf0 & 0xcc) | 0xaa>(i4s >> 12, MASK, EX);
+
+  nv_bfloat162* res = reinterpret_cast<nv_bfloat162*>(h);
+  res[0] = __hfma2(
+      *reinterpret_cast<nv_bfloat162*>(&lo0),
+      *reinterpret_cast<const nv_bfloat162*>(&MUL),
+      *reinterpret_cast<const nv_bfloat162*>(&ADD));
+  res[1] = __hfma2(
+      *reinterpret_cast<nv_bfloat162*>(&hi0),
+      *reinterpret_cast<const nv_bfloat162*>(&MUL),
+      *reinterpret_cast<const nv_bfloat162*>(&ADD));
+  res[2] = __hfma2(
+      *reinterpret_cast<nv_bfloat162*>(&lo1),
+      *reinterpret_cast<const nv_bfloat162*>(&MUL),
+      *reinterpret_cast<const nv_bfloat162*>(&ADD));
+  res[3] = __hfma2(
+      *reinterpret_cast<nv_bfloat162*>(&hi1),
+      *reinterpret_cast<const nv_bfloat162*>(&MUL),
+      *reinterpret_cast<const nv_bfloat162*>(&ADD));
+
+  return result;
+#else
+  assert(false);
+  return {};
+#endif
+}
+
+template <typename OutputT>
 __global__ void __launch_bounds__(256) dequantize_weights(
     int* __restrict__ qweight,
-    half* __restrict__ scales,
+    OutputT* __restrict__ scales,
     int* __restrict__ qzeros,
-    half* __restrict__ output,
+    OutputT* __restrict__ output,
     int group_size,
     int qweight_cols) {
   int col = blockIdx.x * blockDim.x + threadIdx.x;
   int row = blockIdx.y * blockDim.y + threadIdx.y;
 
-  uint4 zeros = dequantize_s4_to_fp16x2(qzeros[col + (row / group_size) * qweight_cols]);
+  int group_idx = row / group_size;
-  uint4 loaded_scale = *(uint4*)(scales + 8 * col + (row / group_size) * qweight_cols * 8);
+  int scale_offset = 8 * col + group_idx * qweight_cols * 8;
+  uint4 loaded_scale = *(uint4*)(scales + scale_offset);
 
-  uint4 weight_fp16 = dequantize_s4_to_fp16x2(qweight[col + row * qweight_cols]);
+  // Handle different data types
+  if constexpr (std::is_same<OutputT, half>::value) {
+    // FP16 path
+    uint4 zeros = dequantize_s4_to_fp16x2(qzeros[col + group_idx * qweight_cols]);
+    uint4 weight_fp16 = dequantize_s4_to_fp16x2(qweight[col + row * qweight_cols]);
 
-  asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(weight_fp16.x) : "r"(weight_fp16.x), "r"(zeros.x));
+    // Use PTX assembly for FP16 operations
-  asm volatile("mul.rn.f16x2 %0, %1, %2;\n" : "=r"(weight_fp16.x) : "r"(weight_fp16.x), "r"(loaded_scale.x));
+    asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(weight_fp16.x) : "r"(weight_fp16.x), "r"(zeros.x));
-  asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(weight_fp16.y) : "r"(weight_fp16.y), "r"(zeros.y));
+    asm volatile("mul.rn.f16x2 %0, %1, %2;\n" : "=r"(weight_fp16.x) : "r"(weight_fp16.x), "r"(loaded_scale.x));
-  asm volatile("mul.rn.f16x2 %0, %1, %2;\n" : "=r"(weight_fp16.y) : "r"(weight_fp16.y), "r"(loaded_scale.y));
+    asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(weight_fp16.y) : "r"(weight_fp16.y), "r"(zeros.y));
-  asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(weight_fp16.z) : "r"(weight_fp16.z), "r"(zeros.z));
+    asm volatile("mul.rn.f16x2 %0, %1, %2;\n" : "=r"(weight_fp16.y) : "r"(weight_fp16.y), "r"(loaded_scale.y));
-  asm volatile("mul.rn.f16x2 %0, %1, %2;\n" : "=r"(weight_fp16.z) : "r"(weight_fp16.z), "r"(loaded_scale.z));
+    asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(weight_fp16.z) : "r"(weight_fp16.z), "r"(zeros.z));
-  asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(weight_fp16.w) : "r"(weight_fp16.w), "r"(zeros.w));
+    asm volatile("mul.rn.f16x2 %0, %1, %2;\n" : "=r"(weight_fp16.z) : "r"(weight_fp16.z), "r"(loaded_scale.z));
-  asm volatile("mul.rn.f16x2 %0, %1, %2;\n" : "=r"(weight_fp16.w) : "r"(weight_fp16.w), "r"(loaded_scale.w));
+    asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(weight_fp16.w) : "r"(weight_fp16.w), "r"(zeros.w));
+    asm volatile("mul.rn.f16x2 %0, %1, %2;\n" : "=r"(weight_fp16.w) : "r"(weight_fp16.w), "r"(loaded_scale.w));
 
-  half* output_ptr = output + 8 * col + 8 * row * qweight_cols;
+    OutputT* output_ptr = output + 8 * col + 8 * row * qweight_cols;
-  *(uint4*)output_ptr = weight_fp16;
+    *(uint4*)output_ptr = weight_fp16;
+  } else if constexpr (std::is_same<OutputT, __nv_bfloat16>::value) {
+    uint4 weight_raw = dequantize_s4_to_bf16x2(qweight[col + row * qweight_cols]);
+    uint4 zero_raw = dequantize_s4_to_bf16x2(qzeros[col + group_idx * qweight_cols]);
+    uint4 scale_raw = *reinterpret_cast<uint4*>(scales + scale_offset);
+
+    // Vectorized processing (each uint4 contains 4 nv_bfloat162)
+    nv_bfloat162* weight_vec = reinterpret_cast<nv_bfloat162*>(&weight_raw);
+    nv_bfloat162* zero_vec = reinterpret_cast<nv_bfloat162*>(&zero_raw);
+    nv_bfloat162* scale_vec = reinterpret_cast<nv_bfloat162*>(&scale_raw);
+
+// Single instruction dual-channel operation
+#pragma unroll
+    for (int i = 0; i < 4; ++i) {  // uint4 = 4 * nv_bfloat162
+      weight_vec[i] = __hmul2(__hsub2(weight_vec[i], zero_vec[i]), scale_vec[i]);
+    }
+
+    // Directly store to OutputT array (guaranteed contiguous memory)
+    OutputT* output_ptr = output + 8 * col + row * qweight_cols * 8;
+    static_assert(sizeof(uint4) == 8 * sizeof(OutputT), "Memory layout mismatch");
+    *reinterpret_cast<uint4*>(output_ptr) = weight_raw;
+  }
 }
 
 torch::Tensor awq_dequantize(torch::Tensor qweight, torch::Tensor scales, torch::Tensor qzeros) {
@@ -112,16 +192,23 @@ torch::Tensor awq_dequantize(torch::Tensor qweight, torch::Tensor scales, torch:
   at::Tensor output = torch::empty({qweight_rows, qweight_cols * 8}, output_tensor_options);
 
   auto _qweight = reinterpret_cast<int*>(qweight.data_ptr<int>());
-  auto _scales = reinterpret_cast<half*>(scales.data_ptr<at::Half>());
   auto _zeros = reinterpret_cast<int*>(qzeros.data_ptr<int>());
-  auto _output = reinterpret_cast<half*>(output.data_ptr<at::Half>());
 
   dim3 num_blocks(x_blocks, y_blocks);
   dim3 threads_per_block(x_num_threads, y_num_threads);
-
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  dequantize_weights<<<num_blocks, threads_per_block, 0, stream>>>(
+
-      _qweight, _scales, _zeros, _output, group_size, qweight_cols);
+  if (scales.scalar_type() == at::ScalarType::Half) {
+    auto _scales = reinterpret_cast<half*>(scales.data_ptr<at::Half>());
+    auto _output = reinterpret_cast<half*>(output.data_ptr<at::Half>());
+    dequantize_weights<half>
+        <<<num_blocks, threads_per_block, 0, stream>>>(_qweight, _scales, _zeros, _output, group_size, qweight_cols);
+  } else {
+    auto _scales = reinterpret_cast<__nv_bfloat16*>(scales.data_ptr<at::BFloat16>());
+    auto _output = reinterpret_cast<__nv_bfloat16*>(output.data_ptr<at::BFloat16>());
+    dequantize_weights<__nv_bfloat16>
+        <<<num_blocks, threads_per_block, 0, stream>>>(_qweight, _scales, _zeros, _output, group_size, qweight_cols);
+  }
 
   return output;
 }

sgl-kernel/tests/test_awq_dequant.py

@@ -7,6 +7,57 @@ from sgl_kernel import awq_dequantize
 from vllm import _custom_ops as ops
 
 
+def reverse_awq_order(t: torch.Tensor):
+    bits = 4
+    AWQ_REVERSE_ORDER = [0, 4, 1, 5, 2, 6, 3, 7]
+    reverse_order_tensor = torch.arange(
+        t.shape[-1],
+        dtype=torch.int32,
+        device=t.device,
+    )
+    reverse_order_tensor = reverse_order_tensor.view(-1, 32 // bits)
+    reverse_order_tensor = reverse_order_tensor[:, AWQ_REVERSE_ORDER]
+    reverse_order_tensor = reverse_order_tensor.view(-1)
+
+    t = t[:, reverse_order_tensor] & 0xF
+    return t
+
+
+# qweights - [R     , C // 8], int32
+# scales   - [R // G, C     ], float16
+# zeros    - [R // G, C // 8], int32
+def awq_dequantize_torch(
+    qweight: torch.Tensor, scales: torch.Tensor, qzeros: torch.Tensor, group_size: int
+) -> torch.Tensor:
+
+    if group_size == -1:
+        group_size = qweight.shape[0]
+
+    bits = 4
+    shifts = torch.arange(0, 32, bits, device=qzeros.device)
+
+    iweights = torch.bitwise_right_shift(qweight[:, :, None], shifts[None, None, :]).to(
+        torch.int8
+    )
+
+    iweights = iweights.view(iweights.shape[0], -1)
+
+    zeros = torch.bitwise_right_shift(qzeros[:, :, None], shifts[None, None, :]).to(
+        torch.int8
+    )
+    zeros = zeros.view(qzeros.shape[0], -1)
+    zeros = reverse_awq_order(zeros)
+
+    iweights = reverse_awq_order(iweights)
+
+    iweights = torch.bitwise_and(iweights, (2**bits) - 1)
+    zeros = torch.bitwise_and(zeros, (2**bits) - 1)
+
+    scales = scales.repeat_interleave(group_size, dim=0)
+    zeros = zeros.repeat_interleave(group_size, dim=0)
+    return (iweights - zeros) * scales
+
+
 def vllm_awq_dequantize(
     qweight: torch.Tensor, scales: torch.Tensor, qzeros: torch.Tensor
 ) -> torch.Tensor:
@@ -20,16 +71,17 @@ def sglang_awq_dequantize(
 
 
 @pytest.mark.parametrize(
-    "qweight_row,qweight_col",
+    "qweight_row,qweight_col,is_bf16_act",
     list(
         itertools.product(
-            [3584, 18944, 128, 256, 512, 1024], [448, 576, 4736, 16, 32, 64, 128]
+            [3584, 18944, 128, 256, 512, 1024],
+            [448, 576, 4736, 16, 32, 64, 128],
+            [True, False],
         )
     ),
 )
 def test_awq_dequant_compare_implementations(
-    qweight_row: int,
+    qweight_row: int, qweight_col: int, is_bf16_act: bool
-    qweight_col: int,
 ):
     device = torch.device("cuda")
 
@@ -43,7 +95,12 @@ def test_awq_dequant_compare_implementations(
     group_size = qweight_row
     scales_row = qweight_row // group_size
     scales_col = qweight_col * 8
-    scales = torch.rand(scales_row, scales_col, dtype=torch.float16, device=device)
+
+    if is_bf16_act:
+        scales = torch.rand(scales_row, scales_col, dtype=torch.bfloat16, device=device)
+    else:
+        scales = torch.rand(scales_row, scales_col, dtype=torch.float16, device=device)
+
     qzeros = torch.randint(
         0,
         torch.iinfo(torch.int32).max,
@@ -53,13 +110,21 @@ def test_awq_dequant_compare_implementations(
     )
 
     # Run both implementations
-    vllm_out = vllm_awq_dequantize(qweight, scales, qzeros)
+    vllm_out = vllm_awq_dequantize(qweight, scales.to(torch.float16), qzeros)
+    torch_out = awq_dequantize_torch(qweight, scales, qzeros, group_size)
     sglang_out = sglang_awq_dequantize(qweight, scales, qzeros)
 
     # Compare results
     torch.testing.assert_close(
-        vllm_out.to(torch.float32), sglang_out.to(torch.float32), rtol=1e-3, atol=1e-5
+        torch_out.to(torch.float32), sglang_out.to(torch.float32), rtol=1e-3, atol=1e-5
     )
+    if not is_bf16_act:
+        torch.testing.assert_close(
+            vllm_out.to(torch.float32),
+            sglang_out.to(torch.float32),
+            rtol=1e-3,
+            atol=1e-5,
+        )
 
 
 if __name__ == "__main__":