[Kernel] AQ AZP 3/4: Asymmetric quantization kernels

csrc/ops.h

@@ -165,10 +165,12 @@ torch::Tensor marlin_qqq_gemm(torch::Tensor const& a,
 #endif
 
 void static_scaled_int8_quant(torch::Tensor& out, torch::Tensor const& input,
-                              torch::Tensor const& scale);
+                              torch::Tensor const& scale,
+                              c10::optional<torch::Tensor> const& azp);
 
 void dynamic_scaled_int8_quant(torch::Tensor& out, torch::Tensor const& input,
-                               torch::Tensor& scales);
+                               torch::Tensor& scales,
+                               c10::optional<torch::Tensor> const& azp);
 
 void squeezellm_gemm(torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
                      torch::Tensor lookup_table);

csrc/quantization/compressed_tensors/int8_quant_kernels.cu

@@ -14,9 +14,9 @@
 
 static inline __device__ int8_t float_to_int8_rn(float x) {
 #ifdef USE_ROCM
-  static const float i8_min =
+  static const auto i8_min =
       static_cast<float>(std::numeric_limits<int8_t>::min());
-  static const float i8_max =
+  static const auto i8_max =
       static_cast<float>(std::numeric_limits<int8_t>::max());
   // round
   float dst = std::nearbyint(x);
@@ -31,6 +31,43 @@ static inline __device__ int8_t float_to_int8_rn(float x) {
 #endif
 }
 
+static inline __device__ int32_t float_to_int32_rn(float x) {
+#ifdef USE_ROCM
+  static const auto i32_min =
+      static_cast<float>(std::numeric_limits<int32_t>::min());
+  static const auto i32_max =
+      static_cast<float>(std::numeric_limits<int32_t>::max());
+  // round
+  float dst = std::nearbyint(x);
+  // saturate
+  dst = std::clamp(dst, i32_min, i32_max);
+  return static_cast<int32_t>(dst);
+#else
+  // CUDA path
+  uint32_t dst;
+  asm volatile("cvt.rni.sat.s32.f32 %0, %1;" : "=r"(dst) : "f"(x));
+  return reinterpret_cast<const int32_t&>(dst);
+#endif
+}
+
+static inline __device__ int8_t int32_to_int8(int32_t x) {
+#ifdef USE_ROCM
+  static const auto i8_min =
+      static_cast<int32_t>(std::numeric_limits<int8_t>::min());
+  static const auto i8_max =
+      static_cast<int32_t>(std::numeric_limits<int8_t>::max());
+
+  // saturate
+  int32_t dst = std::clamp(x, i8_min, i8_max);
+  return static_cast<int8_t>(dst);
+#else
+  // CUDA path
+  uint32_t dst;
+  asm volatile("cvt.sat.s8.s32 %0, %1;" : "=r"(dst) : "r"(x));
+  return reinterpret_cast<const int8_t&>(dst);
+#endif
+}
+
 namespace vllm {
 
 template <typename scalar_t, typename scale_type>
@@ -47,6 +84,23 @@ __global__ void static_scaled_int8_quant_kernel(
   }
 }
 
+template <typename scalar_t, typename scale_type, typename azp_type>
+__global__ void static_scaled_int8_azp_quant_kernel(
+    scalar_t const* __restrict__ input, int8_t* __restrict__ out,
+    scale_type const* scale_ptr, azp_type const* azp_ptr,
+    const int hidden_size) {
+  int const tid = threadIdx.x;
+  int const token_idx = blockIdx.x;
+  scale_type const scale = *scale_ptr;
+  azp_type const azp = *azp_ptr;
+
+  for (int i = tid; i < hidden_size; i += blockDim.x) {
+    auto const val = static_cast<float>(input[token_idx * hidden_size + i]);
+    auto const quant_val = int32_to_int8(float_to_int32_rn(val / scale) + azp);
+    out[token_idx * hidden_size + i] = quant_val;
+  }
+}
+
 template <typename scalar_t, typename scale_type>
 __global__ void dynamic_scaled_int8_quant_kernel(
     scalar_t const* __restrict__ input, int8_t* __restrict__ out,
@@ -80,14 +134,68 @@ __global__ void dynamic_scaled_int8_quant_kernel(
   }
 }
 
+template <typename scalar_t, typename scale_type, typename azp_type>
+__global__ void dynamic_scaled_int8_azp_quant_kernel(
+    scalar_t const* __restrict__ input, int8_t* __restrict__ out,
+    scale_type* scale, azp_type* azp, const int hidden_size) {
+  int const token_idx = blockIdx.x;
+
+  // Scan for the min and max value for this token
+  float max_val = std::numeric_limits<float>::min();
+  float min_val = std::numeric_limits<float>::max();
+  for (int i = threadIdx.x; i < hidden_size; i += blockDim.x) {
+    auto val = static_cast<float>(input[token_idx * hidden_size + i]);
+    max_val = std::max(max_val, val);
+    min_val = std::min(min_val, val);
+  }
+
+  // Reduce the max and min values across the block
+  using BlockReduce = cub::BlockReduce<float, 1024>;
+  __shared__ typename BlockReduce::TempStorage reduceStorage;
+  max_val = BlockReduce(reduceStorage).Reduce(max_val, cub::Max{}, blockDim.x);
+  __syncthreads();  // Make sure min doesn't mess with max shared memory
+  min_val = BlockReduce(reduceStorage).Reduce(min_val, cub::Min{}, blockDim.x);
+
+  __shared__ scale_type scale_sh;
+  __shared__ azp_type azp_sh;
+
+  // Compute the scale and zero point and store them, only on the first thread
+  if (threadIdx.x == 0) {
+    float const scale_val = (max_val - min_val) / 255.0f;
+    // Use rounding to even (same as torch.round)
+    auto const azp_float = std::nearbyint(-128.0f - min_val / scale_val);
+    auto const azp_val = static_cast<azp_type>(azp_float);
+
+    // Store the scale and azp into shared and global
+    scale[token_idx] = scale_sh = scale_val;
+    azp[token_idx] = azp_sh = azp_val;
+  }
+
+  // Wait for the scale and azp to be computed
+  __syncthreads();
+
+  float const scale_val = scale_sh;
+  azp_type const azp_val = azp_sh;
+
+  // Quantize the values
+  for (int i = threadIdx.x; i < hidden_size; i += blockDim.x) {
+    auto const val = static_cast<float>(input[token_idx * hidden_size + i]);
+    auto const quant_val =
+        int32_to_int8(float_to_int32_rn(val / scale_val) + azp_val);
+    out[token_idx * hidden_size + i] = quant_val;
+  }
+}
+
 }  // namespace vllm
 
 void static_scaled_int8_quant(torch::Tensor& out,          // [..., hidden_size]
                               torch::Tensor const& input,  // [..., hidden_size]
-                              torch::Tensor const& scale) {
+                              torch::Tensor const& scale,
+                              c10::optional<torch::Tensor> const& azp) {
   TORCH_CHECK(input.is_contiguous());
   TORCH_CHECK(out.is_contiguous());
   TORCH_CHECK(scale.numel() == 1);
+  TORCH_CHECK(!azp || azp->numel() == 1);
 
   int const hidden_size = input.size(-1);
   int const num_tokens = input.numel() / hidden_size;
@@ -96,19 +204,29 @@ void static_scaled_int8_quant(torch::Tensor& out,          // [..., hidden_size]
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
   VLLM_DISPATCH_FLOATING_TYPES(
       input.scalar_type(), "static_scaled_int8_quant_kernel", [&] {
-        vllm::static_scaled_int8_quant_kernel<scalar_t, float>
-            <<<grid, block, 0, stream>>>(input.data_ptr<scalar_t>(),
-                                         out.data_ptr<int8_t>(),
-                                         scale.data_ptr<float>(), hidden_size);
+        if (!azp) {
+          vllm::static_scaled_int8_quant_kernel<scalar_t, float>
+              <<<grid, block, 0, stream>>>(
+                  input.data_ptr<scalar_t>(), out.data_ptr<int8_t>(),
+                  scale.data_ptr<float>(), hidden_size);
+        } else {
+          vllm::static_scaled_int8_azp_quant_kernel<scalar_t, float, int32_t>
+              <<<grid, block, 0, stream>>>(
+                  input.data_ptr<scalar_t>(), out.data_ptr<int8_t>(),
+                  scale.data_ptr<float>(), azp->data_ptr<int32_t>(),
+                  hidden_size);
+        }
       });
 }
 
 void dynamic_scaled_int8_quant(
     torch::Tensor& out,          // [..., hidden_size]
     torch::Tensor const& input,  // [..., hidden_size]
-    torch::Tensor& scales) {
+    torch::Tensor& scales, c10::optional<torch::Tensor> const& azp) {
   TORCH_CHECK(input.is_contiguous());
   TORCH_CHECK(out.is_contiguous());
+  TORCH_CHECK(scales.is_contiguous());
+  TORCH_CHECK(!azp || azp->is_contiguous());
 
   int const hidden_size = input.size(-1);
   int const num_tokens = input.numel() / hidden_size;
@@ -117,9 +235,17 @@ void dynamic_scaled_int8_quant(
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
   VLLM_DISPATCH_FLOATING_TYPES(
       input.scalar_type(), "dynamic_scaled_int8_quant_kernel", [&] {
-        vllm::dynamic_scaled_int8_quant_kernel<scalar_t, float>
-            <<<grid, block, 0, stream>>>(input.data_ptr<scalar_t>(),
-                                         out.data_ptr<int8_t>(),
-                                         scales.data_ptr<float>(), hidden_size);
+        if (!azp) {
+          vllm::dynamic_scaled_int8_quant_kernel<scalar_t, float>
+              <<<grid, block, 0, stream>>>(
+                  input.data_ptr<scalar_t>(), out.data_ptr<int8_t>(),
+                  scales.data_ptr<float>(), hidden_size);
+        } else {
+          vllm::dynamic_scaled_int8_azp_quant_kernel<scalar_t, float, int32_t>
+              <<<grid, block, 0, stream>>>(
+                  input.data_ptr<scalar_t>(), out.data_ptr<int8_t>(),
+                  scales.data_ptr<float>(), azp->data_ptr<int32_t>(),
+                  hidden_size);
+        }
       });
 }

csrc/torch_bindings.cpp

@@ -248,14 +248,14 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
 
   // Compute int8 quantized tensor for given scaling factor.
   ops.def(
-      "static_scaled_int8_quant(Tensor! out, Tensor input, Tensor scale) -> "
-      "()");
+      "static_scaled_int8_quant(Tensor! out, Tensor input, Tensor scale,"
+      "Tensor? azp) -> ()");
   ops.impl("static_scaled_int8_quant", torch::kCUDA, &static_scaled_int8_quant);
 
   // Compute int8 quantized tensor and scaling factor
   ops.def(
-      "dynamic_scaled_int8_quant(Tensor! out, Tensor input, Tensor! scale) -> "
-      "()");
+      "dynamic_scaled_int8_quant(Tensor! out, Tensor input, Tensor! scale, "
+      "Tensor!? azp) -> ()");
   ops.impl("dynamic_scaled_int8_quant", torch::kCUDA,
            &dynamic_scaled_int8_quant);
 }

tests/kernels/test_int8_quant.py

@@ -12,6 +12,18 @@
 SCALE = [0.1, 0.5, 0.8, 1.2, 2.1]
 
 
+def allclose_int(input, other, atol: int = 0, rtol: float = 1e-5):
+    INT_DTYPES = [
+        torch.int8, torch.int16, torch.int32, torch.int64, torch.uint8,
+        torch.uint16, torch.uint32, torch.uint64
+    ]
+    assert input.dtype in INT_DTYPES and other.dtype in INT_DTYPES
+    diff = torch.abs(input.to(torch.int64) - other.to(torch.int64))
+    return torch.all(
+        diff <= atol +
+        torch.ceil(rtol * torch.abs(other).to(torch.float32)).to(torch.int64))
+
+
 @pytest.mark.parametrize("num_tokens", NUM_TOKENS)
 @pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
 @pytest.mark.parametrize("dtype", DTYPES)
@@ -27,14 +39,52 @@ def test_dynamic_scaled_int8_quant(num_tokens: int, hidden_size: int,
     # reference
     ref_out, ref_scales = ref_dynamic_per_token_quant(x, torch.int8)
     # kernel
-    ops_out, ops_scales = scaled_int8_quant(x)
+    ops_out, ops_scales, _ = scaled_int8_quant(x)
 
     torch.testing.assert_close(ops_scales, ref_scales)
     torch.testing.assert_close(
         ops_out, ref_out, atol=1,
         rtol=0.0)  # big atol to account for rounding errors
 
 
+@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
+@pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+@torch.inference_mode()
+def test_dynamic_scaled_int8_azp_quant(num_tokens: int, hidden_size: int,
+                                       dtype: torch.dtype, seed: int) -> None:
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    int8_traits = torch.iinfo(torch.int8)
+
+    x = torch.rand(num_tokens, hidden_size, dtype=dtype,
+                   device="cuda") * 1000 - 300
+
+    x_token_max, _ = x.to(dtype=torch.float32).max(dim=1, keepdim=True)
+    x_token_min, _ = x.to(dtype=torch.float32).min(dim=1, keepdim=True)
+
+    # calculate scale and azp, and adjust the range
+    scales = (x_token_max - x_token_min) / torch.tensor(255.0)
+    azps = torch.round(-128.0 - x_token_min / scales).to(torch.int32)
+
+    torch_out = ((x / scales).round() + azps).clamp(
+        int8_traits.min, int8_traits.max).to(torch.int8)
+    assert torch_out.min() >= int8_traits.min and torch_out.max(
+    ) <= int8_traits.max
+
+    ops_out = torch.empty_like(x, dtype=torch.int8)
+    scales_out = torch.empty_like(scales, dtype=torch.float32)
+    azp_out = torch.empty_like(azps, dtype=torch.int32)
+    torch.ops._C.dynamic_scaled_int8_quant(ops_out, x, scales_out, azp_out)
+
+    if (not torch.allclose(scales_out, scales)):
+        print(torch.argmax(torch.abs(scales_out - scales)))
+    torch.testing.assert_close(scales_out, scales)
+    assert allclose_int(azp_out, azps, atol=1)  # azp rounding error
+    assert allclose_int(torch_out, ops_out, atol=1)  # azp rounding error
+
+
 @pytest.mark.parametrize("num_tokens", NUM_TOKENS)
 @pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
 @pytest.mark.parametrize("dtype", DTYPES)
@@ -53,8 +103,37 @@ def test_static_scaled_int8_quant(num_tokens: int, hidden_size: int,
 
     out1 = (x / scale).round().clamp(int8_traits.min,
                                      int8_traits.max).to(torch.int8)
-    out2, _ = scaled_int8_quant(x, scale)
+    out2, _, _ = scaled_int8_quant(x, scale)
 
-    torch.testing.assert_close(
-        out1, out2, atol=1,
-        rtol=0.0)  # big atol to account for rounding errors
+    # big atol to account for rounding errors
+    torch.testing.assert_close(out1, out2, atol=1, rtol=0.0)
+
+
+@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
+@pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+@pytest.mark.parametrize("scale", SCALE[2:])  # Reduce test time
+@pytest.mark.parametrize("azp", [-255, 54])
+@torch.inference_mode()
+def test_static_scaled_int8_azp_quant(num_tokens: int, hidden_size: int,
+                                      dtype: torch.dtype, seed: int,
+                                      scale: float, azp: int) -> None:
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    int8_traits = torch.iinfo(torch.int8)
+
+    x = torch.rand(num_tokens, hidden_size, dtype=dtype,
+                   device="cuda") * 1000 - 300
+
+    out1 = ((x / scale).round() + azp).clamp(int8_traits.min,
+                                             int8_traits.max).to(torch.int8)
+    out2 = torch.empty_like(x, dtype=torch.int8)
+    scale_argument = torch.tensor([scale], dtype=torch.float32, device="cuda")
+    azp_argument = torch.tensor([azp], dtype=torch.int32, device="cuda")
+
+    torch.ops._C.static_scaled_int8_quant(out2, x, scale_argument,
+                                          azp_argument)
+
+    # big atol to account for rounding errors
+    torch.testing.assert_close(out1, out2, atol=1, rtol=0.0)