[Kernel] Dynamic Per-Token Activation Quantization

csrc/ops.h

@@ -97,6 +97,9 @@ int cutlass_scaled_mm_dq(torch::Tensor& out, torch::Tensor const& a,
 void static_scaled_int8_quant(torch::Tensor& out, torch::Tensor const& input,
                               torch::Tensor const& scale);
 
+void dynamic_scaled_int8_quant(torch::Tensor& out, torch::Tensor& input,
+                               torch::Tensor& scales);
+
 void squeezellm_gemm(torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
                      torch::Tensor lookup_table);
 

csrc/pybind.cpp

@@ -70,6 +70,9 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
   ops.def("static_scaled_int8_quant", &static_scaled_int8_quant,
           "Compute int8 quantized tensor for given scaling factor");
 
+  ops.def("dynamic_scaled_int8_quant", &dynamic_scaled_int8_quant,
+          "Compute int8 quantized tensor and scaling factor");
+
   // Cache ops
   pybind11::module cache_ops = m.def_submodule("cache_ops", "vLLM cache ops");
   cache_ops.def("swap_blocks", &swap_blocks,

csrc/quantization/compressed_tensors/int8_quant_kernels.cu

@@ -3,6 +3,7 @@
 #include <cmath>
 
 #include "../../dispatch_utils.h"
+#include "../../reduction_utils.cuh"
 
 static inline __device__ int8_t float_to_int8_rn(float x) {
 #ifdef USE_ROCM
@@ -38,6 +39,38 @@ __global__ void static_scaled_int8_quant_kernel(
         float_to_int8_rn(((float)input[token_idx * hidden_size + i]) / scale);
   }
 }
+
+template <typename scalar_t, typename scale_type>
+__global__ void dynamic_scaled_int8_quant_kernel(
+    const scalar_t* __restrict__ input, int8_t* __restrict__ out,
+    scale_type scale, const int hidden_size) {
+  const int tid = threadIdx.x;
+  const int token_idx = blockIdx.x;
+
+  float amax_val = 0.0f;
+  const float zero = 0.0f;
+
+  for (int i = tid; i < hidden_size; i += blockDim.x) {
+    float val = (float)input[token_idx * hidden_size + i];
+    val = val > zero ? val : -val;
+    if (val > amax_val) amax_val = val;
+  }
+
+  __shared__ float s_amax;
+  const float block_amax_val = blockReduceMax(amax_val);
+  if (tid == 0) {
+    s_amax = block_amax_val;
+    scale[token_idx] = block_amax_val / 127.0f;
+  }
+  __syncthreads();
+
+  float tmp_scale = 127.0f / s_amax;
+  for (int i = tid; i < hidden_size; i += blockDim.x) {
+    out[token_idx * hidden_size + i] = float_to_int8_rn(
+        ((float)input[token_idx * hidden_size + i]) * tmp_scale);
+  }
+}
+
 }  // namespace vllm
 
 void static_scaled_int8_quant(torch::Tensor& out,          // [..., hidden_size]
@@ -60,3 +93,22 @@ void static_scaled_int8_quant(torch::Tensor& out,          // [..., hidden_size]
                                          scale.data_ptr<float>(), hidden_size);
       });
 }
+
+void dynamic_scaled_int8_quant(torch::Tensor& out,    // [..., hidden_size]
+                               torch::Tensor& input,  // [..., hidden_size]
+                               torch::Tensor& scales) {
+  TORCH_CHECK(input.is_contiguous());
+  TORCH_CHECK(out.is_contiguous());
+  int hidden_size = input.size(-1);
+  int num_tokens = input.numel() / hidden_size;
+  dim3 grid(num_tokens);
+  dim3 block(std::min(hidden_size, 1024));
+  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);
+      });
+}

csrc/reduction_utils.cuh

@@ -21,29 +21,47 @@
 #include "cuda_compat.h"
 
 namespace vllm {
+
+namespace detail {
+
+template <typename T>
+__inline__ __device__ T _max(T a, T b) {
+  return max(a, b);
+}
+
+template <typename T>
+__inline__ __device__ T _sum(T a, T b) {
+  return a + b;
+}
+
+}  // namespace detail
+
+template <typename T>
+using ReduceFnType = T (*)(T, T);
+
+// Helper function to return the next largest power of 2
+static constexpr int _nextPow2(unsigned int num) {
+  if (num <= 1) return num;
+  return 1 << (CHAR_BIT * sizeof(num) - __builtin_clz(num - 1));
+}
+
 template <typename T, int numLanes = WARP_SIZE>
-__inline__ __device__ T warpReduceSum(T val) {
+__inline__ __device__ T warpReduce(T val, ReduceFnType<T> fn) {
   static_assert(numLanes > 0 && (numLanes & (numLanes - 1)) == 0,
                 "numLanes is not a positive power of 2!");
   static_assert(numLanes <= WARP_SIZE);
 #pragma unroll
   for (int mask = numLanes >> 1; mask > 0; mask >>= 1)
-    val += VLLM_SHFL_XOR_SYNC(val, mask);
-  return val;
-}
+    val = fn(val, VLLM_SHFL_XOR_SYNC(val, mask));
 
-// Helper function to return the next largest power of 2
-static constexpr int _nextPow2(unsigned int num) {
-  if (num <= 1) return num;
-  return 1 << (CHAR_BIT * sizeof(num) - __builtin_clz(num - 1));
+  return val;
 }
 
-/* Calculate the sum of all elements in a block */
 template <typename T, int maxBlockSize = 1024>
-__inline__ __device__ T blockReduceSum(T val) {
+__inline__ __device__ T blockReduce(T val, ReduceFnType<T> fn) {
   static_assert(maxBlockSize <= 1024);
   if constexpr (maxBlockSize > WARP_SIZE) {
-    val = warpReduceSum<T>(val);
+    val = warpReduce<T>(val, fn);
     // Calculates max number of lanes that need to participate in the last
     // warpReduce
     constexpr int maxActiveLanes = (maxBlockSize + WARP_SIZE - 1) / WARP_SIZE;
@@ -56,12 +74,22 @@ __inline__ __device__ T blockReduceSum(T val) {
 
     val = (threadIdx.x < blockDim.x / float(WARP_SIZE)) ? shared[lane]
                                                         : (T)(0.0f);
-    val = warpReduceSum<T, _nextPow2(maxActiveLanes)>(val);
+    val = warpReduce<T, _nextPow2(maxActiveLanes)>(val, fn);
   } else {
     // A single warpReduce is equal to blockReduce
-    val = warpReduceSum<T, _nextPow2(maxBlockSize)>(val);
+    val = warpReduce<T, _nextPow2(maxBlockSize)>(val, fn);
   }
   return val;
 }
 
+template <typename T, int maxBlockSize = 1024>
+__inline__ __device__ T blockReduceMax(T val) {
+  return blockReduce<T, maxBlockSize>(val, detail::_max<T>);
+}
+
+template <typename T, int maxBlockSize = 1024>
+__inline__ __device__ T blockReduceSum(T val) {
+  return blockReduce<T, maxBlockSize>(val, detail::_sum<T>);
+}
+
 }  // namespace vllm

tests/kernels/test_int8_quant.py

@@ -10,21 +10,52 @@
 SCALE = [0.1, 0.5, 0.8, 1.2, 2.1]
 
 
+@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_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
+
+    x_token_max, _ = x.max(dim=1)
+    x_token_max = x_token_max.to(dtype=torch.float32)
+    scales = (x_token_max / float(127.0))[:, None].to(device="cuda",
+                                                      dtype=torch.float32)
+    torch_out = (x / scales).round().clamp(int8_traits.min,
+                                           int8_traits.max).to(torch.int8)
+
+    ops_out = torch.empty_like(x, dtype=torch.int8, device="cuda")
+    scales_out = torch.empty_like(scales, dtype=torch.float32, device="cuda")
+    ops.dynamic_scaled_int8_quant(ops_out, x, scales_out)
+
+    assert torch.allclose(scales_out, scales)
+    assert torch.allclose(torch_out, ops_out,
+                          atol=1)  # 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)
 @pytest.mark.parametrize("scale", SCALE)
 @torch.inference_mode()
-def test_quant(num_tokens: int, hidden_size: int, dtype: torch.dtype,
-               seed: int, scale: float) -> None:
+def test_static_scaled_int8_quant(num_tokens: int, hidden_size: int,
+                                  dtype: torch.dtype, seed: int,
+                                  scale: float) -> 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
 
-    out1 = (x / scale).round().clamp(
-        torch.iinfo(torch.int8).min,
-        torch.iinfo(torch.int8).max).to(torch.int8)
+    out1 = (x / scale).round().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")
 

tests/quantization/test_compressed_tensors.py

@@ -6,7 +6,8 @@
 import torch
 
 from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors import (  # noqa: E501
-    CompressedTensorsLinearMethod, CompressedTensorsW8A8StaticTensor)
+    CompressedTensorsLinearMethod, CompressedTensorsW8A8DynamicToken,
+    CompressedTensorsW8A8StaticTensor)
 
 
 def test_compressed_tensors_w8a8_static_setup(vllm_runner):
@@ -34,3 +35,19 @@ def test_compressed_tensors_w8a8_static_setup(vllm_runner):
     assert qkv_proj.weight_scale.shard_splitter is not None
     assert qkv_proj.weight_scale.logical_widths is not None
     assert qkv_proj.input_scale.dtype is torch.float32
+
+
+def test_compressed_tensors_w8a8_dynanmic_per_token(vllm_runner):
+    model_path = "nm-testing/tinyllama-one-shot-dynamic-test"
+    llm = vllm_runner(model_path,
+                      quantization="sparseml",
+                      enforce_eager=True,
+                      dtype=torch.float16)
+    model = llm.model.llm_engine.model_executor.driver_worker.model_runner.model
+    layer = model.model.layers[0]
+
+    qkv_proj = layer.self_attn.qkv_proj
+
+    assert isinstance(qkv_proj.quant_method, CompressedTensorsLinearMethod)
+    assert isinstance(qkv_proj.scheme, CompressedTensorsW8A8DynamicToken)
+    assert qkv_proj.weight.dtype is torch.int8

vllm/_custom_ops.py

@@ -264,21 +264,33 @@ def scaled_fp8_quant(
 
 
 # int8
-def static_scaled_int8_quant(input: torch.Tensor,
-                             scale: torch.Tensor) -> torch.Tensor:
+def scaled_int8_quant(
+        input: torch.Tensor,
+        scale: Optional[torch.Tensor] = None
+) -> Tuple[torch.Tensor, torch.Tensor]:
     """
-    Quantize the input tensor to int8 and return the quantized tensor.
+    Quantize the input tensor to int8 and return the quantized tensor and scale.
 
     Args:
         input: The input tensor to be quantized to int8.
-        scale: Scaling factor for the int8 quantization.
+        scale: Optional scaling factor for the int8 quantization.
+            When not provided, we invoke dynamic-per-token quantization.
 
     Returns:
-        torch.Tensor: Output tensor in int8.
+      Tuple[Torch.Tensor, Torch.Tensor] : Output int8 tensor and scales.
     """
-    q = torch.empty_like(input, dtype=torch.int8)
-    vllm_ops.static_scaled_int8_quant(q, input, scale)
-    return q
+    output = torch.empty_like(input, dtype=torch.int8)
+    if scale is not None:
+        # static-per-tensor quantization.
+        vllm_ops.static_scaled_int8_quant(output, input, scale)
+        return output, scale
+
+    # dynamic-per-token quantization.
+    input_scales = torch.empty((input.numel() // input.shape[-1], 1),
+                               device=input.device,
+                               dtype=torch.float32)
+    vllm_ops.dynamic_scaled_int8_quant(output, input, input_scales)
+    return output, input_scales
 
 
 # moe