Add sgl_per_token_quant_fp8 (#4089)

sgl-kernel/benchmark/bench_per_token_quant_fp8.py

+import itertools
+from typing import Optional, Tuple
+
+import torch
+import triton
+import triton.testing
+from sgl_kernel import sgl_per_token_quant_fp8
+from vllm import _custom_ops as ops
+
+from sglang.srt.utils import is_hip
+
+is_hip_ = is_hip()
+fp8_type_ = torch.float8_e4m3fnuz if is_hip_ else torch.float8_e4m3fn
+
+
+def vllm_per_token_quant_fp8(
+    input: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    return ops.scaled_fp8_quant(input, use_per_token_if_dynamic=True)
+
+
+def sglang_per_token_quant_fp8(
+    input: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    scale = torch.zeros(input.size(0), device=input.device, dtype=torch.float32)
+    output = torch.empty_like(input, device=input.device, dtype=fp8_type_)
+    sgl_per_token_quant_fp8(input, output, scale)
+
+    return output, scale
+
+
+def calculate_diff(batch_size: int, seq_len: int):
+    """Calculate difference between VLLM and SGLang implementations."""
+    device = torch.device("cuda")
+    x = torch.rand((batch_size, seq_len), dtype=torch.float16, device=device)
+
+    vllm_out, vllm_scale = vllm_per_token_quant_fp8(x)
+    sglang_out, sglang_scale = sglang_per_token_quant_fp8(x)
+
+    scale_diff = torch.abs(vllm_scale - sglang_scale).mean().item()
+    output_diff = torch.abs(vllm_out.float() - sglang_out.float()).mean().item()
+
+    print(f"Scale difference: {scale_diff}")
+    print(f"Output difference: {output_diff}")
+
+    if torch.allclose(
+        vllm_out.to(torch.float32), sglang_out.to(torch.float32), rtol=1e-3, atol=1e-5
+    ) and torch.allclose(vllm_scale, sglang_scale, rtol=1e-3, atol=1e-5):
+        print("✅ All implementations match")
+    else:
+        print("❌ Implementations differ")
+
+
+batch_size_range = [16, 32, 64, 128]
+seq_len_range = [64, 128, 256, 512, 1024, 2048, 4096]
+
+configs = list(itertools.product(batch_size_range, seq_len_range))
+
+
+@triton.testing.perf_report(
+    triton.testing.Benchmark(
+        x_names=["batch_size", "seq_len"],
+        x_vals=configs,
+        line_arg="provider",
+        line_vals=["vllm", "sglang"],
+        line_names=["VLLM", "SGL Kernel"],
+        styles=[("blue", "-"), ("green", "-")],
+        ylabel="us",
+        plot_name="per-token-dynamic-quant-fp8-performance",
+        args={},
+    )
+)
+def benchmark_quantization(batch_size, seq_len, provider):
+    dtype = torch.float16
+    device = torch.device("cuda")
+
+    x = torch.randn(batch_size * seq_len, 4096, device=device, dtype=dtype)
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    if provider == "vllm":
+        fn = lambda: vllm_per_token_quant_fp8(x.clone())
+    elif provider == "sglang":
+        fn = lambda: sglang_per_token_quant_fp8(x.clone())
+
+    ms, min_ms, max_ms = triton.testing.do_bench(fn, quantiles=quantiles)
+
+    return 1000 * ms, 1000 * max_ms, 1000 * min_ms
+
+
+if __name__ == "__main__":
+    calculate_diff(batch_size=4, seq_len=4096)
+    benchmark_quantization.run(print_data=True)

sgl-kernel/setup.py

@@ -106,6 +106,7 @@ sources = [
     "src/sgl-kernel/csrc/gemm/fp8_blockwise_gemm_kernel.cu",
     "src/sgl-kernel/csrc/gemm/int8_gemm_kernel.cu",
     "src/sgl-kernel/csrc/gemm/per_token_group_quant_fp8.cu",
+    "src/sgl-kernel/csrc/gemm/per_token_quant_fp8.cu",
     "src/sgl-kernel/csrc/gemm/per_tensor_quant_fp8.cu",
     "src/sgl-kernel/csrc/moe/moe_align_kernel.cu",
     "src/sgl-kernel/csrc/speculative/eagle_utils.cu",

sgl-kernel/src/sgl-kernel/__init__.py

@@ -29,6 +29,7 @@ from sgl_kernel.ops.gemm import (
     int8_scaled_mm,
     sgl_per_tensor_quant_fp8,
     sgl_per_token_group_quant_fp8,
+    sgl_per_token_quant_fp8,
 )
 from sgl_kernel.ops.moe import moe_align_block_size
 from sgl_kernel.ops.sampling import (

sgl-kernel/src/sgl-kernel/csrc/gemm/per_token_quant_fp8.cu

+#include <ATen/cuda/CUDAContext.h>
+#include <c10/util/Float8_e4m3fn.h>
+
+#include <cmath>
+#include <cub/block/block_reduce.cuh>
+#include <flashinfer/vec_dtypes.cuh>
+
+#include "utils.h"
+
+#define WARP_SIZE 32
+
+#ifndef USE_ROCM
+#include <c10/util/Float8_e4m3fn.h>
+using FP8_TYPE = c10::Float8_e4m3fn;
+C10_HOST_DEVICE constexpr auto FP8_E4M3_MAX = std::numeric_limits<FP8_TYPE>::max();
+#else
+#include <c10/util/Float8_e4m3fnuz.h>
+
+#include "amd/quant_utils.cuh"
+using FP8_TYPE = c10::Float8_e4m3fnuz;
+// Using the default max value from pytorch (240.0) will cause accuracy
+// issue when running dynamic quantization. Here use 224.0f for rocm.
+constexpr auto FP8_E4M3_MAX = 224.0f;
+#endif
+
+__device__ __forceinline__ float warpReduceMax(float max_value) {
+  max_value = fmaxf(max_value, __shfl_xor_sync(0xffffffff, max_value, 16));
+  max_value = fmaxf(max_value, __shfl_xor_sync(0xffffffff, max_value, 8));
+  max_value = fmaxf(max_value, __shfl_xor_sync(0xffffffff, max_value, 4));
+  max_value = fmaxf(max_value, __shfl_xor_sync(0xffffffff, max_value, 2));
+  max_value = fmaxf(max_value, __shfl_xor_sync(0xffffffff, max_value, 1));
+  return max_value;
+}
+
+template <typename T>
+__global__ void per_token_quant_fp8_kernel(const T* __restrict__ input, FP8_TYPE* __restrict__ output_q,
+                                           float* __restrict__ output_s, const int64_t hidden_dim,
+                                           const int64_t num_tokens) {
+  const int token_idx = blockIdx.x;
+
+  if (token_idx >= num_tokens) return;
+
+  const int tid = threadIdx.x;
+  const int block_dim = blockDim.x;
+
+  const T* token_input = input + token_idx * hidden_dim;
+  FP8_TYPE* token_output = output_q + token_idx * hidden_dim;
+
+  float max_value = 0.0f;
+
+  for (int i = tid; i < hidden_dim; i += block_dim) {
+    float val = static_cast<float>(token_input[i]);
+    max_value = fmaxf(max_value, fabsf(val));
+  }
+
+  max_value = warpReduceMax(max_value);
+
+  static __shared__ float warpLevelMaxs[WARP_SIZE];
+  const int laneId = threadIdx.x % WARP_SIZE;
+  const int warpId = threadIdx.x / WARP_SIZE;
+
+  if (laneId == 0) warpLevelMaxs[warpId] = max_value;
+  __syncthreads();
+
+  if (warpId == 0) {
+    max_value = (threadIdx.x < blockDim.x / WARP_SIZE) ? warpLevelMaxs[laneId] : 0;
+    max_value = warpReduceMax(max_value);
+  }
+
+  __shared__ float block_max;
+  if (tid == 0) {
+    block_max = max_value / FP8_E4M3_MAX;
+    output_s[token_idx] = block_max;
+  }
+  __syncthreads();
+
+  const float scale_val = 1.0f / block_max;
+
+  constexpr uint32_t vec_size = 16 / sizeof(T);
+  using vec_t = flashinfer::vec_t<T, vec_size>;
+
+  const int32_t num_vec_elems = hidden_dim / vec_size;
+
+  for (int32_t i = tid; i < num_vec_elems; i += block_dim) {
+    vec_t input_vec;
+    input_vec.cast_load(token_input + i * vec_size);
+
+    FP8_TYPE output_arr[vec_size];
+#pragma unroll
+    for (uint32_t j = 0; j < vec_size; ++j) {
+      float val = fmax(fmin(static_cast<float>(input_vec[j]) * scale_val, FP8_E4M3_MAX), -FP8_E4M3_MAX);
+#ifndef USE_ROCM
+      output_arr[j] = static_cast<FP8_TYPE>(val);
+#else
+      output_arr[j] = c10::Float8_e4m3fnuz(
+          __hip_cvt_float_to_fp8(val, fp8::fp8_type::__default_saturation, fp8::fp8_type::__default_interpret),
+          c10::Float8_e4m3fnuz::from_bits());
+#endif
+    }
+
+#pragma unroll
+    for (uint32_t j = 0; j < vec_size; ++j) {
+      token_output[i * vec_size + j] = output_arr[j];
+    }
+  }
+
+  const int32_t remaining_start = num_vec_elems * vec_size;
+  for (int32_t idx = remaining_start + tid; idx < hidden_dim; idx += block_dim) {
+    float val = fmax(-FP8_E4M3_MAX, fmin(static_cast<float>(token_input[idx]) * scale_val, FP8_E4M3_MAX));
+#ifndef USE_ROCM
+    token_output[idx] = static_cast<FP8_TYPE>(val);
+#else
+    token_output[idx] = c10::Float8_e4m3fnuz(
+        __hip_cvt_float_to_fp8(val, fp8::fp8_type::__default_saturation, fp8::fp8_type::__default_interpret),
+        c10::Float8_e4m3fnuz::from_bits());
+#endif
+  }
+}
+
+void sgl_per_token_quant_fp8(torch::Tensor input, torch::Tensor output_q, torch::Tensor output_s) {
+  CHECK_INPUT(input);
+  CHECK_INPUT(output_q);
+  CHECK_INPUT(output_s);
+
+  const auto input_sizes = input.sizes();
+  const int64_t num_tokens = input_sizes[0];
+  const int64_t hidden_dim = input_sizes[1];
+
+  const int block_size = 128;
+  const int num_blocks = num_tokens;
+
+  dim3 grid(num_blocks);
+  dim3 block(block_size);
+
+  cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+
+  DISPATCH_PYTORCH_DTYPE_TO_CTYPE_FLOAT_FP16(input.scalar_type(), scalar_t, [&] {
+    per_token_quant_fp8_kernel<scalar_t><<<grid, block, 0, stream>>>(
+        static_cast<scalar_t*>(input.data_ptr()), static_cast<FP8_TYPE*>(output_q.data_ptr()),
+        static_cast<float*>(output_s.data_ptr()), hidden_dim, num_tokens);
+    return true;
+  });
+}

sgl-kernel/src/sgl-kernel/include/sgl_kernels_ops.h

@@ -160,3 +160,6 @@ void apply_rope_pos_ids_cos_sin_cache(at::Tensor q, at::Tensor k, at::Tensor q_r
 void lightning_attention_decode(const torch::Tensor& q, const torch::Tensor& k, const torch::Tensor& v,
                                 const torch::Tensor& past_kv, const torch::Tensor& slope, torch::Tensor output,
                                 torch::Tensor new_kv);
+
+// sgl_per_token_quant_fp8
+void sgl_per_token_quant_fp8(at::Tensor input, at::Tensor output_q, at::Tensor output_s);

sgl-kernel/src/sgl-kernel/ops/gemm.py

@@ -118,3 +118,11 @@ def cublas_grouped_gemm(
         cublas_handle,
         get_cuda_stream(),
     )
+
+
+def sgl_per_token_quant_fp8(
+    input: torch.Tensor,
+    output_q: torch.Tensor,
+    output_s: torch.Tensor,
+) -> None:
+    torch.ops.sgl_kernels.sgl_per_token_quant_fp8(input, output_q, output_s)

sgl-kernel/src/sgl-kernel/torch_extension.cc

@@ -171,6 +171,9 @@ TORCH_LIBRARY_EXPAND(sgl_kernels, m) {
       "apply_rope_pos_ids_cos_sin_cache(Tensor q, Tensor k, Tensor! q_rope, Tensor! k_rope, Tensor cos_sin_cache, "
       "Tensor pos_ids, bool interleave, int cuda_stream) -> ()");
   m.impl("apply_rope_pos_ids_cos_sin_cache", torch::kCUDA, &apply_rope_pos_ids_cos_sin_cache);
+
+  m.def("sgl_per_token_quant_fp8(Tensor input, Tensor output_q, Tensor output_s) -> ()");
+  m.impl("sgl_per_token_quant_fp8", torch::kCUDA, &sgl_per_token_quant_fp8);
 }
 
 REGISTER_EXTENSION(_kernels)

sgl-kernel/tests/test_per_token_quant_fp8.py

+import itertools
+from typing import Optional, Tuple
+
+import pytest
+import torch
+from sgl_kernel import sgl_per_token_quant_fp8
+from vllm import _custom_ops as ops
+
+from sglang.srt.utils import is_hip
+
+is_hip_ = is_hip()
+fp8_type_ = torch.float8_e4m3fnuz if is_hip_ else torch.float8_e4m3fn
+
+
+def vllm_per_token_quant_fp8(
+    input: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    return ops.scaled_fp8_quant(input, use_per_token_if_dynamic=True)
+
+
+def sglang_per_token_quant_fp8(
+    input: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    scale = torch.zeros(input.size(0), device=input.device, dtype=torch.float32)
+    output = torch.empty_like(input, device=input.device, dtype=fp8_type_)
+
+    sgl_per_token_quant_fp8(input, output, scale)
+    scale = scale.reshape(-1, 1)
+
+    return output, scale
+
+
+@pytest.mark.parametrize(
+    "num_tokens,hidden_dim",
+    list(itertools.product([128, 256, 512], [512, 2048, 4096])),
+)
+def test_per_token_quant_compare_implementations(
+    num_tokens: int,
+    hidden_dim: int,
+):
+    device = torch.device("cuda")
+    x = torch.rand((num_tokens, hidden_dim), dtype=torch.float16, device=device)
+
+    vllm_out, vllm_scale = vllm_per_token_quant_fp8(x)
+    sglang_out, sglang_scale = sglang_per_token_quant_fp8(x)
+
+    torch.testing.assert_close(vllm_scale, sglang_scale, rtol=1e-3, atol=1e-3)
+    torch.testing.assert_close(
+        vllm_out.float(), sglang_out.float(), rtol=1e-3, atol=1e-3
+    )
+
+
+if __name__ == "__main__":
+    # Run the specific test function directly
+    pytest.main([__file__])