[Perf] Deepgemm fused layout kernel for activations, 4.3% throughput improvement, 10.7% TTFT improvement. (#29546)

Signed-off-by: yewentao256 <zhyanwentao@126.com>

Author: yewentao256

csrc/ops.h

@@ -299,6 +299,14 @@ void per_token_group_quant_int8(const torch::Tensor& input,
                                 torch::Tensor& output_q,
                                 torch::Tensor& output_s, int64_t group_size,
                                 double eps, double int8_min, double int8_max);
+
+// Fused activation quantisation + DeepGEMM-compatible UE8M0-packed scales.
+void per_token_group_quant_8bit_packed(const torch::Tensor& input,
+                                       torch::Tensor& output_q,
+                                       torch::Tensor& output_s_packed,
+                                       int64_t group_size, double eps,
+                                       double min_8bit, double max_8bit);
+
 #endif
 
 void static_scaled_int8_quant(torch::Tensor& out, torch::Tensor const& input,

csrc/quantization/w8a8/fp8/per_token_group_quant.cu

@@ -206,6 +206,191 @@ void per_token_group_quant_8bit(const torch::Tensor& input,
 #undef LAUNCH_KERNEL
 }
 
+template <typename T, typename DST_DTYPE>
+__global__ void per_token_group_quant_8bit_packed_kernel(
+    const T* __restrict__ input, void* __restrict__ output_q,
+    unsigned int* __restrict__ output_s_packed, const int group_size,
+    const int num_groups, const int groups_per_block, const int groups_per_row,
+    const int mn, const int tma_aligned_mn, const float eps,
+    const float min_8bit, const float max_8bit) {
+  const int threads_per_group = 16;
+  const int64_t local_group_id = threadIdx.x / threads_per_group;
+  const int lane_id = threadIdx.x % threads_per_group;
+
+  const int64_t block_group_id = blockIdx.x * groups_per_block;
+  const int64_t global_group_id = block_group_id + local_group_id;
+  if (global_group_id >= num_groups) {
+    return;
+  }
+
+  const int64_t block_group_offset = global_group_id * group_size;
+
+  float local_absmax = eps;
+
+  const T* group_input = input + block_group_offset;
+  DST_DTYPE* group_output =
+      static_cast<DST_DTYPE*>(output_q) + block_group_offset;
+
+  // shared memory to cache each group's data to avoid double DRAM reads.
+  extern __shared__ __align__(16) char smem_raw[];
+  T* smem = reinterpret_cast<T*>(smem_raw);
+  T* smem_group = smem + local_group_id * group_size;
+
+  constexpr int vec_size = 16 / sizeof(T);
+  using vec_t = vllm::vec_n_t<T, vec_size>;
+
+  // copy global -> shared & compute absmax
+  auto scalar_op_cache = [&] __device__(T & dst, const T& src) {
+    float abs_v = fabsf(static_cast<float>(src));
+    local_absmax = fmaxf(local_absmax, abs_v);
+    dst = src;
+  };
+
+  vllm::vectorize_with_alignment<vec_size>(
+      group_input,        // in
+      smem_group,         // out (shared)
+      group_size,         // elements per group
+      lane_id,            // thread id
+      threads_per_group,  // stride in group
+      scalar_op_cache);   // scalar handler
+
+  local_absmax = GroupReduceMax(local_absmax);
+
+  float y_s = local_absmax / max_8bit;
+  y_s = exp2f(ceilf(log2f(fmaxf(fabsf(y_s), 1e-10f))));
+
+  // pack 4 scales into a uint32
+  if (lane_id == 0) {
+    // map flat group id to 2D indices (mn_idx, sf_k_idx)
+    const int sf_k_idx = static_cast<int>(global_group_id % groups_per_row);
+    const int mn_idx = static_cast<int>(global_group_id / groups_per_row);
+
+    if (mn_idx < mn) {
+      // each uint32 in output_s_packed stores 4 packed scales
+      const int sf_k_pack_idx = sf_k_idx / 4;
+      const int pos = sf_k_idx % 4;
+
+      // reinterpret the UE8M0 scale y_s as IEEE bits, extract the 8-bit
+      // exponent, and place it into the correct byte of the 32-bit word.
+      const unsigned int bits = __float_as_uint(y_s);
+      const unsigned int exponent = (bits >> 23u) & 0xffu;
+      const unsigned int contrib = exponent << (pos * 8u);
+
+      const int out_idx = sf_k_pack_idx * tma_aligned_mn + mn_idx;
+      // atomically OR 8-bit exponent into the packed scales buffer
+      atomicOr(output_s_packed + out_idx, contrib);
+    }
+  }
+
+  __syncthreads();
+
+  // quantize shared -> global 8-bit
+  auto scalar_op_quant = [&] __device__(DST_DTYPE & dst, const T& src) {
+    float q = fminf(fmaxf(static_cast<float>(src) / y_s, min_8bit), max_8bit);
+    dst = DST_DTYPE(q);
+  };
+
+  vllm::vectorize_with_alignment<vec_size>(
+      smem_group,         // in (shared)
+      group_output,       // out (global quant tensor)
+      group_size,         // elements
+      lane_id,            // tid
+      threads_per_group,  // stride
+      scalar_op_quant);   // scalar handler
+}
+
+void per_token_group_quant_8bit_packed(const torch::Tensor& input,
+                                       torch::Tensor& output_q,
+                                       torch::Tensor& output_s_packed,
+                                       int64_t group_size, double eps,
+                                       double min_8bit, double max_8bit) {
+  TORCH_CHECK(input.is_contiguous());
+  TORCH_CHECK(output_q.is_contiguous());
+
+  const int64_t k = input.size(-1);
+  TORCH_CHECK(k % group_size == 0, "Last dimension (", k,
+              ") must be divisible by group_size (", group_size, ").");
+
+  const int64_t mn = input.numel() / k;
+  const int64_t groups_per_row = k / group_size;
+  const int64_t num_groups = mn * groups_per_row;
+
+  TORCH_CHECK(output_s_packed.dim() == 2,
+              "output_s_packed must be 2D, got dim=", output_s_packed.dim(),
+              ".");
+
+  const int64_t k_num_packed_sfk = (groups_per_row + 3) / 4;
+  const int64_t tma_aligned_mn = ((mn + 3) / 4) * 4;
+
+  TORCH_CHECK(output_s_packed.scalar_type() == at::ScalarType::Int,
+              "output_s_packed must have dtype int32 for UE8M0-packed scales.");
+  // DeepGEMM expects SFA scales in MN-major form with shape
+  // [mn, ceil_div(K, 128 * 4)] and TMA-aligned stride on the last
+  // dimension.
+  TORCH_CHECK(output_s_packed.size(0) == mn &&
+                  output_s_packed.size(1) == k_num_packed_sfk,
+              "output_s_packed shape must be [", mn, ", ", k_num_packed_sfk,
+              "], but got [", output_s_packed.size(0), ", ",
+              output_s_packed.size(1), "].");
+
+  cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+
+  constexpr int THREADS_PER_GROUP = 16;
+
+  int groups_per_block = 1;
+
+  if (num_groups % 16 == 0) {
+    groups_per_block = 16;
+  } else if (num_groups % 8 == 0) {
+    groups_per_block = 8;
+  } else if (num_groups % 4 == 0) {
+    groups_per_block = 4;
+  } else if (num_groups % 2 == 0) {
+    groups_per_block = 2;
+  }
+
+  auto dst_type = output_q.scalar_type();
+  const int num_blocks = num_groups / groups_per_block;
+  const int num_threads = groups_per_block * THREADS_PER_GROUP;
+
+  // zero-initialize packed scales, since we use atomicOr to accumulate
+  // exponents from different groups.
+  output_s_packed.zero_();
+
+#define LAUNCH_PACKED_KERNEL(T, DST_DTYPE)                                \
+  do {                                                                    \
+    dim3 grid(num_blocks);                                                \
+    dim3 block(num_threads);                                              \
+    size_t smem_bytes =                                                   \
+        static_cast<size_t>(groups_per_block) * group_size * sizeof(T);   \
+    per_token_group_quant_8bit_packed_kernel<T, DST_DTYPE>                \
+        <<<grid, block, smem_bytes, stream>>>(                            \
+            static_cast<const T*>(input.data_ptr()), output_q.data_ptr(), \
+            reinterpret_cast<unsigned int*>(output_s_packed.data_ptr()),  \
+            static_cast<int>(group_size), static_cast<int>(num_groups),   \
+            groups_per_block, static_cast<int>(groups_per_row),           \
+            static_cast<int>(mn), static_cast<int>(tma_aligned_mn),       \
+            static_cast<float>(eps), static_cast<float>(min_8bit),        \
+            static_cast<float>(max_8bit));                                \
+  } while (0)
+
+  VLLM_DISPATCH_FLOATING_TYPES(
+      input.scalar_type(), "per_token_group_quant_8bit_packed", ([&] {
+        if (dst_type == at::ScalarType::Float8_e4m3fn) {
+          LAUNCH_PACKED_KERNEL(scalar_t, __nv_fp8_e4m3);
+        } else if (dst_type == at::ScalarType::Char) {
+          LAUNCH_PACKED_KERNEL(scalar_t, int8_t);
+        } else {
+          TORCH_CHECK(
+              false,
+              "per_token_group_quant_8bit_packed only supports FP8/INT8 "
+              "outputs.");
+        }
+      }));
+
+#undef LAUNCH_PACKED_KERNEL
+}
+
 void per_token_group_quant_fp8(const torch::Tensor& input,
                                torch::Tensor& output_q, torch::Tensor& output_s,
                                int64_t group_size, double eps, double fp8_min,

csrc/torch_bindings.cpp

@@ -617,6 +617,15 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.impl("per_token_group_fp8_quant", torch::kCUDA,
            &per_token_group_quant_fp8);
 
+  // Compute per-token-group 8-bit quantized tensor and UE8M0-packed,
+  // TMA-aligned scales for DeepGEMM.
+  ops.def(
+      "per_token_group_fp8_quant_packed(Tensor input, Tensor! output_q, "
+      "Tensor! output_s_packed, int group_size, float eps, float fp8_min, "
+      "float fp8_max) -> ()");
+  ops.impl("per_token_group_fp8_quant_packed", torch::kCUDA,
+           &per_token_group_quant_8bit_packed);
+
   // Compute per-token-group INT8 quantized tensor and scaling factor.
   ops.def(
       "per_token_group_quant_int8(Tensor input, Tensor! output_q, Tensor! "

vllm/model_executor/layers/fused_moe/deep_gemm_moe.py

@@ -23,9 +23,11 @@
 from vllm.model_executor.layers.fused_moe.utils import _resize_cache
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     per_token_group_quant_fp8,
+    per_token_group_quant_fp8_packed_for_deepgemm,
     silu_mul_per_token_group_quant_fp8_colmajor,
 )
 from vllm.utils.deep_gemm import (
+    DeepGemmQuantScaleFMT,
     get_mk_alignment_for_contiguous_layout,
     m_grouped_fp8_gemm_nt_contiguous,
 )
@@ -157,23 +159,40 @@ def workspace_shapes(
     def _act_mul_quant(
         self, input: torch.Tensor, output: torch.Tensor, activation: str
     ) -> tuple[torch.Tensor, torch.Tensor]:
-        if activation == "silu":
-            return silu_mul_per_token_group_quant_fp8_colmajor(
-                input=input, output=output
-            )
-        else:
-            # This is a fallback path. If we find ourselves using any activation other
-            # than silu, we should add that activation to
-            # silu_mul_per_token_group_quant_fp8_colmajor kernel as it is much faster.
+        assert self.block_shape is not None
+        block_k = self.block_shape[1]
+        scale_fmt = DeepGemmQuantScaleFMT.from_oracle()
+
+        # 1. DeepGemm UE8M0: use packed per-token-group quant
+        if scale_fmt == DeepGemmQuantScaleFMT.UE8M0:
             M_sum, N = input.size()
             act_out = torch.empty(
                 (M_sum, N // 2), dtype=input.dtype, device=input.device
             )
             self.activation(activation, act_out, input)
-            assert self.block_shape is not None
-            return per_token_group_quant_fp8(
-                act_out, self.block_shape[1], column_major_scales=True, out_q=output
+            a2q, a2q_scale = per_token_group_quant_fp8_packed_for_deepgemm(
+                act_out,
+                block_k,
+                out_q=output,
             )
+            return a2q, a2q_scale
+
+        # 2. Hopper / non‑E8M0: prefer the fused SiLU+mul+quant kernel
+        if activation == "silu":
+            use_ue8m0 = scale_fmt == DeepGemmQuantScaleFMT.FLOAT32_CEIL_UE8M0
+            return silu_mul_per_token_group_quant_fp8_colmajor(
+                input=input,
+                output=output,
+                use_ue8m0=use_ue8m0,
+            )
+
+        # 3. fallback path for non-SiLU activations in non‑UE8M0 cases.
+        M_sum, N = input.size()
+        act_out = torch.empty((M_sum, N // 2), dtype=input.dtype, device=input.device)
+        self.activation(activation, act_out, input)
+        return per_token_group_quant_fp8(
+            act_out, block_k, column_major_scales=True, out_q=output
+        )
 
     def apply(
         self,

vllm/model_executor/layers/quantization/utils/fp8_utils.py

@@ -269,7 +269,11 @@ def _run_deepgemm(
         weight_scale: torch.Tensor,
     ) -> torch.Tensor:
         assert self.deepgemm_input_quant_op is not None
-        q_input, input_scale = self.deepgemm_input_quant_op(input_2d)
+        q_input, input_scale = per_token_group_quant_fp8_packed_for_deepgemm(
+            input_2d,
+            group_size=self.act_quant_group_shape.col,
+            use_ue8m0=True,
+        )
         output = torch.empty(
             (q_input.shape[0], weight.shape[0]),
             dtype=torch.bfloat16,
@@ -791,6 +795,80 @@ def per_token_group_quant_fp8(
     return x_q, x_s
 
 
+def per_token_group_quant_fp8_packed_for_deepgemm(
+    x: torch.Tensor,
+    group_size: int,
+    eps: float = 1e-10,
+    use_ue8m0: bool | None = None,
+    out_q: torch.Tensor | None = None,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """FP8 per-token-group quantization for DeepGEMM.
+
+    Returns:
+        (x_q, x_s_packed)
+            x_q: FP8 activations, same shape as `x`.
+            x_s_packed: Int32 tensor with logical shape
+                        [mn, ceil(num_groups_per_row / 4)], laid out with
+                        TMA-aligned stride along the packed-K dimension
+    """
+    if use_ue8m0 is None:
+        use_ue8m0 = is_deep_gemm_e8m0_used()
+    # for DeepGEMM UE8M0-packed layout we *require* UE8M0 scales.
+    assert use_ue8m0, (
+        "per_token_group_quant_fp8_packed_for_deepgemm requires UE8M0 scales."
+    )
+
+    dtype = current_platform.fp8_dtype()
+    assert x.shape[-1] % group_size == 0, (
+        f"the last dimension of `x` {x.shape[-1]} must be divisible "
+        f"by `group_size` {group_size}"
+    )
+    assert x.stride(-1) == 1, "`x` groups must be contiguous"
+
+    finfo = torch.finfo(dtype)
+    fp8_min, fp8_max = finfo.min, finfo.max
+
+    # compute DeepGEMM-style packed scale tensor shape.
+    hidden_dim = x.shape[-1]
+    mn = x.numel() // hidden_dim
+    num_groups_per_row = hidden_dim // group_size
+    k_num_packed_sf_k = (num_groups_per_row + 3) // 4
+    tma_aligned_mn = ((mn + 3) // 4) * 4
+
+    x_s_packed = torch.empty_strided(
+        (mn, k_num_packed_sf_k),
+        (1, tma_aligned_mn),
+        device=x.device,
+        dtype=torch.int32,
+    )
+
+    # CUDA kernel path only (DeepGEMM + E8M0 is CUDA-specific).
+    assert current_platform.is_cuda(), (
+        "per_token_group_quant_fp8_packed_for_deepgemm is only valid on CUDA "
+        "platforms using DeepGEMM."
+    )
+
+    x_contiguous = x.contiguous()
+    if out_q is not None:
+        x_q_local = out_q
+    else:
+        x_q_local = torch.empty_like(x_contiguous, device=x.device, dtype=dtype)
+
+    torch.ops._C.per_token_group_fp8_quant_packed(
+        x_contiguous,
+        x_q_local,
+        x_s_packed,
+        group_size,
+        eps,
+        fp8_min,
+        fp8_max,
+    )
+
+    # return a tensor with the original logical shape.
+    x_q = x_q_local.view_as(x)
+    return x_q, x_s_packed
+
+
 @triton.jit
 def _w8a8_triton_block_scaled_mm(
     # Pointers to inputs and output