From d3dfb8f2dbf325acb2727fffaa33f312f7204d56 Mon Sep 17 00:00:00 2001 From: xgqdut2016 Date: Wed, 24 Jun 2026 09:24:32 +0800 Subject: [PATCH 1/4] issue/1319: success moe_wna16_marlin_gemm kernel --- include/infiniop/ops/moe_wna16_marlin_gemm.h | 57 + .../nvidia/gptq_marlin_gemm_nvidia.cu | 2 +- src/infiniop/ops/moe_wna16_marlin_gemm/info.h | 84 + .../moe_wna16_marlin_gemm/marlin/dequant.h | 508 +++++ .../ops/moe_wna16_marlin_gemm/marlin/kernel.h | 37 + .../moe_wna16_marlin_gemm/marlin/marlin.cuh | 85 + .../marlin/marlin_dtypes.cuh | 78 + .../marlin/marlin_template.h | 1912 +++++++++++++++++ .../moe_wna16_marlin_gemm.h | 83 + .../moe_wna16_marlin_gemm/nvidia/kernel.cuh | 834 +++++++ .../nvidia/moe_wna16_marlin_gemm_nvidia.cu | 401 ++++ .../nvidia/moe_wna16_marlin_gemm_nvidia.cuh | 8 + .../ops/moe_wna16_marlin_gemm/operator.cc | 144 ++ .../sgl_kernel/scalar_type.hpp | 332 +++ .../sgl_kernel/source_location.h | 40 + .../moe_wna16_marlin_gemm/sgl_kernel/tensor.h | 547 +++++ .../sgl_kernel/utils.cuh | 312 +++ .../moe_wna16_marlin_gemm/sgl_kernel/utils.h | 245 +++ test/infiniop/libinfiniop/op_register.py | 63 + 19 files changed, 5771 insertions(+), 1 deletion(-) create mode 100644 include/infiniop/ops/moe_wna16_marlin_gemm.h create mode 100644 src/infiniop/ops/moe_wna16_marlin_gemm/info.h create mode 100644 src/infiniop/ops/moe_wna16_marlin_gemm/marlin/dequant.h create mode 100644 src/infiniop/ops/moe_wna16_marlin_gemm/marlin/kernel.h create mode 100644 src/infiniop/ops/moe_wna16_marlin_gemm/marlin/marlin.cuh create mode 100644 src/infiniop/ops/moe_wna16_marlin_gemm/marlin/marlin_dtypes.cuh create mode 100644 src/infiniop/ops/moe_wna16_marlin_gemm/marlin/marlin_template.h create mode 100644 src/infiniop/ops/moe_wna16_marlin_gemm/moe_wna16_marlin_gemm.h create mode 100644 src/infiniop/ops/moe_wna16_marlin_gemm/nvidia/kernel.cuh create mode 100644 src/infiniop/ops/moe_wna16_marlin_gemm/nvidia/moe_wna16_marlin_gemm_nvidia.cu create mode 100644 src/infiniop/ops/moe_wna16_marlin_gemm/nvidia/moe_wna16_marlin_gemm_nvidia.cuh create mode 100644 src/infiniop/ops/moe_wna16_marlin_gemm/operator.cc create mode 100644 src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/scalar_type.hpp create mode 100644 src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/source_location.h create mode 100644 src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/tensor.h create mode 100644 src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/utils.cuh create mode 100644 src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/utils.h diff --git a/include/infiniop/ops/moe_wna16_marlin_gemm.h b/include/infiniop/ops/moe_wna16_marlin_gemm.h new file mode 100644 index 000000000..eaf487bc0 --- /dev/null +++ b/include/infiniop/ops/moe_wna16_marlin_gemm.h @@ -0,0 +1,57 @@ +#ifndef __INFINIOP_MOE_WNA16_MARLIN_GEMM_API_H__ +#define __INFINIOP_MOE_WNA16_MARLIN_GEMM_API_H__ + +#include "../operator_descriptor.h" +#include + +typedef struct InfiniopDescriptor *infiniopMoeWna16MarlinGemmDescriptor_t; + +__INFINI_C __export infiniStatus_t infiniopCreateMoeWna16MarlinGemmDescriptor(infiniopHandle_t handle, + infiniopMoeWna16MarlinGemmDescriptor_t *desc_ptr, + infiniopTensorDescriptor_t c_desc, + infiniopTensorDescriptor_t a_desc, + infiniopTensorDescriptor_t b_q_weight_desc, + infiniopTensorDescriptor_t b_bias_desc, + infiniopTensorDescriptor_t b_scales_desc, + infiniopTensorDescriptor_t global_scales_desc, + infiniopTensorDescriptor_t b_zeros_desc, + infiniopTensorDescriptor_t g_idx_desc, + infiniopTensorDescriptor_t perm_desc, + infiniopTensorDescriptor_t sorted_token_desc, + infiniopTensorDescriptor_t expert_ids_desc, + infiniopTensorDescriptor_t num_tokens_post_padded_desc, + infiniopTensorDescriptor_t topk_weights_desc, + int size_m, int size_n, int size_k, + int top_k, int moe_block_size); +; + +__INFINI_C __export infiniStatus_t infiniopGetMoeWna16MarlinGemmWorkspaceSize(infiniopMoeWna16MarlinGemmDescriptor_t desc, size_t *size); + +__INFINI_C __export infiniStatus_t infiniopMoeWna16MarlinGemm(infiniopMoeWna16MarlinGemmDescriptor_t desc, + void *workspace, + size_t workspace_size, + void *c, + const void *a, + const void *b_q_weight, + void *b_bias, + void *b_scales, + void *global_scales, + void *b_zeros, + void *g_idx, + void *perm, + void *sorted_token_ids, + void *expert_ids, + void *num_tokens_post_padded, + void *topk_weights, + bool mul_topk_weights, + bool is_ep, + int64_t b_q_type_id, + bool is_k_full, + bool use_atomic_add, + bool use_fp32_reduce, + bool is_zp_float, + void *stream); + +__INFINI_C __export infiniStatus_t infiniopDestroyMoeWna16MarlinGemmDescriptor(infiniopMoeWna16MarlinGemmDescriptor_t desc); + +#endif diff --git a/src/infiniop/ops/gptq_marlin_gemm/nvidia/gptq_marlin_gemm_nvidia.cu b/src/infiniop/ops/gptq_marlin_gemm/nvidia/gptq_marlin_gemm_nvidia.cu index edbe5fc8f..ef2696848 100644 --- a/src/infiniop/ops/gptq_marlin_gemm/nvidia/gptq_marlin_gemm_nvidia.cu +++ b/src/infiniop/ops/gptq_marlin_gemm/nvidia/gptq_marlin_gemm_nvidia.cu @@ -1041,7 +1041,7 @@ infiniStatus_t gptq_marlin_gemm_kernel(void *c, return INFINI_STATUS_SUCCESS; } #endif -int getCudaDeviceSMCount() { +static int getCudaDeviceSMCount() { int dev; cudaGetDevice(&dev); cudaDeviceProp prop; diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/info.h b/src/infiniop/ops/moe_wna16_marlin_gemm/info.h new file mode 100644 index 000000000..94e3bf71b --- /dev/null +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/info.h @@ -0,0 +1,84 @@ +#ifndef __MOE_WNA16_MARLIN_GEMM_INFO_H__ +#define __MOE_WNA16_MARLIN_GEMM_INFO_H__ + +#include "../../../utils.h" +#include "../../tensor.h" +#include + +#include + +namespace op::moe_wna16_marlin_gemm { + +class MoeWna16MarlinGemmInfo { + MoeWna16MarlinGemmInfo() = default; + +public: + infiniDtype_t dtype; + int size_m, size_n, size_k, top_k, moe_block_size; + size_t num_groups, sorted_token_ids_size_0, b_q_weight_size_1, b_q_weight_size_2, b_zeros_size_1, b_zeros_size_2, c_size_0; + bool has_act_order, has_bias, has_zp; + + static utils::Result create( + infiniopTensorDescriptor_t c_desc, + infiniopTensorDescriptor_t a_desc, + infiniopTensorDescriptor_t b_q_weight_desc, + infiniopTensorDescriptor_t b_bias_desc, + infiniopTensorDescriptor_t b_scales_desc, + infiniopTensorDescriptor_t global_scales_desc, + infiniopTensorDescriptor_t b_zeros_desc, + infiniopTensorDescriptor_t g_idx_desc, + infiniopTensorDescriptor_t perm_desc, + infiniopTensorDescriptor_t sorted_token_desc, + infiniopTensorDescriptor_t expert_ids_desc, + infiniopTensorDescriptor_t num_tokens_post_padded_desc, + infiniopTensorDescriptor_t topk_weights_desc, int size_m, int size_n, int size_k, int top_k, int moe_block_size) { + CHECK_OR_RETURN( + c_desc != nullptr && a_desc != nullptr && b_q_weight_desc != nullptr && b_scales_desc != nullptr, + INFINI_STATUS_NULL_POINTER); + const infiniDtype_t dtype = a_desc->dtype(); + + CHECK_OR_RETURN(a_desc->dim(0) == static_cast(size_m) + && a_desc->dim(1) == static_cast(size_k) + && c_desc->dim(1) == static_cast(size_n), + INFINI_STATUS_BAD_TENSOR_SHAPE); + CHECK_OR_RETURN(b_scales_desc->ndim() == 3 + && b_scales_desc->dim(2) == static_cast(size_n), + INFINI_STATUS_BAD_TENSOR_SHAPE); + size_t num_groups = b_scales_desc->dim(1); + bool has_act_order = false; + bool has_bias = (b_bias_desc != nullptr); + bool has_zp = (b_zeros_desc != nullptr); + if (g_idx_desc != nullptr && perm_desc != nullptr) { + CHECK_OR_RETURN(g_idx_desc->dim(g_idx_desc->ndim() - 1) == static_cast(size_k) + && perm_desc->dim(perm_desc->ndim() - 1) == static_cast(size_k), + INFINI_STATUS_BAD_TENSOR_SHAPE); + has_act_order = true; + } + if (num_groups > 1) { + CHECK_OR_RETURN(static_cast(size_k) % num_groups == 0, + INFINI_STATUS_BAD_TENSOR_SHAPE); + } + if (b_bias_desc != nullptr) { + CHECK_OR_RETURN(b_bias_desc->dim(1) == static_cast(size_n) + && b_bias_desc->strides()[1] == 1, + INFINI_STATUS_BAD_TENSOR_SHAPE); + } + + size_t sorted_token_ids_size_0 = sorted_token_desc->dim(0); + size_t b_q_weight_size_1 = b_q_weight_desc->dim(1); + size_t b_q_weight_size_2 = b_q_weight_desc->dim(2); + size_t b_zeros_size_1 = 0; + size_t b_zeros_size_2 = 0; + if (b_zeros_desc != nullptr) { + b_zeros_size_1 = b_zeros_desc->dim(1); + b_zeros_size_2 = b_zeros_desc->dim(2); + } + size_t c_size_0 = c_desc->dim(0); + return utils::Result( + MoeWna16MarlinGemmInfo{dtype, size_m, size_n, size_k, top_k, moe_block_size, num_groups, sorted_token_ids_size_0, b_q_weight_size_1, b_q_weight_size_2, b_zeros_size_1, b_zeros_size_2, has_act_order, has_bias, has_zp}); + } +}; + +} // namespace op::moe_wna16_marlin_gemm + +#endif // __MOE_WNA16_MARLIN_GEMM_INFO_H__ diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/marlin/dequant.h b/src/infiniop/ops/moe_wna16_marlin_gemm/marlin/dequant.h new file mode 100644 index 000000000..6a0d90e5d --- /dev/null +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/marlin/dequant.h @@ -0,0 +1,508 @@ +/* +Fast Dequantization (Converting INT4/INT8/FP4/FP8 to FP16/BF16) + +The process of fast dequantization can be summarized as a combination +of bitwise operations and floating-point computations: + +weight =>(bit_op / bitwise operations)=> +f16_value =>(flop / floating-point computation)=> +dequantized_weight + +Since the dequantized weights typically require subtracting the zero point and +applying a scale factor, the floating-point computation step can be fused with +the zero-point subtraction and scaling operations. + +The following are the parts that need to be modified for the fused operation +of zero-point subtraction and scaling. + +## INT4 => FP16/BF16 or INT8 => FP16 + +The floating-point computation is `__hsub2` + +If has zero points: + + flop(bit_op(weight)) - flop(bit_op(zp)) + = sub(bit_op(weight), bias) - sub(bit_op(zp), bias) + = bit_op(weight) - bit_op(zp) + +so we don't need additional modification. + +If has float zero points: + + flop(bit_op(weight)) - fzp + = sub(bit_op(weight), bias) - fzp + = bit_op(weight) - (fzp + bias) + +where the `fzp + bias` can be computed at weight loading. But this +may have accuracy issue, so we should not use this in most cases. + +If has not zero points: + + scale(flop(bit_op(weight))) + = scale(sub(bit_op(weight), bias)) + = scale(bit_op(weight)) - scale(bias) + = fma(bit_op(weight), scale_factor, scale(bias)) + +where the `scale(bias)` can be cached. But this may have accuracy issue, +so we should not use this in most cases. + + +## INT8 => BF16 + +INT8 => BF16 is a special case, it use byte_perm instead of flop. +We cannot fused byte_perm with scaling. + + +## FP4/FP8 => FP16/BF16 + + scale(flop(bit_op(weight))) + = scale(mul(bit_op(weight), multiplier)) + = mul(bit_op(weight), scale_factor * multiplier) + +where `scale_factor * multiplier` can be computed at weight loading. + +*/ + +#include "marlin_dtypes.cuh" + +namespace device::marlin { + +#if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 800 +// Lookup-table based 3-input logical operation; explicitly used for +// dequantization as the compiler does not seem to automatically recognize it in +// all cases. +template +__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; +} + +// Constructs destination register by taking bytes from 2 sources (based on +// mask) +template +__device__ inline uint32_t prmt(uint32_t a) { + uint32_t res; + asm volatile("prmt.b32 %0, %1, %2, %3;\n" + : "=r"(res) + : "r"(a), "n"(start_byte), "n"(mask)); + return res; +} + +template +__device__ inline void dequant(int q, scalar_t2 *frag_b); + +// +// Efficiently dequantize 4bit values packed in an int32 value into a full +// B-fragment of 4 fp16 values. We mostly follow the strategy in the link below, +// with some small changes: +// - FP16: +// https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/cutlass_extensions/include/cutlass_extensions/interleaved_numeric_conversion.h#L215-L287 +// - BF16: +// https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/cutlass_extensions/include/cutlass_extensions/interleaved_numeric_conversion.h#L327-L385 +// +template <> +__device__ inline void dequant(int q, half2 *frag_b) { + const int MASK = 0x000f000f; + const int EX = 0x64006400; + // Guarantee that the `(a & b) | c` operations are LOP3s. + int lo = lop3<(0xf0 & 0xcc) | 0xaa>(q, MASK, EX); + q >>= 4; + int hi = lop3<(0xf0 & 0xcc) | 0xaa>(q, MASK, EX); + + frag_b[0] = *reinterpret_cast(&lo); + frag_b[1] = *reinterpret_cast(&hi); +} + +template <> +__device__ inline void dequant(int q, half2 *frag_b) { + const int LO = 0x000f000f; + const int HI = 0x00f000f0; + const int EX = 0x64006400; + // Guarantee that the `(a & b) | c` operations are LOP3s. + // clang-format off + int lo = lop3<(0xf0 & 0xcc) | 0xaa>(q, LO, EX); + int hi = lop3<(0xf0 & 0xcc) | 0xaa>(q, HI, EX); + // clang-format on + // We want signed int4 outputs, hence we fuse the `-8` symmetric zero point + // directly into `SUB` and `ADD`. + const int SUB = 0x64086408; + const int MUL = 0x2c002c00; + const int ADD = 0xd480d480; + frag_b[0] = __hsub2(*reinterpret_cast(&lo), *reinterpret_cast(&SUB)); + frag_b[1] = __hfma2( + *reinterpret_cast(&hi), *reinterpret_cast(&MUL), *reinterpret_cast(&ADD)); +} + +template <> +__device__ inline void dequant(int q, half2 *frag_b) { + dequant(q, frag_b); +} + +template <> +__device__ inline void dequant(int q, half2 *frag_b) { + const int LO = 0x000f000f; + const int HI = 0x00f000f0; + const int EX = 0x64006400; + // Guarantee that the `(a & b) | c` operations are LOP3s. + // clang-format off + int lo = lop3<(0xf0 & 0xcc) | 0xaa>(q, LO, EX); + int hi = lop3<(0xf0 & 0xcc) | 0xaa>(q, HI, EX); + // clang-format on + // We want signed int4 outputs, hence we fuse the `-8` symmetric zero point + // directly into `SUB` and `ADD`. + const int SUB = 0x64006400; + const int MUL = 0x2c002c00; + const int ADD = 0xd400d400; + frag_b[0] = __hsub2(*reinterpret_cast(&lo), *reinterpret_cast(&SUB)); + frag_b[1] = __hfma2( + *reinterpret_cast(&hi), *reinterpret_cast(&MUL), *reinterpret_cast(&ADD)); +} + +template <> +__device__ inline void dequant(int q, nv_bfloat162 *frag_b) { + static constexpr uint32_t MASK = 0x000f000f; + static constexpr uint32_t EX = 0x43004300; + + // Guarantee that the `(a & b) | c` operations are LOP3s. + // clang-format off + int lo = lop3<(0xf0 & 0xcc) | 0xaa>(q, MASK, EX); + q >>= 4; + int hi = lop3<(0xf0 & 0xcc) | 0xaa>(q, MASK, EX); + // clang-format on + + frag_b[0] = *reinterpret_cast(&lo); + frag_b[1] = *reinterpret_cast(&hi); +} + +template <> +__device__ inline void dequant(int q, nv_bfloat162 *frag_b) { + dequant(q, frag_b); + + static constexpr uint32_t SUB = 0x43084308; + + frag_b[0] = __hsub2(frag_b[0], *reinterpret_cast(&SUB)); + frag_b[1] = __hsub2(frag_b[1], *reinterpret_cast(&SUB)); +} + +template <> +__device__ inline void dequant(int q, nv_bfloat162 *frag_b) { + dequant(q, frag_b); +} + +template <> +__device__ inline void dequant(int q, nv_bfloat162 *frag_b) { + dequant(q, frag_b); + + static constexpr uint32_t SUB = 0x43004300; + + frag_b[0] = __hsub2(frag_b[0], *reinterpret_cast(&SUB)); + frag_b[1] = __hsub2(frag_b[1], *reinterpret_cast(&SUB)); +} + +// +// Fast Int8ToFp16/Int8ToBf16: Efficiently dequantize 8bit int values to fp16 or +// bf16 Reference: +// - FP16: +// https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/cutlass_extensions/include/cutlass_extensions/interleaved_numeric_conversion.h#L53-L85 +// - BF16: +// https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/cutlass_extensions/include/cutlass_extensions/interleaved_numeric_conversion.h#L125-L175 +// +template <> +__device__ inline void dequant(int q, half2 *frag_b) { + static constexpr uint32_t mask_for_elt_01 = 0x5250; + static constexpr uint32_t mask_for_elt_23 = 0x5351; + static constexpr uint32_t start_byte_for_fp16 = 0x64646464; + + uint32_t lo = prmt(q); + uint32_t hi = prmt(q); + + frag_b[0] = *reinterpret_cast(&lo); + frag_b[1] = *reinterpret_cast(&hi); +} + +template <> +__device__ inline void dequant(int q, half2 *frag_b) { + dequant(q, frag_b); + + static constexpr uint32_t I8s_TO_F16s_MAGIC_NUM = 0x64806480; + frag_b[0] = __hsub2(frag_b[0], *reinterpret_cast(&I8s_TO_F16s_MAGIC_NUM)); + frag_b[1] = __hsub2(frag_b[1], *reinterpret_cast(&I8s_TO_F16s_MAGIC_NUM)); +} + +template <> +__device__ inline void dequant(int q, half2 *frag_b) { + dequant(q, frag_b); +} + +template <> +__device__ inline void dequant(int q, half2 *frag_b) { + dequant(q, frag_b); + + static constexpr uint32_t I8s_TO_F16s_MAGIC_NUM = 0x64006400; + frag_b[0] = __hsub2(frag_b[0], *reinterpret_cast(&I8s_TO_F16s_MAGIC_NUM)); + frag_b[1] = __hsub2(frag_b[1], *reinterpret_cast(&I8s_TO_F16s_MAGIC_NUM)); +} + +template <> +__device__ inline void dequant(int q, nv_bfloat162 *frag_b) { + float fp32_intermediates[4]; + uint32_t *fp32_intermediates_casted = reinterpret_cast(fp32_intermediates); + + static constexpr uint32_t fp32_base = 0x4B000000; + fp32_intermediates_casted[0] = __byte_perm(q, fp32_base, 0x7650); + fp32_intermediates_casted[1] = __byte_perm(q, fp32_base, 0x7652); + fp32_intermediates_casted[2] = __byte_perm(q, fp32_base, 0x7651); + fp32_intermediates_casted[3] = __byte_perm(q, fp32_base, 0x7653); + + fp32_intermediates[0] -= 8388736.f; + fp32_intermediates[1] -= 8388736.f; + fp32_intermediates[2] -= 8388736.f; + fp32_intermediates[3] -= 8388736.f; + + uint32_t *bf16_result_ptr = reinterpret_cast(frag_b); + bf16_result_ptr[0] = __byte_perm(fp32_intermediates_casted[0], fp32_intermediates_casted[1], 0x7632); + bf16_result_ptr[1] = __byte_perm(fp32_intermediates_casted[2], fp32_intermediates_casted[3], 0x7632); +} + +template <> +__device__ inline void dequant(int q, nv_bfloat162 *frag_b) { + float fp32_intermediates[4]; + uint32_t *fp32_intermediates_casted = reinterpret_cast(fp32_intermediates); + + static constexpr uint32_t fp32_base = 0x4B000000; + fp32_intermediates_casted[0] = __byte_perm(q, fp32_base, 0x7650); + fp32_intermediates_casted[1] = __byte_perm(q, fp32_base, 0x7652); + fp32_intermediates_casted[2] = __byte_perm(q, fp32_base, 0x7651); + fp32_intermediates_casted[3] = __byte_perm(q, fp32_base, 0x7653); + + fp32_intermediates[0] -= 8388608.f; + fp32_intermediates[1] -= 8388608.f; + fp32_intermediates[2] -= 8388608.f; + fp32_intermediates[3] -= 8388608.f; + + uint32_t *bf16_result_ptr = reinterpret_cast(frag_b); + bf16_result_ptr[0] = __byte_perm(fp32_intermediates_casted[0], fp32_intermediates_casted[1], 0x7632); + bf16_result_ptr[1] = __byte_perm(fp32_intermediates_casted[2], fp32_intermediates_casted[3], 0x7632); +} + +template <> +__device__ inline void dequant(int q, half2 *frag_b) { + // Constants for FP8 (E4M3) and FP16 formats + constexpr int FP8_EXPONENT = 4, FP16_EXPONENT = 5; + constexpr int RIGHT_SHIFT = FP16_EXPONENT - FP8_EXPONENT; + constexpr int MASK = 0x7F007F00; + + // Extract and shift FP8 values to FP16 format + int Out1 = (q & 0x80008000) | ((q & MASK) >> RIGHT_SHIFT); + q <<= 8; + int Out2 = (q & 0x80008000) | ((q & MASK) >> RIGHT_SHIFT); + + // Note: reverse indexing is intentional because weights are permuted + frag_b[1] = *reinterpret_cast(&Out1); + frag_b[0] = *reinterpret_cast(&Out2); +} + +template <> +__device__ inline void dequant(int q, half2 *frag_b) { + dequant(q, frag_b); + + // Constants for FP8 (E4M3) and FP16 formats + constexpr int FP8_EXPONENT = 4, FP16_EXPONENT = 5; + + // Construct and apply exponent bias + constexpr int BIAS_OFFSET = (1 << (FP16_EXPONENT - 1)) - (1 << (FP8_EXPONENT - 1)); + const half2 bias_reg = __float2half2_rn(float(1 << BIAS_OFFSET)); + + // Convert to half2 and apply bias + frag_b[1] = __hmul2(frag_b[1], bias_reg); + frag_b[0] = __hmul2(frag_b[0], bias_reg); +} + +template <> +__device__ inline void dequant(int q, nv_bfloat162 *frag_b) { + // Constants for FP8 (E4M3) and BF16 formats + constexpr int FP8_EXPONENT = 4, BF16_EXPONENT = 8; + constexpr int RIGHT_SHIFT = BF16_EXPONENT - FP8_EXPONENT; + + constexpr int MASK = 0x7F007F00; + + // Extract and shift FP8 values to BF16 format + int Out1 = (q & 0x80008000) | ((q & MASK) >> RIGHT_SHIFT); + q <<= 8; + int Out2 = (q & 0x80008000) | ((q & MASK) >> RIGHT_SHIFT); + + // Note: reverse indexing is intentional because weights are permuted + frag_b[1] = *reinterpret_cast(&Out1); + frag_b[0] = *reinterpret_cast(&Out2); +} + +template <> +__device__ inline void dequant(int q, nv_bfloat162 *frag_b) { + dequant(q, frag_b); + + // Constants for FP8 (E4M3) and BF16 formats + constexpr int FP8_EXPONENT = 4, BF16_EXPONENT = 8; + + // Construct and apply exponent bias + constexpr int BIAS_OFFSET = (1 << (BF16_EXPONENT - 1)) - (1 << (FP8_EXPONENT - 1)); + // Add 127 (float exponent bias) to BIAS_OFFSET and shift to float exponent + // position + constexpr uint32_t BIAS = (BIAS_OFFSET + 127) << 23; + const nv_bfloat162 bias_reg = __float2bfloat162_rn(*reinterpret_cast(&BIAS)); + + // Convert to bfloat162 and apply bias + frag_b[1] = __hmul2(frag_b[1], bias_reg); + frag_b[0] = __hmul2(frag_b[0], bias_reg); +} + +template <> +__device__ inline void dequant(int q, half2 *frag_b) { + // Constants for FP4 (E2M1) and FP16 formats + constexpr int FP4_EXPONENT = 2, FP16_EXPONENT = 5; + constexpr int RIGHT_SHIFT = FP16_EXPONENT - FP4_EXPONENT; + constexpr int MASK = 0x70007000; + + // Extract and shift FP4 values to FP16 format + int Out1 = (q & 0x80008000) | ((q & MASK) >> RIGHT_SHIFT); + q <<= 4; + int Out2 = (q & 0x80008000) | ((q & MASK) >> RIGHT_SHIFT); + + // Note: reverse indexing is intentional because weights are permuted + frag_b[1] = *reinterpret_cast(&Out1); + frag_b[0] = *reinterpret_cast(&Out2); +} + +template <> +__device__ inline void dequant(int q, half2 *frag_b) { + dequant(q, frag_b); + + // Constants for FP4 (E2M1) and FP16 formats + constexpr int FP4_EXPONENT = 2, FP16_EXPONENT = 5; + + // Construct and apply exponent bias + constexpr int BIAS_OFFSET = (1 << (FP16_EXPONENT - 1)) - (1 << (FP4_EXPONENT - 1)); + const half2 bias_reg = __float2half2_rn(float(1 << BIAS_OFFSET)); + + // Convert to half2 and apply bias + frag_b[1] = __hmul2(frag_b[1], bias_reg); + frag_b[0] = __hmul2(frag_b[0], bias_reg); +} + +template <> +__device__ inline void dequant(int q, nv_bfloat162 *frag_b) { + // Constants for FP4 (E2M1) and FP16 formats + constexpr int FP4_EXPONENT = 2, BF16_EXPONENT = 8; + constexpr int RIGHT_SHIFT = BF16_EXPONENT - FP4_EXPONENT; + constexpr int MASK = 0x70007000; + + // Extract and shift FP4 values to FP16 format + int Out1 = (q & 0x80008000) | ((q & MASK) >> RIGHT_SHIFT); + q <<= 4; + int Out2 = (q & 0x80008000) | ((q & MASK) >> RIGHT_SHIFT); + + // Note: reverse indexing is intentional because weights are permuted + frag_b[1] = *reinterpret_cast(&Out1); + frag_b[0] = *reinterpret_cast(&Out2); +} + +template <> +__device__ inline void dequant(int q, nv_bfloat162 *frag_b) { + dequant(q, frag_b); + + // Constants for FP4 (E2M1) and BF16 formats + constexpr int FP4_EXPONENT = 2, BF16_EXPONENT = 8; + + // Construct and apply exponent bias + constexpr int BIAS_OFFSET = (1 << (BF16_EXPONENT - 1)) - (1 << (FP4_EXPONENT - 1)); + // Add 127 (float exponent bias) to BIAS_OFFSET and shift to float exponent + // position + constexpr uint32_t BIAS = (BIAS_OFFSET + 127) << 23; + const nv_bfloat162 bias_reg = __float2bfloat162_rn(*reinterpret_cast(&BIAS)); + + // Convert to half2 and apply bias + frag_b[1] = __hmul2(frag_b[1], bias_reg); + frag_b[0] = __hmul2(frag_b[0], bias_reg); +} + +template +__device__ inline void dequant_fp8_scales(int q, scalar_t2 *frag_b); + +template <> +__device__ inline void dequant_fp8_scales(int q, half2 *frag_b) { + int Out1 = (q & 0xFF00FF00) >> 1; + ; + q <<= 8; + int Out2 = (q & 0xFF00FF00) >> 1; + + // Note: reverse indexing is intentional because weights are permuted + frag_b[1] = *reinterpret_cast(&Out1); + frag_b[0] = *reinterpret_cast(&Out2); +}; + +template <> +__device__ inline void dequant_fp8_scales(int q, nv_bfloat162 *frag_b) { + constexpr int FP8_EXPONENT = 4, BF16_EXPONENT = 8; + constexpr int RIGHT_SHIFT = BF16_EXPONENT - FP8_EXPONENT; + constexpr int MASK = 0x7F007F00; + + // Extract and shift FP8 values to BF16 format + int Out1 = ((q & 0x80008000) >> 1) | ((q & MASK) >> RIGHT_SHIFT); + q <<= 8; + int Out2 = ((q & 0x80008000) >> 1) | ((q & MASK) >> RIGHT_SHIFT); + + // Note: reverse indexing is intentional because weights are permuted + frag_b[1] = *reinterpret_cast(&Out1); + frag_b[0] = *reinterpret_cast(&Out2); +}; + +// New version with s_type_id parameter for marlin_moe_wna16_v2 +template +__device__ inline void dequant_fp8_scales(int q, scalar_t2 *frag_b); + +template <> +__device__ inline void dequant_fp8_scales(int q, half2 *frag_b) { + int Out1 = (q & 0xFF00FF00) >> 1; + ; + q <<= 8; + int Out2 = (q & 0xFF00FF00) >> 1; + + // Note: reverse indexing is intentional because weights are permuted + frag_b[1] = *reinterpret_cast(&Out1); + frag_b[0] = *reinterpret_cast(&Out2); +}; + +template <> +__device__ inline void dequant_fp8_scales(int q, nv_bfloat162 *frag_b) { + constexpr int FP8_EXPONENT = 4, BF16_EXPONENT = 8; + constexpr int RIGHT_SHIFT = BF16_EXPONENT - FP8_EXPONENT; + constexpr int MASK = 0x7F007F00; + + // Extract and shift FP8 values to BF16 format + int Out1 = ((q & 0x80008000) >> 1) | ((q & MASK) >> RIGHT_SHIFT); + q <<= 8; + int Out2 = ((q & 0x80008000) >> 1) | ((q & MASK) >> RIGHT_SHIFT); + + // Note: reverse indexing is intentional because weights are permuted + frag_b[1] = *reinterpret_cast(&Out1); + frag_b[0] = *reinterpret_cast(&Out2); +} + +template <> +__device__ inline void dequant_fp8_scales(int q, nv_bfloat162 *frag_b) { + // In this conversion, 2 ** -127 in FP8E8M0 would become 0 in BF16, + // but we assume that such a extreme value would not occur in real models. + int Out1 = (q & 0xFF00FF00) >> 1; + q <<= 7; + int Out2 = q & 0x7F807F80; + + // Note: reverse indexing is intentional because weights are permuted + frag_b[1] = *reinterpret_cast(&Out1); + frag_b[0] = *reinterpret_cast(&Out2); +} + +#endif + +} // namespace device::marlin diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/marlin/kernel.h b/src/infiniop/ops/moe_wna16_marlin_gemm/marlin/kernel.h new file mode 100644 index 000000000..b629d781e --- /dev/null +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/marlin/kernel.h @@ -0,0 +1,37 @@ + +#include "../sgl_kernel/scalar_type.hpp" + +#include "marlin.cuh" +#include "marlin_dtypes.cuh" + +#define MARLIN_KERNEL_PARAMS \ + const int4 *__restrict__ A, const int4 *__restrict__ B, int4 *__restrict__ C, int4 *__restrict__ C_tmp, \ + const int4 *__restrict__ b_bias_ptr, const int4 *__restrict__ scales_ptr, \ + const uint16_t *__restrict__ scale2_ptr, const int4 *__restrict__ zp_ptr, const int *__restrict__ g_idx, \ + const int32_t *__restrict__ sorted_token_ids_ptr, const int32_t *__restrict__ expert_ids_ptr, \ + const int32_t *__restrict__ num_tokens_past_padded_ptr, const float *__restrict__ topk_weights_ptr, int top_k, \ + bool mul_topk_weights, bool is_ep, int num_groups, int prob_m, int prob_n, int prob_k, int *locks, \ + bool has_bias, bool use_atomic_add, bool use_fp32_reduce, int max_shared_mem + +namespace device::marlin_moe { +template < + typename scalar_t, // compute dtype, half or nv_float16 + const host::ScalarTypeId w_type_id, // weight ScalarType id + const host::ScalarTypeId s_type_id, // weight scale ScalarType id + const int threads, // number of threads in a threadblock + const int thread_m_blocks, // number of 16x16 blocks in the m + // dimension (batchsize) of the + // threadblock + const int thread_n_blocks, // same for n dimension (output) + const int thread_k_blocks, // same for k dimension (reduction) + const bool m_block_size_8, // whether m_block_size == 8 + // only works when thread_m_blocks == 1 + const int stages, // number of stages for the async global->shared + // fetch pipeline + const int group_blocks, // number of consecutive 16x16 blocks + // with a separate quantization scale + const bool is_zp_float // is zero point of float16 type? + > +__global__ void Marlin(MARLIN_KERNEL_PARAMS); + +} // namespace device::marlin_moe diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/marlin/marlin.cuh b/src/infiniop/ops/moe_wna16_marlin_gemm/marlin/marlin.cuh new file mode 100644 index 000000000..944ca3522 --- /dev/null +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/marlin/marlin.cuh @@ -0,0 +1,85 @@ +#pragma once + +#include "../sgl_kernel/utils.cuh" + +#include + +namespace device::marlin { +// Marlin params + +// 8 warps are a good choice since every SM has 4 schedulers and having more +// than 1 warp per schedule allows some more latency hiding. At the same time, +// we want relatively few warps to have many registers per warp and small tiles. +static constexpr int default_threads = 256; + +static constexpr int pipe_stages = 4; // 4 pipeline stages fit into shared memory + +static constexpr int min_thread_n = 64; +static constexpr int min_thread_k = 64; +static constexpr int max_thread_n = 256; + +static constexpr int tile_size = 16; +static constexpr int max_par = 16; + +// Repack params +static constexpr int repack_stages = 8; + +static constexpr int repack_threads = 256; + +static constexpr int tile_k_size = tile_size; +static constexpr int tile_n_size = tile_k_size * 4; + +// Helpers +template +struct Vec { + T elems[n]; + __device__ T &operator[](int i) { + return elems[i]; + } +}; + +using I4 = Vec; + +using host::div_ceil; + +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800 +// No support for async +#else + +__device__ inline void cp_async4_pred(void *smem_ptr, const void *glob_ptr, bool pred = true) { + const int BYTES = 16; + uint32_t smem = static_cast(__cvta_generic_to_shared(smem_ptr)); + asm volatile( + "{\n" + " .reg .pred p;\n" + " setp.ne.b32 p, %0, 0;\n" + " @p cp.async.cg.shared.global [%1], [%2], %3;\n" + "}\n" ::"r"((int)pred), + "r"(smem), + "l"(glob_ptr), + "n"(BYTES)); +} + +__device__ inline void cp_async4(void *smem_ptr, const void *glob_ptr) { + const int BYTES = 16; + uint32_t smem = static_cast(__cvta_generic_to_shared(smem_ptr)); + asm volatile( + "{\n" + " cp.async.cg.shared.global [%0], [%1], %2;\n" + "}\n" ::"r"(smem), + "l"(glob_ptr), + "n"(BYTES)); +} + +__device__ inline void cp_async_fence() { + asm volatile("cp.async.commit_group;\n" ::); +} + +template +__device__ inline void cp_async_wait() { + asm volatile("cp.async.wait_group %0;\n" ::"n"(n)); +} + +#endif + +} // namespace device::marlin diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/marlin/marlin_dtypes.cuh b/src/infiniop/ops/moe_wna16_marlin_gemm/marlin/marlin_dtypes.cuh new file mode 100644 index 000000000..783374ff2 --- /dev/null +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/marlin/marlin_dtypes.cuh @@ -0,0 +1,78 @@ +#ifndef _data_types_cuh +#define _data_types_cuh +#include "../sgl_kernel/utils.cuh" + +#include "marlin.cuh" + +namespace device::marlin { + +template +class ScalarType { +}; + +template <> +class ScalarType<__half> { +public: + using scalar_t = __half; + using scalar_t2 = fp16x2_t; + + // Matrix fragments for tensor core instructions; their precise layout is + // documented here: + // https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#matrix-fragments-for-mma-m16n8k16-with-floating-point-type + using FragA = Vec; + using FragB = Vec; + using FragC = Vec; + using FragS = Vec; + using FragZP = Vec; + + static __device__ float inline num2float(const __half x) { + return __half2float(x); + } + + static __device__ fp16x2_t inline num2num2(const __half x) { + return __half2half2(x); + } + + static __device__ fp16x2_t inline nums2num2(const __half x1, const __half x2) { + return __halves2half2(x1, x2); + } + + static __host__ __device__ __half inline float2num(const float x) { + return __float2half(x); + } +}; + +template <> +class ScalarType<__nv_bfloat16> { +public: + using scalar_t = __nv_bfloat16; + using scalar_t2 = bf16x2_t; + + using FragA = Vec; + using FragB = Vec; + using FragC = Vec; + using FragS = Vec; + using FragZP = Vec; + +#if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 800 + static __device__ float inline num2float(const __nv_bfloat16 x) { + return __bfloat162float(x); + } + + static __device__ bf16x2_t inline num2num2(const __nv_bfloat16 x) { + return __bfloat162bfloat162(x); + } + + static __device__ bf16x2_t inline nums2num2(const __nv_bfloat16 x1, const __nv_bfloat16 x2) { + return __halves2bfloat162(x1, x2); + } + + static __host__ __device__ __nv_bfloat16 inline float2num(const float x) { + return __float2bfloat16(x); + } +#endif +}; + +} // namespace device::marlin + +#endif diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/marlin/marlin_template.h b/src/infiniop/ops/moe_wna16_marlin_gemm/marlin/marlin_template.h new file mode 100644 index 000000000..9fba3b18f --- /dev/null +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/marlin/marlin_template.h @@ -0,0 +1,1912 @@ +/* + * Modified by Neural Magic + * Copyright (C) Marlin.2024 Elias Frantar + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* + * Adapted from https://github.com/IST-DASLab/marlin + */ + +#include "../sgl_kernel/scalar_type.hpp" + +#include "dequant.h" +#include "marlin.cuh" +#include "marlin_dtypes.cuh" +#include + +#define STATIC_ASSERT_SCALAR_TYPE_VALID(scalar_t) \ + static_assert( \ + std::is_same::value || std::is_same::value, \ + "only float16 and bfloat16 is supported"); + +namespace device::marlin_moe { +using namespace device::marlin; + +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800 + +template < + typename scalar_t, // compute dtype, half or nv_float16 + const host::ScalarTypeId w_type_id, // weight ScalarType id + const int threads, // number of threads in a threadblock + const int thread_m_blocks, // number of 16x16 blocks in the m + // dimension (batchsize) of the + // threadblock + const int thread_n_blocks, // same for n dimension (output) + const int thread_k_blocks, // same for k dimension (reduction) + const bool m_block_size_8, // whether m_block_size == 8 + // only works when thread_m_blocks == 1 + const int stages, // number of stages for the async global->shared + // fetch pipeline + const int group_blocks, // number of consecutive 16x16 blocks + // with a separate quantization scale + const bool is_zp_float // is zero point of float16 type? + > +__global__ void Marlin( + const int4 *__restrict__ A, // fp16 input matrix of shape mxk + const int4 *__restrict__ B, // 4bit quantized weight matrix of shape kxn + int4 *__restrict__ C, // fp16 output buffer of shape mxn + int4 *__restrict__ C_tmp, // fp32 tmp output buffer (for reduce) + const int4 *__restrict__ scales_ptr, // fp16 quantization scales of shape + // (k/groupsize)xn + const int4 *__restrict__ zp_ptr, // 4bit packed zero-points of shape + // (k/groupsize)x(n/pack_factor) + const int *__restrict__ g_idx, // int32 group indices of shape k + const int32_t *__restrict__ sorted_token_ids_ptr, // moe sorted_ids + const int32_t *__restrict__ expert_ids_ptr, // moe expert ids + const int32_t *__restrict__ num_tokens_past_padded_ptr, // moe num tokens + const float *__restrict__ topk_weights_ptr, // moe top weights + int top_k, // num of experts per token + bool mul_topk_weights, // mul topk weights or not + bool is_ep, // expert parallelism + int num_groups, // number of scale groups per output channel + int prob_m, // batch dimension m + int prob_n, // output dimension n + int prob_k, // reduction dimension k + int *locks, // extra global storage for barrier synchronization + bool use_atomic_add, // whether to use atomic add to reduce + bool use_fp32_reduce, // whether to use fp32 global reduce + int max_shared_mem) {} + +} // namespace device::marlin_moe + +#else + +// m16n8k16 tensor core mma instruction with fp16 inputs and fp32 +// output/accumulation. +template +__device__ inline void +mma(const typename ScalarType::FragA &a_frag, + const typename ScalarType::FragB &frag_b, + typename ScalarType::FragC &frag_c) { + const uint32_t *a = reinterpret_cast(&a_frag); + const uint32_t *b = reinterpret_cast(&frag_b); + float *c = reinterpret_cast(&frag_c); + if constexpr (std::is_same::value) { + asm volatile( + "mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32 " + "{%0,%1,%2,%3}, {%4,%5,%6,%7}, {%8,%9}, {%10,%11,%12,%13};\n" + : "=f"(c[0]), "=f"(c[1]), "=f"(c[2]), "=f"(c[3]) + : "r"(a[0]), "r"(a[1]), "r"(a[2]), "r"(a[3]), "r"(b[0]), "r"(b[1]), "f"(c[0]), "f"(c[1]), "f"(c[2]), "f"(c[3])); + } else if constexpr (std::is_same::value) { + asm volatile( + "mma.sync.aligned.m16n8k16.row.col.f32.bf16.bf16.f32 " + "{%0,%1,%2,%3}, {%4,%5,%6,%7}, {%8,%9}, {%10,%11,%12,%13};\n" + : "=f"(c[0]), "=f"(c[1]), "=f"(c[2]), "=f"(c[3]) + : "r"(a[0]), "r"(a[1]), "r"(a[2]), "r"(a[3]), "r"(b[0]), "r"(b[1]), "f"(c[0]), "f"(c[1]), "f"(c[2]), "f"(c[3])); + } else { + STATIC_ASSERT_SCALAR_TYPE_VALID(scalar_t); + } +} + +template +__device__ inline void mma_trans( + const typename ScalarType::FragA &a_frag, + const typename ScalarType::FragB &frag_b, + const typename ScalarType::FragB &frag_b2, + typename ScalarType::FragC &frag_c) { + const uint32_t *a = reinterpret_cast(&a_frag); + const uint32_t *b = reinterpret_cast(&frag_b); + const uint32_t *b2 = reinterpret_cast(&frag_b2); + float *c = reinterpret_cast(&frag_c); + if constexpr (std::is_same::value) { + asm volatile( + "mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32 " + "{%0,%1,%2,%3}, {%4,%5,%6,%7}, {%8,%9}, {%10,%11,%12,%13};\n" + : "=f"(c[0]), "=f"(c[1]), "=f"(c[2]), "=f"(c[3]) + : "r"(b[0]), + "r"(b2[0]), + "r"(b[1]), + "r"(b2[1]), + "r"(a[0]), + "r"(a[1]), + "f"(c[0]), + "f"(c[1]), + "f"(c[2]), + "f"(c[3])); + } else if constexpr (std::is_same::value) { + asm volatile( + "mma.sync.aligned.m16n8k16.row.col.f32.bf16.bf16.f32 " + "{%0,%1,%2,%3}, {%4,%5,%6,%7}, {%8,%9}, {%10,%11,%12,%13};\n" + : "=f"(c[0]), "=f"(c[1]), "=f"(c[2]), "=f"(c[3]) + : "r"(b[0]), + "r"(b2[0]), + "r"(b[1]), + "r"(b2[1]), + "r"(a[0]), + "r"(a[1]), + "f"(c[0]), + "f"(c[1]), + "f"(c[2]), + "f"(c[3])); + } else { + STATIC_ASSERT_SCALAR_TYPE_VALID(scalar_t); + } +} + +// Instruction for loading a full 16x16 matrix fragment of operand A from shared +// memory, directly in tensor core layout. +template +__device__ inline void ldsm(typename ScalarType::FragA &frag_a, const void *smem_ptr) { + uint32_t *a = reinterpret_cast(&frag_a); + uint32_t smem = static_cast(__cvta_generic_to_shared(smem_ptr)); + if constexpr (count == 4) { + asm volatile("ldmatrix.sync.aligned.m8n8.x4.shared.b16 {%0,%1,%2,%3}, [%4];\n" + : "=r"(a[0]), "=r"(a[1]), "=r"(a[2]), "=r"(a[3]) + : "r"(smem)); + } else if constexpr (count == 2) { + asm volatile("ldmatrix.sync.aligned.m8n8.x2.shared.b16 {%0,%1}, [%2];\n" : "=r"(a[0]), "=r"(a[1]) : "r"(smem)); + } else if constexpr (count == 1) { + asm volatile("ldmatrix.sync.aligned.m8n8.x1.shared.b16 {%0}, [%1];\n" : "=r"(a[0]) : "r"(smem)); + } else { + static_assert(count == 1 || count == 2 || count == 4, "invalid count"); + } +} + +// Multiply dequantized values by the corresponding quantization scale; used +// only for grouped quantization. +template +__device__ inline void +scale(typename ScalarType::FragB &frag_b, typename ScalarType::FragS &frag_s, int i) { + using scalar_t2 = typename ScalarType::scalar_t2; + scalar_t2 s = ScalarType::num2num2(reinterpret_cast(&frag_s)[i]); + frag_b[0] = __hmul2(frag_b[0], s); + frag_b[1] = __hmul2(frag_b[1], s); +} + +template +__device__ inline void scale_and_sub(typename ScalarType::FragB &frag_b, scalar_t s, scalar_t zp) { + using scalar_t2 = typename ScalarType::scalar_t2; + scalar_t2 s2 = ScalarType::num2num2(s); + scalar_t2 zp2 = ScalarType::num2num2(zp); + frag_b[0] = __hfma2(frag_b[0], s2, __hneg2(zp2)); + frag_b[1] = __hfma2(frag_b[1], s2, __hneg2(zp2)); +} + +template +__device__ inline void +sub_zp(typename ScalarType::FragB &frag_b, typename ScalarType::scalar_t2 &frag_zp, int i) { + using scalar_t2 = typename ScalarType::scalar_t2; + scalar_t2 zp = ScalarType::num2num2(reinterpret_cast(&frag_zp)[i]); + frag_b[0] = __hsub2(frag_b[0], zp); + frag_b[1] = __hsub2(frag_b[1], zp); +} + +// Same as above, but for act_order (each K is multiplied individually) +template +__device__ inline void scale4( + typename ScalarType::FragB &frag_b, + typename ScalarType::FragS &frag_s_1, + typename ScalarType::FragS &frag_s_2, + typename ScalarType::FragS &frag_s_3, + typename ScalarType::FragS &frag_s_4, + int i) { + using scalar_t2 = typename ScalarType::scalar_t2; + scalar_t2 s_val_1_2; + s_val_1_2.x = reinterpret_cast(&frag_s_1)[i]; + s_val_1_2.y = reinterpret_cast(&frag_s_2)[i]; + + scalar_t2 s_val_3_4; + s_val_3_4.x = reinterpret_cast(&frag_s_3)[i]; + s_val_3_4.y = reinterpret_cast(&frag_s_4)[i]; + + frag_b[0] = __hmul2(frag_b[0], s_val_1_2); + frag_b[1] = __hmul2(frag_b[1], s_val_3_4); +} + +// Given 2 floats multiply by 2 scales (halves) +template +__device__ inline void scale_float(float *c, typename ScalarType::FragS &s) { + scalar_t *s_ptr = reinterpret_cast(&s); + c[0] = __fmul_rn(c[0], ScalarType::num2float(s_ptr[0])); + c[1] = __fmul_rn(c[1], ScalarType::num2float(s_ptr[1])); +} + +// Wait until barrier reaches `count`, then lock for current threadblock. +__device__ inline void barrier_acquire(int *lock, int count) { + if (threadIdx.x == 0) { + int state = -1; + do { + // Guarantee that subsequent writes by this threadblock will be visible + // globally. + asm volatile("ld.global.acquire.gpu.b32 %0, [%1];\n" : "=r"(state) : "l"(lock)); + } while (state != count); + } + __syncthreads(); +} + +// Release barrier and increment visitation count. +__device__ inline void barrier_release(int *lock, bool reset = false) { + __syncthreads(); + if (threadIdx.x == 0) { + if (reset) { + lock[0] = 0; + return; + } + int val = 1; + // Make sure that all writes since acquiring this barrier are visible + // globally, while releasing the barrier. + asm volatile("fence.acq_rel.gpu;\n"); + asm volatile("red.relaxed.gpu.global.add.s32 [%0], %1;\n" : : "l"(lock), "r"(val)); + } +} + +// Wait until value of lock to be negative, and then add 1 +__device__ inline void wait_negative_and_add(int *lock) { + if (threadIdx.x == 0) { + int state = 0; + do { + // Guarantee that subsequent writes by this threadblock will be visible + // globally. + asm volatile("ld.global.acquire.gpu.b32 %0, [%1];\n" : "=r"(state) : "l"(lock)); + } while (state >= 0); + atomicAdd(lock, 1); + } + __syncthreads(); +} + +template < + typename scalar_t, // compute dtype, half or nv_float16 + const host::ScalarTypeId w_type_id, // weight ScalarType id + const host::ScalarTypeId s_type_id, // weight scale ScalarType id + const int threads, // number of threads in a threadblock + const int thread_m_blocks, // number of 16x16 blocks in the m + // dimension (batchsize) of the + // threadblock + const int thread_n_blocks, // same for n dimension (output) + const int thread_k_blocks, // same for k dimension (reduction) + const bool m_block_size_8, // whether m_block_size == 8 + // only works when thread_m_blocks == 1 + const int stages, // number of stages for the async global->shared + // fetch pipeline + const int group_blocks, // number of consecutive 16x16 blocks + // with a separate quantization scale + const bool is_zp_float // is zero point of float16 type? + > +__global__ void Marlin( + const int4 *__restrict__ A, // fp16 input matrix of shape mxk + const int4 *__restrict__ B, // 4bit quantized weight matrix of shape kxn + int4 *__restrict__ C, // fp16 output buffer of shape mxn + int4 *__restrict__ C_tmp, // fp32 tmp output buffer (for reduce) + const int4 *__restrict__ b_bias_ptr, + const int4 *__restrict__ scales_ptr, // fp16 quantization scales of shape + // (k/groupsize)xn + const uint16_t *__restrict__ scale2_ptr, // fp16 global scale (for nvfp4 + // only) + const int4 *__restrict__ zp_ptr, // 4bit packed zero-points of shape + // (k/groupsize)x(n/pack_factor) + const int *__restrict__ g_idx, // int32 group indices of shape k + const int32_t *__restrict__ sorted_token_ids_ptr, // moe sorted_ids + const int32_t *__restrict__ expert_ids_ptr, // moe expert ids + const int32_t *__restrict__ num_tokens_past_padded_ptr, // moe num tokens + const float *__restrict__ topk_weights_ptr, // moe top weights + int top_k, // num of experts per token + bool mul_topk_weights, // mul topk weights or not + bool is_ep, // expert parallelism + int num_groups, // number of scale groups per output channel + int prob_m, // batch dimension m + int prob_n, // output dimension n + int prob_k, // reduction dimension k + int *locks, // extra global storage for barrier synchronization + bool has_bias, + bool use_atomic_add, // whether to use atomic add to reduce + bool use_fp32_reduce, // whether to use fp32 global reduce + int max_shared_mem) { + // Each threadblock processes one "stripe" of the B matrix with (roughly) the + // same size, which might involve multiple column "slices" (of width 16 * + // `thread_n_blocks`). Stripes are defined as shown in the 3x3 matrix 5 SM + // example: + // 0 1 3 + // 0 2 3 + // 1 2 4 + // While this kind of partitioning makes things somewhat more complicated, it + // ensures good utilization of all SMs for many kinds of shape and GPU + // configurations, while requiring as few slow global cross-threadblock + // reductions as possible. + using Dtype = ScalarType; + using scalar_t2 = typename ScalarType::scalar_t2; + using FragA = typename ScalarType::FragA; + using FragB = typename ScalarType::FragB; + using FragC = typename ScalarType::FragC; + using FragS = typename ScalarType::FragS; + using FragZP = typename ScalarType::FragZP; + + extern __shared__ int4 sh[]; + static constexpr auto w_type = host::ScalarType::from_id(w_type_id); + static constexpr auto s_type = host::ScalarType::from_id(s_type_id); + if constexpr (w_type == host::kFE2M1f) { + static_assert(s_type == host::kFE4M3fn && group_blocks == 1 || s_type == host::kFE8M0fnu && group_blocks == 2); + } else if constexpr (std::is_same::value) { + static_assert(s_type == host::kBFloat16); + } else if constexpr (std::is_same::value) { + static_assert(s_type == host::kFloat16); + } + + constexpr bool has_zp = w_type == host::kU4 || w_type == host::kU8; + constexpr bool is_int_type = w_type == host::kU4 || w_type == host::kU8 || w_type == host::kU4B8 || w_type == host::kU8B128; + constexpr bool is_8bit_scale = s_type.size_bits() == 8; + // see comments of dequant.h for more details + constexpr bool dequant_skip_flop = w_type == host::kFE4M3fn || w_type == host::kFE2M1f && s_type == host::kFE4M3fn || has_zp && !is_zp_float && !std::is_same::value || has_zp && !is_zp_float && !(w_type == host::kU8); + + scalar_t2 global_scale; + + constexpr bool has_act_order = group_blocks == 0; + + constexpr int pack_factor = 32 / w_type.size_bits(); + static_assert(thread_m_blocks == 1 || !m_block_size_8); + constexpr int moe_block_size = m_block_size_8 ? 8 : (16 * thread_m_blocks); + const int group_size = (!has_act_order && group_blocks == -1) ? prob_k : prob_k / num_groups; + const int scales_expert_stride = prob_n * prob_k / group_size / (is_8bit_scale ? 16 : 8); + const int zp_expert_stride = is_zp_float ? prob_n * prob_k / group_size / 8 : prob_n * prob_k / group_size / (pack_factor * 4); + const int b_bias_expert_stride = prob_n / 8; + + // parallel: num valid moe blocks + int num_tokens_past_padded = num_tokens_past_padded_ptr[0]; + int parallel = num_tokens_past_padded / moe_block_size; + int num_valid_blocks = parallel; + if (is_ep) { + for (int i = 0; i < parallel; i++) { + if (expert_ids_ptr[i] == -1) { + num_valid_blocks--; + } + } + } + int num_invalid_blocks = parallel - num_valid_blocks; + parallel = num_valid_blocks; + + int k_tiles = prob_k / 16 / thread_k_blocks; + int n_tiles = prob_n / 16 / thread_n_blocks; + int iters = device::div_ceil(k_tiles * n_tiles * parallel, gridDim.x); + + if constexpr (!has_act_order && group_blocks != -1) { + if (group_blocks >= thread_k_blocks) { + // Ensure that the number of tiles in each stripe is a multiple of the + // groupsize; this avoids an annoying special case where a stripe starts + // in the middle of group. + iters = (group_blocks / thread_k_blocks) * device::div_ceil(iters, (group_blocks / thread_k_blocks)); + } + } + + int slice_row = (iters * blockIdx.x) % k_tiles; + int slice_col_par = (iters * blockIdx.x) / k_tiles; + int slice_col = slice_col_par; + int slice_iters; // number of threadblock tiles in the current slice + int slice_count = 0; // total number of active threadblocks in the current slice + int slice_idx; // index of threadblock in current slice; numbered bottom to + // top + + int par_id = 0; + int block_id = -1; + int64_t expert_id = 0; // use int64 to avoid computation result overflow + int old_expert_id = 0; + int64_t B_expert_off = 0; + + int4 *sh_block_sorted_ids_int4 = sh; + int4 *sh_rd_block_sorted_ids_int4 = sh_block_sorted_ids_int4 + moe_block_size / 4; + int4 *sh_block_topk_weights_int4 = sh_rd_block_sorted_ids_int4 + moe_block_size / 4; + // sh_block_topk_weights_int4 only need (moe_block_size / 4); + // but we pad to align to 256 bytes + int4 *sh_new = sh_block_topk_weights_int4 + moe_block_size / 2 + moe_block_size; + int32_t *sh_block_sorted_ids = reinterpret_cast(sh_block_sorted_ids_int4); + int32_t *sh_rd_block_sorted_ids = reinterpret_cast(sh_rd_block_sorted_ids_int4); + scalar_t2 *sh_block_topk_weights = reinterpret_cast(sh_block_topk_weights_int4); + + int32_t block_num_valid_tokens = 0; + int32_t locks_off = 0; + + // We can easily implement parallel problem execution by just remapping + // indices and advancing global pointers + if (slice_col_par >= n_tiles) { + slice_col = slice_col_par % n_tiles; + par_id = slice_col_par / n_tiles; + } + if (parallel * n_tiles >= gridDim.x) { + // when parallel * n_tiles >= sms + // then there are at most $sms$ conflict tile blocks + locks_off = blockIdx.x; + } else { + locks_off = (iters * blockIdx.x) / k_tiles - 1; + } + + int prob_m_top_k = prob_m * top_k; + // read moe block data given block_id + // block_sorted_ids / block_num_valid_tokens / block_topk_weights + auto read_moe_block_data = [&](int block_id) { + block_num_valid_tokens = moe_block_size; + + cp_async4_pred( + sh_block_sorted_ids_int4 + threadIdx.x, + reinterpret_cast(sorted_token_ids_ptr) + (block_id * moe_block_size / 4 + threadIdx.x), + threadIdx.x < moe_block_size / 4); + + cp_async_fence(); + cp_async_wait<0>(); + + __syncthreads(); + + if (threadIdx.x >= threads - 32) { + constexpr int size_per_thread = device::div_ceil(moe_block_size, 32); + int lane_id = threadIdx.x - (threads - 32); + + int local_count = 0; +#pragma unroll + for (int i = 0; i < size_per_thread; i++) { + int j = lane_id * size_per_thread + i; + if (j < moe_block_size) { + int idx = sh_block_sorted_ids[j]; + if (idx < prob_m_top_k) + local_count++; + } + } + +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ == 750 + if constexpr (moe_block_size >= 16) + local_count += __shfl_down_sync(0xFFFFFFFF, local_count, 16); + if constexpr (moe_block_size >= 8) + local_count += __shfl_down_sync(0xFFFFFFFF, local_count, 8); + if constexpr (moe_block_size >= 4) + local_count += __shfl_down_sync(0xFFFFFFFF, local_count, 4); + if constexpr (moe_block_size >= 2) + local_count += __shfl_down_sync(0xFFFFFFFF, local_count, 2); + + local_count += __shfl_down_sync(0xFFFFFFFF, local_count, 1); + block_num_valid_tokens = local_count; +#else + block_num_valid_tokens = __reduce_add_sync(0xffffffff, local_count); +#endif + + if (lane_id == 0) + reinterpret_cast(sh_new)[0] = block_num_valid_tokens; + } + + if (threadIdx.x < moe_block_size) { + int idx = sh_block_sorted_ids[threadIdx.x]; + sh_rd_block_sorted_ids[threadIdx.x] = idx / top_k; + + if (mul_topk_weights) { + idx = idx < prob_m_top_k ? idx : 0; + scalar_t topk_weight_tmp = Dtype::float2num(topk_weights_ptr[idx]); + if constexpr (w_type == host::kFE2M1f && s_type == host::kFE4M3fn) { + sh_block_topk_weights[threadIdx.x] = __hmul2(global_scale, Dtype::num2num2(topk_weight_tmp)); + } else { + sh_block_topk_weights[threadIdx.x] = Dtype::num2num2(topk_weight_tmp); + } + } + } + + __syncthreads(); + + block_num_valid_tokens = reinterpret_cast(sh_new)[0]; + __syncthreads(); + }; + + // when move to next moe block, find the next block_id and expert_id + // and then read moe block data + auto update_next_moe_block_data = [&]() { + if (par_id >= parallel) { + return; + } + + old_expert_id = expert_id; + if (num_invalid_blocks > 0) { + int skip_count = block_id == -1 ? par_id : 0; + block_id++; + for (int i = block_id; i < num_tokens_past_padded / moe_block_size; i++) { + expert_id = expert_ids_ptr[i]; + if (expert_id != -1) { + if (skip_count == 0) { + block_id = i; + break; + }; + skip_count--; + }; + } + } else { + block_id = par_id; + expert_id = expert_ids_ptr[block_id]; + } + + if constexpr (w_type == host::kFE2M1f && s_type == host::kFE4M3fn) { + uint16_t val = scale2_ptr[expert_id]; + global_scale = Dtype::num2num2(*reinterpret_cast(&val)); + } + + B_expert_off = expert_id * prob_n * prob_k / (pack_factor * 4); + scales_ptr += (expert_id - old_expert_id) * scales_expert_stride; + if constexpr (has_zp) { + zp_ptr += (expert_id - old_expert_id) * zp_expert_stride; + } + if constexpr (has_act_order) { + g_idx += (expert_id - old_expert_id) * prob_k; + } + if (has_bias) { + b_bias_ptr += (expert_id - old_expert_id) * b_bias_expert_stride; + } + + read_moe_block_data(block_id); + }; + + // Compute all information about the current slice which is required for + // synchronization. + auto init_slice = [&](bool first_init = false) { + slice_iters = iters * (blockIdx.x + 1) - (k_tiles * slice_col_par + slice_row); + if (slice_iters < 0 || slice_col_par >= n_tiles * parallel) { + slice_iters = 0; + } + if (slice_iters == 0) { + return; + } + if (slice_row + slice_iters > k_tiles) { + slice_iters = k_tiles - slice_row; + } + slice_count = 1; + slice_idx = 0; + int col_first = iters * device::div_ceil(k_tiles * slice_col_par, iters); + if (col_first <= k_tiles * (slice_col_par + 1)) { + int col_off = col_first - k_tiles * slice_col_par; + slice_count = device::div_ceil(k_tiles - col_off, iters); + if (col_off > 0) { + slice_count++; + } + int delta_first = iters * blockIdx.x - col_first; + if (delta_first < 0 || (col_off == 0 && delta_first == 0)) { + slice_idx = slice_count - 1; + } else { + slice_idx = slice_count - 1 - delta_first / iters; + if (col_off > 0) { + slice_idx--; + } + } + } + if (parallel * n_tiles >= gridDim.x) { + if (slice_count > 1 && slice_idx == slice_count - 1) { + locks_off++; + } + } else { + locks_off++; + } + + if (first_init && use_atomic_add && slice_count > 1 && slice_idx == 0) { + constexpr int threads_per_m = 16 * thread_n_blocks / 8; + int m_per_thread = device::div_ceil(block_num_valid_tokens, threads / threads_per_m); + for (int i = 0; i < m_per_thread; i++) { + int row = threads / threads_per_m * i + threadIdx.x / threads_per_m; + if (row < block_num_valid_tokens) { + int64_t sorted_row = sh_block_sorted_ids[row]; + int col = slice_col * 16 * thread_n_blocks / 8 + threadIdx.x % threads_per_m; + C[sorted_row * prob_n / 8 + col] = {0, 0, 0, 0}; + } + } + // After write zero to output, write a negative value to lock. + // Every SM that processes the same slice would wait for + // the negative value, and then atomicAdd 1 to it. + // After all SMs are processed, the lock value would back to 0 again. + __syncthreads(); + if (threadIdx.x == 0) { + locks[locks_off] = 1 - slice_count; + } + } + + if (slice_col == n_tiles) { + slice_col = 0; + par_id++; + update_next_moe_block_data(); + } + }; + + update_next_moe_block_data(); + init_slice(true); + + // A sizes/strides + + // stride of the A matrix in global memory + int a_gl_stride = prob_k / 8; + // stride of an A matrix tile in shared memory + constexpr int a_sh_stride = 16 * thread_k_blocks / 8; + // delta between subsequent A tiles in global memory + constexpr int a_gl_rd_delta_o = 16 * thread_k_blocks / 8; + // between subsequent accesses within a tile + int a_gl_rd_delta_i = a_gl_stride * (threads / a_gl_rd_delta_o); + // between shared memory writes + constexpr int a_sh_wr_delta = a_sh_stride * (threads / a_gl_rd_delta_o); + // between shared memory tile reads + constexpr int a_sh_rd_delta_o = 2 * ((threads / 32) / (thread_n_blocks / 4)); + // within a shared memory tile + constexpr int a_sh_rd_delta_i = a_sh_stride * 16; + // overall size of a tile + constexpr int a_sh_stage = a_sh_stride * (16 * thread_m_blocks); + // number of shared write iterations for a tile + constexpr int a_sh_wr_iters = device::div_ceil(a_sh_stage, a_sh_wr_delta); + + // B sizes/strides + int b_gl_stride = 16 * prob_n / (pack_factor * 4); + constexpr int b_sh_stride = ((thread_n_blocks * 16) * 16 / pack_factor) / 4; + constexpr int b_thread_vecs = w_type.size_bits() == 4 ? 1 : 2; + constexpr int b_sh_stride_threads = b_sh_stride / b_thread_vecs; + + int b_gl_rd_delta_o = b_gl_stride * thread_k_blocks; + int b_gl_rd_delta_i = b_gl_stride * (threads / b_sh_stride_threads); + constexpr int b_sh_wr_delta = threads * b_thread_vecs; + constexpr int b_sh_rd_delta = threads * b_thread_vecs; + constexpr int b_sh_stage = b_sh_stride * thread_k_blocks; + constexpr int b_sh_wr_iters = b_sh_stage / b_sh_wr_delta; + + // Scale sizes/strides without act_order + int s_gl_stride = prob_n / (is_8bit_scale ? 16 : 8); + constexpr int s_sh_stride = 16 * thread_n_blocks / (is_8bit_scale ? 16 : 8); + constexpr int s_tb_groups = !has_act_order && group_blocks != -1 && group_blocks < thread_k_blocks ? thread_k_blocks / group_blocks : 1; + constexpr int s_sh_stage = s_tb_groups * s_sh_stride; + int s_gl_rd_delta = s_gl_stride; + + // Scale size/strides with act_order + constexpr int tb_k = 16 * thread_k_blocks; + constexpr int g_idx_stage = has_act_order ? (tb_k * sizeof(int)) / 16 : 0; + // constexpr int act_s_row_stride = 1; + // int act_s_col_stride = act_s_row_stride * num_groups; + constexpr int act_s_max_num_groups = 32; + int act_s_col_stride = 1; + int act_s_col_warp_stride = act_s_col_stride * 8; + int tb_n_warps = thread_n_blocks / 4; + int act_s_col_tb_stride = act_s_col_warp_stride * tb_n_warps; + + // Zero-points sizes/strides + int zp_gl_stride = is_zp_float ? prob_n / 8 : (prob_n / pack_factor) / 4; + constexpr int zp_sh_stride = is_zp_float ? 16 * thread_n_blocks / 8 : ((16 * thread_n_blocks) / pack_factor) / 4; + constexpr int zp_tb_groups = s_tb_groups; + constexpr int zp_sh_stage = has_zp ? zp_tb_groups * zp_sh_stride : 0; + int zp_gl_rd_delta = zp_gl_stride; + + // Global A read index of current thread. + int a_gl_rd_row = threadIdx.x / a_gl_rd_delta_o; + int a_gl_rd_col = a_gl_rd_delta_o * slice_row + threadIdx.x % a_gl_rd_delta_o; + + // Shared write index of current thread. + int a_sh_wr = a_sh_stride * (threadIdx.x / a_gl_rd_delta_o) + (threadIdx.x % a_gl_rd_delta_o); + // Shared read index. + int a_sh_rd = a_sh_stride * ((threadIdx.x % 32) % (16 / (m_block_size_8 ? 2 : 1))) + (threadIdx.x % 32) / (16 / (m_block_size_8 ? 2 : 1)); + a_sh_rd += 2 * ((threadIdx.x / 32) / (thread_n_blocks / 4)); + + int b_gl_rd = b_gl_stride * (threadIdx.x / b_sh_stride_threads) + (threadIdx.x % b_sh_stride_threads) * b_thread_vecs; + b_gl_rd += b_sh_stride * slice_col; + b_gl_rd += b_gl_rd_delta_o * slice_row; + auto b_sh_wr = threadIdx.x * b_thread_vecs; + auto b_sh_rd = threadIdx.x * b_thread_vecs; + + // For act_order + constexpr int k_iter_size = tb_k / b_sh_wr_iters; + int slice_k_start = tb_k * slice_row; + int slice_k_finish = slice_k_start + tb_k * slice_iters; + int slice_k_start_shared_fetch = slice_k_start; + int slice_n_offset = act_s_col_tb_stride * slice_col; + + // No act_order + int s_gl_rd; + if constexpr (!has_act_order) { + if constexpr (group_blocks == -1) { + s_gl_rd = s_sh_stride * slice_col + threadIdx.x; + } else if constexpr (group_blocks >= thread_k_blocks) { + s_gl_rd = s_gl_stride * ((thread_k_blocks * slice_row) / group_blocks) + s_sh_stride * slice_col + threadIdx.x; + } else { + s_gl_rd = s_gl_stride * ((thread_k_blocks * slice_row) / group_blocks + threadIdx.x / s_sh_stride) + s_sh_stride * slice_col + threadIdx.x % s_sh_stride; + } + } + auto s_sh_wr = threadIdx.x; + bool s_sh_wr_pred = threadIdx.x < s_sh_stage; + + // Zero-points + int zp_gl_rd; + if constexpr (has_zp) { + if constexpr (group_blocks == -1) { + zp_gl_rd = zp_sh_stride * slice_col + threadIdx.x; + } else { + zp_gl_rd = zp_gl_stride * ((thread_k_blocks * slice_row) / group_blocks) + zp_sh_stride * slice_col + threadIdx.x; + } + } + auto zp_sh_wr = threadIdx.x; + bool zp_sh_wr_pred = threadIdx.x < zp_sh_stride; + + // We use a different scale layout for grouped and column-wise quantization as + // we scale a `half2` tile in column-major layout in the former and in + // row-major in the latter case. + int s_sh_rd; + if constexpr (group_blocks != -1) { + s_sh_rd = 8 * ((threadIdx.x / 32) % (thread_n_blocks / 4)) + (threadIdx.x % 32) / 4; + } else if constexpr (group_blocks == -1 && (m_block_size_8 || (has_zp && !dequant_skip_flop))) { + s_sh_rd = 8 * ((threadIdx.x / 32) % (thread_n_blocks / 4)) + (threadIdx.x % 32) / 8; + } else { + s_sh_rd = 8 * ((threadIdx.x / 32) % (thread_n_blocks / 4)) + (threadIdx.x % 32) % 4; + } + + int bias_sh_rd; + if constexpr (m_block_size_8) { + bias_sh_rd = 8 * ((threadIdx.x / 32) % (thread_n_blocks / 4)) + (threadIdx.x % 32) / 8; + } else { + bias_sh_rd = 8 * ((threadIdx.x / 32) % (thread_n_blocks / 4)) + (threadIdx.x % 32) % 4; + } + + int bias_sh_wr = threadIdx.x; + int bias_gl_rd = (thread_n_blocks * 16 / 8) * slice_col + threadIdx.x; + + // Zero-points have the same read layout as the scales + // (without column-wise case) + constexpr int num_col_threads = 8; + constexpr int num_row_threads = 4; + constexpr int num_ints_per_thread = 8 / pack_factor; + int zp_sh_rd; + if constexpr (has_zp) { + if constexpr (is_zp_float) { + if constexpr (group_blocks != -1) { + zp_sh_rd = 8 * ((threadIdx.x / 32) % (thread_n_blocks / 4)) + (threadIdx.x % 32) / 4; + } + } else { + zp_sh_rd = num_ints_per_thread * num_col_threads * ((threadIdx.x / 32) % (thread_n_blocks / 4)) + num_ints_per_thread * ((threadIdx.x % 32) / num_row_threads); + } + } + + // To ensure that writing and reading A tiles to/from shared memory, the + // latter in fragment format, is fully bank conflict free, we need to use a + // rather fancy XOR-based layout. The key here is that neither reads nor + // writes of the 16-byte `int4` blocks of 8 consecutive threads involve the + // same shared memory banks. Further, it seems (based on NSight-Compute) that + // each warp must also write a consecutive memory segment? + auto transform_a = [&](int i) { + int row = i / a_gl_rd_delta_o; + return a_gl_rd_delta_o * row + (i % a_gl_rd_delta_o) ^ (row % 8); + }; + // Since the computation of this remapping is non-trivial and, due to our main + // loop unrolls, all shared memory accesses are static, we simply precompute + // both transformed reads and writes. + int a_sh_wr_trans[a_sh_wr_iters]; +#pragma unroll + for (int i = 0; i < a_sh_wr_iters; i++) { + a_sh_wr_trans[i] = transform_a(a_sh_wr_delta * i + a_sh_wr); + } + int a_sh_rd_trans[b_sh_wr_iters][thread_m_blocks]; +#pragma unroll + for (int i = 0; i < b_sh_wr_iters; i++) { +#pragma unroll + for (int j = 0; j < thread_m_blocks; j++) { + a_sh_rd_trans[i][j] = transform_a(a_sh_rd_delta_o * i + a_sh_rd_delta_i * j + a_sh_rd); + } + } + + // Since B-accesses have non-constant stride they have to be computed at + // runtime; we break dependencies between subsequent accesses with a tile by + // maintining multiple pointers (we have enough registers), a tiny + // optimization. + const int4 *B_ptr[b_sh_wr_iters]; +#pragma unroll + for (int i = 0; i < b_sh_wr_iters; i++) { + B_ptr[i] = B + b_gl_rd_delta_i * i + b_gl_rd; + } + + // Shared memory storage for global fetch pipelines. + constexpr int sh_red_size = (2 * thread_n_blocks + 1) * 16 * thread_m_blocks; + constexpr int sh_b_size = stages * b_sh_stage; + int4 *sh_b = sh_new; + int4 *sh_red = sh_new; + + constexpr int sh_size_b_red_min = (sh_red_size < sh_b_size ? sh_red_size : sh_b_size); + constexpr int sh_size_b_red_max = (sh_red_size > sh_b_size ? sh_red_size : sh_b_size); + constexpr int sh_bias_size = (thread_n_blocks * 16 / 8); + constexpr int sh_b_red_bias_size = sh_size_b_red_max > (sh_size_b_red_min + sh_bias_size) ? sh_size_b_red_max : (sh_size_b_red_min + sh_bias_size); + + int4 *sh_bias = sh_new + sh_size_b_red_min; + int4 *sh_g_idx = sh_new + sh_b_red_bias_size; + int4 *sh_zp = sh_g_idx + (stages * g_idx_stage); + constexpr int sh_s_size = has_act_order ? (act_s_max_num_groups * s_sh_stride) : (stages * s_sh_stage); + int4 *sh_s = sh_zp + (stages * zp_sh_stage); + // shared memory reused by reduction should be smaller than + // shared memory used by weight. + static_assert(thread_m_blocks * 16 * thread_n_blocks * 16 / 8 <= stages * b_sh_stage); + int4 *sh_a = sh_s + sh_s_size; + constexpr int shm_size_used = moe_block_size + stages * (g_idx_stage + zp_sh_stage) + sh_s_size + sh_b_red_bias_size; + + // all remaining shared memory is used to cache A (input) + // sh_a_max_row is at least ` stages * 16 * thread_m_blocks ` + int sh_a_max_row = ((max_shared_mem - 1024) / 16 - shm_size_used) / (thread_k_blocks * 2); + + // Register storage for double buffer of shared memory reads. + FragA frag_a[2][thread_m_blocks]; + I4 frag_b_quant[2][b_thread_vecs]; + FragC frag_c[thread_m_blocks][4][2]; + FragS frag_s[2][4]; // No act-order + FragS frag_bias[2][4]; + FragS act_frag_s[2][4][4]; // For act-order + int frag_qzp[2][num_ints_per_thread]; // Zero-points + FragZP frag_zp; // Zero-points in fp16 + FragZP frag_zpf[2]; // Zero-points in fp16 in HQQ + + // Zero accumulators. + auto zero_accums = [&]() { +#pragma unroll + for (int i = 0; i < thread_m_blocks * 4 * 2 * 4; i++) + reinterpret_cast(frag_c)[i] = 0; + }; + + int sh_first_group_id = -1; + int sh_num_groups = -1; + + auto fetch_act_order_scales_to_shared = [&](bool is_async, int first_group_id, int last_group_id) { + sh_first_group_id = first_group_id; + sh_num_groups = last_group_id - first_group_id + 1; + + if (sh_num_groups > act_s_max_num_groups) { + sh_num_groups = act_s_max_num_groups; + } + + if (sh_first_group_id + sh_num_groups > num_groups) { + sh_num_groups = num_groups - sh_first_group_id; + } + + int row_offset = first_group_id * s_gl_stride; + + if (is_async) { + for (int i = 0; i < sh_num_groups; i++) { + if (threadIdx.x < s_sh_stride) { + cp_async4_pred( + &sh_s[(i * s_sh_stride) + threadIdx.x], + &scales_ptr[row_offset + (i * s_gl_stride) + slice_n_offset + threadIdx.x]); + } + } + } else { + for (int i = 0; i < sh_num_groups; i++) { + if (threadIdx.x < s_sh_stride) { + sh_s[(i * s_sh_stride) + threadIdx.x] = scales_ptr[row_offset + (i * s_gl_stride) + slice_n_offset + threadIdx.x]; + } + } + } + }; + + // Asynchronously fetch the next A, B and s tile from global to the next + // shared memory pipeline location. + bool should_load_a = true; + int max_num_stage_groups = ((sh_a_max_row - moe_block_size) / moe_block_size + 1) / stages; + max_num_stage_groups = max(max_num_stage_groups, 1); + auto fetch_to_shared = [&](int pipe, int a_off, bool pred = true, int pipe_a = 0) { + if (pred) { + if (should_load_a) { + int4 *sh_a_stage = sh_a + moe_block_size * a_sh_stride * pipe_a; +#pragma unroll + for (int i = 0; i < a_sh_wr_iters; i++) { + int row = a_gl_rd_delta_i / a_gl_stride * i + a_gl_rd_row; + int64_t sorted_row = 0; + if (!m_block_size_8 || row < 8) + sorted_row = sh_rd_block_sorted_ids[row]; + int64_t true_idx = sorted_row * a_gl_stride + a_gl_rd_col + a_gl_rd_delta_o * a_off; + cp_async4_pred(&sh_a_stage[a_sh_wr_trans[i]], &A[true_idx], row < block_num_valid_tokens); + } + } + + int4 *sh_b_stage = sh_b + b_sh_stage * pipe; +#pragma unroll + for (int i = 0; i < b_sh_wr_iters; i++) { +#pragma unroll + for (int j = 0; j < b_thread_vecs; j++) { + cp_async4(&sh_b_stage[b_sh_wr_delta * i + b_sh_wr + j], B_ptr[i] + j + B_expert_off); + } + + B_ptr[i] += b_gl_rd_delta_o; + } + + if constexpr (has_act_order) { + // Fetch g_idx thread-block portion + int full_pipe = a_off; + int cur_k = slice_k_start_shared_fetch + tb_k * full_pipe; + if (cur_k < prob_k && cur_k < slice_k_finish) { + int4 *sh_g_idx_stage = sh_g_idx + g_idx_stage * pipe; + + int4 const *cur_g_idx_stage_ptr = reinterpret_cast(&g_idx[cur_k]); + + if (threadIdx.x < g_idx_stage) { + cp_async4_pred(&sh_g_idx_stage[threadIdx.x], &cur_g_idx_stage_ptr[threadIdx.x]); + } + } + } else { + if constexpr (group_blocks != -1) { + int4 *sh_s_stage = sh_s + s_sh_stage * pipe; + if (pipe % device::div_ceil(group_blocks, thread_k_blocks) == 0) { + if (s_sh_wr_pred) { + cp_async4(&sh_s_stage[s_sh_wr], &scales_ptr[s_gl_rd]); + } + s_gl_rd += s_gl_rd_delta * s_tb_groups; + } + } + + if constexpr (has_zp && group_blocks != -1) { + int4 *sh_zp_stage = sh_zp + zp_sh_stage * pipe; + if (pipe % device::div_ceil(group_blocks, thread_k_blocks) == 0) { + if (zp_sh_wr_pred) { + cp_async4(&sh_zp_stage[zp_sh_wr], &zp_ptr[zp_gl_rd]); + } + zp_gl_rd += zp_gl_rd_delta * zp_tb_groups; + } + } + } + } + // Insert a fence even when we are winding down the pipeline to ensure that + // waiting is also correct at this point. + cp_async_fence(); + }; + + auto fetch_col_zp_to_shared = [&]() { + if (zp_sh_wr_pred) { + cp_async4(&sh_zp[zp_sh_wr], &zp_ptr[zp_gl_rd]); + } + }; + + auto fetch_col_scale_to_shared = [&]() { + if (s_sh_wr_pred) { + cp_async4(&sh_s[s_sh_wr], &scales_ptr[s_gl_rd]); + } + }; + + // Wait until the next thread tile has been loaded to shared memory. + auto wait_for_stage = [&]() { + // We only have `stages - 2` active fetches since we are double buffering + // and can only issue the next fetch when it is guaranteed that the previous + // shared memory load is fully complete (as it may otherwise be + // overwritten). + cp_async_wait(); + __syncthreads(); + }; + + // Load the next sub-tile from the current location in the shared memory pipe + // into the current register buffer. + auto fetch_to_registers = [&](int k, int pipe, int pipe_a = 0) { + int4 *sh_a_stage = sh_a + moe_block_size * a_sh_stride * pipe_a; +#pragma unroll + for (int i = 0; i < thread_m_blocks; i++) + ldsm(frag_a[k % 2][i], &sh_a_stage[a_sh_rd_trans[k % b_sh_wr_iters][i]]); + int4 *sh_b_stage = sh_b + b_sh_stage * pipe; + +#pragma unroll + for (int i = 0; i < b_thread_vecs; i++) { + frag_b_quant[k % 2][i] = *reinterpret_cast(&sh_b_stage[b_sh_rd_delta * (k % b_sh_wr_iters) + b_sh_rd + i]); + } + }; + + bool is_same_group[stages]; + int same_group_id[stages]; + + auto init_same_group = [&](int pipe) { + if constexpr (!has_act_order) { + return; + } + + int4 *sh_g_idx_stage = sh_g_idx + g_idx_stage * pipe; + int *sh_g_idx_int_ptr = reinterpret_cast(sh_g_idx_stage); + + int group_id_1 = sh_g_idx_int_ptr[0]; + int group_id_2 = sh_g_idx_int_ptr[tb_k - 1]; + + is_same_group[pipe] = group_id_1 == group_id_2; + same_group_id[pipe] = group_id_1; + }; + + auto fetch_scales_to_registers = [&](int k, int full_pipe) { + int pipe = full_pipe % stages; + + if constexpr (!has_act_order) { + // No act-order case + if constexpr (group_blocks == -1) { + // load only when starting a new slice + if (k == 0 && full_pipe == 0) { + reinterpret_cast(&frag_s)[0] = sh_s[s_sh_rd]; + reinterpret_cast(&frag_s)[1] = sh_s[s_sh_rd + 4]; + } + } else if constexpr (group_blocks != -1) { + if constexpr (group_blocks >= thread_k_blocks) { + constexpr int g = group_blocks / thread_k_blocks; + if (pipe % g == 0) { + if (k % b_sh_wr_iters == 0) { + int4 *sh_s_stage = sh_s + s_sh_stage * (g * (pipe / g)); + reinterpret_cast(&frag_s[k % 2])[0] = sh_s_stage[s_sh_rd]; + } else { + reinterpret_cast(&frag_s[1])[0] = reinterpret_cast(&frag_s[0])[0]; + } + } + } else { + auto warp_id = threadIdx.x / 32; + int n_warps = thread_n_blocks / 4; + + int warp_row = warp_id / n_warps; + int cur_k = warp_row * 16; + cur_k += k_iter_size * (k % b_sh_wr_iters); + int k_blocks = cur_k / 16; + int cur_group_id = k_blocks / group_blocks; + + int4 *sh_s_stage = sh_s + s_sh_stage * pipe; + + if constexpr (!is_8bit_scale) { + reinterpret_cast(&frag_s[k % 2])[0] = sh_s_stage[s_sh_rd + cur_group_id * s_sh_stride]; + } else { + reinterpret_cast(&frag_s[k % 2])[0] = reinterpret_cast(sh_s_stage)[s_sh_rd + cur_group_id * (2 * s_sh_stride)]; + } + } + } + + return; + } + + // Act-order case + + // Determine K of the "current" thread-block + int cur_k = slice_k_start + tb_k * full_pipe; + if (cur_k >= prob_k || cur_k >= slice_k_finish) { + return; + } + + // Reset (to current thread-block) since we read g_idx portion from the + // shared memory + cur_k = 0; + + // Progress to current iteration + cur_k += k_iter_size * (k % b_sh_wr_iters); + + // Determine "position" inside the thread-block (based on warp and + // thread-id) + auto warp_id = threadIdx.x / 32; + int n_warps = thread_n_blocks / 4; // Each warp processes 4 16-size tiles over N + + int warp_row = warp_id / n_warps; + int warp_col = warp_id % n_warps; + + cur_k += warp_row * 16; + + auto th_id = threadIdx.x % 32; + cur_k += (th_id % 4) * 2; // Due to tensor-core layout for fp16 B matrix + + int s_col_shift = + /*slice_n_offset +*/ (act_s_col_warp_stride * warp_col) + (th_id / 4) * act_s_col_stride; + + if (is_same_group[pipe]) { + if (k % 2 == 0) { + *(reinterpret_cast(&(act_frag_s[k % 2][0][0]))) = sh_s[(same_group_id[pipe] - sh_first_group_id) * s_sh_stride + s_col_shift]; + } else { + *(reinterpret_cast(&(act_frag_s[k % 2][0][0]))) = *(reinterpret_cast(&(act_frag_s[(k - 1) % 2][0][0]))); + } + + for (int i = 1; i < 4; i++) { + *(reinterpret_cast(&(act_frag_s[k % 2][i][0]))) = *(reinterpret_cast(&(act_frag_s[k % 2][0][0]))); + } + return; + } + + int4 *sh_g_idx_stage = sh_g_idx + g_idx_stage * pipe; + int *sh_g_idx_int_ptr = reinterpret_cast(sh_g_idx_stage); + + constexpr int k_frag_offsets[4] = {0, 1, 8, 9}; // Tensor core offsets per thread + +#pragma unroll + for (int i = 0; i < 4; i++) { + int actual_k = cur_k + k_frag_offsets[i]; + + int group_id = sh_g_idx_int_ptr[actual_k]; + int rel_group_id = group_id - sh_first_group_id; + + *(reinterpret_cast(&(act_frag_s[k % 2][i][0]))) = sh_s[rel_group_id * s_sh_stride + s_col_shift]; + } + }; + + auto fetch_zp_to_registers = [&](int k, int full_pipe) { + // This code does not handle group_blocks == 0, + // which signifies act_order. + // has_zp implies AWQ, which doesn't have act_order, + static_assert(!has_zp || group_blocks != 0); + + if constexpr (has_zp && !is_zp_float) { + int pipe = full_pipe % stages; + + if constexpr (group_blocks == -1) { + // load only when starting a new slice + if (k == 0 && full_pipe == 0) { +#pragma unroll + for (int i = 0; i < num_ints_per_thread; i++) { + frag_qzp[k % 2][i] = (reinterpret_cast(sh_zp))[zp_sh_rd + i]; + } + } + + } else if constexpr (group_blocks >= thread_k_blocks) { + if (k % b_sh_wr_iters == 0) { + int4 *sh_zp_stage = sh_zp + zp_sh_stage * ((group_blocks / thread_k_blocks) * (pipe / (group_blocks / thread_k_blocks))); +#pragma unroll + for (int i = 0; i < num_ints_per_thread; i++) { + frag_qzp[k % 2][i] = (reinterpret_cast(sh_zp_stage))[zp_sh_rd + i]; + } + } + } else { + auto warp_id = threadIdx.x / 32; + int n_warps = thread_n_blocks / 4; + + int warp_row = warp_id / n_warps; + + int cur_k = warp_row * 16; + cur_k += k_iter_size * (k % b_sh_wr_iters); + + int k_blocks = cur_k / 16; + int cur_group_id = 0; + + // Suppress bogus and persistent divide-by-zero warning +#pragma nv_diagnostic push +#pragma nv_diag_suppress divide_by_zero + cur_group_id = k_blocks / group_blocks; +#pragma nv_diagnostic pop + + int4 *sh_zp_stage = sh_zp + zp_sh_stage * pipe; + + sh_zp_stage += cur_group_id * zp_sh_stride; + +#pragma unroll + for (int i = 0; i < num_ints_per_thread; i++) { + frag_qzp[k % 2][i] = (reinterpret_cast(sh_zp_stage))[zp_sh_rd + i]; + } + } + } + + else if constexpr (has_zp && is_zp_float) { + int pipe = full_pipe % stages; + + if constexpr (group_blocks != -1) { + if constexpr (group_blocks >= thread_k_blocks) { + if (k % b_sh_wr_iters == 0) { + int4 *sh_zp_stage = sh_zp + zp_sh_stage * ((group_blocks / thread_k_blocks) * (pipe / (group_blocks / thread_k_blocks))); + reinterpret_cast(&frag_zpf[k % 2])[0] = sh_zp_stage[zp_sh_rd]; + } + } else { + auto warp_id = threadIdx.x / 32; + int n_warps = thread_n_blocks / 4; + + int warp_row = warp_id / n_warps; + + int cur_k = warp_row * 16; + cur_k += k_iter_size * (k % b_sh_wr_iters); + + int k_blocks = cur_k / 16; + // Suppress bogus and persistent divide-by-zero warning +#pragma nv_diagnostic push +#pragma nv_diag_suppress divide_by_zero + int cur_group_id = k_blocks / group_blocks; +#pragma nv_diagnostic pop + + int4 *sh_zp_stage = sh_zp + zp_sh_stage * pipe; + + reinterpret_cast(&frag_zpf[k % 2])[0] = sh_zp_stage[zp_sh_rd + cur_group_id * zp_sh_stride]; + } + } + } + }; + + auto dequant_data = [&](int q, scalar_t2 *frag_b_ptr) { + dequant(q, frag_b_ptr); + }; + + // Execute the actual tensor core matmul of a sub-tile. + bool is_first_matmul_in_slice = true; + auto matmul = [&](int k) { + int k2 = k % 2; + const bool is_new_zp = ((group_blocks != -1) && (group_blocks < thread_k_blocks || k == 0)) || (group_blocks == -1 && is_first_matmul_in_slice); + if constexpr (has_zp && !is_zp_float) { + if (is_new_zp) { + if constexpr (group_blocks == -1) { + is_first_matmul_in_slice = false; + } + int zp_quant_0, zp_quant_1; + + if constexpr (w_type.size_bits() == 4) { + zp_quant_0 = frag_qzp[k2][0]; + zp_quant_1 = zp_quant_0 >> 8; + } else { + static_assert(w_type.size_bits() == 8); + zp_quant_0 = frag_qzp[k2][0]; + zp_quant_1 = frag_qzp[k2][1]; + } + + dequant_data(zp_quant_0, reinterpret_cast(&frag_zp)); + dequant_data(zp_quant_1, reinterpret_cast(&frag_zp) + 2); + } + } + if constexpr (!dequant_skip_flop && has_zp && is_zp_float) { + if (is_new_zp) { + reinterpret_cast(&frag_zp)[0] = reinterpret_cast(&frag_zpf[k2])[0]; + } + } + + // FP4/FP8 scale dequantization (E4M3 for NVFP4 and E8M0 for MXFP4). + if constexpr ( + (s_type == host::kFE4M3fn || s_type == host::kFE8M0fnu) && !(std::is_same::value && s_type == host::kFE8M0fnu)) { + int s_quant_0 = reinterpret_cast(frag_s[k2])[0]; + int s_quant_1 = reinterpret_cast(frag_s[k2])[1]; + + dequant_fp8_scales(s_quant_0, reinterpret_cast(&frag_s[k2])); + dequant_fp8_scales(s_quant_1, reinterpret_cast(&frag_s[k2]) + 2); + } + +// We have the m dimension as the inner loop in order to encourage overlapping +// dequantization and matmul operations. +#pragma unroll + for (int j = 0; j < 4; j++) { + FragB frag_b0; + FragB frag_b1; + int b_quant_0, b_quant_1; + + if constexpr (w_type_id == host::kFE2M1f.id()) { + b_quant_1 = frag_b_quant[k2][0][j]; + b_quant_0 = b_quant_1 << 8; + } else if constexpr (w_type.size_bits() == 4) { + b_quant_0 = frag_b_quant[k2][0][j]; + b_quant_1 = b_quant_0 >> 8; + } else { + static_assert(w_type.size_bits() == 8); + int *frag_b_quant_ptr = reinterpret_cast(frag_b_quant[k2]); + b_quant_0 = frag_b_quant_ptr[j * 2 + 0]; + b_quant_1 = frag_b_quant_ptr[j * 2 + 1]; + } + + dequant_data(b_quant_0, reinterpret_cast(&frag_b0)); + dequant_data(b_quant_1, reinterpret_cast(&frag_b1)); + + if constexpr (dequant_skip_flop && has_zp && !is_zp_float) { + sub_zp(frag_b0, frag_zp[j], 0); + sub_zp(frag_b1, frag_zp[j], 1); + } + + // Apply scale to frag_b0 + if constexpr (has_act_order) { + static_assert(group_blocks != -1); + scale4( + frag_b0, act_frag_s[k2][0][j], act_frag_s[k2][1][j], act_frag_s[k2][2][j], act_frag_s[k2][3][j], 0); + scale4( + frag_b1, act_frag_s[k2][0][j], act_frag_s[k2][1][j], act_frag_s[k2][2][j], act_frag_s[k2][3][j], 1); + } else if constexpr (!dequant_skip_flop && has_zp && !is_zp_float && group_blocks == -1) { + int idx = (threadIdx.x / 4) % 2; + scalar_t2 s2 = Dtype::nums2num2( + reinterpret_cast(&frag_s[j / 2][j % 2 * 2 + 0])[idx], + reinterpret_cast(&frag_s[j / 2][j % 2 * 2 + 1])[idx]); + if (is_new_zp) + frag_zp[j] = __hmul2(frag_zp[j], s2); + scale_and_sub(frag_b0, s2.x, frag_zp[j].x); + scale_and_sub(frag_b1, s2.y, frag_zp[j].y); + } else if constexpr (!dequant_skip_flop && has_zp && group_blocks != -1) { + if (is_new_zp) + frag_zp[j] = __hmul2(frag_zp[j], *reinterpret_cast(&frag_s[k2][j])); + scale_and_sub(frag_b0, frag_s[k2][j][0].x, frag_zp[j].x); + scale_and_sub(frag_b1, frag_s[k2][j][0].y, frag_zp[j].y); + } else if constexpr (group_blocks != -1) { + scale(frag_b0, frag_s[k2][j], 0); + scale(frag_b1, frag_s[k2][j], 1); + } + +#pragma unroll + for (int i = 0; i < thread_m_blocks; i++) { + if constexpr (m_block_size_8) { + mma_trans(frag_a[k2][i], frag_b0, frag_b1, frag_c[i][j][0]); + } else { + mma(frag_a[k2][i], frag_b0, frag_c[i][j][0]); + mma(frag_a[k2][i], frag_b1, frag_c[i][j][1]); + } + } + } + }; + + // Since we slice across the k dimension of a tile in order to increase the + // number of warps while keeping the n dimension of a tile reasonable, we have + // multiple warps that accumulate their partial sums of the same output + // location; which we have to reduce over in the end. We do in shared memory. + auto thread_block_reduce = [&]() { + constexpr int red_off = threads / b_sh_stride_threads / 2; + if (red_off >= 1) { + auto red_idx = threadIdx.x / b_sh_stride_threads; + constexpr int red_sh_stride = b_sh_stride_threads * 4 * 2; + constexpr int red_sh_delta = b_sh_stride_threads; + int red_sh_rd = red_sh_stride * (threadIdx.x / b_sh_stride_threads) + (threadIdx.x % b_sh_stride_threads); + + // Parallel logarithmic shared memory reduction. We make sure to avoid any + // unnecessary read or write iterations, e.g., for two warps we write only + // once by warp 1 and read only once by warp 0. + +#pragma unroll + for (int m_block = 0; m_block < thread_m_blocks; m_block++) { +#pragma unroll + for (int i = red_off; i > 0; i /= 2) { + if (i <= red_idx && red_idx < 2 * i) { +#pragma unroll + for (int j = 0; j < 4 * 2; j += (m_block_size_8 ? 2 : 1)) { + int red_sh_wr = red_sh_delta * j + (red_sh_rd - red_sh_stride * i); + if (i < red_off) { + float *c_rd = reinterpret_cast(&sh_red[red_sh_delta * j + red_sh_rd]); + float *c_wr = reinterpret_cast(&sh_red[red_sh_wr]); +#pragma unroll + for (int k = 0; k < 4; k++) + reinterpret_cast(frag_c)[4 * 2 * m_block + j][k] += c_rd[k] + c_wr[k]; + } + sh_red[red_sh_wr] = reinterpret_cast(&frag_c)[4 * 2 * m_block + j]; + } + } + __syncthreads(); + } + if (red_idx == 0) { +#pragma unroll + for (int i = 0; i < 4 * 2; i += (m_block_size_8 ? 2 : 1)) { + float *c_rd = reinterpret_cast(&sh_red[red_sh_delta * i + red_sh_rd]); +#pragma unroll + for (int j = 0; j < 4; j++) + reinterpret_cast(frag_c)[4 * 2 * m_block + i][j] += c_rd[j]; + } + } + __syncthreads(); + } + } + }; + + // Since multiple threadblocks may process parts of the same column slice, we + // finally have to globally reduce over the results. As the striped + // partitioning minimizes the number of such reductions and our outputs are + // usually rather small, we perform this reduction serially in L2 cache. + auto global_reduce_fp16 = [&](bool first = false, bool last = false) { + // We are very careful here to reduce directly in the output buffer to + // maximize L2 cache utilization in this step. To do this, we write out + // results in FP16 (but still reduce with FP32 compute). + constexpr int active_threads = 32 * thread_n_blocks / 4; + bool is_th_active = threadIdx.x < active_threads; + if (!is_th_active) { + return; + } + + int c_gl_stride = prob_n / 8; + int c_gl_wr_delta_o = 8 * c_gl_stride; + int c_gl_wr_delta_i = 4 * (active_threads / 32); + int c_gl_wr; + if constexpr (m_block_size_8) { + c_gl_wr = c_gl_stride * ((threadIdx.x % 4) * 2) + 4 * (threadIdx.x / 32) + (threadIdx.x % 32) / 8; + c_gl_wr += (2 * thread_n_blocks) * slice_col; + } else { + c_gl_wr = c_gl_stride * ((threadIdx.x % 32) / 4) + 4 * (threadIdx.x / 32) + threadIdx.x % 4; + c_gl_wr += (2 * thread_n_blocks) * slice_col; + } + constexpr int c_sh_wr_delta = active_threads; + int c_sh_wr = threadIdx.x; + + if (!first) { + +#pragma unroll + for (int i = 0; i < (m_block_size_8 ? 2 : thread_m_blocks * 4); i++) { + int c_idx; + if constexpr (m_block_size_8) + c_idx = c_gl_wr + i * c_gl_stride + (threadIdx.x % 8) / 4 * c_gl_wr_delta_i; + else + c_idx = c_gl_wr + c_gl_wr_delta_o * (i / 2) + c_gl_wr_delta_i * (i % 2); + if (c_idx / c_gl_stride < block_num_valid_tokens) { + int64_t sorted_row = sh_block_sorted_ids[c_idx / c_gl_stride]; + int64_t true_idx = sorted_row * c_gl_stride + c_idx % c_gl_stride; + sh_red[c_sh_wr + c_sh_wr_delta * i] = C[true_idx]; + } + } + } + +#pragma unroll + for (int i = 0; i < (m_block_size_8 ? 2 : thread_m_blocks * 4); i++) { + if (!first) { + int4 c_red = sh_red[c_sh_wr + i * c_sh_wr_delta]; +#pragma unroll + for (int j = 0; j < 2 * 4; j++) { + int delta = 0; + if constexpr (m_block_size_8) { + delta = j % 2 == 1 ? -2 : 0; + } + reinterpret_cast(&frag_c)[4 * 2 * 4 * (i / 4) + 4 * j + (i % 4) + delta] += Dtype::num2float(reinterpret_cast(&c_red)[j]); + } + } + if (!last) { + int4 c; +#pragma unroll + for (int j = 0; j < 2 * 4; j++) { + int delta = 0; + if constexpr (m_block_size_8) { + delta = j % 2 == 1 ? -2 : 0; + } + reinterpret_cast(&c)[j] = Dtype::float2num(reinterpret_cast(&frag_c)[4 * 2 * 4 * (i / 4) + 4 * j + (i % 4) + delta]); + } + + int c_idx; + if constexpr (m_block_size_8) + c_idx = c_gl_wr + i * c_gl_stride + (threadIdx.x % 8) / 4 * c_gl_wr_delta_i; + else + c_idx = c_gl_wr + c_gl_wr_delta_o * (i / 2) + c_gl_wr_delta_i * (i % 2); + if (c_idx / c_gl_stride < block_num_valid_tokens) { + int64_t sorted_row = sh_block_sorted_ids[c_idx / c_gl_stride]; + int64_t true_idx = sorted_row * c_gl_stride + c_idx % c_gl_stride; + C[true_idx] = c; + } + } + } + }; + + // Globally reduce over threadblocks that compute the same column block. + // We use a tmp C buffer to reduce in full fp32 precision. + auto global_reduce_fp32 = [&](bool first = false, bool last = false) { + constexpr int tb_m = thread_m_blocks * 16; + constexpr int tb_n = thread_n_blocks * 16; + + constexpr int c_size = tb_m * tb_n * sizeof(float) / 16; + + constexpr int active_threads = 32 * thread_n_blocks / 4; + bool is_th_active = threadIdx.x < active_threads; + + constexpr int num_floats = thread_m_blocks * 4 * 2 * 4; + constexpr int th_size = num_floats * sizeof(float) / 16; + + int c_cur_offset = locks_off * c_size; + + if (!is_th_active) { + return; + } + + if (!first) { + float *frag_c_ptr = reinterpret_cast(&frag_c); +#pragma unroll + for (int k = 0; k < th_size; k++) { + if constexpr (m_block_size_8) { + if (k % 2) + continue; + } else { + if (k / 8 * 16 + (threadIdx.x % 32) / 4 >= block_num_valid_tokens) + continue; + } + + sh_red[threadIdx.x] = C_tmp[c_cur_offset + active_threads * k + threadIdx.x]; + + float *sh_c_ptr = reinterpret_cast(&sh_red[threadIdx.x]); +#pragma unroll + for (int f = 0; f < 4; f++) { + frag_c_ptr[k * 4 + f] += sh_c_ptr[f]; + } + } + } + + if (!last) { + int4 *frag_c_ptr = reinterpret_cast(&frag_c); +#pragma unroll + for (int k = 0; k < th_size; k++) { + if constexpr (m_block_size_8) { + if (k % 2) + continue; + } else { + if (k / 8 * 16 + (threadIdx.x % 32) / 4 >= block_num_valid_tokens) + continue; + } + + C_tmp[c_cur_offset + active_threads * k + threadIdx.x] = frag_c_ptr[k]; + } + } + }; + + // Write out the reduce final result in the correct layout. We only actually + // reshuffle matrix fragments in this step, the reduction above is performed + // in fragment layout. + auto write_result = [&](bool last) { + int c_gl_stride = prob_n / 8; + constexpr int c_sh_stride = 2 * thread_n_blocks + 1; + int c_gl_wr_delta = c_gl_stride * (threads / (2 * thread_n_blocks)); + constexpr int c_sh_rd_delta = c_sh_stride * (threads / (2 * thread_n_blocks)); + + int c_gl_wr = c_gl_stride * (threadIdx.x / (2 * thread_n_blocks)) + (threadIdx.x % (2 * thread_n_blocks)); + c_gl_wr += (2 * thread_n_blocks) * slice_col; + int c_sh_wr; + if constexpr (m_block_size_8) { + c_sh_wr = (8 * c_sh_stride) * ((threadIdx.x % 32) % 4 * 2) + (threadIdx.x % 32) / 4; + c_sh_wr += 64 * (threadIdx.x / 32); + } else { + c_sh_wr = (4 * c_sh_stride) * ((threadIdx.x % 32) / 4) + (threadIdx.x % 32) % 4; + c_sh_wr += 32 * (threadIdx.x / 32); + } + + int c_sh_rd = c_sh_stride * (threadIdx.x / (2 * thread_n_blocks)) + (threadIdx.x % (2 * thread_n_blocks)); + + // We first reorder in shared memory to guarantee the most efficient final + // global write patterns + auto write = [&](int idx, float c0, float c1, FragS &s, FragS &b_bias) { + scalar_t2 res = Dtype::nums2num2(Dtype::float2num(c0), Dtype::float2num(c1)); + + // For per-column quantization we finally apply the scale here (only for + // 4-bit) + if constexpr ( + !has_act_order && group_blocks == -1 && w_type.size_bits() == 4 && (has_zp && dequant_skip_flop || !has_zp)) { + scalar_t2 tmp_scale = s[0]; + if constexpr (m_block_size_8) { + tmp_scale = Dtype::num2num2(reinterpret_cast(&s[0])[(threadIdx.x % 8) / 4]); + } + res = __hmul2(res, tmp_scale); + } + + if constexpr (w_type == host::kFE2M1f && s_type == host::kFE4M3fn) { + if (!mul_topk_weights) { + res = __hmul2(res, global_scale); + } + } + if (has_bias && last) { + scalar_t2 tmp_bias = b_bias[0]; + if constexpr (m_block_size_8) { + tmp_bias = Dtype::num2num2(reinterpret_cast(&b_bias[0])[(threadIdx.x % 8) / 4]); + } + res = __hadd2(res, tmp_bias); + } + + if constexpr (m_block_size_8) { + ((scalar_t *)sh_red)[idx] = res.x; + ((scalar_t *)sh_red)[idx + 8 * c_sh_stride] = res.y; + } else { + ((scalar_t2 *)sh_red)[idx] = res; + } + }; + + if (threadIdx.x / 32 < thread_n_blocks / 4) { +#pragma unroll + for (int i = 0; i < thread_m_blocks; i++) { +#pragma unroll + for (int j = 0; j < 4; j++) { + if constexpr (m_block_size_8) { + int wr = c_sh_wr + 16 * j; + write( + wr, + frag_c[i][j][0][0], + frag_c[i][j][0][1], + frag_s[j / 2][2 * (j % 2) + 0], + frag_bias[j / 2][2 * (j % 2) + 0]); + write( + wr + 8, + frag_c[i][j][0][2], + frag_c[i][j][0][3], + frag_s[j / 2][2 * (j % 2) + 1], + frag_bias[j / 2][2 * (j % 2) + 1]); + } else { + int wr = c_sh_wr + 8 * j; + write( + wr + (4 * c_sh_stride) * 0 + 0, + frag_c[i][j][0][0], + frag_c[i][j][0][1], + frag_s[j / 2][2 * (j % 2) + 0], + frag_bias[j / 2][2 * (j % 2) + 0]); + write( + wr + (4 * c_sh_stride) * 8 + 0, + frag_c[i][j][0][2], + frag_c[i][j][0][3], + frag_s[j / 2][2 * (j % 2) + 0], + frag_bias[j / 2][2 * (j % 2) + 0]); + write( + wr + (4 * c_sh_stride) * 0 + 4, + frag_c[i][j][1][0], + frag_c[i][j][1][1], + frag_s[j / 2][2 * (j % 2) + 1], + frag_bias[j / 2][2 * (j % 2) + 1]); + write( + wr + (4 * c_sh_stride) * 8 + 4, + frag_c[i][j][1][2], + frag_c[i][j][1][3], + frag_s[j / 2][2 * (j % 2) + 1], + frag_bias[j / 2][2 * (j % 2) + 1]); + } + } + c_sh_wr += 16 * (4 * c_sh_stride); + } + } + __syncthreads(); + +#pragma unroll + for (int i = 0; i < device::div_ceil(16 * thread_m_blocks, threads / (2 * thread_n_blocks)); i++) { + int row = c_gl_wr / c_gl_stride; + if (row < block_num_valid_tokens) { + int64_t sorted_row = sh_block_sorted_ids[row]; + int64_t true_idx = sorted_row * c_gl_stride + c_gl_wr % c_gl_stride; + scalar_t2 topk_weight_score; + if (mul_topk_weights) + topk_weight_score = sh_block_topk_weights[row]; + if (use_atomic_add && slice_count > 1 || mul_topk_weights) { + scalar_t2 *C_half2 = reinterpret_cast(&C[true_idx]); + scalar_t2 *sh_red_half2 = reinterpret_cast(&sh_red[c_sh_rd]); +#pragma unroll + for (int a = 0; a < 4; a++) { + scalar_t2 res = sh_red_half2[a]; + if (mul_topk_weights) { + res = __hmul2(res, topk_weight_score); + } + + if (use_atomic_add && slice_count > 1) { + atomicAdd(&C_half2[a], res); + } else { + C_half2[a] = res; + }; + } + } else { + C[true_idx] = sh_red[c_sh_rd]; + } + c_gl_wr += c_gl_wr_delta; + c_sh_rd += c_sh_rd_delta; + } + } + __syncthreads(); + }; + + // Start global fetch and register load pipelines. + auto start_pipes = [&]() { + +#pragma unroll + for (int i = 0; i < stages - 1; i++) { + if (has_act_order && i == 0) { + int last_g_idx = slice_k_start + stages * tb_k * 2; + if (last_g_idx >= prob_k) { + last_g_idx = prob_k - 1; + } + fetch_act_order_scales_to_shared(true, g_idx[slice_k_start], g_idx[last_g_idx]); + } + + if constexpr (has_zp && !is_zp_float && group_blocks == -1) { + if (i == 0) { + fetch_col_zp_to_shared(); + if constexpr (!dequant_skip_flop) { + fetch_col_scale_to_shared(); + } + } + } + fetch_to_shared(i, i, i < slice_iters, i); + } + + zero_accums(); + wait_for_stage(); + init_same_group(0); + fetch_to_registers(0, 0); + fetch_scales_to_registers(0, 0); + fetch_zp_to_registers(0, 0); + a_gl_rd_col += a_gl_rd_delta_o * (stages - 1); + if constexpr (has_act_order) { + slice_k_start_shared_fetch += tb_k * (stages - 1); + } + }; + if (slice_iters) { + start_pipes(); + } + + // Main loop. + while (slice_iters) { + // We unroll over both the global fetch and the register load pipeline to + // ensure all shared memory accesses are static. Note that both pipelines + // have even length meaning that the next iteration will always start at + // index 0. + + for (int stage_group_id = 0; stage_group_id < max_num_stage_groups; stage_group_id++) { +#pragma unroll + for (int pipe = 0; pipe < stages;) { +#pragma unroll + for (int k = 0; k < b_sh_wr_iters; k++) { + int idx = (pipe >= stages && stage_group_id == max_num_stage_groups - 1) ? (pipe - stages) + : (pipe + stage_group_id * stages); + fetch_to_registers(k + 1, pipe % stages, idx); + fetch_scales_to_registers(k + 1, pipe); + fetch_zp_to_registers(k + 1, pipe); + if (k == b_sh_wr_iters - 2) { + int idx = (pipe >= 1 && stage_group_id == max_num_stage_groups - 1) + ? (pipe - 1) + : (pipe + (stage_group_id + 1) * stages - 1); + fetch_to_shared((pipe + stages - 1) % stages, pipe, slice_iters >= stages, idx); + pipe++; + wait_for_stage(); + init_same_group(pipe % stages); + } + matmul(k); + } + slice_iters--; + if (slice_iters == 0) { + break; + } + } + + a_gl_rd_col += a_gl_rd_delta_o * stages; + + if constexpr (has_act_order) { + slice_k_start += tb_k * stages; + + if (slice_k_start < prob_k) { + slice_k_start_shared_fetch += tb_k * stages; + int first_group_id = g_idx[slice_k_start]; + int last_g_idx = slice_k_start + stages * tb_k * 2; + if (last_g_idx >= prob_k) { + last_g_idx = prob_k - 1; + } + int last_group_id = g_idx[last_g_idx]; + if (last_group_id >= sh_first_group_id + sh_num_groups) { + fetch_act_order_scales_to_shared(false, first_group_id, last_group_id); + __syncthreads(); + } + } + } + if (slice_iters == 0) { + break; + } + } + + // Process results and, if necessary, proceed to the next column slice. + // While this pattern may not be the most readable, other ways of writing + // the loop seemed to noticeably worse performance after compilation. + if (slice_iters == 0) { + cp_async_wait<0>(); + bool last = slice_idx == slice_count - 1; + // For per-column scales, we only fetch them here in the final step before + // write-out + if constexpr (!has_act_order && group_blocks == -1 && (has_zp && dequant_skip_flop || !has_zp)) { + if (w_type.size_bits() == 8 || (last || use_atomic_add)) { + if (s_sh_wr_pred) { + cp_async4(&sh_s[s_sh_wr], &scales_ptr[s_gl_rd]); + } + cp_async_fence(); + } + } + + thread_block_reduce(); + + if (has_bias && last) { + __syncthreads(); + cp_async4_pred(&sh_bias[bias_sh_wr], &b_bias_ptr[bias_gl_rd], threadIdx.x < 16 * thread_n_blocks / 8); + cp_async_fence(); + } + + if constexpr (!has_act_order && group_blocks == -1 && (has_zp && dequant_skip_flop || !has_zp)) { + if (w_type.size_bits() == 8 || (last || use_atomic_add)) { + cp_async_wait<0>(); + __syncthreads(); + if (threadIdx.x / 32 < thread_n_blocks / 4) { + reinterpret_cast(&frag_s)[0] = sh_s[s_sh_rd + 0]; + reinterpret_cast(&frag_s)[1] = sh_s[s_sh_rd + 4]; + if constexpr (m_block_size_8) { + int idx = (threadIdx.x / 4) % 2; + scalar_t2 *frag_s_half2 = reinterpret_cast(frag_s); +#pragma unroll + for (int i = 0; i < 8; i++) { + frag_s_half2[i] = Dtype::num2num2(reinterpret_cast(&frag_s_half2[i])[idx]); + } + } + } + } + } + + // For 8-bit channelwise, we apply the scale before the global reduction + // that converts the fp32 results to fp16 (so that we avoid possible + // overflow in fp16) + if constexpr ( + !has_act_order && group_blocks == -1 && w_type.size_bits() == 8 && (has_zp && dequant_skip_flop || !has_zp)) { + if (threadIdx.x / 32 < thread_n_blocks / 4) { +#pragma unroll + for (int i = 0; i < thread_m_blocks; i++) { +#pragma unroll + for (int j = 0; j < 4; j++) { + scale_float(reinterpret_cast(&frag_c[i][j][0][0]), frag_s[j / 2][2 * (j % 2) + 0]); + scale_float( + reinterpret_cast(&frag_c[i][j][0][2]), frag_s[j / 2][2 * (j % 2) + (m_block_size_8 ? 1 : 0)]); + + if constexpr (!m_block_size_8) { + scale_float(reinterpret_cast(&frag_c[i][j][1][0]), frag_s[j / 2][2 * (j % 2) + 1]); + scale_float(reinterpret_cast(&frag_c[i][j][1][2]), frag_s[j / 2][2 * (j % 2) + 1]); + } + } + } + } + } + + if (slice_count > 1 && !use_atomic_add) { + // only globally reduce if there is more than one block in a slice + barrier_acquire(&locks[locks_off], slice_idx); + if (use_fp32_reduce) { + global_reduce_fp32(slice_idx == 0, last); + } else { + global_reduce_fp16(slice_idx == 0, last); + } + barrier_release(&locks[locks_off], last); + } + + if (has_bias && last) { + cp_async_wait<0>(); + __syncthreads(); + reinterpret_cast(&frag_bias)[0] = sh_bias[bias_sh_rd]; + reinterpret_cast(&frag_bias)[1] = sh_bias[bias_sh_rd + 4]; + __syncthreads(); + } + + if (use_atomic_add && slice_count > 1 && slice_idx != 0) { + wait_negative_and_add(&locks[locks_off]); + } + if (last || use_atomic_add) { + // only the last block in a slice actually writes the result + write_result(last); + } + int old_slice_row = slice_row; + slice_row = 0; + slice_col_par++; + slice_col++; + is_first_matmul_in_slice = true; + init_slice(); + + // Should we load A matrix in next slice? + // `slice_col == 0`: when move to a new moe block + // `old_slice_row > 0`: + // when the last slice is not starting from k_index == 0 + // (only happen when it is the first slice of a threadblock) + // `prob_k > thread_k_blocks * 16 * stages * max_num_stage_groups`: + // when the required shared memory size is larger than + // the remaining shared memory + if (slice_col == 0 || old_slice_row || prob_k > thread_k_blocks * 16 * stages * max_num_stage_groups) { + should_load_a = true; + } else { + should_load_a = false; + } + + if (slice_iters) { + a_gl_rd_col = (threadIdx.x % a_gl_rd_delta_o); +#pragma unroll + for (int i = 0; i < b_sh_wr_iters; i++) { + B_ptr[i] += b_sh_stride - b_gl_rd_delta_o * k_tiles; + } + if (slice_col == 0) { +#pragma unroll + for (int i = 0; i < b_sh_wr_iters; i++) { + B_ptr[i] -= b_gl_stride; + } + } + + bias_gl_rd = (thread_n_blocks * 16 / 8) * slice_col + threadIdx.x; + // Update slice k/n for scales loading + if constexpr (has_act_order) { + slice_k_start = tb_k * slice_row; + slice_k_finish = slice_k_start + tb_k * slice_iters; + slice_k_start_shared_fetch = slice_k_start; + slice_n_offset = act_s_col_tb_stride * slice_col; + } else { + if constexpr (group_blocks == -1) { + s_gl_rd = s_sh_stride * slice_col + threadIdx.x; + zp_gl_rd = zp_sh_stride * slice_col + threadIdx.x; + } else if constexpr (group_blocks >= thread_k_blocks) { + s_gl_rd = s_gl_stride * ((thread_k_blocks * slice_row) / group_blocks) + s_sh_stride * slice_col + threadIdx.x; + zp_gl_rd = zp_gl_stride * ((thread_k_blocks * slice_row) / group_blocks) + zp_sh_stride * slice_col + threadIdx.x; + } else { + s_gl_rd = s_gl_stride * ((thread_k_blocks * slice_row) / group_blocks + threadIdx.x / s_sh_stride) + s_sh_stride * slice_col + threadIdx.x % s_sh_stride; + zp_gl_rd = zp_gl_stride * ((thread_k_blocks * slice_row) / group_blocks + threadIdx.x / zp_sh_stride) + zp_sh_stride * slice_col + threadIdx.x % zp_sh_stride; + } + } + start_pipes(); + } + } + } +} + +} // namespace device::marlin_moe + +#endif diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/moe_wna16_marlin_gemm.h b/src/infiniop/ops/moe_wna16_marlin_gemm/moe_wna16_marlin_gemm.h new file mode 100644 index 000000000..97028610b --- /dev/null +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/moe_wna16_marlin_gemm.h @@ -0,0 +1,83 @@ +#ifndef __MOE_WNA16_MARLIN_GEMM_H__ +#define __MOE_WNA16_MARLIN_GEMM_H__ + +#include "../../../utils.h" +#include "../../operator.h" +#include "../../tensor.h" +#include "info.h" + +#define DESCRIPTOR(NAMESPACE) \ + \ + namespace op::moe_wna16_marlin_gemm::NAMESPACE { \ + class Descriptor final : public InfiniopDescriptor { \ + struct Opaque; \ + Opaque *_opaque; \ + MoeWna16MarlinGemmInfo _info; \ + size_t _workspace_size; \ + \ + Descriptor( \ + size_t workspace_size_, \ + Opaque *opaque, \ + MoeWna16MarlinGemmInfo info, \ + infiniDevice_t device_type, \ + int device_id) \ + : InfiniopDescriptor{device_type, device_id}, \ + _opaque(opaque), \ + _info(info), \ + _workspace_size(workspace_size_) {} \ + \ + public: \ + ~Descriptor(); \ + \ + size_t workspaceSize() const { return _workspace_size; } \ + \ + static infiniStatus_t create( \ + infiniopHandle_t handle, \ + Descriptor **desc_ptr, \ + infiniopTensorDescriptor_t c_desc, \ + infiniopTensorDescriptor_t a_desc, \ + infiniopTensorDescriptor_t b_q_weight_desc, \ + infiniopTensorDescriptor_t b_bias_desc, \ + infiniopTensorDescriptor_t b_scales_desc, \ + infiniopTensorDescriptor_t global_scales_desc, \ + infiniopTensorDescriptor_t b_zeros_desc, \ + infiniopTensorDescriptor_t g_idx_desc, \ + infiniopTensorDescriptor_t perm_desc, \ + infiniopTensorDescriptor_t sorted_token_desc, \ + infiniopTensorDescriptor_t expert_ids_desc, \ + infiniopTensorDescriptor_t num_tokens_post_padded_desc, \ + infiniopTensorDescriptor_t topk_weights_desc, \ + int size_m, \ + int size_n, \ + int size_k, \ + int top_k, \ + int moe_block_size); \ + \ + infiniStatus_t calculate( \ + void *workspace, \ + size_t workspace_size, \ + void *c, \ + const void *a, \ + const void *b_q_weight, \ + void *b_bias, \ + void *b_scales, \ + void *global_scales, \ + void *b_zeros, \ + void *g_idx, \ + void *perm, \ + void *sorted_token_ids, \ + void *expert_ids, \ + void *num_tokens_post_padded, \ + void *topk_weights, \ + bool mul_topk_weights, \ + bool is_ep, \ + int64_t b_q_type_id, \ + bool is_k_full, \ + bool use_atomic_add, \ + bool use_fp32_reduce, \ + bool is_zp_float, \ + void *stream) const; \ + }; \ + } + +#endif //__MOE_WNA16_MARLIN_GEMM_H__ diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/nvidia/kernel.cuh b/src/infiniop/ops/moe_wna16_marlin_gemm/nvidia/kernel.cuh new file mode 100644 index 000000000..ee0502a94 --- /dev/null +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/nvidia/kernel.cuh @@ -0,0 +1,834 @@ +/* + * Modified by Neural Magic + * Copyright (C) Marlin.2024 Elias Frantar + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* + * Adapted from https://github.com/IST-DASLab/marlin + */ + +#pragma once + +#include "../sgl_kernel/tensor.h" + +#include "../sgl_kernel/scalar_type.hpp" + +#include "../marlin/kernel.h" +#include "../marlin/marlin_template.h" + +namespace device::marlin_moe { + +__global__ void MarlinDefault(MARLIN_KERNEL_PARAMS){}; + +using MarlinFuncPtr = void (*)(MARLIN_KERNEL_PARAMS); + +#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800 + +template +__global__ void permute_cols_kernel( + int4 const *__restrict__ a_int4_ptr, + int const *__restrict__ perm_int_ptr, + int4 *__restrict__ out_int4_ptr, + const int32_t *__restrict__ sorted_token_ids_ptr, + const int32_t *__restrict__ expert_ids_ptr, + const int32_t *__restrict__ num_tokens_past_padded_ptr, + int size_m, + int size_k, + int top_k) {}; + +#else + +// For a given "a" of size [M,K] performs a permutation of the K columns based +// on the given "perm" indices. +template +__global__ void permute_cols_kernel( + int4 const *__restrict__ a_int4_ptr, + int const *__restrict__ perm_int_ptr, + int4 *__restrict__ out_int4_ptr, + const int32_t *__restrict__ sorted_token_ids_ptr, + const int32_t *__restrict__ expert_ids_ptr, + const int32_t *__restrict__ num_tokens_past_padded_ptr, + int size_m, + int size_k, + int top_k) { + int num_tokens_past_padded = num_tokens_past_padded_ptr[0]; + int num_moe_blocks = device::div_ceil(num_tokens_past_padded, moe_block_size); + int32_t block_sorted_ids[moe_block_size]; + int block_num_valid_tokens = 0; + int64_t old_expert_id = 0; + int64_t expert_id = 0; + int row_stride = size_k * sizeof(half) / 16; + + auto read_moe_block_data = [&](int block_id) { + block_num_valid_tokens = moe_block_size; + int4 *tmp_block_sorted_ids = reinterpret_cast(block_sorted_ids); + for (int i = 0; i < moe_block_size / 4; i++) { + tmp_block_sorted_ids[i] = ((int4 *)sorted_token_ids_ptr)[block_id * moe_block_size / 4 + i]; + } + for (int i = 0; i < moe_block_size; i++) { + if (block_sorted_ids[i] >= size_m * top_k) { + block_num_valid_tokens = i; + break; + }; + } + }; + + auto permute_row = [&](int row) { + int iters = size_k / default_threads; + int rest = size_k % default_threads; + + int in_offset = (row / top_k) * row_stride; + int out_offset = row * row_stride; + + half const *a_row_half = reinterpret_cast(a_int4_ptr + in_offset); + half *out_half = reinterpret_cast(out_int4_ptr + out_offset); + + int base_k = 0; + + for (int i = 0; i < iters; i++) { + auto cur_k = base_k + threadIdx.x; + int src_pos = perm_int_ptr[cur_k]; + + out_half[cur_k] = a_row_half[src_pos]; + + base_k += default_threads; + } + + if (rest) { + if (threadIdx.x < rest) { + auto cur_k = base_k + threadIdx.x; + int src_pos = perm_int_ptr[cur_k]; + + out_half[cur_k] = a_row_half[src_pos]; + } + } + }; + + for (int index = blockIdx.x; index < num_moe_blocks; index += gridDim.x) { + old_expert_id = expert_id; + int tmp_expert_id = expert_ids_ptr[index]; + if (tmp_expert_id == -1) { + continue; + } + expert_id = tmp_expert_id; + perm_int_ptr += (expert_id - old_expert_id) * size_k; + read_moe_block_data(index); + + for (int i = 0; i < block_num_valid_tokens; i++) { + permute_row(block_sorted_ids[i]); + } + } +} + +typedef struct +{ + int thread_k; + int thread_n; + int num_threads; +} thread_config_t; + +thread_config_t small_batch_thread_configs[] = { + // Ordered by priority + + // thread_k, thread_n, num_threads + {128, 128, 256}, + {64, 128, 128}}; + +thread_config_t large_batch_thread_configs[] = { + // Ordered by priority + + // thread_k, thread_n, num_threads + {64, 256, 256}, + {64, 128, 128}}; + +typedef struct +{ + int blocks_per_sm; + thread_config_t tb_cfg; +} exec_config_t; + +int get_scales_cache_size( + thread_config_t const &th_config, + int prob_m, + int prob_n, + int prob_k, + int num_bits, + int group_size, + bool has_act_order, + bool is_k_full) { + bool cache_scales_chunk = has_act_order && !is_k_full; + + int tb_n = th_config.thread_n; + int tb_k = th_config.thread_k; + + // Get max scale groups per thread-block + int tb_groups; + if (group_size == -1) { + tb_groups = 1; + } else if (group_size == 0) { + tb_groups = device::div_ceil(tb_k, 32); // Worst case is 32 group size + } else { + tb_groups = device::div_ceil(tb_k, group_size); + } + + if (cache_scales_chunk) { + int load_groups = tb_groups * pipe_stages * 2; // Chunk size is 2x pipeline over dim K + load_groups = max(load_groups, 32); // We load at least 32 scale groups + return load_groups * tb_n * 2; + } else { + int tb_scales = tb_groups * tb_n * 2; + + return tb_scales * pipe_stages; + } +} + +int get_kernel_cache_size( + thread_config_t const &th_config, + bool m_block_size_8, + int thread_m_blocks, + int prob_m, + int prob_n, + int prob_k, + int num_bits, + int group_size, + bool has_act_order, + bool is_k_full, + int has_zp, + int is_zp_float) { + int pack_factor = 32 / num_bits; + + // Get B size + int tb_k = th_config.thread_k; + int tb_n = th_config.thread_n; + int tb_m = thread_m_blocks * 16; + + // shm size for block_sorted_ids/rd_block_sorted_ids/block_topk_weights + // both of them requires tb_m * 4 bytes (tb_m * int32 or tb_m * float32) + int sh_block_meta_size = tb_m * 4; + int sh_a_size = pipe_stages * (tb_m * tb_k) * 2; + int sh_b_size = pipe_stages * (tb_k * tb_n / pack_factor) * 4; + int sh_red_size = tb_m * (tb_n + 8) * 2; + int sh_bias_size = tb_n * 2; + int tmp_size = (sh_b_size > sh_red_size ? sh_red_size : sh_b_size) + sh_bias_size; + tmp_size = max(max(sh_b_size, sh_red_size), tmp_size); + + int sh_s_size = get_scales_cache_size(th_config, prob_m, prob_n, prob_k, num_bits, group_size, has_act_order, is_k_full); + int sh_g_idx_size = has_act_order && !is_k_full ? pipe_stages * tb_k / 4 : 0; + int sh_zp_size = 0; + if (has_zp) { + if (is_zp_float) { + sh_zp_size = sh_s_size; + } else if (num_bits == 4) { + sh_zp_size = sh_s_size / 4; + } else if (num_bits == 8) { + sh_zp_size = sh_s_size / 2; + } + } + + int total_size = tmp_size + sh_a_size + sh_s_size + sh_zp_size + sh_g_idx_size + sh_block_meta_size; + + return total_size; +} + +bool is_valid_config( + thread_config_t const &th_config, + bool m_block_size_8, + int thread_m_blocks, + int prob_m, + int prob_n, + int prob_k, + int num_bits, + int group_size, + bool has_act_order, + bool is_k_full, + int has_zp, + int is_zp_float, + int max_shared_mem) { + // Sanity + if (th_config.thread_k == -1 || th_config.thread_n == -1 || th_config.num_threads == -1) { + return false; + } + + // Verify K/N are divisible by thread K/N + if (prob_k % th_config.thread_k != 0 || prob_n % th_config.thread_n != 0) { + return false; + } + + // Verify min for thread K/N + if (th_config.thread_n < min_thread_n || th_config.thread_k < min_thread_k) { + return false; + } + + // num_threads must be at least 128 (= 4 warps) + if (th_config.num_threads < 128) { + return false; + } + + // Check that pipeline fits into cache + int cache_size = get_kernel_cache_size( + th_config, + m_block_size_8, + thread_m_blocks, + prob_m, + prob_n, + prob_k, + num_bits, + group_size, + has_act_order, + is_k_full, + has_zp, + is_zp_float); + return cache_size + 512 <= max_shared_mem; +} + +#define _GET_IF( \ + W_TYPE, THREAD_M_BLOCKS, THREAD_N_BLOCKS, THREAD_K_BLOCKS, M_BLOCK_SIZE_8, GROUP_BLOCKS, NUM_THREADS, IS_ZP_FLOAT) \ + else if ( \ + q_type == W_TYPE && thread_m_blocks == THREAD_M_BLOCKS && thread_n_blocks == THREAD_N_BLOCKS && thread_k_blocks == THREAD_K_BLOCKS && m_block_size_8 == M_BLOCK_SIZE_8 && group_blocks == GROUP_BLOCKS && num_threads == NUM_THREADS && is_zp_float == IS_ZP_FLOAT) { \ + constexpr auto S_TYPE = W_TYPE == host::kFE2M1f \ + ? (GROUP_BLOCKS == 1 ? host::kFE4M3fn : host::kFE8M0fnu) \ + : (std::is_same::value ? host::kFloat16 : host::kBFloat16); \ + kernel = Marlin< \ + scalar_t, \ + W_TYPE.id(), \ + S_TYPE.id(), \ + NUM_THREADS, \ + THREAD_M_BLOCKS, \ + THREAD_N_BLOCKS, \ + THREAD_K_BLOCKS, \ + M_BLOCK_SIZE_8, \ + pipe_stages, \ + GROUP_BLOCKS, \ + IS_ZP_FLOAT>; \ + } + +// COMMON: cases for (group_blocks in [-1, 2, 4, 8] and is_zp_float == false) +// this is the most common cases +// BIGGROUP: cases for big group size (group_blocks in [-1, 8]) +// FZP: cases for float-zero-point (is_zp_float = true) +// ACT: cases for act order case (group_blocks == 0) +// NVFP4: cases for nvfp4(e2m1) (group_blocks == 1) +// MXFP4: cases for mxfp4(e2m1) (group_blocks == 2) +#define COMMON_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS) \ + _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, -1, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, 2, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, 4, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, 8, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, -1, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, 2, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, 4, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, 8, NUM_THREADS, false) + +#define COMMON_GET_IF_M234(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS) \ + _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, -1, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, 2, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, 4, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, 8, NUM_THREADS, false) \ + \ + _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, -1, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, 2, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, 4, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, 8, NUM_THREADS, false) \ + \ + _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, -1, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, 2, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, 4, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, 8, NUM_THREADS, false) + +#define COMMON_GET_IF(W_TYPE) \ + COMMON_GET_IF_M1(W_TYPE, 8, 8, 256) \ + COMMON_GET_IF_M1(W_TYPE, 8, 4, 128) \ + COMMON_GET_IF_M234(W_TYPE, 16, 4, 256) \ + COMMON_GET_IF_M234(W_TYPE, 8, 4, 128) + +#define BIGGROUP_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS) \ + _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, -1, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, 8, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, -1, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, 8, NUM_THREADS, false) + +#define BIGGROUP_GET_IF_M234(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS) \ + _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, -1, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, 8, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, -1, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, 8, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, -1, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, 8, NUM_THREADS, false) + +#define BIGGROUP_GET_IF(W_TYPE) \ + BIGGROUP_GET_IF_M1(W_TYPE, 8, 8, 256) \ + BIGGROUP_GET_IF_M1(W_TYPE, 8, 4, 128) \ + BIGGROUP_GET_IF_M234(W_TYPE, 16, 4, 256) \ + BIGGROUP_GET_IF_M234(W_TYPE, 8, 4, 128) + +#define NVFP4_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS) \ + _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, 1, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, 1, NUM_THREADS, false) + +#define NVFP4_GET_IF_M234(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS) \ + _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, 1, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, 1, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, 1, NUM_THREADS, false) + +#define NVFP4_GET_IF(W_TYPE) \ + NVFP4_GET_IF_M1(W_TYPE, 8, 8, 256) \ + NVFP4_GET_IF_M1(W_TYPE, 8, 4, 128) \ + NVFP4_GET_IF_M234(W_TYPE, 16, 4, 256) \ + NVFP4_GET_IF_M234(W_TYPE, 8, 4, 128) + +#define MXFP4_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS) \ + _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, 2, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, 2, NUM_THREADS, false) + +#define MXFP4_GET_IF_M234(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS) \ + _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, 2, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, 2, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, 2, NUM_THREADS, false) + +#define MXFP4_GET_IF(W_TYPE) \ + MXFP4_GET_IF_M1(W_TYPE, 8, 8, 256) \ + MXFP4_GET_IF_M1(W_TYPE, 8, 4, 128) \ + MXFP4_GET_IF_M234(W_TYPE, 16, 4, 256) \ + MXFP4_GET_IF_M234(W_TYPE, 8, 4, 128) + +// We currently have 4-bit models only with group_blocks == 4 +#define FZP_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS) \ + _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, 4, NUM_THREADS, true) \ + _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, 4, NUM_THREADS, true) + +#define FZP_GET_IF_M234(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS) \ + _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, 4, NUM_THREADS, true) \ + _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, 4, NUM_THREADS, true) \ + _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, 4, NUM_THREADS, true) + +#define FZP_GET_IF(W_TYPE) \ + FZP_GET_IF_M1(W_TYPE, 8, 8, 256) \ + FZP_GET_IF_M1(W_TYPE, 8, 4, 128) \ + FZP_GET_IF_M234(W_TYPE, 16, 4, 256) \ + FZP_GET_IF_M234(W_TYPE, 8, 4, 128) + +// We currently have 4-bit models only with group_blocks == 4 +#define ACT_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS) \ + _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, 0, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, 0, NUM_THREADS, false) + +#define ACT_GET_IF_M234(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS) \ + _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, 0, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, 0, NUM_THREADS, false) \ + _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, 0, NUM_THREADS, false) + +#define ACT_GET_IF(W_TYPE) \ + ACT_GET_IF_M1(W_TYPE, 8, 8, 256) \ + ACT_GET_IF_M1(W_TYPE, 8, 4, 128) \ + ACT_GET_IF_M234(W_TYPE, 16, 4, 256) \ + ACT_GET_IF_M234(W_TYPE, 8, 4, 128) + +template +MarlinFuncPtr get_marlin_kernel( + const host::ScalarType q_type, + int thread_m_blocks, + int thread_n_blocks, + int thread_k_blocks, + bool m_block_size_8, + bool has_act_order, + bool has_zp, + int group_blocks, + int num_threads, + bool is_zp_float) { + int num_bits = q_type.size_bits(); + auto kernel = MarlinDefault; + if (false) { + } + + COMMON_GET_IF(host::kU4) + COMMON_GET_IF(host::kU4B8) + COMMON_GET_IF(host::kU8B128) + + NVFP4_GET_IF(host::kFE2M1f) + + BIGGROUP_GET_IF(host::kFE4M3fn) + + ACT_GET_IF(host::kU4B8) + ACT_GET_IF(host::kU8B128) + if (std::is_same::value) { + if (false) { + } + MXFP4_GET_IF(host::kFE2M1f) + } + + return kernel; +} + +template +exec_config_t determine_exec_config( + const host::ScalarType &q_type, + int prob_m, + int prob_n, + int prob_k, + int thread_m_blocks, + bool m_block_size_8, + int num_bits, + int group_size, + bool has_act_order, + bool is_k_full, + bool has_zp, + bool is_zp_float, + int max_shared_mem) { + exec_config_t exec_cfg = exec_config_t{1, thread_config_t{-1, -1, -1}}; + thread_config_t *thread_configs = thread_m_blocks > 1 ? large_batch_thread_configs : small_batch_thread_configs; + int thread_configs_size = thread_m_blocks > 1 ? sizeof(large_batch_thread_configs) / sizeof(thread_config_t) + : sizeof(small_batch_thread_configs) / sizeof(thread_config_t); + + int count = 0; + constexpr int device_max_reg_size = 255 * 1024; + for (int i = 0; i < thread_configs_size; i++) { + thread_config_t th_config = thread_configs[i]; + + if (!is_valid_config( + th_config, + m_block_size_8, + thread_m_blocks, + prob_m, + prob_n, + prob_k, + num_bits, + group_size, + has_act_order, + is_k_full, + has_zp, + is_zp_float, + max_shared_mem)) { + continue; + } + + int cache_size = get_kernel_cache_size( + th_config, + m_block_size_8, + thread_m_blocks, + prob_m, + prob_n, + prob_k, + num_bits, + group_size, + has_act_order, + is_k_full, + has_zp, + is_zp_float); + + int group_blocks = 0; + if (!has_act_order) { + group_blocks = group_size == -1 ? -1 : (group_size / 16); + } + + auto kernel = get_marlin_kernel( + q_type, + thread_m_blocks, + th_config.thread_n / 16, + th_config.thread_k / 16, + m_block_size_8, + has_act_order, + has_zp, + group_blocks, + th_config.num_threads, + is_zp_float); + + if (kernel == MarlinDefault) { + continue; + } + + if (thread_m_blocks > 1) { + exec_cfg = {1, th_config}; + break; + } else { + cudaFuncAttributes attr; + cudaFuncGetAttributes(&attr, kernel); + int reg_size = max(attr.numRegs, 1) * th_config.num_threads * 4; + int allow_count = min(device_max_reg_size / reg_size, max_shared_mem / (cache_size + 1024)); + allow_count = max(min(allow_count, 4), 1); + if (allow_count > count) { + count = allow_count; + exec_cfg = {count, th_config}; + }; + } + } + + return exec_cfg; +} + +template +void marlin_mm( + const void *A, + const void *B, + void *C, + void *C_tmp, + void *b_bias, + void *s, + void *s2, + void *zp, + void *g_idx, + void *perm, + void *a_tmp, + void *sorted_token_ids, + void *expert_ids, + void *num_tokens_past_padded, + void *topk_weights, + int moe_block_size, + int top_k, + bool mul_topk_weights, + bool is_ep, + int prob_m, + int prob_n, + int prob_k, + void *workspace, + host::ScalarType const &q_type, + bool has_bias, + bool has_act_order, + bool is_k_full, + bool has_zp, + int num_groups, + int group_size, + int dev, + cudaStream_t stream, + int thread_k, + int thread_n, + int sms, + bool use_atomic_add, + bool use_fp32_reduce, + bool is_zp_float) { + int thread_m_blocks = device::div_ceil(moe_block_size, 16); + bool m_block_size_8 = moe_block_size == 8; + + if (has_zp) { + host::RuntimeCheck( + q_type == host::kU4 || q_type == host::kU8, "q_type must be u4 or u8 when has_zp = True. Got = ", q_type.str()); + } else { + host::RuntimeCheck( + q_type == host::kU4B8 || q_type == host::kU8B128 || q_type == host::kFE4M3fn || q_type == host::kFE2M1f, + "q_type must be uint4b8, uint8b128, float8_e4m3fn or float4_e2m1f when " + "has_zp = False. Got = ", + q_type.str()); + } + + host::RuntimeCheck( + prob_m > 0 && prob_n > 0 && prob_k > 0, "Invalid MNK = [", prob_m, ", ", prob_n, ", ", prob_k, "]"); + + int group_blocks = 0; + if (has_act_order) { + if (is_k_full) { + host::RuntimeCheck(group_size != -1); + group_blocks = group_size / 16; + host::RuntimeCheck( + prob_k % group_blocks == 0, "prob_k = ", prob_k, " is not divisible by group_blocks = ", group_blocks); + } else { + host::RuntimeCheck(group_size == 0); + group_blocks = 0; + } + } else { + if (group_size == -1) { + group_blocks = -1; + } else { + group_blocks = group_size / 16; + host::RuntimeCheck( + prob_k % group_blocks == 0, "prob_k = ", prob_k, " is not divisible by group_blocks = ", group_blocks); + } + } + + int num_bits = q_type.size_bits(); + const int4 *A_ptr = (const int4 *)A; + const int4 *B_ptr = (const int4 *)B; + int4 *C_ptr = (int4 *)C; + int4 *C_tmp_ptr = (int4 *)C_tmp; + const int4 *bias_ptr = (const int4 *)b_bias; + const int4 *s_ptr = (const int4 *)s; + const uint16_t *s2_ptr = (const uint16_t *)s2; + const int4 *zp_ptr = (const int4 *)zp; + const int *g_idx_ptr = (const int *)g_idx; + const int *perm_ptr = (const int *)perm; + int4 *a_tmp_ptr = (int4 *)a_tmp; + const int32_t *sorted_token_ids_ptr = (const int32_t *)sorted_token_ids; + const int32_t *expert_ids_ptr = (const int32_t *)expert_ids; + const int32_t *num_tokens_past_padded_ptr = (const int32_t *)num_tokens_past_padded; + const float *topk_weights_ptr = (const float *)topk_weights; + int *locks = (int *)workspace; + + if (has_act_order) { + // Permute A columns + auto perm_kernel = permute_cols_kernel<8>; + if (moe_block_size == 8) { + } else if (moe_block_size == 16) { + perm_kernel = permute_cols_kernel<16>; + } else if (moe_block_size == 32) { + perm_kernel = permute_cols_kernel<32>; + } else if (moe_block_size == 48) { + perm_kernel = permute_cols_kernel<48>; + } else if (moe_block_size == 64) { + perm_kernel = permute_cols_kernel<64>; + } else { + host::Panic("unsupported moe_block_size ", moe_block_size); + } + + // clang-format off + perm_kernel<<>>( + A_ptr, perm_ptr, a_tmp_ptr, sorted_token_ids_ptr, expert_ids_ptr, + num_tokens_past_padded_ptr, prob_m, prob_k, top_k); + // clang-format on + A_ptr = a_tmp_ptr; + prob_m = prob_m * top_k; + top_k = 1; + + // If we have a full K, then we can run the non-act-order version of Marlin + // (since the weight rows are reordered by increasing group ids, and by + // having a full K, we have full original groups) + if (is_k_full) { + has_act_order = false; + } + } + + int max_shared_mem = 0; + host::RuntimeDeviceCheck(cudaDeviceGetAttribute(&max_shared_mem, cudaDevAttrMaxSharedMemoryPerBlockOptin, dev)); + host::RuntimeCheck(max_shared_mem > 0); + + // Set thread config + exec_config_t exec_cfg; + thread_config_t thread_tfg; + if (thread_k != -1 && thread_n != -1) { + thread_tfg = thread_config_t{thread_k, thread_n, default_threads}; + exec_cfg = exec_config_t{1, thread_tfg}; + host::RuntimeCheck(prob_n % thread_n == 0, "prob_n = ", prob_n, " is not divisible by thread_n = ", thread_n); + host::RuntimeCheck(prob_k % thread_k == 0, "prob_k = ", prob_k, " is not divisible by thread_k = ", thread_k); + } else { + // Auto config + exec_cfg = determine_exec_config( + q_type, + prob_m, + prob_n, + prob_k, + thread_m_blocks, + m_block_size_8, + num_bits, + group_size, + has_act_order, + is_k_full, + has_zp, + is_zp_float, + max_shared_mem); + thread_tfg = exec_cfg.tb_cfg; + } + + int num_threads = thread_tfg.num_threads; + thread_k = thread_tfg.thread_k; + thread_n = thread_tfg.thread_n; + int blocks = sms * exec_cfg.blocks_per_sm; + if (exec_cfg.blocks_per_sm > 1) { + max_shared_mem = max_shared_mem / exec_cfg.blocks_per_sm - 1024; + } + + int thread_k_blocks = thread_k / 16; + int thread_n_blocks = thread_n / 16; + + host::RuntimeCheck( + is_valid_config( + thread_tfg, + m_block_size_8, + thread_m_blocks, + prob_m, + prob_n, + prob_k, + num_bits, + group_size, + has_act_order, + is_k_full, + has_zp, + is_zp_float, + max_shared_mem), + "Invalid thread config: thread_m_blocks = ", + thread_m_blocks, + ", thread_k = ", + thread_tfg.thread_k, + ", thread_n = ", + thread_tfg.thread_n, + ", num_threads = ", + thread_tfg.num_threads, + " for MKN = [", + prob_m, + ", ", + prob_k, + ", ", + prob_n, + "] and num_bits = ", + num_bits, + ", group_size = ", + group_size, + ", has_act_order = ", + has_act_order, + ", is_k_full = ", + is_k_full, + ", has_zp = ", + has_zp, + ", is_zp_float = ", + is_zp_float, + ", max_shared_mem = ", + max_shared_mem); + + auto kernel = get_marlin_kernel( + q_type, + thread_m_blocks, + thread_n_blocks, + thread_k_blocks, + m_block_size_8, + has_act_order, + has_zp, + group_blocks, + num_threads, + is_zp_float); + + if (kernel == MarlinDefault) { + host::Panic( + "Unsupported shapes: MNK = [", + prob_m, + ", ", + prob_n, + ", ", + prob_k, + "]", + ", has_act_order = ", + has_act_order, + ", num_groups = ", + num_groups, + ", group_size = ", + group_size, + ", thread_m_blocks = ", + thread_m_blocks, + ", thread_n_blocks = ", + thread_n_blocks, + ", thread_k_blocks = ", + thread_k_blocks, + ", num_bits = ", + num_bits); + } + + host::RuntimeDeviceCheck(cudaFuncSetAttribute(kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, max_shared_mem)); + // clang-format off + kernel<<>>( + A_ptr, B_ptr, C_ptr, C_tmp_ptr, bias_ptr, s_ptr, s2_ptr, zp_ptr, g_idx_ptr, + sorted_token_ids_ptr, expert_ids_ptr, num_tokens_past_padded_ptr, + topk_weights_ptr, top_k, mul_topk_weights, is_ep, num_groups, prob_m, + prob_n, prob_k, locks, has_bias, use_atomic_add, use_fp32_reduce, max_shared_mem); + // clang-format on +} + +#endif + +} // namespace device::marlin_moe diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/nvidia/moe_wna16_marlin_gemm_nvidia.cu b/src/infiniop/ops/moe_wna16_marlin_gemm/nvidia/moe_wna16_marlin_gemm_nvidia.cu new file mode 100644 index 000000000..d64f43c2b --- /dev/null +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/nvidia/moe_wna16_marlin_gemm_nvidia.cu @@ -0,0 +1,401 @@ +#if defined ENABLE_NVIDIA_API + +#include "../../../devices/nvidia/nvidia_handle.cuh" +#include "../../../devices/nvidia/nvidia_kernel_common.cuh" +#include "../moe_wna16_marlin_gemm.h" +#include "kernel.cuh" +#include "moe_wna16_marlin_gemm_nvidia.cuh" + +// #if defined ENABLE_TVM_API +// #include "../marlin/kernel.h" +// #include "../marlin/marlin_template.h" +// #include "../sgl_kernel/scalar_type.hpp" +// #include "../sgl_kernel/tensor.h" +// #endif + +template +infiniStatus_t sglang_moe_wna16_marlin_gemm( + void *c, + const void *a, + const void *b_q_weight, + void *b_bias, + void *b_scales, + void *global_scales, + void *b_zeros, + void *g_idx, + void *perm, + void *sorted_token_ids, + void *expert_ids, + void *num_tokens_post_padded, + void *topk_weights, + int moe_block_size, + int top_k, + bool mul_topk_weights, + bool is_ep, + int64_t b_q_type_id, + int size_m, + int size_n, + int size_k, + bool has_act_order, + bool has_bias, + bool is_k_full, + bool has_zp, + int num_groups, + int group_size, + bool use_atomic_add, + bool use_fp32_reduce, + bool is_zp_float, + int sorted_token_ids_size_0, + int b_q_weight_size_1, + int b_q_weight_size_2, + int b_zeros_size_1, + int b_zeros_size_2, + int c_size_0, + void *total_buffer, + cudaStream_t stream) { + using namespace host; + + ScalarType const b_q_type = ScalarType::from_id(b_q_type_id); + int pack_factor = 32 / b_q_type.size_bits(); + + if (moe_block_size != 8) { + if (moe_block_size % 16 != 0) { + printf("unsupported moe_block_size=%d\n", moe_block_size); + return INFINI_STATUS_BAD_PARAM; + } + if (moe_block_size < 16 || moe_block_size > 64) { + printf("unsupported moe_block_size=%d\n", moe_block_size); + return INFINI_STATUS_BAD_PARAM; + } + } + if (size_k % device::marlin::tile_size != 0 || (size_k / device::marlin::tile_size) != b_q_weight_size_1 || b_q_weight_size_2 % device::marlin::tile_size != 0) { + return INFINI_STATUS_BAD_PARAM; + } + int actual_size_n = (b_q_weight_size_2 / device::marlin::tile_size) * pack_factor; + if (actual_size_n != size_n) { + printf("size_n =%d, actual_size_n = %d\n", size_n, actual_size_n); + return INFINI_STATUS_BAD_TENSOR_SHAPE; + } + if (c_size_0 != size_m * top_k) { + printf("Shape mismatch: c.size(0) = %d, top_k * size_m = %d\n", c_size_0, size_m * top_k); + return INFINI_STATUS_BAD_TENSOR_SHAPE; + } + + // Verify device and strides + + // thread_k, thread_n, sms + int thread_k = -1; + int thread_n = -1; + int sms = -1; + + int device_id = 0; + + RuntimeDeviceCheck(cudaDeviceGetAttribute(&sms, cudaDevAttrMultiProcessorCount, device_id)); + + // Verify global_scale (Optional unwrap done in Python) + // int64_t global_scale_size = global_scale.size(0); + if (global_scales != nullptr) { + RuntimeCheck(b_q_type == kFE2M1f && group_size == 16, "global_scales can only be used for nvfp4 format."); + } else { + RuntimeCheck( + !(b_q_type == kFE2M1f && group_size == 16), "the global_scales parameter must be passed for nvfp4 format."); + } + + // b_zeros Optional unwrap + has_zp derivation: SKIP (done in Python) + + // Verify b_q_type vs has_zp + if (has_zp) { + + RuntimeCheck( + b_q_type == kU4 || b_q_type == kU8, "b_q_type must be u4 or u8 when has_zp = True. Got = ", b_q_type.str()); + } else { + RuntimeCheck( + b_q_type == kU4B8 || b_q_type == kU8B128 || b_q_type == kFE4M3fn || b_q_type == kFE2M1f, + "b_q_type must be uint4b8, uint8b128, float8_e4m3fn or " + "float4_e2m1f when " + "has_zp = False. Got = ", + b_q_type.str()); + } + + if (has_zp && is_zp_float) { + RuntimeCheck( + std::is_same::value, + "Computation type must be float16 (half) when using float zero " + "points."); + } + + if (b_q_type == kFE2M1f) { + RuntimeCheck( + group_size == 16 || group_size == 32, + "float4_e2m1f only supports group_size == 16 (NVFP4) or group_size == 32 (MXFP4). Got group_size = ", + group_size); + RuntimeCheck( + group_size != 32 || std::is_same::value, + "MXFP4 Marlin with E8M0 scales is only instantiated for bfloat16 activations."); + } + + // Verify b_zeros + if (has_zp) { + // b_zeros.ndim = 3 + + if (is_zp_float) { + RuntimeCheck(b_zeros_size_2 == size_n, "b_zeros dim 2 = ", b_zeros_size_2, " is not size_n = ", size_n); + RuntimeCheck( + num_groups == b_zeros_size_1, "b_zeros dim 1 = ", b_zeros_size_1, " is not num_groups = ", num_groups); + RuntimeCheck(num_groups != -1, "num_groups must be != -1"); + } else { + RuntimeCheck( + b_zeros_size_1 == num_groups, "b_zeros dim 1 = ", b_zeros_size_1, " is not num_groups = ", num_groups); + RuntimeCheck( + b_zeros_size_2 == size_n / pack_factor, + "b_zeros dim 2 = ", + b_zeros_size_2, + " is not size_n / pack_factor = ", + size_n / pack_factor); + } + } + + // Alloc C tmp buffer that is going to be used for the global reduce + if (size_n % device::marlin::min_thread_n != 0) { + return INFINI_STATUS_BAD_TENSOR_SHAPE; + } + int c_tmp_bytes = 0; + if (use_fp32_reduce && !use_atomic_add) { + // max num of threadblocks is sms * 4 + int max_c_tmp_size = min( + size_n * sorted_token_ids_size_0, + sms * 4 * moe_block_size * device::marlin::max_thread_n); + if (moe_block_size == 8) { + max_c_tmp_size *= 2; + } + + c_tmp_bytes = max_c_tmp_size * sizeof(float); + } + + int a_tmp_bytes = 0; + if (has_act_order) { + a_tmp_bytes = size_m * top_k * size_k * sizeof(scalar_t); + if (is_k_full) { + RuntimeCheck(num_groups > 1, "For act_order, num_groups must be > 1"); + RuntimeCheck(size_k % num_groups == 0, "size_k = ", size_k, + ", is not divisible by num_groups = ", num_groups); + group_size = size_k / num_groups; + } else { + group_size = 0; + } + } else { + + if (num_groups > 1) { + // RuntimeCheck( + // size_k % num_groups == 0, "size_k = ", size_k, + // ", is not divisible by b_scales.size(1) = ", b_scales_size_1); + group_size = size_k / num_groups; + } else { + group_size = -1; + } + } + + int64_t max_n_tiles = size_n / device::marlin::min_thread_n; + int64_t min_workspace_size = std::min(max_n_tiles * (sorted_token_ids_size_0 / moe_block_size), static_cast(sms) * 4); + const int total_bytes = c_tmp_bytes + a_tmp_bytes + min_workspace_size; + + void *workspace = nullptr; + void *c_tmp = nullptr; + void *a_tmp = nullptr; + uint8_t *ptr = reinterpret_cast(total_buffer); + if (use_fp32_reduce && !use_atomic_add) { + c_tmp = reinterpret_cast(ptr); + ptr += c_tmp_bytes; + } + if (has_act_order && a_tmp_bytes > 0) { + a_tmp = ptr; + ptr += a_tmp_bytes; + } + if (min_workspace_size > 0) { + workspace = ptr; + cudaMemset(workspace, 0, min_workspace_size); + ptr += min_workspace_size; + } + + // Early return for zero-size M (moved after all validation) + if (size_m == 0) { + return INFINI_STATUS_BAD_TENSOR_SHAPE; + } + + device::marlin_moe::marlin_mm( + a, + b_q_weight, + c, + c_tmp, + b_bias, + b_scales, + global_scales, + b_zeros, + g_idx, + perm, + a_tmp, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + topk_weights, + moe_block_size, + top_k, + mul_topk_weights, + is_ep, + size_m, + size_n, + size_k, + workspace, + b_q_type, + has_bias, + has_act_order, + is_k_full, + has_zp, + num_groups, + group_size, + device_id, + stream, + thread_k, + thread_n, + sms, + use_atomic_add, + use_fp32_reduce, + is_zp_float); + return INFINI_STATUS_SUCCESS; +} + +static int getCudaDeviceSMCount() { + int dev; + cudaGetDevice(&dev); + cudaDeviceProp prop; + cudaGetDeviceProperties(&prop, dev); + + return prop.multiProcessorCount; +} + +namespace op::moe_wna16_marlin_gemm::nvidia { + +struct Descriptor::Opaque { + std::shared_ptr internal; +}; + +Descriptor::~Descriptor() { delete _opaque; } + +infiniStatus_t Descriptor::create( + infiniopHandle_t handle_, + Descriptor **desc_ptr, + infiniopTensorDescriptor_t c_desc, + infiniopTensorDescriptor_t a_desc, + infiniopTensorDescriptor_t b_q_weight_desc, + infiniopTensorDescriptor_t b_bias_desc, + infiniopTensorDescriptor_t b_scales_desc, + infiniopTensorDescriptor_t global_scales_desc, + infiniopTensorDescriptor_t b_zeros_desc, + infiniopTensorDescriptor_t g_idx_desc, + infiniopTensorDescriptor_t perm_desc, + infiniopTensorDescriptor_t sorted_token_desc, + infiniopTensorDescriptor_t expert_ids_desc, + infiniopTensorDescriptor_t num_tokens_post_padded_desc, + infiniopTensorDescriptor_t topk_weights_desc, int size_m, int size_n, int size_k, + int top_k, int moe_block_size) { + + auto handle = reinterpret_cast(handle_); + auto result = MoeWna16MarlinGemmInfo::create(c_desc, a_desc, b_q_weight_desc, b_bias_desc, b_scales_desc, global_scales_desc, b_zeros_desc, g_idx_desc, perm_desc, sorted_token_desc, expert_ids_desc, num_tokens_post_padded_desc, topk_weights_desc, size_m, size_n, size_k, top_k, moe_block_size); + int sms = getCudaDeviceSMCount(); + + int sorted_token_size_0 = static_cast(sorted_token_desc->dim(0)); + int max_c_tmp_size = std::min( + size_n * sorted_token_size_0, + sms * 4 * moe_block_size * device::marlin::max_thread_n); + if (moe_block_size == 8) { + max_c_tmp_size *= 2; + } + + size_t c_tmp_bytes = max_c_tmp_size * sizeof(float); + + size_t a_tmp_bytes = size_m * top_k * size_k * infiniSizeOf(a_desc->dtype()); + int max_n_tiles = size_n / device::marlin::min_thread_n; + const size_t workspace_bytes = std::min(max_n_tiles * (sorted_token_size_0 / moe_block_size), sms * 4); + size_t workspace_size = c_tmp_bytes + a_tmp_bytes + workspace_bytes; + + *desc_ptr = new Descriptor( + workspace_size, + new Opaque{handle->internal()}, + result.take(), + handle->device, handle->device_id); + return INFINI_STATUS_SUCCESS; +} + +infiniStatus_t +Descriptor::calculate( + void *workspace, size_t workspace_size, + void *c, + const void *a, + const void *b_q_weight, + void *b_bias, + void *b_scales, + void *global_scales, + void *b_zeros, + void *g_idx, + void *perm, + void *sorted_token_ids, + void *expert_ids, + void *num_tokens_post_padded, + void *topk_weights, + bool mul_topk_weights, + bool is_ep, + int64_t b_q_type_id, + bool is_k_full, + bool use_atomic_add, + bool use_fp32_reduce, + bool is_zp_float, + void *stream_) const { + + cudaStream_t stream = (cudaStream_t)stream_; + int size_m = _info.size_m; + int size_n = _info.size_n; + int size_k = _info.size_k; + int top_k = _info.top_k; + int moe_block_size = _info.moe_block_size; + int num_groups = static_cast(_info.num_groups); + int group_size = 0; + if (g_idx != nullptr && perm != nullptr) { + if (is_k_full) { + group_size = size_k / num_groups; + } else { + group_size = 0; + } + } else { + if (num_groups > 1) { + group_size = size_k / num_groups; + } else { + group_size = -1; + } + } + int sorted_token_ids_size_0 = static_cast(_info.sorted_token_ids_size_0); + int b_q_weight_size_1 = static_cast(_info.b_q_weight_size_1); + int b_q_weight_size_2 = static_cast(_info.b_q_weight_size_2); + int b_zeros_size_1 = static_cast(_info.b_zeros_size_1); + int b_zeros_size_2 = static_cast(_info.b_zeros_size_2); + int c_size_0 = static_cast(_info.c_size_0); + bool has_act_order = _info.has_act_order; + bool has_bias = _info.has_bias; + bool has_zp = _info.has_zp; + +#define MARLIN(TDATA) \ + sglang_moe_wna16_marlin_gemm(c, a, b_q_weight, b_bias, b_scales, global_scales, b_zeros, g_idx, perm, sorted_token_ids, expert_ids, num_tokens_post_padded, topk_weights, moe_block_size, top_k, mul_topk_weights, is_ep, b_q_type_id, size_m, size_n, size_k, has_act_order, has_bias, is_k_full, has_zp, num_groups, group_size, use_atomic_add, use_fp32_reduce, is_zp_float, sorted_token_ids_size_0, b_q_weight_size_1, b_q_weight_size_2, b_zeros_size_1, b_zeros_size_2, c_size_0, workspace, stream) + + if (_info.dtype == INFINI_DTYPE_F16) { + return MARLIN(half); + } else if (_info.dtype == INFINI_DTYPE_BF16) { + return MARLIN(__nv_bfloat16); + } else { + return INFINI_STATUS_BAD_TENSOR_DTYPE; + } + + return INFINI_STATUS_SUCCESS; +} + +} // namespace op::moe_wna16_marlin_gemm::nvidia + +#endif diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/nvidia/moe_wna16_marlin_gemm_nvidia.cuh b/src/infiniop/ops/moe_wna16_marlin_gemm/nvidia/moe_wna16_marlin_gemm_nvidia.cuh new file mode 100644 index 000000000..fa90fe563 --- /dev/null +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/nvidia/moe_wna16_marlin_gemm_nvidia.cuh @@ -0,0 +1,8 @@ +#ifndef __MOE_WNA16_MARLIN_GEMM_CUDA_CUH__ +#define __MOE_WNA16_MARLIN_GEMM_CUDA_CUH__ + +#include "../moe_wna16_marlin_gemm.h" + +DESCRIPTOR(nvidia) + +#endif // __MOE_WNA16_MARLIN_GEMM_CUDA_CUH__ diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/operator.cc b/src/infiniop/ops/moe_wna16_marlin_gemm/operator.cc new file mode 100644 index 000000000..84c913bf2 --- /dev/null +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/operator.cc @@ -0,0 +1,144 @@ +#include "../../operator.h" +#include "../../handle.h" +#include "infiniop/ops/moe_wna16_marlin_gemm.h" + +#if defined ENABLE_NVIDIA_API +#include "nvidia/moe_wna16_marlin_gemm_nvidia.cuh" +#endif + +__INFINI_C infiniStatus_t infiniopCreateMoeWna16MarlinGemmDescriptor( + infiniopHandle_t handle, + infiniopMoeWna16MarlinGemmDescriptor_t *desc_ptr, + infiniopTensorDescriptor_t c_desc, + infiniopTensorDescriptor_t a_desc, + infiniopTensorDescriptor_t b_q_weight_desc, + infiniopTensorDescriptor_t b_bias_desc, + infiniopTensorDescriptor_t b_scales_desc, + infiniopTensorDescriptor_t global_scales_desc, + infiniopTensorDescriptor_t b_zeros_desc, + infiniopTensorDescriptor_t g_idx_desc, + infiniopTensorDescriptor_t perm_desc, + infiniopTensorDescriptor_t sorted_token_desc, + infiniopTensorDescriptor_t expert_ids_desc, + infiniopTensorDescriptor_t num_tokens_post_padded_desc, + infiniopTensorDescriptor_t topk_weights_desc, + int size_m, int size_n, int size_k, + int top_k, int moe_block_size) { +#define CREATE(CASE, NAMESPACE) \ + case CASE: \ + return op::moe_wna16_marlin_gemm::NAMESPACE::Descriptor::create( \ + handle, \ + reinterpret_cast(desc_ptr), \ + c_desc, \ + a_desc, \ + b_q_weight_desc, \ + b_bias_desc, \ + b_scales_desc, \ + global_scales_desc, \ + b_zeros_desc, \ + g_idx_desc, \ + perm_desc, \ + sorted_token_desc, \ + expert_ids_desc, \ + num_tokens_post_padded_desc, \ + topk_weights_desc, \ + size_m, \ + size_n, \ + size_k, \ + top_k, \ + moe_block_size) + + switch (handle->device) { +#ifdef ENABLE_NVIDIA_API + CREATE(INFINI_DEVICE_NVIDIA, nvidia); +#endif + + default: + return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED; + } + +#undef CREATE +} + +__INFINI_C infiniStatus_t infiniopGetMoeWna16MarlinGemmWorkspaceSize(infiniopMoeWna16MarlinGemmDescriptor_t desc, + size_t *size) { +#define GET(CASE, NAMESPACE) \ + case CASE: \ + *size = reinterpret_cast(desc)->workspaceSize(); \ + return INFINI_STATUS_SUCCESS + + switch (desc->device_type) { +#ifdef ENABLE_NVIDIA_API + GET(INFINI_DEVICE_NVIDIA, nvidia); +#endif + + default: + return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED; + } +#undef GET +} + +__INFINI_C infiniStatus_t infiniopMoeWna16MarlinGemm( + infiniopMoeWna16MarlinGemmDescriptor_t desc, + void *workspace, + size_t workspace_size, + void *c, + const void *a, + const void *b_q_weight, + void *b_bias, + void *b_scales, + void *global_scales, + void *b_zeros, + void *g_idx, + void *perm, + void *sorted_token_ids, + void *expert_ids, + void *num_tokens_post_padded, + void *topk_weights, + bool mul_topk_weights, + bool is_ep, + int64_t b_q_type_id, + bool is_k_full, + bool use_atomic_add, + bool use_fp32_reduce, + bool is_zp_float, + void *stream) { + +#define CALCULATE(CASE, NAMESPACE) \ + case CASE: \ + return reinterpret_cast(desc) \ + ->calculate(workspace, workspace_size, c, a, b_q_weight, b_bias, b_scales, global_scales, b_zeros, g_idx, perm, sorted_token_ids, expert_ids, num_tokens_post_padded, topk_weights, mul_topk_weights, is_ep, b_q_type_id, is_k_full, use_atomic_add, use_fp32_reduce, is_zp_float, stream) + + switch (desc->device_type) { +#ifdef ENABLE_NVIDIA_API + CALCULATE(INFINI_DEVICE_NVIDIA, nvidia); +#endif + + default: + return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED; + } + +#undef CALCULATE +} + +__INFINI_C infiniStatus_t +infiniopDestroyMoeWna16MarlinGemmDescriptor(infiniopMoeWna16MarlinGemmDescriptor_t desc) { + +#define DELETE(CASE, NAMESPACE) \ + case CASE: \ + delete reinterpret_cast(desc); \ + return INFINI_STATUS_SUCCESS; + + switch (desc->device_type) { +#ifdef ENABLE_NVIDIA_API + DELETE(INFINI_DEVICE_NVIDIA, nvidia); +#endif + + default: + return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED; + } + +#undef DELETE +} + +// #endif diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/scalar_type.hpp b/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/scalar_type.hpp new file mode 100644 index 000000000..04ce7a537 --- /dev/null +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/scalar_type.hpp @@ -0,0 +1,332 @@ +#pragma once + +#include +#include +#ifndef __CUDACC__ +#include +#endif + +namespace host { + +// +// ScalarType can represent a wide range of floating point and integer types, +// in particular it can be used to represent sub-byte data types (something +// that torch.dtype currently does not support). +// +// The type definitions on the Python side can be found in: vllm/scalar_type.py +// these type definitions should be kept up to date with any Python API changes +// here. +// +class ScalarType { +public: + enum NanRepr : uint8_t { + NAN_NONE = 0, // nans are not supported + NAN_IEEE_754 = 1, // nans are: exp all 1s, mantissa not all 0s + NAN_EXTD_RANGE_MAX_MIN = 2, // nans are: exp all 1s, mantissa all 1s + + NAN_REPR_ID_MAX + }; + + constexpr ScalarType( + uint8_t exponent, + uint8_t mantissa, + bool signed_, + int32_t bias, + bool finite_values_only = false, + NanRepr nan_repr = NAN_IEEE_754) + : exponent(exponent), + mantissa(mantissa), + signed_(signed_), + bias(bias), + finite_values_only(finite_values_only), + nan_repr(nan_repr){}; + + static constexpr ScalarType int_(uint8_t size_bits, int32_t bias = 0) { + return ScalarType(0, size_bits - 1, true, bias); + } + + static constexpr ScalarType uint(uint8_t size_bits, int32_t bias = 0) { + return ScalarType(0, size_bits, false, bias); + } + + // IEEE 754 compliant floating point type + static constexpr ScalarType float_IEEE754(uint8_t exponent, uint8_t mantissa) { + assert(mantissa > 0 && exponent > 0); + return ScalarType(exponent, mantissa, true, 0, false, NAN_IEEE_754); + } + + // IEEE 754 non-compliant floating point type + static constexpr ScalarType float_(uint8_t exponent, uint8_t mantissa, bool finite_values_only, NanRepr nan_repr) { + assert(nan_repr < NAN_REPR_ID_MAX); + assert(mantissa > 0 && exponent > 0); + assert(nan_repr != NAN_IEEE_754); + return ScalarType(exponent, mantissa, true, 0, finite_values_only, nan_repr); + } + + uint8_t const exponent; // size of the exponent field (0 for integer types) + uint8_t const mantissa; // size of the mantissa field (size of the integer + // excluding the sign bit for integer types) + bool const signed_; // flag if the type supports negative numbers (i.e. has a + // sign bit) + int32_t const bias; // stored values equal value + bias, + // used for quantized type + + // Extra Floating point info + bool const finite_values_only; // i.e. no +/-inf if true + NanRepr const nan_repr; // how NaNs are represented + // (not applicable for integer types) + + using Id = int64_t; + +private: + // Field size in id + template + static constexpr size_t member_id_field_width() { + using T = std::decay_t; + return std::is_same_v ? 1 : sizeof(T) * 8; + } + + template + static constexpr auto reduce_members_helper(Fn f, Init val, Member member, Rest... rest) { + auto new_val = f(val, member); + if constexpr (sizeof...(rest) > 0) { + return reduce_members_helper(f, new_val, rest...); + } else { + return new_val; + }; + } + + template + constexpr auto reduce_members(Fn f, Init init) const { + // Should be in constructor order for `from_id` + return reduce_members_helper(f, init, exponent, mantissa, signed_, bias, finite_values_only, nan_repr); + }; + + template + static constexpr auto reduce_member_types(Fn f, Init init) { + constexpr auto dummy_type = ScalarType(0, 0, false, 0, false, NAN_NONE); + return dummy_type.reduce_members(f, init); + }; + + static constexpr auto id_size_bits() { + return reduce_member_types( + [](int acc, auto member) -> int { return acc + member_id_field_width(); }, 0); + } + +public: + // unique id for this scalar type that can be computed at compile time for + // c++17 template specialization this is not needed once we migrate to + // c++20 and can pass literal classes as template parameters + constexpr Id id() const { + static_assert(id_size_bits() <= sizeof(Id) * 8, "ScalarType id is too large to be stored"); + + auto or_and_advance = [](std::pair result, auto member) -> std::pair { + auto [id, bit_offset] = result; + auto constexpr bits = member_id_field_width(); + return {id | (int64_t(member) & ((uint64_t(1) << bits) - 1)) << bit_offset, bit_offset + bits}; + }; + return reduce_members(or_and_advance, std::pair{}).first; + } + + // create a ScalarType from an id, for c++17 template specialization, + // this is not needed once we migrate to c++20 and can pass literal + // classes as template parameters + static constexpr ScalarType from_id(Id id) { + auto extract_and_advance = [id](auto result, auto member) { + using T = decltype(member); + auto [tuple, bit_offset] = result; + auto constexpr bits = member_id_field_width(); + auto extracted_val = static_cast((int64_t(id) >> bit_offset) & ((uint64_t(1) << bits) - 1)); + auto new_tuple = std::tuple_cat(tuple, std::make_tuple(extracted_val)); + return std::pair{new_tuple, bit_offset + bits}; + }; + + auto [tuple_args, _] = reduce_member_types(extract_and_advance, std::pair, int>{}); + return std::apply([](auto... args) { return ScalarType(args...); }, tuple_args); + } + + constexpr int64_t size_bits() const { + return mantissa + exponent + is_signed(); + } + constexpr bool is_signed() const { + return signed_; + } + constexpr bool is_integer() const { + return exponent == 0; + } + constexpr bool is_floating_point() const { + return exponent > 0; + } + constexpr bool is_ieee_754() const { + return is_floating_point() && finite_values_only == false && nan_repr == NAN_IEEE_754; + } + constexpr bool has_nans() const { + return is_floating_point() && nan_repr != NAN_NONE; + } + constexpr bool has_infs() const { + return is_floating_point() && finite_values_only == false; + } + constexpr bool has_bias() const { + return bias != 0; + } + +#ifndef __CUDACC__ +private: + double _floating_point_max() const { + assert(mantissa <= 52 && exponent <= 11); + + uint64_t max_mantissa = (uint64_t(1) << mantissa) - 1; + if (nan_repr == NAN_EXTD_RANGE_MAX_MIN) { + max_mantissa -= 1; + } + + uint64_t max_exponent = (uint64_t(1) << exponent) - 2; + if (nan_repr == NAN_EXTD_RANGE_MAX_MIN || nan_repr == NAN_NONE) { + assert(exponent < 11); + max_exponent += 1; + } + + // adjust the exponent to match that of a double + // for now we assume the exponent bias is the standard 2^(e-1) -1, (where e + // is the exponent bits), there is some precedent for non-standard biases, + // example `float8_e4m3b11fnuz` here: https://github.com/jax-ml/ml_dtypes + // but to avoid premature over complication we are just assuming the + // standard exponent bias until there is a need to support non-standard + // biases + uint64_t exponent_bias = (uint64_t(1) << (exponent - 1)) - 1; + uint64_t exponent_bias_double = (uint64_t(1) << 10) - 1; // double e = 11 + + uint64_t max_exponent_double = max_exponent - exponent_bias + exponent_bias_double; + + // shift the mantissa into the position for a double and + // the exponent + uint64_t double_raw = (max_mantissa << (52 - mantissa)) | (max_exponent_double << 52); + + return *reinterpret_cast(&double_raw); + } + + constexpr std::variant _raw_max() const { + if (is_floating_point()) { + return {_floating_point_max()}; + } else { + assert(size_bits() < 64 || (size_bits() == 64 && is_signed())); + return {(int64_t(1) << mantissa) - 1}; + } + } + + constexpr std::variant _raw_min() const { + if (is_floating_point()) { + assert(is_signed()); + constexpr uint64_t sign_bit_double = (uint64_t(1) << 63); + + double max = _floating_point_max(); + uint64_t max_raw = *reinterpret_cast(&max); + uint64_t min_raw = max_raw | sign_bit_double; + return {*reinterpret_cast(&min_raw)}; + } else { + assert(!is_signed() || size_bits() <= 64); + if (is_signed()) { + // set the top bit to 1 (i.e. INT64_MIN) and the rest to 0 + // then perform an arithmetic shift right to set all the bits above + // (size_bits() - 1) to 1 + return {INT64_MIN >> (64 - size_bits())}; + } else { + return {int64_t(0)}; + } + } + } + +public: + // Max representable value for this scalar type. + // (accounting for bias if there is one) + constexpr std::variant max() const { + return std::visit([this](auto x) -> std::variant { return {x - bias}; }, _raw_max()); + } + + // Min representable value for this scalar type. + // (accounting for bias if there is one) + constexpr std::variant min() const { + return std::visit([this](auto x) -> std::variant { return {x - bias}; }, _raw_min()); + } +#endif // __CUDACC__ + +public: + std::string str() const { + /* naming generally follows: https://github.com/jax-ml/ml_dtypes + * for floating point types (leading f) the scheme is: + * `float_em[flags]` + * flags: + * - no-flags: means it follows IEEE 754 conventions + * - f: means finite values only (no infinities) + * - n: means nans are supported (non-standard encoding) + * for integer types the scheme is: + * `[u]int[b]` + * - if bias is not present it means its zero + */ + if (is_floating_point()) { + auto ret = "float" + std::to_string(size_bits()) + "_e" + std::to_string(exponent) + "m" + std::to_string(mantissa); + if (!is_ieee_754()) { + if (finite_values_only) { + ret += "f"; + } + if (nan_repr != NAN_NONE) { + ret += "n"; + } + } + return ret; + } else { + auto ret = ((is_signed()) ? "int" : "uint") + std::to_string(size_bits()); + if (has_bias()) { + ret += "b" + std::to_string(bias); + } + return ret; + } + } + + constexpr bool operator==(ScalarType const &other) const { + return mantissa == other.mantissa && exponent == other.exponent && bias == other.bias && signed_ == other.signed_ && finite_values_only == other.finite_values_only && nan_repr == other.nan_repr; + } +}; + +using ScalarTypeId = ScalarType::Id; + +// "rust style" names generally following: +// https://github.com/pytorch/pytorch/blob/6d9f74f0af54751311f0dd71f7e5c01a93260ab3/torch/csrc/api/include/torch/types.h#L60-L70 +static inline constexpr auto kS4 = ScalarType::int_(4); +static inline constexpr auto kU4 = ScalarType::uint(4); +static inline constexpr auto kU4B8 = ScalarType::uint(4, 8); +static inline constexpr auto kS8 = ScalarType::int_(8); +static inline constexpr auto kU8 = ScalarType::uint(8); +static inline constexpr auto kU8B128 = ScalarType::uint(8, 128); + +static inline constexpr auto kFE2M1f = ScalarType::float_(2, 1, true, ScalarType::NAN_NONE); +static inline constexpr auto kFE3M2f = ScalarType::float_(3, 2, true, ScalarType::NAN_NONE); +static inline constexpr auto kFE4M3fn = ScalarType::float_(4, 3, true, ScalarType::NAN_EXTD_RANGE_MAX_MIN); +static inline constexpr auto kFE8M0fnu = ScalarType(8, 0, false, 0, true, ScalarType::NAN_EXTD_RANGE_MAX_MIN); +static inline constexpr auto kFE5M2 = ScalarType::float_IEEE754(5, 2); +static inline constexpr auto kFE8M7 = ScalarType::float_IEEE754(8, 7); +static inline constexpr auto kFE5M10 = ScalarType::float_IEEE754(5, 10); + +// Fixed width style names, generally following: +// https://github.com/pytorch/pytorch/blob/6d9f74f0af54751311f0dd71f7e5c01a93260ab3/torch/csrc/api/include/torch/types.h#L47-L57 +static inline constexpr auto kInt4 = kS4; +static inline constexpr auto kUint4 = kU4; +static inline constexpr auto kUint4b8 = kU4B8; +static inline constexpr auto kInt8 = kS8; +static inline constexpr auto kUint8 = kU8; +static inline constexpr auto kUint8b128 = kU8B128; + +static inline constexpr auto kFloat4_e2m1f = kFE2M1f; +static inline constexpr auto kFloat6_e3m2f = kFE3M2f; +static inline constexpr auto kFloat8_e4m3fn = kFE4M3fn; +static inline constexpr auto kFloat8_e5m2 = kFE5M2; +static inline constexpr auto kFloat16_e8m7 = kFE8M7; +static inline constexpr auto kFloat16_e5m10 = kFE5M10; + +// colloquial names +static inline constexpr auto kHalf = kFE5M10; +static inline constexpr auto kFloat16 = kHalf; +static inline constexpr auto kBFloat16 = kFE8M7; + +static inline constexpr auto kFloat16Id = kFloat16.id(); +} // namespace host diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/source_location.h b/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/source_location.h new file mode 100644 index 000000000..9a06fb380 --- /dev/null +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/source_location.h @@ -0,0 +1,40 @@ +/// \file source_location.h +/// \brief Portable `source_location` wrapper. +/// +/// Uses `std::source_location` when available (C++20), otherwise falls +/// back to a minimal stub that returns empty/zero values. + +#pragma once +#include + +/// NOTE: fallback to a minimal source_location implementation +#if defined(__cpp_lib_source_location) +#include + +using source_location_t = std::source_location; + +#else + +struct source_location_fallback { +public: + static constexpr source_location_fallback current() noexcept { + return source_location_fallback{}; + } + constexpr source_location_fallback() noexcept = default; + constexpr unsigned line() const noexcept { + return 0; + } + constexpr unsigned column() const noexcept { + return 0; + } + constexpr const char *file_name() const noexcept { + return ""; + } + constexpr const char *function_name() const noexcept { + return ""; + } +}; + +using source_location_t = source_location_fallback; + +#endif diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/tensor.h b/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/tensor.h new file mode 100644 index 000000000..48c22c3f1 --- /dev/null +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/tensor.h @@ -0,0 +1,547 @@ +/// \file tensor.h +/// \brief Tensor validation and symbolic matching utilities. +#pragma once +#include "utils.h" + +#include +#include +#include + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#ifdef __CUDACC__ +#include "utils.cuh" +#endif + +namespace host { +struct SymbolicSize; +struct SymbolicDType; +struct SymbolicDevice; + +namespace details { +inline constexpr auto kAnyDeviceID = -1; +inline constexpr auto kAnySize = static_cast(-1); +inline constexpr auto kNullSize = static_cast(-1); +inline constexpr auto kNullDType = static_cast(18u); +inline constexpr auto kNullDevice = static_cast(-1); + +template +struct ArrayView { + const T *data; + size_t size; + + __host__ __device__ ArrayView() : data(nullptr), size(0) {} + __host__ __device__ ArrayView(const T *d, size_t s) : data(d), size(s) {} + + template + __host__ __device__ ArrayView(const std::array &arr) + : data(arr.data()), size(arr.size()) {} + + __host__ __device__ const T &operator[](size_t i) const { return data[i]; } + __host__ __device__ bool empty() const { return size == 0; } +}; + +template +struct PrintAbleSpan { + const T *data; + size_t length; + + PrintAbleSpan(const T *p, size_t l) : data(p), length(l) {} + size_t size() const { return length; } + const T &operator[](size_t i) const { return data[i]; } +}; + +inline constexpr const char *kDeviceStringMap[] = { + "", // 0 + "cpu", // 1 + "cuda", // 2 + "cuda_host", // 3 + "opencl", // 4 + "vulkan", // 5 + "metal", // 6 + "vpi", // 7 + "rocm", // 8 + "rocm_host", // 9 + "ext_dev", // 10 + "cuda_managed", // 11 + "oneapi", // 12 + "webgpu", // 13 + "hexagon", // 14 + "maia", // 15 + "trn", // 16 +}; + +constexpr int kMaxDeviceType = 16; + +template +struct _dtype_trait; + +template <> +struct _dtype_trait { + static constexpr DLDataType value = {kDLInt, 8, 1}; +}; +template <> +struct _dtype_trait { + static constexpr DLDataType value = {kDLInt, 16, 1}; +}; +template <> +struct _dtype_trait { + static constexpr DLDataType value = {kDLInt, 32, 1}; +}; +template <> +struct _dtype_trait { + static constexpr DLDataType value = {kDLInt, 64, 1}; +}; +template <> +struct _dtype_trait { + static constexpr DLDataType value = {kDLUInt, 8, 1}; +}; +template <> +struct _dtype_trait { + static constexpr DLDataType value = {kDLUInt, 16, 1}; +}; +template <> +struct _dtype_trait { + static constexpr DLDataType value = {kDLUInt, 32, 1}; +}; +template <> +struct _dtype_trait { + static constexpr DLDataType value = {kDLUInt, 64, 1}; +}; +template <> +struct _dtype_trait { + static constexpr DLDataType value = {kDLFloat, 32, 1}; +}; +template <> +struct _dtype_trait { + static constexpr DLDataType value = {kDLFloat, 64, 1}; +}; + +#ifdef __CUDACC__ +template <> +struct _dtype_trait { + static constexpr DLDataType value = {kDLFloat, 16, 1}; +}; +template <> +struct _dtype_trait { + static constexpr DLDataType value = {kDLBfloat, 16, 1}; +}; +template <> +struct _dtype_trait { + static constexpr DLDataType value = {kDLFloat8_e4m3fn, 8, 1}; +}; +#endif + +template +struct _device_trait { + static constexpr DLDevice value = {Code, kAnyDeviceID}; +}; + +template +inline constexpr std::array kDTypeList = { + _dtype_trait::value...}; + +template +inline constexpr std::array kDeviceList = { + _device_trait::value...}; + +} // namespace details + +inline std::ostream &operator<<(std::ostream &os, DLDevice device) { + int code = static_cast(device.device_type); + if (code < 1 || code > details::kMaxDeviceType) { + RuntimeCheck(false, "Unknown device: ", code); + } + os << details::kDeviceStringMap[code]; + if (device.device_id != details::kAnyDeviceID && device.device_type != kDLCPU) { + os << ":" << device.device_id; + } + return os; +} + +template +inline std::ostream &operator<<(std::ostream &os, const details::PrintAbleSpan &span) { + os << "["; + for (size_t i = 0; i < span.size(); ++i) { + if (i > 0) { + os << ", "; + } + os << span[i]; + } + os << "]"; + return os; +} + +// ============================================== +// SymbolicSize 完整定义 +// ============================================== +struct SymbolicSize { +public: + explicit SymbolicSize(std::string_view ann = {}) + : m_value(details::kNullSize), m_ann(ann) {} + + SymbolicSize(const SymbolicSize &) = delete; + SymbolicSize &operator=(const SymbolicSize &) = delete; + + std::string_view get_name() const { return m_ann; } + bool has_value() const { return m_value != details::kNullSize; } + + void set_value(int64_t v) { + RuntimeCheck(!has_value(), "Size already set"); + m_value = v; + } + + std::optional get_value() const { + return has_value() ? std::optional(m_value) : std::nullopt; + } + + int64_t unwrap(DebugInfo info = {}) const { + RuntimeCheck(info, has_value(), "Size not set"); + return m_value; + } + + void verify(int64_t v, const char *prefix, int64_t dim) { + if (has_value()) { + if (m_value != v) [[unlikely]] { + Panic("Size mismatch for ", m_name_str(prefix, dim), ": expected ", m_value, " got ", v); + } + } else { + set_value(v); + } + } + + std::string value_or_name(const char *prefix, int64_t dim) const { + if (auto v = get_value()) { + return std::to_string(*v); + } + return m_name_str(prefix, dim); + } + +private: + std::string m_name_str(const char *prefix, int64_t dim) const { + std::ostringstream os; + os << prefix << '#' << dim; + if (!m_ann.empty()) { + os << "('" << m_ann << "')"; + } + return os.str(); + } + + int64_t m_value; + std::string_view m_ann; +}; + +inline bool operator==(DLDevice a, DLDevice b) { + return a.device_type == b.device_type && a.device_id == b.device_id; +} + +// ============================================== +// SymbolicDType 完整定义 +// ============================================== +struct SymbolicDType { +public: + SymbolicDType() : m_value({details::kNullDType, 0, 0}) {} + SymbolicDType(const SymbolicDType &) = delete; + SymbolicDType &operator=(const SymbolicDType &) = delete; + + bool has_value() const { return m_value.code != details::kNullDType; } + + void set_value(DLDataType v) { + RuntimeCheck(!has_value(), "DType already set"); + RuntimeCheck(m_check(v), "DType not allowed: ", v); + m_value = v; + } + + std::optional get_value() const { + return has_value() ? std::optional(m_value) : std::nullopt; + } + + DLDataType unwrap(DebugInfo info = {}) const { + RuntimeCheck(info, has_value(), "DType not set"); + return m_value; + } + + void set_options(details::ArrayView opts) { m_opts = opts; } + + template + void set_options() { + m_opts = details::ArrayView(details::kDTypeList.data(), details::kDTypeList.size()); + } + + void verify(DLDataType dtype) { + if (has_value()) { + RuntimeCheck(m_value == dtype, "DType mismatch: expected ", m_value, " got ", dtype); + } else { + set_value(dtype); + } + } + + template + bool is_type() const { + return m_value == details::_dtype_trait::value; + } + +private: + bool m_check(DLDataType v) const { + if (m_opts.empty()) { + return true; + } + for (size_t i = 0; i < m_opts.size; ++i) { + if (m_opts[i] == v) { + return true; + } + } + return false; + } + + details::ArrayView m_opts; + DLDataType m_value; +}; + +// ============================================== +// SymbolicDevice 完整定义 +// ============================================== +struct SymbolicDevice { +public: + SymbolicDevice() : m_value({details::kNullDevice, details::kAnyDeviceID}) {} + SymbolicDevice(const SymbolicDevice &) = delete; + SymbolicDevice &operator=(const SymbolicDevice &) = delete; + + bool has_value() const { return m_value.device_type != details::kNullDevice; } + + void set_value(DLDevice v) { + RuntimeCheck(!has_value(), "Device already set"); + RuntimeCheck(m_check(v), "Device not allowed: ", details::PrintableDevice{v}); + m_value = v; + } + + std::optional get_value() const { + return has_value() ? std::optional(m_value) : std::nullopt; + } + + DLDevice unwrap(DebugInfo info = {}) const { + RuntimeCheck(info, has_value(), "Device not set"); + return m_value; + } + + void set_options(details::ArrayView opts) { m_opts = opts; } + + template + void set_options() { + m_opts = details::ArrayView(details::kDeviceList.data(), details::kDeviceList.size()); + } + + void verify(DLDevice dev) { + if (has_value()) { + RuntimeCheck(m_value == dev, "Device mismatch: expected ", + details::PrintableDevice{m_value}, " got ", details::PrintableDevice{dev}); + } else { + set_value(dev); + } + } + +private: + bool m_check(DLDevice v) const { + if (m_opts.empty()) { + return true; + } + for (size_t i = 0; i < m_opts.size; ++i) { + auto o = m_opts[i]; + if (o.device_type != v.device_type) { + continue; + } + if (o.device_id == details::kAnyDeviceID || o.device_id == v.device_id) { + return true; + } + } + return false; + } + + details::ArrayView m_opts; + DLDevice m_value; +}; + +// ============================================== +// BaseRef / Ref 类型（现在类型已完整定义） +// ============================================== +namespace details { +template +struct BaseRef { + BaseRef() : m_ref(&m_cache) {} + explicit BaseRef(T &r) : m_ref(&r) {} + BaseRef(const BaseRef &) = delete; + BaseRef &operator=(const BaseRef &) = delete; + + T *operator->() const { return m_ref; } + T &operator*() const { return *m_ref; } + void rebind(T &r) { m_ref = &r; } + +private: + T *m_ref; + T m_cache; +}; + +struct SizeRef : public BaseRef { + using BaseRef::BaseRef; + SizeRef(int64_t v); +}; + +struct DTypeRef : public BaseRef { + using BaseRef::BaseRef; + DTypeRef(DLDataType); + DTypeRef(std::initializer_list); + DTypeRef(ArrayView); +}; + +struct DeviceRef : public BaseRef { + using BaseRef::BaseRef; + DeviceRef(DLDevice); + DeviceRef(std::initializer_list); + DeviceRef(ArrayView); +}; + +inline SizeRef::SizeRef(int64_t v) { + if (v != kAnySize) { + (**this).set_value(v); + } +} +inline DTypeRef::DTypeRef(DLDataType v) { (**this).set_value(v); } +inline DTypeRef::DTypeRef(std::initializer_list l) : DTypeRef(ArrayView(l.begin(), l.size())) {} +inline DTypeRef::DTypeRef(ArrayView v) { (**this).set_options(v); } +inline DeviceRef::DeviceRef(DLDevice v) { (**this).set_value(v); } +inline DeviceRef::DeviceRef(std::initializer_list l) : DeviceRef(ArrayView(l.begin(), l.size())) {} +inline DeviceRef::DeviceRef(ArrayView v) { (**this).set_options(v); } + +} // namespace details + +template +inline bool is_type(DLDataType dtype) { + return dtype == details::_dtype_trait::value; +} + +// ============================================== +// TensorMatcher +// ============================================== +struct TensorMatcher { + using SizeRef = details::SizeRef; + using DTypeRef = details::DTypeRef; + using DeviceRef = details::DeviceRef; + + TensorMatcher(const TensorMatcher &) = delete; + TensorMatcher &operator=(const TensorMatcher &) = delete; + + explicit TensorMatcher(std::initializer_list s) + : m_shape(s.begin(), s.size()), m_strides(nullptr, 0) {} + + TensorMatcher &&with_strides(std::initializer_list s) && { + RuntimeCheck(m_strides.empty(), "Strides already set"); + RuntimeCheck(m_shape.size == s.size(), "Stride/shape size mismatch"); + m_strides = details::ArrayView(s.begin(), s.size()); + return std::move(*this); + } + + template + TensorMatcher &&with_dtype(DTypeRef &&d) && { + m_dtype.rebind(*d); + m_dtype->template set_options(); + return std::move(*this); + } + + template + TensorMatcher &&with_dtype() && { + m_dtype->template set_options(); + return std::move(*this); + } + + template + TensorMatcher &&with_device(DeviceRef &&d) && { + m_device.rebind(*d); + m_device->template set_options(); + return std::move(*this); + } + + template + TensorMatcher &&with_device() && { + m_device->template set_options(); + return std::move(*this); + } + + const TensorMatcher &&verify(tvm::ffi::TensorView, DebugInfo = {}) const &&; + +private: + static void s_print_tensor(std::ostringstream &, tvm::ffi::TensorView); + void m_verify_impl(tvm::ffi::TensorView) const; + + details::ArrayView m_shape; + details::ArrayView m_strides; + DTypeRef m_dtype; + DeviceRef m_device; +}; + +inline void TensorMatcher::s_print_tensor(std::ostringstream &os, tvm::ffi::TensorView v) { + os << "Tensor<"; + size_t d = 0; + for (int64_t s : v.shape()) { + if (d++) { + os << ", "; + } + os << s; + } + os << ">[strides=<"; + d = 0; + for (int64_t s : v.strides()) { + if (d++) { + os << ", "; + } + os << s; + } + os << ">, dtype=" << v.dtype(); + os << ", device=" << details::PrintableDevice{v.device()} << "]"; +} + +inline const TensorMatcher &&TensorMatcher::verify(tvm::ffi::TensorView v, DebugInfo info) const && { + try { + m_verify_impl(v); + } catch (PanicError &e) { + std::ostringstream os; + os << "Tensor match failed: "; + s_print_tensor(os, v); + os << " @ " << info.file_name() << ":" << info.line() << "\n- cause: " << e.root_cause(); + throw PanicError(os.str()); + } + return std::move(*this); +} + +inline void TensorMatcher::m_verify_impl(tvm::ffi::TensorView v) const { + size_t dim = static_cast(v.dim()); + RuntimeCheck(dim == m_shape.size, "Dim mismatch: expected ", m_shape.size, " got ", dim); + + for (size_t i = 0; i < dim; ++i) { + m_shape[i]->verify(v.size(i), "shape", (int64_t)i); + } + + if (!m_strides.empty()) { + for (size_t i = 0; i < dim; ++i) { + if (v.size(i) != 1 || !m_strides[i]->has_value()) { + m_strides[i]->verify(v.stride(i), "stride", (int64_t)i); + } + } + } else { + RuntimeCheck(v.is_contiguous(), "Tensor not contiguous"); + } + + m_dtype->verify(v.dtype()); + m_device->verify(v.device()); +} + +} // namespace host diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/utils.cuh b/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/utils.cuh new file mode 100644 index 000000000..18d5da7c3 --- /dev/null +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/utils.cuh @@ -0,0 +1,312 @@ +/// \file utils.cuh +/// \brief Core CUDA/device utilities: type aliases, PDL helpers, +/// typed pointer access, kernel launch wrapper, and error checking. +/// +/// This header is included (directly or transitively) by nearly every +/// JIT kernel. It provides: +/// - Scalar/packed type aliases (`fp16_t`, `bf16_t`, `fp8_e4m3_t`, ...). +/// - `SGL_DEVICE` macro (forced-inline device function qualifier). +/// - `kWarpThreads` constant (32). +/// - PDL (Programmatic Dependent Launch) helpers for Hopper (sm_90+). +/// - Typed `load_as` / `store_as` for void-pointer access. +/// - `pointer::offset` for safe void-pointer arithmetic. +/// - `host::LaunchKernel` - kernel launcher with optional PDL. +/// - `host::RuntimeDeviceCheck` - CUDA error checking. + +#pragma once + +#include "utils.h" + +#include +#include + +#include +#include +#include +#ifndef USE_ROCM +#include +#include +#include +#include +#else +#include +#include +#include +#ifndef __grid_constant__ +#define __grid_constant__ +#endif +using cudaError_t = hipError_t; +using cudaStream_t = hipStream_t; +using cudaLaunchConfig_t = hipLaunchConfig_t; +using cudaLaunchAttribute = hipLaunchAttribute; +inline constexpr auto cudaSuccess = hipSuccess; +#define cudaStreamPerThread hipStreamPerThread +#define cudaGetErrorString hipGetErrorString +#define cudaGetLastError hipGetLastError +#define cudaLaunchKernel hipLaunchKernel +#endif + +#ifndef USE_ROCM +using fp32_t = float; +// using fp16_t = __half; +// using bf16_t = __nv_bfloat16; +using fp8_e4m3_t = __nv_fp8_e4m3; +using fp8_e5m2_t = __nv_fp8_e5m2; + +using fp32x2_t = float2; +using fp16x2_t = __half2; +using bf16x2_t = __nv_bfloat162; +using fp8x2_e4m3_t = __nv_fp8x2_e4m3; +using fp8x2_e5m2_t = __nv_fp8x2_e5m2; + +using fp32x4_t = float4; +#else +using fp32_t = float; +using fp16_t = __half; +using bf16_t = __hip_bfloat16; +using fp8_e4m3_t = uint8_t; +using fp8_e5m2_t = uint8_t; +using fp32x2_t = float2; +using fp16x2_t = half2; +using bf16x2_t = __hip_bfloat162; +using fp8x2_e4m3_t = uint16_t; +using fp8x2_e5m2_t = uint16_t; +using fp32x4_t = float4; +#endif + +/* + * LDG Support + */ +#ifndef USE_ROCM +#define SGLANG_LDG(arg) __ldg(arg) +#else +#define SGLANG_LDG(arg) *(arg) +#endif + +namespace device { + +/// \brief Macro: forced-inline device function qualifier. +#define SGL_DEVICE __forceinline__ __device__ + +// Architecture detection: SGL_CUDA_ARCH is injected by load_jit() and is +// available in both host and device compilation passes, whereas __CUDA_ARCH__ +// is only defined by nvcc during the device pass. +#if !defined(USE_ROCM) +#if !defined(SGL_CUDA_ARCH) +#error "SGL_CUDA_ARCH is not defined. JIT compilation must inject -DSGL_CUDA_ARCH via load_jit()." +#endif +#if defined(__CUDA_ARCH__) +static_assert( + __CUDA_ARCH__ == SGL_CUDA_ARCH, "SGL_CUDA_ARCH mismatch: injected arch flag does not match device target"); +#endif +#define SGL_ARCH_HOPPER_OR_GREATER (SGL_CUDA_ARCH >= 900) +#define SGL_ARCH_BLACKWELL_OR_GREATER ((SGL_CUDA_ARCH >= 1000) && (CUDA_VERSION >= 12090)) +#else // USE_ROCM +#define SGL_ARCH_HOPPER_OR_GREATER 0 +#define SGL_ARCH_BLACKWELL_OR_GREATER 0 +#endif + +// Maximum vector size in bytes supported by current architecture. +// Pre-Blackwell / AMD: 128-bit (16 bytes) +// Blackwell or greater: 256-bit (32 bytes) +inline constexpr std::size_t kMaxVecBytes = SGL_ARCH_BLACKWELL_OR_GREATER ? 32 : 16; + +/// \brief Number of threads per warp (always 32 on NVIDIA/AMD GPUs). +inline constexpr auto kWarpThreads = 32u; +/// \brief Full warp active mask (all 32 lanes). +inline constexpr auto kFullMask = 0xffffffffu; + +/** + * \brief PDL (Programmatic Dependent Launch): wait for the primary kernel. + * + * On Hopper (sm_90+), inserts a `griddepcontrol.wait` instruction to + * synchronize with a preceding kernel in the same stream. On older + * architectures or ROCm this is a no-op. + */ +template +SGL_DEVICE void PDLWaitPrimary() { +#if SGL_ARCH_HOPPER_OR_GREATER + if constexpr (kUsePDL) { + asm volatile("griddepcontrol.wait;" :: + : "memory"); + } +#endif +} + +/** + * \brief PDL: trigger dependent (secondary) kernel launch. + * + * On Hopper (sm_90+), inserts a `griddepcontrol.launch_dependents` + * instruction. On older architectures or ROCm this is a no-op. + */ +template +SGL_DEVICE void PDLTriggerSecondary() { +#if SGL_ARCH_HOPPER_OR_GREATER + if constexpr (kUsePDL) { + asm volatile("griddepcontrol.launch_dependents;" :: + :); + } +#endif +} + +template +SGL_DEVICE constexpr auto div_ceil(T a, U b) { + static_assert(std::is_integral::value && std::is_integral::value, + "div_ceil requires integer types"); + return (a + b - 1) / b; +} + +/** + * \brief Load data with the specified type and offset from a void pointer. + * \tparam T The type to load. + * \param ptr The base pointer. + * \param offset The offset in number of elements of type T. + */ +template +SGL_DEVICE T load_as(const void *ptr, int64_t offset = 0) { + return static_cast(ptr)[offset]; +} + +/** + * \brief Store data with the specified type and offset to a void pointer. + * \tparam T The type to store. + * \param ptr The base pointer. + * \param val The value to store. + * \param offset The offset in number of elements of type T. + * \note we use type_identity_t to force the caller to explicitly specify + * the template parameter `T`, which can avoid accidentally using the wrong type. + */ +template +SGL_DEVICE void store_as(void *ptr, T val, int64_t offset = 0) { + static_cast(ptr)[offset] = val; +} + +/// \brief Safe void-pointer arithmetic (byte-level by default). +namespace pointer { +// we only allow void * pointer arithmetic for safety + +template +SGL_DEVICE auto offset(void *ptr, U... offset) -> void * { + return static_cast(ptr) + (offset + ...); +} + +template +SGL_DEVICE auto offset(const void *ptr, U... offset) -> const void * { + return static_cast(ptr) + (offset + ...); +} + +} // namespace pointer + +} // namespace device + +namespace host { + +/** + * \brief Check the CUDA error code and panic with location info on failure. + */ +inline void RuntimeDeviceCheck(::cudaError_t error, DebugInfo location = {}) { + if (error != ::cudaSuccess) { + [[unlikely]]; + ::host::panic(location, "CUDA error: ", ::cudaGetErrorString(error)); + } +} + +/// \brief Check the last CUDA error (calls `cudaGetLastError`). +inline void RuntimeDeviceCheck(DebugInfo location = {}) { + return RuntimeDeviceCheck(::cudaGetLastError(), location); +} + +/** + * \brief Kernel launcher with automatic stream resolution and PDL support. + * + * Usage: + * \code + * host::LaunchKernel(grid, block, device) + * .enable_pdl(true) + * (my_kernel, arg1, arg2); + * \endcode + * + * The constructor resolves the CUDA stream from a `DLDevice` (via + * `TVMFFIEnvGetStream`) or accepts a raw `cudaStream_t`. The call + * operator launches the kernel and checks for errors. + */ +struct LaunchKernel { +public: + explicit LaunchKernel( + dim3 grid_dim, + dim3 block_dim, + DLDevice device, + std::size_t dynamic_shared_mem_bytes = 0, + DebugInfo location = {}) noexcept + : m_config(s_make_config(grid_dim, block_dim, resolve_device(device), dynamic_shared_mem_bytes)), + m_location(location) {} + + explicit LaunchKernel( + dim3 grid_dim, + dim3 block_dim, + cudaStream_t stream, + std::size_t dynamic_shared_mem_bytes = 0, + DebugInfo location = {}) noexcept + : m_config(s_make_config(grid_dim, block_dim, stream, dynamic_shared_mem_bytes)), m_location(location) {} + + LaunchKernel(const LaunchKernel &) = delete; + LaunchKernel &operator=(const LaunchKernel &) = delete; + + static auto resolve_device(DLDevice device) -> cudaStream_t { + return static_cast(::TVMFFIEnvGetStream(device.device_type, device.device_id)); + } + + auto enable_pdl(bool enabled = true) -> LaunchKernel & { +#ifdef USE_ROCM + (void)enabled; + m_config.numAttrs = 0; +#else + if (enabled) { + m_attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization; + m_attrs[0].val.programmaticStreamSerializationAllowed = true; + m_config.numAttrs = 1; + m_config.attrs = m_attrs; + } else { + m_config.numAttrs = 0; + } +#endif + return *this; + } + + template + auto operator()(T &&kernel, Args &&...args) const -> void { +#ifdef USE_ROCM + hipLaunchKernelGGL( + std::forward(kernel), + m_config.gridDim, + m_config.blockDim, + m_config.dynamicSmemBytes, + m_config.stream, + std::forward(args)...); + RuntimeDeviceCheck(m_location); +#else + RuntimeDeviceCheck(::cudaLaunchKernelEx(&m_config, kernel, std::forward(args)...), m_location); +#endif + } + +private: + static auto s_make_config( // Make a config for kernel launch + dim3 grid_dim, + dim3 block_dim, + cudaStream_t stream, + std::size_t smem) -> cudaLaunchConfig_t { + auto config = ::cudaLaunchConfig_t{}; + config.gridDim = grid_dim; + config.blockDim = block_dim; + config.dynamicSmemBytes = smem; + config.stream = stream; + config.numAttrs = 0; + return config; + } + + cudaLaunchConfig_t m_config; + const DebugInfo m_location; + cudaLaunchAttribute m_attrs[1]; +}; + +} // namespace host diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/utils.h b/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/utils.h new file mode 100644 index 000000000..855b32c36 --- /dev/null +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/utils.h @@ -0,0 +1,245 @@ +/// \file utils.h +/// \brief Host-side C++ utilities used by JIT kernel wrappers. +/// +/// Provides: +/// - `DebugInfo` - wraps `std::source_location` for error reporting. +/// - `RuntimeCheck` - runtime assertion with formatted error messages. +/// - `Panic` - unconditional abort with formatted error messages. +/// - `pointer::offset` - safe void-pointer arithmetic (host side). +/// - `div_ceil` - integer ceiling division. +/// - `dtype_bytes` - byte width of a `DLDataType`. +/// - `irange` - Python-style integer range for range-for loops. +/// - `PrintableDevice` - wrapper for outputting device info to streams (NVIDIA only). + +#pragma once + +// ref: https://forums.developer.nvidia.com/t/c-20s-source-location-compilation-error-when-using-nvcc-12-1/258026/3 +#ifdef __CUDACC__ +#include +#if CUDA_VERSION <= 12010 + +#pragma push_macro("__cpp_consteval") +#pragma push_macro("_NODISCARD") +#pragma push_macro("__builtin_LINE") + +#pragma clang diagnostic push +#pragma clang diagnostic ignored "-Wbuiltin-macro-redefined" +#define __cpp_consteval 201811L +#pragma clang diagnostic pop + +#ifdef _NODISCARD +#undef _NODISCARD +#define _NODISCARD +#endif + +#define consteval constexpr + +#include "source_location.h" + +#undef consteval +#pragma pop_macro("__cpp_consteval") +#pragma pop_macro("_NODISCARD") +#else // __CUDACC__ && CUDA_VERSION > 12010 +#include "source_location.h" +#endif +#else // no __CUDACC__ +#include "source_location.h" +#endif + +#include +#include +#include +#include +#include +#include +#include +#include + +namespace host { + +template +inline constexpr bool dependent_false_v = false; + +#if defined(ENABLE_NVIDIA_API) + +namespace details { + +/// \brief PrintableDevice wrapper for outputting device info to streams. +struct PrintableDevice { + DLDevice device; +}; + +/// \brief Stream output operator for PrintableDevice. +inline std::ostream &operator<<(std::ostream &os, PrintableDevice pd) { + int code = static_cast(pd.device.device_type); + static const char *device_names[] = { + "", "cpu", "cuda", "cuda_host", "opencl", "vulkan", + "metal", "vpi", "rocm", "rocm_host", "ext_dev", + "cuda_managed", "oneapi", "webgpu", "hexagon", "maia", "trn"}; + if (code >= 0 && code <= 16) { + os << device_names[code]; + if (pd.device.device_id >= 0 && code != 1 && code != 0) { + os << ":" << pd.device.device_id; + } + } else { + os << "device(" << code << "," << pd.device.device_id << ")"; + } + return os; +} + +} // namespace details + +#endif // ENABLE_NVIDIA_API + +/// \brief Source-location wrapper for debug/error messages. +struct DebugInfo : public source_location_t { + DebugInfo(source_location_t loc = source_location_t::current()) : source_location_t(loc) {} +}; + +/// \brief Exception type thrown by `RuntimeCheck` and `Panic`. +struct PanicError : public std::runtime_error { +public: + explicit PanicError(std::string msg) : runtime_error(msg), m_message(std::move(msg)) {} + auto root_cause() const -> std::string_view { + const auto str = std::string_view{m_message}; + const auto pos = str.find(": "); + return pos == std::string_view::npos ? str : str.substr(pos + 2); + } + +private: + std::string m_message; +}; + +/// \brief Unconditionally abort with a formatted error message. +template +[[noreturn]] inline auto panic(DebugInfo location, Args &&...args) -> void { + std::ostringstream os; + os << "Runtime check failed at " << location.file_name() << ":" << location.line(); + if constexpr (sizeof...(args) > 0) { + os << ": "; + (os << ... << std::forward(args)); + } else { + os << " in " << location.function_name(); + } + throw PanicError(std::move(os).str()); +} + +/** + * \brief Runtime assertion: panics with a formatted message when `condition` + * is false. Extra `args` are streamed to the error message. + * + * Example: + * \code + * RuntimeCheck(n > 0, "n must be positive, got ", n); + * \endcode + */ +template +struct RuntimeCheck { + template + explicit RuntimeCheck(Cond &&condition, Args &&...args, DebugInfo location = {}) { + if (condition) { + return; + } + [[unlikely]] ::host::panic(location, std::forward(args)...); + } + template + explicit RuntimeCheck(DebugInfo location, Cond &&condition, Args &&...args) { + if (condition) { + return; + } + [[unlikely]] ::host::panic(location, std::forward(args)...); + } +}; + +template +struct Panic { + explicit Panic(Args &&...args, DebugInfo location = {}) { + ::host::panic(location, std::forward(args)...); + } + explicit Panic(DebugInfo location, Args &&...args) { + ::host::panic(location, std::forward(args)...); + } + [[noreturn]] ~Panic() { + std::terminate(); + } +}; + +template +explicit RuntimeCheck(Cond &&, Args &&...) -> RuntimeCheck; + +template +explicit RuntimeCheck(DebugInfo, Cond &&, Args &&...) -> RuntimeCheck; + +template +explicit Panic(Args &&...) -> Panic; + +template +explicit Panic(DebugInfo, Args &&...) -> Panic; + +namespace pointer { + +// we only allow void * pointer arithmetic for safety + +template ::value && ...)>> +inline auto offset(void *ptr, U... offset) -> void * { + return static_cast(ptr) + (... + offset); +} + +template ::value && ...)>> +inline auto offset(const void *ptr, U... offset) -> const void * { + return static_cast(ptr) + (... + offset); +} + +} // namespace pointer + +/// \brief Integer ceiling division: ceil(a / b). +template +inline constexpr auto div_ceil(T a, U b) { + static_assert(std::is_integral::value, "T must be integral"); + static_assert(std::is_integral::value, "U must be integral"); + return (a + b - 1) / b; +} + +/// \brief Returns the byte width of a DLPack data type. +inline auto dtype_bytes(DLDataType dtype) -> std::size_t { + return static_cast(dtype.bits / 8); +} + +// Pure C++11 compatible irange - removes std::ranges/std::integral for older CUDA compilers +template +struct IntegerRange { + T start_; + T end_; + + struct Iterator { + T value; + + T operator*() const { return value; } + Iterator &operator++() { + ++value; + return *this; + } + bool operator!=(const Iterator &other) const { + return value != other.value; + } + }; + + Iterator begin() const { return {start_}; } + Iterator end() const { return {end_}; } +}; + +/// Python-style integer range: irange(n) -> [0, n) +template +IntegerRange irange(T end) { + return {0, end}; +} + +/// Python-style integer range: irange(start, end) -> [start, end) +template +IntegerRange irange(T start, T end) { + return {start, end}; +} + +} // namespace host diff --git a/test/infiniop/libinfiniop/op_register.py b/test/infiniop/libinfiniop/op_register.py index 3ec20a72e..4da1d4ae3 100644 --- a/test/infiniop/libinfiniop/op_register.py +++ b/test/infiniop/libinfiniop/op_register.py @@ -1359,6 +1359,69 @@ def gptq_marlin_gemm_(lib): ] +@OpRegister.operator +def moe_wna16_marlin_gemm_(lib): + lib.infiniopCreateMoeWna16MarlinGemmDescriptor.restype = c_int32 + lib.infiniopCreateMoeWna16MarlinGemmDescriptor.argtypes = [ + infiniopHandle_t, + POINTER(infiniopOperatorDescriptor_t), + infiniopTensorDescriptor_t, + infiniopTensorDescriptor_t, + infiniopTensorDescriptor_t, + infiniopTensorDescriptor_t, + infiniopTensorDescriptor_t, + infiniopTensorDescriptor_t, + infiniopTensorDescriptor_t, + infiniopTensorDescriptor_t, + infiniopTensorDescriptor_t, + infiniopTensorDescriptor_t, + infiniopTensorDescriptor_t, + infiniopTensorDescriptor_t, + infiniopTensorDescriptor_t, + c_int32, + c_int32, + c_int32, + c_int32, + c_int32, + ] + lib.infiniopGetMoeWna16MarlinGemmWorkspaceSize.restype = c_int32 + lib.infiniopGetMoeWna16MarlinGemmWorkspaceSize.argtypes = [ + infiniopOperatorDescriptor_t, + POINTER(c_size_t), + ] + lib.infiniopMoeWna16MarlinGemm.restype = c_int32 + lib.infiniopMoeWna16MarlinGemm.argtypes = [ + infiniopOperatorDescriptor_t, + c_void_p, + c_size_t, + c_void_p, + c_void_p, + c_void_p, + c_void_p, + c_void_p, + c_void_p, + c_void_p, + c_void_p, + c_void_p, + c_void_p, + c_void_p, + c_void_p, + c_void_p, + c_bool, + c_bool, + c_int64, + c_bool, + c_bool, + c_bool, + c_bool, + c_void_p, + ] + lib.infiniopDestroyMoeWna16MarlinGemmDescriptor.restype = c_int32 + lib.infiniopDestroyMoeWna16MarlinGemmDescriptor.argtypes = [ + infiniopOperatorDescriptor_t, + ] + + @OpRegister.operator def gptq_qyblas_gemm_(lib): lib.infiniopCreateGptqQyblasGemmDescriptor.restype = c_int32 From 72bf07f6b3712f36b4c123bd0f1ca19c188d391c Mon Sep 17 00:00:00 2001 From: xgqdut2016 Date: Wed, 24 Jun 2026 09:49:01 +0800 Subject: [PATCH 2/4] issue/1319: modified format --- include/infiniop/ops/moe_wna16_marlin_gemm.h | 2 +- .../marlin/marlin_template.h | 20 ++++++++++++++----- .../moe_wna16_marlin_gemm/nvidia/kernel.cuh | 2 +- .../ops/moe_wna16_marlin_gemm/operator.cc | 2 +- 4 files changed, 18 insertions(+), 8 deletions(-) diff --git a/include/infiniop/ops/moe_wna16_marlin_gemm.h b/include/infiniop/ops/moe_wna16_marlin_gemm.h index eaf487bc0..eeabd9c4a 100644 --- a/include/infiniop/ops/moe_wna16_marlin_gemm.h +++ b/include/infiniop/ops/moe_wna16_marlin_gemm.h @@ -20,7 +20,7 @@ __INFINI_C __export infiniStatus_t infiniopCreateMoeWna16MarlinGemmDescriptor(in infiniopTensorDescriptor_t sorted_token_desc, infiniopTensorDescriptor_t expert_ids_desc, infiniopTensorDescriptor_t num_tokens_post_padded_desc, - infiniopTensorDescriptor_t topk_weights_desc, + infiniopTensorDescriptor_t topk_weights_desc, int size_m, int size_n, int size_k, int top_k, int moe_block_size); ; diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/marlin/marlin_template.h b/src/infiniop/ops/moe_wna16_marlin_gemm/marlin/marlin_template.h index 9fba3b18f..52dcdd969 100644 --- a/src/infiniop/ops/moe_wna16_marlin_gemm/marlin/marlin_template.h +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/marlin/marlin_template.h @@ -166,9 +166,13 @@ __device__ inline void ldsm(typename ScalarType::FragA &frag_a, const : "=r"(a[0]), "=r"(a[1]), "=r"(a[2]), "=r"(a[3]) : "r"(smem)); } else if constexpr (count == 2) { - asm volatile("ldmatrix.sync.aligned.m8n8.x2.shared.b16 {%0,%1}, [%2];\n" : "=r"(a[0]), "=r"(a[1]) : "r"(smem)); + asm volatile("ldmatrix.sync.aligned.m8n8.x2.shared.b16 {%0,%1}, [%2];\n" + : "=r"(a[0]), "=r"(a[1]) + : "r"(smem)); } else if constexpr (count == 1) { - asm volatile("ldmatrix.sync.aligned.m8n8.x1.shared.b16 {%0}, [%1];\n" : "=r"(a[0]) : "r"(smem)); + asm volatile("ldmatrix.sync.aligned.m8n8.x1.shared.b16 {%0}, [%1];\n" + : "=r"(a[0]) + : "r"(smem)); } else { static_assert(count == 1 || count == 2 || count == 4, "invalid count"); } @@ -240,7 +244,9 @@ __device__ inline void barrier_acquire(int *lock, int count) { do { // Guarantee that subsequent writes by this threadblock will be visible // globally. - asm volatile("ld.global.acquire.gpu.b32 %0, [%1];\n" : "=r"(state) : "l"(lock)); + asm volatile("ld.global.acquire.gpu.b32 %0, [%1];\n" + : "=r"(state) + : "l"(lock)); } while (state != count); } __syncthreads(); @@ -258,7 +264,9 @@ __device__ inline void barrier_release(int *lock, bool reset = false) { // Make sure that all writes since acquiring this barrier are visible // globally, while releasing the barrier. asm volatile("fence.acq_rel.gpu;\n"); - asm volatile("red.relaxed.gpu.global.add.s32 [%0], %1;\n" : : "l"(lock), "r"(val)); + asm volatile("red.relaxed.gpu.global.add.s32 [%0], %1;\n" + : + : "l"(lock), "r"(val)); } } @@ -269,7 +277,9 @@ __device__ inline void wait_negative_and_add(int *lock) { do { // Guarantee that subsequent writes by this threadblock will be visible // globally. - asm volatile("ld.global.acquire.gpu.b32 %0, [%1];\n" : "=r"(state) : "l"(lock)); + asm volatile("ld.global.acquire.gpu.b32 %0, [%1];\n" + : "=r"(state) + : "l"(lock)); } while (state >= 0); atomicAdd(lock, 1); } diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/nvidia/kernel.cuh b/src/infiniop/ops/moe_wna16_marlin_gemm/nvidia/kernel.cuh index ee0502a94..6035b83a8 100644 --- a/src/infiniop/ops/moe_wna16_marlin_gemm/nvidia/kernel.cuh +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/nvidia/kernel.cuh @@ -46,7 +46,7 @@ __global__ void permute_cols_kernel( const int32_t *__restrict__ num_tokens_past_padded_ptr, int size_m, int size_k, - int top_k) {}; + int top_k){}; #else diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/operator.cc b/src/infiniop/ops/moe_wna16_marlin_gemm/operator.cc index 84c913bf2..276ca1cf1 100644 --- a/src/infiniop/ops/moe_wna16_marlin_gemm/operator.cc +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/operator.cc @@ -21,7 +21,7 @@ __INFINI_C infiniStatus_t infiniopCreateMoeWna16MarlinGemmDescriptor( infiniopTensorDescriptor_t sorted_token_desc, infiniopTensorDescriptor_t expert_ids_desc, infiniopTensorDescriptor_t num_tokens_post_padded_desc, - infiniopTensorDescriptor_t topk_weights_desc, + infiniopTensorDescriptor_t topk_weights_desc, int size_m, int size_n, int size_k, int top_k, int moe_block_size) { #define CREATE(CASE, NAMESPACE) \ From fbe87069754d02522ae126a5e9462e0227a5528a Mon Sep 17 00:00:00 2001 From: xgqdut2016 Date: Wed, 24 Jun 2026 10:26:10 +0800 Subject: [PATCH 3/4] issue/1319: modified moe_wna16_marlin_gemm_nvidia.cu --- .../nvidia/moe_wna16_marlin_gemm_nvidia.cu | 9 +-------- 1 file changed, 1 insertion(+), 8 deletions(-) diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/nvidia/moe_wna16_marlin_gemm_nvidia.cu b/src/infiniop/ops/moe_wna16_marlin_gemm/nvidia/moe_wna16_marlin_gemm_nvidia.cu index d64f43c2b..6c79dec97 100644 --- a/src/infiniop/ops/moe_wna16_marlin_gemm/nvidia/moe_wna16_marlin_gemm_nvidia.cu +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/nvidia/moe_wna16_marlin_gemm_nvidia.cu @@ -1,4 +1,4 @@ -#if defined ENABLE_NVIDIA_API +#if defined(ENABLE_NVIDIA_API) && defined(ENABLE_TVM_API) #include "../../../devices/nvidia/nvidia_handle.cuh" #include "../../../devices/nvidia/nvidia_kernel_common.cuh" @@ -6,13 +6,6 @@ #include "kernel.cuh" #include "moe_wna16_marlin_gemm_nvidia.cuh" -// #if defined ENABLE_TVM_API -// #include "../marlin/kernel.h" -// #include "../marlin/marlin_template.h" -// #include "../sgl_kernel/scalar_type.hpp" -// #include "../sgl_kernel/tensor.h" -// #endif - template infiniStatus_t sglang_moe_wna16_marlin_gemm( void *c, From 7f0b8cae99f646bd096d1a2f09f93aa58f03b8c3 Mon Sep 17 00:00:00 2001 From: xgqdut2016 Date: Mon, 29 Jun 2026 16:48:38 +0800 Subject: [PATCH 4/4] issue/1319: modified sgl_kernel --- .../moe_wna16_marlin_gemm/sgl_kernel/tensor.h | 8 ++++ .../moe_wna16_marlin_gemm/sgl_kernel/utils.h | 37 ++----------------- 2 files changed, 12 insertions(+), 33 deletions(-) diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/tensor.h b/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/tensor.h index 48c22c3f1..f30492621 100644 --- a/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/tensor.h +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/tensor.h @@ -83,6 +83,10 @@ inline constexpr const char *kDeviceStringMap[] = { constexpr int kMaxDeviceType = 16; +struct PrintableDevice { + DLDevice device; +}; + template struct _dtype_trait; @@ -169,6 +173,10 @@ inline std::ostream &operator<<(std::ostream &os, DLDevice device) { return os; } +inline std::ostream &operator<<(std::ostream &os, details::PrintableDevice pd) { + return os << pd.device; +} + template inline std::ostream &operator<<(std::ostream &os, const details::PrintAbleSpan &span) { os << "["; diff --git a/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/utils.h b/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/utils.h index 855b32c36..d6892d0dd 100644 --- a/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/utils.h +++ b/src/infiniop/ops/moe_wna16_marlin_gemm/sgl_kernel/utils.h @@ -9,7 +9,6 @@ /// - `div_ceil` - integer ceiling division. /// - `dtype_bytes` - byte width of a `DLDataType`. /// - `irange` - Python-style integer range for range-for loops. -/// - `PrintableDevice` - wrapper for outputting device info to streams (NVIDIA only). #pragma once @@ -60,37 +59,6 @@ namespace host { template inline constexpr bool dependent_false_v = false; -#if defined(ENABLE_NVIDIA_API) - -namespace details { - -/// \brief PrintableDevice wrapper for outputting device info to streams. -struct PrintableDevice { - DLDevice device; -}; - -/// \brief Stream output operator for PrintableDevice. -inline std::ostream &operator<<(std::ostream &os, PrintableDevice pd) { - int code = static_cast(pd.device.device_type); - static const char *device_names[] = { - "", "cpu", "cuda", "cuda_host", "opencl", "vulkan", - "metal", "vpi", "rocm", "rocm_host", "ext_dev", - "cuda_managed", "oneapi", "webgpu", "hexagon", "maia", "trn"}; - if (code >= 0 && code <= 16) { - os << device_names[code]; - if (pd.device.device_id >= 0 && code != 1 && code != 0) { - os << ":" << pd.device.device_id; - } - } else { - os << "device(" << code << "," << pd.device.device_id << ")"; - } - return os; -} - -} // namespace details - -#endif // ENABLE_NVIDIA_API - /// \brief Source-location wrapper for debug/error messages. struct DebugInfo : public source_location_t { DebugInfo(source_location_t loc = source_location_t::current()) : source_location_t(loc) {} @@ -207,7 +175,9 @@ inline auto dtype_bytes(DLDataType dtype) -> std::size_t { return static_cast(dtype.bits / 8); } -// Pure C++11 compatible irange - removes std::ranges/std::integral for older CUDA compilers +// ====================== 修复开始：纯 C++11 兼容版 irange ====================== +// 移除所有 std::ranges / std::integral，完全兼容旧版 CUDA 编译器 + template struct IntegerRange { T start_; @@ -241,5 +211,6 @@ template IntegerRange irange(T start, T end) { return {start, end}; } +// ====================== 修复结束 ====================== } // namespace host