Skip to content

GitLab

"tests/vscode:/vscode.git/clone" did not exist on "711e912946d23f4ccc1f554b1524c960553c5e28"
Unverified Commit a17cef70 authored by Michael Goin's avatar Michael Goin Committed by GitHub
Browse files

Removed unused marlin cuda code (#17684) 


Signed-off-by: default avatarmgoin <mgoin64@gmail.com>
parent 18dd5e01
Show whitespace changes
Inline Side-by-side
Showing with 0 additions and 3666 deletions
csrc/moe/marlin_kernels/marlin_moe_kernel.h deleted 100644 → 0
View file @ 18dd5e01
#pragma once
#include <torch/all.h>
#include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <cuda.h>
#include <cuda_fp16.h>
#include <cuda_runtime.h>
#include <iostream>
#include "core/scalar_type.hpp"
namespace marlin_moe {
constexpr int ceildiv(int a, int b) { return (a + b - 1) / b; }
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
// Instances of `Vec` are used to organize groups of >>registers<<, as needed
// for instance as inputs to tensor core operations. Consequently, all
// corresponding index accesses must be compile-time constants, which is why we
// extensively use `#pragma unroll` throughout the kernel code to guarantee
// this.
template <typename T, int n>
struct Vec {
T elems[n];
__device__ T& operator[](int i) { return elems[i]; }
};
using I4 = Vec<int, 4>;
// 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<half2, 4>;
using FragB = Vec<half2, 2>;
using FragC = Vec<float, 4>;
using FragS = Vec<half2, 1>; // quantization scales
using FragZP = Vec<half2, 4>;
// Predicated asynchronous global->shared copy; used for inputs A where we apply
// predication to handle batchsizes that are not multiples of 16.
__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<uint32_t>(__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));
}
// Asynchronous global->shared copy
__device__ inline void cp_async4(void* smem_ptr, const void* glob_ptr) {
const int BYTES = 16;
uint32_t smem = static_cast<uint32_t>(__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));
}
// Async copy fence.
__device__ inline void cp_async_fence() {
asm volatile("cp.async.commit_group;\n" ::);
}
// Wait until at most `n` async copy stages are still pending.
template <int n>
__device__ inline void cp_async_wait() {
asm volatile("cp.async.wait_group %0;\n" ::"n"(n));
}
// m16n8k16 tensor core mma instruction with fp16 inputs and fp32
// output/accumulation.
__device__ inline void mma(const FragA& a_frag, const FragB& frag_b,
FragC& frag_c) {
const uint32_t* a = reinterpret_cast<const uint32_t*>(&a_frag);
const uint32_t* b = reinterpret_cast<const uint32_t*>(&frag_b);
float* c = reinterpret_cast<float*>(&frag_c);
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]));
}
// Instruction for loading a full 16x16 matrix fragment of operand A from shared
// memory, directly in tensor core layout.
__device__ inline void ldsm4(FragA& frag_a, const void* smem_ptr) {
uint32_t* a = reinterpret_cast<uint32_t*>(&frag_a);
uint32_t smem = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
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));
}
// 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 <int lut>
__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 <int start_byte, int mask>
__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 <vllm::ScalarTypeId w_type_id>
__device__ inline FragB dequant(int q);
// 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:
// https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/cutlass_extensions/include/cutlass_extensions/interleaved_numeric_conversion.h#L215-L287
template <>
__device__ inline FragB dequant<vllm::kU4B8.id()>(int q) {
const int LO = 0x000f000f;
const int HI = 0x00f000f0;
const int EX = 0x64006400;
// Guarantee that the `(a & b) | c` operations are LOP3s.
int lo = lop3<(0xf0 & 0xcc) | 0xaa>(q, LO, EX);
int hi = lop3<(0xf0 & 0xcc) | 0xaa>(q, HI, EX);
// 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;
FragB frag_b;
frag_b[0] = __hsub2(*reinterpret_cast<half2*>(&lo),
*reinterpret_cast<const half2*>(&SUB));
frag_b[1] = __hfma2(*reinterpret_cast<half2*>(&hi),
*reinterpret_cast<const half2*>(&MUL),
*reinterpret_cast<const half2*>(&ADD));
return frag_b;
}
// Fast Int8ToFp16: Efficiently dequantize 8bit int values to fp16
// Reference:
// https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/cutlass_extensions/include/cutlass_extensions/interleaved_numeric_conversion.h#L53-L85
template <>
__device__ inline FragB dequant<vllm::kU8B128.id()>(int q) {
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<start_byte_for_fp16, mask_for_elt_01>(q);
uint32_t hi = prmt<start_byte_for_fp16, mask_for_elt_23>(q);
static constexpr uint32_t I8s_TO_F16s_MAGIC_NUM = 0x64806480;
FragB frag_b;
frag_b[0] = __hsub2(*reinterpret_cast<half2*>(&lo),
*reinterpret_cast<const half2*>(&I8s_TO_F16s_MAGIC_NUM));
frag_b[1] = __hsub2(*reinterpret_cast<half2*>(&hi),
*reinterpret_cast<const half2*>(&I8s_TO_F16s_MAGIC_NUM));
return frag_b;
}
template <>
__device__ inline FragB dequant<vllm::kU4.id()>(int q) {
const int LO = 0x000f000f;
const int HI = 0x00f000f0;
const int EX = 0x64006400;
// Guarantee that the `(a & b) | c` operations are LOP3s.
int lo = lop3<(0xf0 & 0xcc) | 0xaa>(q, LO, EX);
int hi = lop3<(0xf0 & 0xcc) | 0xaa>(q, HI, EX);
const int SUB = 0x64006400;
const int MUL = 0x2c002c00;
const int ADD = 0xd400d400;
FragB frag_b;
frag_b[0] = __hsub2(*reinterpret_cast<half2*>(&lo),
*reinterpret_cast<const half2*>(&SUB));
frag_b[1] = __hfma2(*reinterpret_cast<half2*>(&hi),
*reinterpret_cast<const half2*>(&MUL),
*reinterpret_cast<const half2*>(&ADD));
return frag_b;
}
template <>
__device__ inline FragB dequant<vllm::kU8.id()>(int q) {
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<start_byte_for_fp16, mask_for_elt_01>(q);
uint32_t hi = prmt<start_byte_for_fp16, mask_for_elt_23>(q);
static constexpr uint32_t I8s_TO_F16s_MAGIC_NUM = 0x64006400;
FragB frag_b;
frag_b[0] = __hsub2(*reinterpret_cast<half2*>(&lo),
*reinterpret_cast<const half2*>(&I8s_TO_F16s_MAGIC_NUM));
frag_b[1] = __hsub2(*reinterpret_cast<half2*>(&hi),
*reinterpret_cast<const half2*>(&I8s_TO_F16s_MAGIC_NUM));
return frag_b;
}
// Multiply dequantized values by the corresponding quantization scale; used
// only for grouped quantization.
__device__ inline void scale(FragB& frag_b, FragS& frag_s, int i) {
half2 s = __half2half2(reinterpret_cast<__half*>(&frag_s)[i]);
frag_b[0] = __hmul2(frag_b[0], s);
frag_b[1] = __hmul2(frag_b[1], s);
}
__device__ inline void sub_zp(FragB& frag_b, half2& frag_zp, int i) {
half2 zp = __half2half2(reinterpret_cast<__half*>(&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)
__device__ inline void scale4(FragB& frag_b, FragS& frag_s_1, FragS& frag_s_2,
FragS& frag_s_3, FragS& frag_s_4, int i) {
__half2 s_val_1_2;
s_val_1_2.x = reinterpret_cast<__half*>(&frag_s_1)[i];
s_val_1_2.y = reinterpret_cast<__half*>(&frag_s_2)[i];
__half2 s_val_3_4;
s_val_3_4.x = reinterpret_cast<__half*>(&frag_s_3)[i];
s_val_3_4.y = reinterpret_cast<__half*>(&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)
__device__ inline void scale_float(float* c, FragS& s) {
__half* s_ptr = reinterpret_cast<__half*>(&s);
c[0] = __fmul_rn(c[0], __half2float(s_ptr[0]));
c[1] = __fmul_rn(c[1], __half2float(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));
}
}
template <const vllm::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 int stages, // number of stages for the async global->shared
// fetch pipeline
const bool has_act_order, // whether act_order is enabled
const bool has_zp, // whether zero-points are enabled
const int group_blocks = -1 // number of consecutive 16x16 blocks
// with a separate quantization scale
>
__device__ void MarlinMoESingle(
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
const int* __restrict__ sorted_ids, // int32 sorted ids of experts
const float* __restrict__ topk_weights, // float topk weights
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 int* __restrict__ expert_offsets,
int num_groups, // number of scale groups per output channel
int expert_idx, // idx of current expert
int num_experts, // number of experts
int topk, // topk parameter of moe
int prob_m, // batch dimension m
int prob_n, // output dimension n
int prob_k, // reduction dimension k
int tot_m, // total number of rows in A and C
int* locks, // extra global storage for barrier synchronization
bool replicate_input, // do we use the same input for each expert?
bool apply_weights, // apply weights to output
int current_m_block // current m block to start kernel computation from
) {
static constexpr auto w_type = vllm::ScalarType::from_id(w_type_id);
constexpr int pack_factor = 32 / w_type.size_bits();
// For larger GEMMs we run multiple batchsize 64 versions in parallel for a
// better partitioning with less reductions
int parallel = 1;
if (prob_m > 16 * thread_m_blocks) {
parallel = prob_m / (16 * thread_m_blocks);
prob_m = 16 * thread_m_blocks;
}
int k_tiles = prob_k / 16 / thread_k_blocks;
int n_tiles = prob_n / 16 / thread_n_blocks;
int iters = ceildiv(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) *
ceildiv(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
// We can easily implement parallel problem execution by just remapping
// indices and advancing global pointers
if (slice_col_par >= n_tiles) {
locks += (slice_col_par / n_tiles) * n_tiles;
slice_col = slice_col_par % n_tiles;
sorted_ids += (slice_col_par / n_tiles) * 16 * thread_m_blocks;
}
// Compute all information about the current slice which is required for
// synchronization.
auto init_slice = [&]() {
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 * ceildiv(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 = ceildiv(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 (slice_col == n_tiles) {
sorted_ids += 16 * thread_m_blocks;
locks += n_tiles;
slice_col = 0;
}
};
init_slice();
// 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 = ceildiv(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 / 8;
constexpr int s_sh_stride = 16 * thread_n_blocks / 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;
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 = (prob_n / pack_factor) / 4;
constexpr int zp_sh_stride = ((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 = a_gl_stride * (threadIdx.x / a_gl_rd_delta_o) +
(threadIdx.x % a_gl_rd_delta_o);
a_gl_rd += a_gl_rd_delta_o * slice_row;
// 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) + (threadIdx.x % 32) / 16;
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;
int b_sh_wr = threadIdx.x * b_thread_vecs;
int 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 {
s_gl_rd = s_gl_stride * ((thread_k_blocks * slice_row) / group_blocks) +
s_sh_stride * slice_col + threadIdx.x;
}
}
int s_sh_wr = threadIdx.x;
bool s_sh_wr_pred = threadIdx.x < s_sh_stride;
// 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;
}
}
int 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
s_sh_rd = 8 * ((threadIdx.x / 32) % (thread_n_blocks / 4)) +
(threadIdx.x % 32) % 4;
// 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) {
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);
}
int sh_first_group_id = -1;
int sh_num_groups = -1;
constexpr int sh_max_num_groups = 32;
extern __shared__ int4 sh[];
// Shared memory storage for global fetch pipelines.
int4* sh_a = sh;
int4* sh_b = sh_a + (stages * a_sh_stage);
int4* sh_g_idx = sh_b + (stages * b_sh_stage);
int4* sh_zp = sh_g_idx + (stages * g_idx_stage);
int4* sh_s = sh_zp + (stages * zp_sh_stage);
// Precompute which thread should not read memory in which iterations; this is
// needed if there are more threads than required for a certain tilesize or
// when the batchsize is not a multiple of 16.
bool a_sh_wr_pred[a_sh_wr_iters];
#pragma unroll
for (int i = 0; i < a_sh_wr_iters; i++) {
int a_idx = a_sh_wr_delta * i + a_sh_wr;
int row = a_idx / a_gl_rd_delta_o;
if (row >= prob_m) {
a_sh_wr_pred[i] = false;
} else {
a_sh_wr_pred[i] = a_sh_wr_delta * i + a_sh_wr < a_sh_stride * prob_m;
}
}
// 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;
};
// 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;
// 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 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
// Zero accumulators.
auto zero_accums = [&]() {
#pragma unroll
for (int i = 0; i < thread_m_blocks * 4 * 2 * 4; i++)
reinterpret_cast<float*>(frag_c)[i] = 0;
};
auto fetch_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 < sh_max_num_groups) {
sh_num_groups = sh_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.
auto fetch_to_shared = [&](int pipe, int a_off, bool pred = true) {
if (pred) {
int4* sh_a_stage = sh_a + a_sh_stage * pipe;
#pragma unroll
for (int i = 0; i < a_sh_wr_iters; i++) {
int a_idx = a_gl_rd_delta_i * i + a_gl_rd + a_gl_rd_delta_o * a_off;
int row = a_idx / a_gl_stride;
int sorted_row =
replicate_input ? sorted_ids[row] / topk : sorted_ids[row];
int new_idx = sorted_row * a_gl_stride + a_idx % a_gl_stride;
if (sorted_row < tot_m * (replicate_input ? 1 : topk) &&
new_idx < a_gl_stride * tot_m * (replicate_input ? 1 : topk)) {
cp_async4_pred(&sh_a_stage[a_sh_wr_trans[i]], &A[new_idx],
a_sh_wr_pred[i]);
}
}
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_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<int4 const*>(&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 constexpr (group_blocks >= thread_k_blocks) {
// Only fetch scales if this tile starts a new group
if (pipe % (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;
}
} else {
for (int i = 0; i < s_tb_groups; i++) {
if (s_sh_wr_pred) {
cp_async4(&sh_s_stage[i * s_sh_stride + s_sh_wr],
&scales_ptr[s_gl_rd]);
}
s_gl_rd += s_gl_rd_delta;
}
}
}
if constexpr (has_zp && group_blocks != -1) {
int4* sh_zp_stage = sh_zp + zp_sh_stage * pipe;
if constexpr (group_blocks >= thread_k_blocks) {
// Only fetch zero-points if this tile starts a new group
if (pipe % (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;
}
} else {
for (int i = 0; i < zp_tb_groups; i++) {
if (zp_sh_wr_pred) {
cp_async4(&sh_zp_stage[i * zp_sh_stride + zp_sh_wr],
&zp_ptr[zp_gl_rd]);
}
zp_gl_rd += zp_gl_rd_delta;
}
}
}
}
}
// 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_zp_to_shared = [&]() {
if (zp_sh_wr_pred) {
cp_async4(&sh_zp[zp_sh_wr], &zp_ptr[zp_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<stages - 2>();
__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) {
int4* sh_a_stage = sh_a + a_sh_stage * pipe;
#pragma unroll
for (int i = 0; i < thread_m_blocks; i++)
ldsm4(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<I4*>(
&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) {
is_same_group[pipe] = false;
same_group_id[pipe] = 0;
return;
}
int4* sh_g_idx_stage = sh_g_idx + g_idx_stage * pipe;
int* sh_g_idx_int_ptr = reinterpret_cast<int*>(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) {
if constexpr (group_blocks >= thread_k_blocks) {
int4* sh_s_stage =
sh_s + s_sh_stage * ((group_blocks / thread_k_blocks) *
(pipe / (group_blocks / thread_k_blocks)));
reinterpret_cast<int4*>(&frag_s[k % 2])[0] = sh_s_stage[s_sh_rd];
} else {
int 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;
reinterpret_cast<int4*>(&frag_s[k % 2])[0] =
sh_s_stage[s_sh_rd + cur_group_id * 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)
int 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;
int 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<int4*>(&(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<int4*>(&(act_frag_s[k % 2][0][0]))) =
*(reinterpret_cast<int4*>(&(act_frag_s[(k - 1) % 2][0][0])));
}
for (int i = 1; i < 4; i++) {
*(reinterpret_cast<int4*>(&(act_frag_s[k % 2][i][0]))) =
*(reinterpret_cast<int4*>(&(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<int*>(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<int4*>(&(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) {
int pipe = full_pipe % stages;
if constexpr (group_blocks == -1) {
for (int i = 0; i < num_ints_per_thread; i++) {
frag_qzp[k % 2][i] = (reinterpret_cast<int*>(sh_zp))[zp_sh_rd + i];
}
} else if constexpr (group_blocks >= thread_k_blocks) {
int4* sh_zp_stage =
sh_zp + zp_sh_stage * ((group_blocks / thread_k_blocks) *
(pipe / (group_blocks / thread_k_blocks)));
for (int i = 0; i < num_ints_per_thread; i++) {
frag_qzp[k % 2][i] =
(reinterpret_cast<int*>(sh_zp_stage))[zp_sh_rd + i];
}
} else {
int 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;
for (int i = 0; i < num_ints_per_thread; i++) {
frag_qzp[k % 2][i] =
(reinterpret_cast<int*>(sh_zp_stage))[zp_sh_rd + i];
}
}
}
};
// Execute the actual tensor core matmul of a sub-tile.
auto matmul = [&](int k) {
if constexpr (has_zp) {
FragB frag_zp_0;
FragB frag_zp_1;
int zp_quant_0, zp_quant_1;
if constexpr (w_type.size_bits() == 4) {
zp_quant_0 = frag_qzp[k % 2][0];
zp_quant_1 = zp_quant_0 >> 8;
} else {
static_assert(w_type.size_bits() == 8);
zp_quant_0 = frag_qzp[k % 2][0];
zp_quant_1 = frag_qzp[k % 2][1];
}
frag_zp_0 = dequant<w_type_id>(zp_quant_0);
frag_zp_1 = dequant<w_type_id>(zp_quant_1);
frag_zp[0] = frag_zp_0[0];
frag_zp[1] = frag_zp_0[1];
frag_zp[2] = frag_zp_1[0];
frag_zp[3] = frag_zp_1[1];
}
// 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++) {
int b_quant_0, b_quant_1;
if constexpr (w_type.size_bits() == 4) {
b_quant_0 = frag_b_quant[k % 2][0][j];
b_quant_1 = b_quant_0 >> 8;
} else {
static_assert(w_type.size_bits() == 8);
int* frag_b_quant_ptr = reinterpret_cast<int*>(frag_b_quant[k % 2]);
b_quant_0 = frag_b_quant_ptr[j * 2 + 0];
b_quant_1 = frag_b_quant_ptr[j * 2 + 1];
}
FragB frag_b0 = dequant<w_type_id>(b_quant_0);
FragB frag_b1 = dequant<w_type_id>(b_quant_1);
// Apply zero-point to frag_b0
if constexpr (has_zp) {
sub_zp(frag_b0, frag_zp[j], 0);
}
// Apply scale to frag_b0
if constexpr (has_act_order) {
scale4(frag_b0, act_frag_s[k % 2][0][j], act_frag_s[k % 2][1][j],
act_frag_s[k % 2][2][j], act_frag_s[k % 2][3][j], 0);
} else {
if constexpr (group_blocks != -1) {
scale(frag_b0, frag_s[k % 2][j], 0);
}
}
// Apply zero-point to frag_b1
if constexpr (has_zp) {
sub_zp(frag_b1, frag_zp[j], 1);
}
// Apply scale to frag_b1
if constexpr (has_act_order) {
scale4(frag_b1, act_frag_s[k % 2][0][j], act_frag_s[k % 2][1][j],
act_frag_s[k % 2][2][j], act_frag_s[k % 2][3][j], 1);
} else {
if constexpr (group_blocks != -1) {
scale(frag_b1, frag_s[k % 2][j], 1);
}
}
#pragma unroll
for (int i = 0; i < thread_m_blocks; i++) {
mma(frag_a[k % 2][i], frag_b0, frag_c[i][j][0]);
mma(frag_a[k % 2][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) {
int 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++) {
int red_sh_wr =
red_sh_delta * j + (red_sh_rd - red_sh_stride * i);
if (i < red_off) {
float* c_rd =
reinterpret_cast<float*>(&sh[red_sh_delta * j + red_sh_rd]);
float* c_wr = reinterpret_cast<float*>(&sh[red_sh_wr]);
#pragma unroll
for (int k = 0; k < 4; k++)
reinterpret_cast<FragC*>(frag_c)[4 * 2 * m_block + j][k] +=
c_rd[k] + c_wr[k];
}
sh[red_sh_wr] =
reinterpret_cast<int4*>(&frag_c)[4 * 2 * m_block + j];
}
}
__syncthreads();
}
if (red_idx == 0) {
#pragma unroll
for (int i = 0; i < 4 * 2; i++) {
float* c_rd =
reinterpret_cast<float*>(&sh[red_sh_delta * i + red_sh_rd]);
#pragma unroll
for (int j = 0; j < 4; j++)
reinterpret_cast<FragC*>(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 = [&](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;
if (threadIdx.x < active_threads) {
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 = 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;
int row = (threadIdx.x % 32) / 4;
if (!first) {
// Interestingly, doing direct global accesses here really seems to mess up
// the compiler and lead to slowdowns, hence we also use async-copies even
// though these fetches are not actually asynchronous.
#pragma unroll
for (int i = 0; i < thread_m_blocks * 4; i++) {
int c_idx =
c_gl_wr + c_gl_wr_delta_o * (i / 2) + c_gl_wr_delta_i * (i % 2);
int sorted_row = sorted_ids[c_idx / c_gl_stride];
int new_idx = sorted_row * c_gl_stride + c_idx % c_gl_stride;
cp_async4_pred(&sh[c_sh_wr + c_sh_wr_delta * i], &C[new_idx],
sorted_row < tot_m * topk &&
(8 * (i / 2) + row < prob_m &&
(i < (thread_m_blocks - 1) * 4 ||
sorted_ids[8 * (i / 2) + row] < tot_m * topk)));
}
cp_async_fence();
cp_async_wait<0>();
}
#pragma unroll
for (int i = 0; i < thread_m_blocks * 4; i++) {
if (8 * (i / 2) + row < prob_m &&
(i < (thread_m_blocks - 1) * 4 ||
sorted_ids[8 * (i / 2) + row] < tot_m * topk)) {
if (!first) {
int4 c_red = sh[c_sh_wr + i * c_sh_wr_delta];
#pragma unroll
for (int j = 0; j < 2 * 4; j++) {
reinterpret_cast<float*>(
&frag_c)[4 * 2 * 4 * (i / 4) + 4 * j + (i % 4)] +=
__half2float(reinterpret_cast<__half*>(&c_red)[j]);
}
}
if (!last) {
int4 c;
#pragma unroll
for (int j = 0; j < 2 * 4; j++) {
reinterpret_cast<__half*>(&c)[j] =
__float2half(reinterpret_cast<float*>(
&frag_c)[4 * 2 * 4 * (i / 4) + 4 * j + (i % 4)]);
}
int c_idx =
c_gl_wr + c_gl_wr_delta_o * (i / 2) + c_gl_wr_delta_i * (i % 2);
int row = sorted_ids[c_idx / c_gl_stride];
if (row < tot_m * topk) {
int new_idx = row * c_gl_stride + c_idx % c_gl_stride;
C[new_idx] = c;
}
}
}
}
}
};
// 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 = [&]() {
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 =
(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));
int c_gl_wr_end = c_gl_stride * prob_m;
// 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) {
half2 res = __halves2half2(__float2half(c0), __float2half(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) {
res = __hmul2(res, s[0]);
}
((half2*)sh)[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++) {
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]);
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]);
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]);
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]);
}
c_sh_wr += 16 * (4 * c_sh_stride);
}
}
__syncthreads();
#pragma unroll
for (int i = 0;
i < ceildiv(16 * thread_m_blocks, threads / (2 * thread_n_blocks));
i++) {
if (c_gl_wr < c_gl_wr_end) {
int row = sorted_ids[c_gl_wr / c_gl_stride];
if (row < tot_m * topk) {
int off = row * c_gl_stride + c_gl_wr % c_gl_stride;
if (!apply_weights) {
C[off] = sh[c_sh_rd];
} else {
__half* ctrg = reinterpret_cast<__half*>(&C[off]);
__half* csrc = reinterpret_cast<__half*>(&sh[c_sh_rd]);
for (int j = 0; j < 8; ++j) {
ctrg[j] = __float2half(topk_weights[row] * __half2float(csrc[j]));
}
}
c_gl_wr += c_gl_wr_delta;
c_sh_rd += c_sh_rd_delta;
}
}
}
};
// 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_scales_to_shared(true, g_idx[slice_k_start], g_idx[last_g_idx]);
}
if constexpr (has_zp && group_blocks == -1) {
if (i == 0) {
fetch_zp_to_shared();
}
}
fetch_to_shared(i, i, i < slice_iters);
}
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 += a_gl_rd_delta_o * (stages - 1);
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.
#pragma unroll
for (int pipe = 0; pipe < stages;) {
#pragma unroll
for (int k = 0; k < b_sh_wr_iters; k++) {
fetch_to_registers(k + 1, pipe % stages);
fetch_scales_to_registers(k + 1, pipe);
fetch_zp_to_registers(k + 1, pipe);
if (k == b_sh_wr_iters - 2) {
fetch_to_shared((pipe + stages - 1) % stages, pipe,
slice_iters >= stages);
pipe++;
wait_for_stage();
init_same_group(pipe % stages);
}
matmul(k);
}
slice_iters--;
if (slice_iters == 0) {
break;
}
}
a_gl_rd += a_gl_rd_delta_o * stages;
slice_k_start += tb_k * stages;
slice_k_start_shared_fetch += tb_k * stages;
if constexpr (has_act_order) {
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_scales_to_shared(false, first_group_id, last_group_id);
__syncthreads();
}
}
// 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;
if constexpr (!has_act_order && group_blocks == -1) {
if constexpr (w_type.size_bits() == 8) {
if (s_sh_wr_pred) {
cp_async4(&sh_s[s_sh_wr], &scales_ptr[s_gl_rd]);
}
cp_async_fence();
} else {
// For 4-bit per-column scales, we only fetch them here in the
// final step before write-out
if (last) {
if (s_sh_wr_pred) {
cp_async4(&sh_s[s_sh_wr], &scales_ptr[s_gl_rd]);
}
cp_async_fence();
}
}
}
thread_block_reduce();
if constexpr (!has_act_order && group_blocks == -1) {
if constexpr (w_type.size_bits() == 8) {
cp_async_wait<0>();
__syncthreads();
if (threadIdx.x / 32 < thread_n_blocks / 4) {
reinterpret_cast<int4*>(&frag_s)[0] = sh_s[s_sh_rd + 0];
reinterpret_cast<int4*>(&frag_s)[1] = sh_s[s_sh_rd + 4];
}
} else {
if (last) {
cp_async_wait<0>();
__syncthreads();
if (threadIdx.x / 32 < thread_n_blocks / 4) {
reinterpret_cast<int4*>(&frag_s)[0] = sh_s[s_sh_rd + 0];
reinterpret_cast<int4*>(&frag_s)[1] = sh_s[s_sh_rd + 4];
}
}
}
}
// 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) {
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<float*>(&frag_c[i][j][0][0]),
frag_s[j / 2][2 * (j % 2) + 0]);
scale_float(reinterpret_cast<float*>(&frag_c[i][j][0][2]),
frag_s[j / 2][2 * (j % 2) + 0]);
scale_float(reinterpret_cast<float*>(&frag_c[i][j][1][0]),
frag_s[j / 2][2 * (j % 2) + 1]);
scale_float(reinterpret_cast<float*>(&frag_c[i][j][1][2]),
frag_s[j / 2][2 * (j % 2) + 1]);
}
}
}
}
if (slice_count > 1) { // only globally reduce if there is more than one
// block in a slice
barrier_acquire(&locks[slice_col], slice_idx);
global_reduce(slice_idx == 0, last);
barrier_release(&locks[slice_col], last);
}
if (last) // only the last block in a slice actually writes the result
write_result();
slice_row = 0;
slice_col_par++;
slice_col++;
init_slice();
if (slice_iters) {
a_gl_rd = a_gl_stride * (threadIdx.x / a_gl_rd_delta_o) +
(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;
}
// 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 {
s_gl_rd = s_sh_stride * slice_col + threadIdx.x;
zp_gl_rd = zp_sh_stride * slice_col + threadIdx.x;
}
start_pipes();
}
}
}
}
template <const vllm::ScalarTypeId w_type_id, // weight ScalarType id
const int threads, // number of threads in a threadblock
const int thread_n_blocks, // same for n dimension (output)
const int thread_k_blocks, // same for k dimension (reduction)
const int stages, // number of stages for the async global->shared
// fetch pipeline
const bool has_act_order, // whether act_order is enabled
const bool has_zp, // whether zero-points are enabled
const int group_blocks = -1 // number of consecutive 16x16 blocks
// with a separate quantization scale
>
__global__ void MarlinMoE(
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
const int* __restrict__ sorted_ids_base, // int32 sorted ids of experts
const float* __restrict__ topk_weights, // float topk weights
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 int* __restrict__ expert_offsets,
int num_groups, // number of scale groups per output channel
int expert_idx, // idx of current expert
int num_experts, // number of experts
int topk, // topk parameter of moe
int prob_m, // batch dimension m
int prob_n, // output dimension n
int prob_k, // reduction dimension k
int tot_m, // total number of rows in A and C
int* locks, // extra global storage for barrier synchronization
bool replicate_input, // do we use the same input for each expert?
bool apply_weights, // apply weights to output
int current_m_block, // current m block to start kernel computation from
int max_par, // maximum parallelism
int cfg_max_m_blocks // upper bound on m blocks
) {
int m_block_ctr = current_m_block;
const int* sorted_ids_expert =
sorted_ids_base + expert_offsets[expert_idx] + m_block_ctr * 4 * max_par;
int tot_its = expert_offsets[expert_idx + 1] - expert_offsets[expert_idx];
if (tot_its == 0) {
return;
}
int tot_m_blocks = ceildiv(tot_its, 16);
int pad = 16 * tot_m_blocks - tot_its;
if (m_block_ctr >= tot_m_blocks) {
return;
}
int max_block = tot_m_blocks - m_block_ctr;
prob_m = tot_its - 16 * m_block_ctr;
int par = 1;
if (max_block > cfg_max_m_blocks) {
// Note that parallel > 1 currently only works for inputs without any
// padding
par = (16 * max_block - pad) / (16 * cfg_max_m_blocks);
if (par > max_par) par = max_par;
prob_m = (16 * cfg_max_m_blocks) * par;
m_block_ctr += cfg_max_m_blocks * (par - 1);
max_block = cfg_max_m_blocks;
}
if (max_block == 1) {
MarlinMoESingle<w_type_id, threads, 1, thread_n_blocks, thread_k_blocks,
stages, has_act_order, has_zp, group_blocks>(
A, B, C, sorted_ids_expert, topk_weights, scales_ptr, zp_ptr, g_idx,
expert_offsets, num_groups, expert_idx, num_experts, topk, prob_m,
prob_n, prob_k, tot_m, locks, replicate_input, apply_weights,
current_m_block);
} else if (max_block == 2) {
MarlinMoESingle<w_type_id, threads, 2, thread_n_blocks, thread_k_blocks,
stages, has_act_order, has_zp, group_blocks>(
A, B, C, sorted_ids_expert, topk_weights, scales_ptr, zp_ptr, g_idx,
expert_offsets, num_groups, expert_idx, num_experts, topk, prob_m,
prob_n, prob_k, tot_m, locks, replicate_input, apply_weights,
current_m_block);
} else if (max_block == 3) {
MarlinMoESingle<w_type_id, threads, 3, thread_n_blocks, thread_k_blocks,
stages, has_act_order, has_zp, group_blocks>(
A, B, C, sorted_ids_expert, topk_weights, scales_ptr, zp_ptr, g_idx,
expert_offsets, num_groups, expert_idx, num_experts, topk, prob_m,
prob_n, prob_k, tot_m, locks, replicate_input, apply_weights,
current_m_block);
} else {
MarlinMoESingle<w_type_id, threads, 4, thread_n_blocks, thread_k_blocks,
stages, has_act_order, has_zp, group_blocks>(
A, B, C, sorted_ids_expert, topk_weights, scales_ptr, zp_ptr, g_idx,
expert_offsets, num_groups, expert_idx, num_experts, topk, prob_m,
prob_n, prob_k, tot_m, locks, replicate_input, apply_weights,
current_m_block);
}
}
#else
template <const vllm::ScalarTypeId w_type_id, // weight ScalarType id
const int threads, // number of threads in a threadblock
const int thread_n_blocks, // same for n dimension (output)
const int thread_k_blocks, // same for k dimension (reduction)
const int stages, // number of stages for the async global->shared
// fetch pipeline
const bool has_act_order, // whether act_order is enabled
const bool has_zp, // whether zero-points are enabled
const int group_blocks = -1 // number of consecutive 16x16 blocks
// with a separate quantization scale
>
__global__ void MarlinMoE(
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
const int* __restrict__ sorted_ids, // int32 sorted ids of experts
const float* __restrict__ topk_weights, // float topk weights
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 int* __restrict__ expert_offsets,
int num_groups, // number of scale groups per output channel
int expert_idx, // idx of current expert
int num_experts, // number of experts
int topk, // topk parameter of moe
int prob_m, // batch dimension m
int prob_n, // output dimension n
int prob_k, // reduction dimension k
int tot_m, // total number of rows in A and C
int* locks, // extra global storage for barrier synchronization
bool replicate_input, // do we use the same input for each expert?
bool apply_weights, // apply weights to output
int current_m_block, // current m block to start kernel computation from
int max_par, // maximum parallelism
int cfg_max_m_blocks // upper bound on m blocks
) {
// Marlin is not implemented yet for SM < 8.0
assert(false);
return;
}
#endif
// 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.
const int USER_THREADS =
256; // Note: This is only used with user-provided thread_k/n
const int STAGES = 4; // 4 pipeline stages fit into shared memory
static constexpr int min_thread_n = 64;
static constexpr int min_thread_k = 64;
#define __CALL_IF_MOE(W_TYPE, THREAD_N_BLOCKS, THREAD_K_BLOCKS, HAS_ACT_ORDER, \
HAS_ZP, GROUP_BLOCKS, NUM_THREADS) \
else if (q_type == W_TYPE && thread_n_blocks == THREAD_N_BLOCKS && \
thread_k_blocks == THREAD_K_BLOCKS && \
has_act_order == HAS_ACT_ORDER && has_zp == HAS_ZP && \
group_blocks == GROUP_BLOCKS && num_threads == NUM_THREADS) { \
cudaFuncSetAttribute( \
MarlinMoE<W_TYPE.id(), NUM_THREADS, THREAD_N_BLOCKS, THREAD_K_BLOCKS, \
STAGES, HAS_ACT_ORDER, HAS_ZP, GROUP_BLOCKS>, \
cudaFuncAttributeMaxDynamicSharedMemorySize, max_shared_mem); \
MarlinMoE<W_TYPE.id(), NUM_THREADS, THREAD_N_BLOCKS, THREAD_K_BLOCKS, \
STAGES, HAS_ACT_ORDER, HAS_ZP, GROUP_BLOCKS> \
<<<blocks, NUM_THREADS, max_shared_mem, stream>>>( \
A_ptr, B_ptr, C_ptr, sorted_ids_ptr, topk_weights_ptr, s_ptr, \
zp_ptr, g_idx_ptr, expert_offsets_ptr, num_groups, expert_idx, \
num_experts, topk, prob_m, prob_n, prob_k, tot_m, locks, \
replicate_input, apply_weights, m_block, max_par, \
cfg_max_m_blocks); \
}
#define GPTQ_CALL_IF_MOE(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS) \
__CALL_IF_MOE(W_TYPE, N_BLOCKS, K_BLOCKS, true, false, 0, NUM_THREADS) \
__CALL_IF_MOE(W_TYPE, N_BLOCKS, K_BLOCKS, false, false, -1, NUM_THREADS) \
__CALL_IF_MOE(W_TYPE, N_BLOCKS, K_BLOCKS, false, false, 2, NUM_THREADS) \
__CALL_IF_MOE(W_TYPE, N_BLOCKS, K_BLOCKS, false, false, 4, NUM_THREADS) \
__CALL_IF_MOE(W_TYPE, N_BLOCKS, K_BLOCKS, false, false, 8, NUM_THREADS)
#define AWQ_CALL_IF_MOE(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS) \
__CALL_IF_MOE(W_TYPE, N_BLOCKS, K_BLOCKS, false, true, -1, NUM_THREADS) \
__CALL_IF_MOE(W_TYPE, N_BLOCKS, K_BLOCKS, false, true, 2, NUM_THREADS) \
__CALL_IF_MOE(W_TYPE, N_BLOCKS, K_BLOCKS, false, true, 4, NUM_THREADS) \
__CALL_IF_MOE(W_TYPE, N_BLOCKS, K_BLOCKS, false, true, 8, NUM_THREADS)
} // namespace marlin_moe
csrc/moe/marlin_kernels/marlin_moe_kernel_ku4.cu deleted 100644 → 0
View file @ 18dd5e01
#include "marlin_moe_kernel_ku4.h"
namespace marlin_moe {
// We return bool so we can create these different kernel calls as a sequence
// of if-elseif's.
bool call_marlin_moe_kernel_ku4(
vllm::ScalarType const& q_type, int thread_n_blocks, int thread_k_blocks,
bool has_act_order, int group_blocks, int num_threads, int blocks,
int max_shared_mem, cudaStream_t stream, const int4* A_ptr,
const int4* B_ptr, int4* C_ptr, const int* sorted_ids_ptr,
const float* topk_weights_ptr, const int4* s_ptr, const int4* zp_ptr,
const int* g_idx_ptr, int* expert_offsets_ptr, int num_groups,
int expert_idx, int num_experts, int topk, int prob_m, int prob_n,
int prob_k, int tot_m, int* locks, bool replicate_input, bool apply_weights,
int m_block, int max_par, int cfg_max_m_blocks) {
bool has_zp = true;
if (false) {
}
AWQ_CALL_IF_MOE(vllm::kU4, 16, 4, 256)
AWQ_CALL_IF_MOE(vllm::kU4, 8, 8, 256)
AWQ_CALL_IF_MOE(vllm::kU4, 8, 4, 128)
AWQ_CALL_IF_MOE(vllm::kU4, 4, 8, 128)
else {
return false;
}
return true;
}
} // namespace marlin_moe
csrc/moe/marlin_kernels/marlin_moe_kernel_ku4.h deleted 100644 → 0
View file @ 18dd5e01
#pragma once
#include "marlin_moe_kernel.h"
namespace marlin_moe {
// We return bool so we can create these different kernel calls as a sequence
// of if-elseif's.
bool call_marlin_moe_kernel_ku4(
vllm::ScalarType const& q_type, int thread_n_blocks, int thread_k_blocks,
bool has_act_order, int group_blocks, int num_threads, int blocks,
int max_shared_mem, cudaStream_t stream, const int4* A_ptr,
const int4* B_ptr, int4* C_ptr, const int* sorted_ids_ptr,
const float* topk_weights_ptr, const int4* s_ptr, const int4* zp_ptr,
const int* g_idx_ptr, int* expert_offsets_ptr, int num_groups,
int expert_idx, int num_experts, int topk, int prob_m, int prob_n,
int prob_k, int tot_m, int* locks, bool replicate_input, bool apply_weights,
int m_block, int max_par, int cfg_max_m_blocks);
} // namespace marlin_moe
csrc/moe/marlin_kernels/marlin_moe_kernel_ku4b8.cu deleted 100644 → 0
View file @ 18dd5e01
#include "marlin_moe_kernel_ku4b8.h"
namespace marlin_moe {
// We return bool so we can create these different kernel calls as a sequence
// of if-elseif's.
bool call_marlin_moe_kernel_ku4b8(
vllm::ScalarType const& q_type, int thread_n_blocks, int thread_k_blocks,
bool has_act_order, int group_blocks, int num_threads, int blocks,
int max_shared_mem, cudaStream_t stream, const int4* A_ptr,
const int4* B_ptr, int4* C_ptr, const int* sorted_ids_ptr,
const float* topk_weights_ptr, const int4* s_ptr, const int4* zp_ptr,
const int* g_idx_ptr, int* expert_offsets_ptr, int num_groups,
int expert_idx, int num_experts, int topk, int prob_m, int prob_n,
int prob_k, int tot_m, int* locks, bool replicate_input, bool apply_weights,
int m_block, int max_par, int cfg_max_m_blocks) {
bool has_zp = false;
if (false) {
}
GPTQ_CALL_IF_MOE(vllm::kU4B8, 16, 4, 256)
GPTQ_CALL_IF_MOE(vllm::kU4B8, 8, 8, 256)
GPTQ_CALL_IF_MOE(vllm::kU4B8, 8, 4, 128)
GPTQ_CALL_IF_MOE(vllm::kU4B8, 4, 8, 128)
else {
return false;
}
return true;
}
} // namespace marlin_moe
csrc/moe/marlin_kernels/marlin_moe_kernel_ku4b8.h deleted 100644 → 0
View file @ 18dd5e01
#pragma once
#include "marlin_moe_kernel.h"
namespace marlin_moe {
// We return bool so we can create these different kernel calls as a sequence
// of if-elseif's.
bool call_marlin_moe_kernel_ku4b8(
vllm::ScalarType const& q_type, int thread_n_blocks, int thread_k_blocks,
bool has_act_order, int group_blocks, int num_threads, int blocks,
int max_shared_mem, cudaStream_t stream, const int4* A_ptr,
const int4* B_ptr, int4* C_ptr, const int* sorted_ids_ptr,
const float* topk_weights_ptr, const int4* s_ptr, const int4* zp_ptr,
const int* g_idx_ptr, int* expert_offsets_ptr, int num_groups,
int expert_idx, int num_experts, int topk, int prob_m, int prob_n,
int prob_k, int tot_m, int* locks, bool replicate_input, bool apply_weights,
int m_block, int max_par, int cfg_max_m_blocks);
} // namespace marlin_moe
csrc/moe/marlin_kernels/marlin_moe_kernel_ku8b128.cu deleted 100644 → 0
View file @ 18dd5e01
#include "marlin_moe_kernel_ku8b128.h"
namespace marlin_moe {
// We return bool so we can create these different kernel calls as a sequence
// of if-elseif's.
bool call_marlin_moe_kernel_ku8b128(
vllm::ScalarType const& q_type, int thread_n_blocks, int thread_k_blocks,
bool has_act_order, int group_blocks, int num_threads, int blocks,
int max_shared_mem, cudaStream_t stream, const int4* A_ptr,
const int4* B_ptr, int4* C_ptr, const int* sorted_ids_ptr,
const float* topk_weights_ptr, const int4* s_ptr, const int4* zp_ptr,
const int* g_idx_ptr, int* expert_offsets_ptr, int num_groups,
int expert_idx, int num_experts, int topk, int prob_m, int prob_n,
int prob_k, int tot_m, int* locks, bool replicate_input, bool apply_weights,
int m_block, int max_par, int cfg_max_m_blocks) {
bool has_zp = false;
if (false) {
}
GPTQ_CALL_IF_MOE(vllm::kU8B128, 16, 4, 256)
GPTQ_CALL_IF_MOE(vllm::kU8B128, 8, 8, 256)
GPTQ_CALL_IF_MOE(vllm::kU8B128, 8, 4, 128)
GPTQ_CALL_IF_MOE(vllm::kU8B128, 4, 8, 128)
else {
return false;
}
return true;
}
} // namespace marlin_moe
csrc/moe/marlin_kernels/marlin_moe_kernel_ku8b128.h deleted 100644 → 0
View file @ 18dd5e01
#pragma once
#include "marlin_moe_kernel.h"
namespace marlin_moe {
bool call_marlin_moe_kernel_ku8b128(
vllm::ScalarType const& q_type, int thread_n_blocks, int thread_k_blocks,
bool has_act_order, int group_blocks, int num_threads, int blocks,
int max_shared_mem, cudaStream_t stream, const int4* A_ptr,
const int4* B_ptr, int4* C_ptr, const int* sorted_ids_ptr,
const float* topk_weights_ptr, const int4* s_ptr, const int4* zp_ptr,
const int* g_idx_ptr, int* expert_offsets_ptr, int num_groups,
int expert_idx, int num_experts, int topk, int prob_m, int prob_n,
int prob_k, int tot_m, int* locks, bool replicate_input, bool apply_weights,
int m_block, int max_par, int cfg_max_m_blocks);
}
csrc/moe/marlin_moe_ops.cu deleted 100644 → 0
View file @ 18dd5e01
/*
* 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.
*/
#include <torch/all.h>
#include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <cuda.h>
#include <cuda_fp16.h>
#include <cuda_runtime.h>
#include <iostream>
#include "core/exception.hpp"
#include "core/scalar_type.hpp"
#include "core/registration.h"
#include "marlin_kernels/marlin_moe_kernel_ku4b8.h"
#include "marlin_kernels/marlin_moe_kernel_ku8b128.h"
#include "marlin_kernels/marlin_moe_kernel_ku4.h"
template <typename T>
inline std::string str(T x) {
return std::to_string(x);
}
namespace marlin_moe {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
// For a given "a" of size [M,K] performs a permutation of the K columns based
// on the given "perm" indices.
__global__ void permute_cols_kernel(int4 const* __restrict__ a_int4_ptr,
int const* __restrict__ perm_int_ptr,
int4* __restrict__ out_int4_ptr, int size_m,
int size_k, int block_rows) {
int start_row = block_rows * blockIdx.x;
int finish_row = start_row + block_rows;
if (finish_row > size_m) {
finish_row = size_m;
}
int cur_block_rows = finish_row - start_row;
int row_stride = size_k * sizeof(half) / 16;
auto permute_row = [&](int row) {
int iters = size_k / blockDim.x;
int rest = size_k % blockDim.x;
int offset = row * row_stride;
half const* a_row_half = reinterpret_cast<half const*>(a_int4_ptr + offset);
half* out_half = reinterpret_cast<half*>(out_int4_ptr + offset);
int base_k = 0;
for (int i = 0; i < iters; i++) {
int 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 += blockDim.x;
}
if (rest) {
if (threadIdx.x < rest) {
int 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 i = 0; i < cur_block_rows; i++) {
int cur_row = start_row + i;
if (cur_row < size_m) {
permute_row(cur_row);
}
}
}
__global__ void compute_expert_offsets(int const* __restrict__ topk_ids,
int* __restrict__ expert_offsets,
int topk_length, int block_size) {
int expert_id = threadIdx.x;
int num_experts = blockDim.x;
int occurrences = 0;
for (int i = 0; i < topk_length; ++i) {
occurrences += (topk_ids[i] == expert_id);
}
expert_offsets[expert_id + 1] = occurrences;
__syncthreads();
if (threadIdx.x == 0) {
int tot_offset = 0;
expert_offsets[0] = 0;
for (int i = 0; i < num_experts; ++i) {
tot_offset += ceildiv(expert_offsets[i + 1], block_size) * block_size;
expert_offsets[i + 1] = tot_offset;
}
}
__syncthreads();
}
#else
__global__ void permute_cols_kernel(int4 const* __restrict__ a_int4_ptr,
int const* __restrict__ perm_int_ptr,
int4* __restrict__ out_int4_ptr, int size_m,
int size_k, int block_rows) {
// Marlin is not implemented yet for SM < 8.0
assert(false);
return;
}
__global__ void compute_expert_offsets(int const* __restrict__ topk_ids,
int* __restrict__ expert_offsets,
int topk_length, int block_size) {
// Marlin is not implemented yet for SM < 8.0
assert(false);
return;
}
#endif
typedef struct {
int thread_k;
int thread_n;
int num_threads;
} thread_config_t;
typedef struct {
int max_m_blocks;
thread_config_t tb_cfg;
} exec_config_t;
thread_config_t small_batch_thread_configs[] = {
// Ordered by priority
// thread_k, thread_n, num_threads
{128, 128, 256}, // Default
{128, 64, 128}, // Reduce N 2X, same K
{64, 256, 256}, // Reduce K 2X, increase N 2X
{64, 128, 128}, // Reduce K 2X, same N
{64, 64, 128}, // Reduce both 2X
};
thread_config_t large_batch_thread_configs[] = {
// Ordered by priority
// thread_k, thread_n, num_threads
{64, 256, 256}, // Default
{128, 128, 256}, // Reduce N 2X, increase K 2X
{64, 128, 128}, // Reduce N 2X, same K
{128, 64, 128}, // Reduce N 4X, increase K 2X
{64, 64, 128}, // Reduce N 4X, same K
};
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 = ceildiv(tb_k, 32); // Worst case is 32 group size
} else {
tb_groups = ceildiv(tb_k, group_size);
}
if (cache_scales_chunk) {
int load_groups =
tb_groups * 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 * 4;
} else {
int tb_scales = tb_groups * tb_n * 2;
return tb_scales * STAGES;
}
}
bool is_valid_cache_size(thread_config_t const& th_config, int max_m_blocks,
int prob_m, int prob_n, int prob_k, int num_bits,
int scales_cache_size, int max_shared_mem) {
int pack_factor = 32 / num_bits;
// Get B size
int tb_k = th_config.thread_k;
int tb_n = th_config.thread_n;
int b_size = (tb_k * tb_n / pack_factor) * 4;
// Get A size
int m_blocks = ceildiv(prob_m, 16);
int tb_max_m = 16;
while (true) {
if (m_blocks >= max_m_blocks) {
tb_max_m *= max_m_blocks;
break;
}
max_m_blocks--;
if (max_m_blocks == 0) {
TORCH_CHECK(false, "Unexpected m_blocks = ", m_blocks);
}
}
int a_size = (tb_max_m * tb_k) * 2;
float pipe_size = (a_size + b_size) * STAGES;
TORCH_CHECK(max_shared_mem / 2 > scales_cache_size); // Sanity
return pipe_size < 0.95f * (max_shared_mem - scales_cache_size);
}
bool is_valid_config(thread_config_t const& th_config, int max_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 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;
}
// thread_k can be only 128 or 64 (because it must be less than groupsize
// which is 128)
if (th_config.thread_k != 128 && th_config.thread_k != 64) {
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;
}
// Determine cache for scales
int scales_cache_size =
get_scales_cache_size(th_config, prob_m, prob_n, prob_k, num_bits,
group_size, has_act_order, is_k_full);
// Check that pipeline fits into cache
if (!is_valid_cache_size(th_config, max_m_blocks, prob_m, prob_n, prob_k,
num_bits, scales_cache_size, max_shared_mem)) {
return false;
}
return true;
}
exec_config_t determine_thread_config(int prob_m, int prob_n, int prob_k,
int num_bits, int group_size,
bool has_act_order, bool is_k_full,
int max_shared_mem) {
int max_m_blocks = 4;
while (max_m_blocks > 0) {
if (prob_m <= 16) {
for (auto th_config : small_batch_thread_configs) {
if (is_valid_config(th_config, max_m_blocks, prob_m, prob_n, prob_k,
num_bits, group_size, has_act_order, is_k_full,
max_shared_mem)) {
return exec_config_t{max_m_blocks, th_config};
}
}
} else {
for (auto th_config : large_batch_thread_configs) {
if (is_valid_config(th_config, max_m_blocks, prob_m, prob_n, prob_k,
num_bits, group_size, has_act_order, is_k_full,
max_shared_mem)) {
return exec_config_t{max_m_blocks, th_config};
}
}
}
max_m_blocks--; // Process less M blocks per invocation to reduce cache
// usage
}
return exec_config_t{0, {-1, -1, -1}};
}
#define CALL_MOE_KERNEL_FUNCTION(KERNEL_FUNCTION) \
else if (KERNEL_FUNCTION( \
q_type, thread_n_blocks, thread_k_blocks, has_act_order, \
group_blocks, num_threads, blocks, max_shared_mem, stream, \
A_ptr, B_ptr, C_ptr, sorted_ids_ptr, topk_weights_ptr, s_ptr, \
zp_ptr, g_idx_ptr, expert_offsets_ptr, num_groups, expert_idx, \
num_experts, topk, prob_m, prob_n, prob_k, tot_m, locks, \
replicate_input, apply_weights, m_block, max_par, \
exec_cfg.max_m_blocks)) { \
}
void marlin_mm_moe(const void* A, const void* B, void* C,
const void* sorted_ids, const void* topk_weights,
const void* topk_ids, const void* s, void* zp,
const void* g_idx, const void* perm, void* a_tmp,
void* expert_offsets, int prob_m, int prob_n, int prob_k,
void* workspace, vllm::ScalarType const& q_type,
bool has_act_order, bool is_k_full, bool has_zp,
int num_groups, int group_size, int num_experts, int topk,
int moe_block_size, int dev, cudaStream_t stream,
int thread_k, int thread_n, int sms, int max_par,
bool replicate_input, bool apply_weights) {
TORCH_CHECK(prob_m > 0 && prob_n > 0 && prob_k > 0, "Invalid MNK = [", prob_m,
", ", prob_n, ", ", prob_k, "]");
if (sms == -1) {
cudaDeviceGetAttribute(&sms, cudaDevAttrMultiProcessorCount, dev);
}
int max_shared_mem = 0;
cudaDeviceGetAttribute(&max_shared_mem,
cudaDevAttrMaxSharedMemoryPerBlockOptin, dev);
TORCH_CHECK(max_shared_mem > 0);
int num_bits = q_type.size_bits();
// Set thread config
exec_config_t exec_cfg;
if (thread_k != -1 && thread_n != -1) {
// User-defined config
exec_cfg =
exec_config_t{4, thread_config_t{thread_k, thread_n, USER_THREADS}};
} else {
// Auto config
exec_cfg =
determine_thread_config(prob_m, prob_n, prob_k, num_bits, group_size,
has_act_order, is_k_full, max_shared_mem);
}
TORCH_CHECK(exec_cfg.max_m_blocks > 0 &&
is_valid_config(exec_cfg.tb_cfg, exec_cfg.max_m_blocks,
prob_m, prob_n, prob_k, num_bits, group_size,
has_act_order, is_k_full, max_shared_mem),
"Invalid thread config: max_m_blocks = ", exec_cfg.max_m_blocks,
", thread_k = ", exec_cfg.tb_cfg.thread_k,
", thread_n = ", exec_cfg.tb_cfg.thread_n,
", num_threads = ", exec_cfg.tb_cfg.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,
", max_shared_mem = ", max_shared_mem);
int num_threads = exec_cfg.tb_cfg.num_threads;
thread_k = exec_cfg.tb_cfg.thread_k;
thread_n = exec_cfg.tb_cfg.thread_n;
int thread_k_blocks = thread_k / 16;
int thread_n_blocks = thread_n / 16;
int blocks = sms;
TORCH_CHECK(prob_n % thread_n == 0, "prob_n = ", prob_n,
" is not divisible by thread_n = ", thread_n);
TORCH_CHECK(prob_k % thread_k == 0, "prob_k = ", prob_k,
" is not divisible by thread_k = ", thread_k);
int group_blocks = 0;
if (has_act_order) {
if (is_k_full) {
TORCH_CHECK(group_size != -1);
group_blocks = group_size / 16;
TORCH_CHECK(prob_k % group_blocks == 0, "prob_k = ", prob_k,
" is not divisible by group_blocks = ", group_blocks);
} else {
TORCH_CHECK(group_size == 0);
group_blocks = 0;
}
} else {
if (group_size == -1) {
group_blocks = -1;
} else {
group_blocks = group_size / 16;
TORCH_CHECK(prob_k % group_blocks == 0, "prob_k = ", prob_k,
" is not divisible by group_blocks = ", group_blocks);
}
}
int tot_m = prob_m;
const int* topk_ids_ptr = (const int*)topk_ids;
int* expert_offsets_ptr = (int*)expert_offsets;
compute_expert_offsets<<<1, num_experts, 0, stream>>>(
topk_ids_ptr, expert_offsets_ptr, tot_m * topk, moe_block_size);
bool do_permute_a = has_act_order;
// 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 pack_factor = 32 / q_type.size_bits();
for (int expert_idx = 0; expert_idx < num_experts; ++expert_idx) {
const int4* A_ptr = (const int4*)A;
int4* a_tmp_ptr = (int4*)a_tmp;
const int4* B_ptr =
(const int4*)B + (prob_n * prob_k / (pack_factor * 4)) * expert_idx;
int4* C_ptr = (int4*)C;
const float* topk_weights_ptr = (const float*)topk_weights;
const int* sorted_ids_ptr = (const int*)sorted_ids;
const int4* s_ptr = (const int4*)s + num_groups * prob_n / 8 * expert_idx;
const int4* zp_ptr =
(const int4*)zp + num_groups * prob_n / (pack_factor * 4) * expert_idx;
const int* g_idx_ptr = (const int*)g_idx + prob_k * expert_idx;
const int* perm_ptr = (const int*)perm + prob_k * expert_idx;
int* locks = (int*)workspace;
if (do_permute_a) {
// Permute A columns
int topk_rows = replicate_input ? tot_m : tot_m * topk;
int block_rows = ceildiv(topk_rows, blocks);
permute_cols_kernel<<<blocks, num_threads, 0, stream>>>(
A_ptr, perm_ptr, a_tmp_ptr, topk_rows, prob_k, block_rows);
A_ptr = a_tmp_ptr;
}
int tot_m_blocks = ceildiv(tot_m, 16);
for (int m_block = 0; m_block < tot_m_blocks;
m_block += 4 * exec_cfg.max_m_blocks) {
if (false) {
}
CALL_MOE_KERNEL_FUNCTION(call_marlin_moe_kernel_ku4b8)
CALL_MOE_KERNEL_FUNCTION(call_marlin_moe_kernel_ku8b128)
CALL_MOE_KERNEL_FUNCTION(call_marlin_moe_kernel_ku4)
else {
TORCH_CHECK(false, "Unsupported shapes: MNK = [" + str(prob_m) + ", " +
str(prob_n) + ", " + str(prob_k) + "]" +
", has_act_order = " + str(has_act_order) +
", num_groups = " + str(num_groups) +
", group_size = " + str(group_size) +
", thread_n_blocks = " + str(thread_n_blocks) +
", thread_k_blocks = " + str(thread_k_blocks));
}
}
}
}
} // namespace marlin_moe
torch::Tensor marlin_gemm_moe(
const torch::Tensor& a, const torch::Tensor& b_q_weights,
const torch::Tensor& sorted_ids, const torch::Tensor& topk_weights,
const torch::Tensor& topk_ids, const torch::Tensor& b_scales,
torch::Tensor& b_zeros, const torch::Tensor& g_idx,
const torch::Tensor& perm, torch::Tensor& workspace,
vllm::ScalarTypeId const b_q_type_id, int64_t size_m, int64_t size_n,
int64_t size_k, bool is_k_full, int64_t num_experts, int64_t topk,
int64_t moe_block_size, bool replicate_input, bool apply_weights) {
vllm::ScalarType const b_q_type = vllm::ScalarType::from_id(b_q_type_id);
bool has_zp = b_zeros.size(1) != 0;
if (has_zp) {
TORCH_CHECK(
b_q_type == vllm::kU4,
"b_q_type must be u4 when has_zp = True. Got = ", b_q_type.str());
} else {
TORCH_CHECK(
b_q_type == vllm::kU4B8 || b_q_type == vllm::kU8B128,
"b_q_type must be uint4b8 or uint8b128. Got = ", b_q_type.str());
}
int pack_factor = 32 / b_q_type.size_bits();
int max_par = 4;
int dev = a.get_device();
auto options_dtype =
torch::TensorOptions().dtype(a.dtype()).device(a.device());
auto options_int =
torch::TensorOptions().dtype(torch::kInt).device(a.device());
torch::Tensor c = torch::zeros({size_m, topk, size_n}, options_dtype);
torch::Tensor a_tmp =
replicate_input ? torch::zeros({size_m, size_k}, options_dtype)
: torch::zeros({size_m, topk, size_k}, options_dtype);
torch::Tensor expert_offsets = torch::empty({num_experts + 1}, options_int);
// thread_k: `k` size of a thread_tile in `weights` (can usually be left as
// auto -1)
int thread_k = -1;
// thread_n: `n` size of a thread_tile in `weights` (can usually be left as
// auto -1)
int thread_n = -1;
// sms: number of SMs to use for the kernel (can usually be left as auto -1)
int sms = -1;
// Detect groupsize and act_order
int num_groups = -1;
int group_size = -1;
bool has_act_order = g_idx.size(1) != 0;
int b_rank = b_scales.sizes().size();
TORCH_CHECK(b_rank == 3, "b_scales rank = ", b_rank, " is not 3");
TORCH_CHECK(b_scales.size(2) == size_n, "b_scales dim 2 = ", b_scales.size(2),
" is not size_n = ", size_n);
num_groups = b_scales.size(1);
TORCH_CHECK(VLLM_IMPLIES(!is_k_full, has_act_order),
"if is_k_full is false, has_act_order must be true");
if (has_act_order) {
if (is_k_full) {
TORCH_CHECK(num_groups > 1, "For act_order, num_groups must be > 1");
TORCH_CHECK(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) {
TORCH_CHECK(
size_k % num_groups == 0, "size_k = ", size_k,
", is not divisible by b_scales.size(0) = ", b_scales.size(0));
group_size = size_k / num_groups;
} else {
group_size = -1;
}
}
// Verify b_zeros
if (has_zp) {
int rank = b_zeros.sizes().size();
TORCH_CHECK(rank == 3, "b_zeros rank = ", rank, " is not 3");
TORCH_CHECK(b_zeros.size(1) == num_groups,
"b_zeros dim 1 = ", b_zeros.size(1),
" is not num_groups = ", num_groups);
TORCH_CHECK(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);
}
marlin_moe::marlin_mm_moe(
a.data_ptr(), b_q_weights.data_ptr(), c.data_ptr(), sorted_ids.data_ptr(),
topk_weights.data_ptr(), topk_ids.data_ptr(), b_scales.data_ptr(),
b_zeros.data_ptr(), g_idx.data_ptr(), perm.data_ptr(), a_tmp.data_ptr(),
expert_offsets.data_ptr(), size_m, size_n, size_k, workspace.data_ptr(),
b_q_type, has_act_order, is_k_full, has_zp, num_groups, group_size,
num_experts, topk, moe_block_size, dev,
at::cuda::getCurrentCUDAStream(dev), thread_k, thread_n, sms, max_par,
replicate_input, apply_weights);
return c;
}
TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
m.impl("marlin_gemm_moe", &marlin_gemm_moe);
}
csrc/quantization/fp8/fp8_marlin.cu deleted 100644 → 0
View file @ 18dd5e01
/*
* 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 "../gptq_marlin/marlin.cuh"
#include "../gptq_marlin/marlin_dtypes.cuh"
#include "core/registration.h"
using namespace marlin;
#define STATIC_ASSERT_SCALAR_TYPE_VALID(scalar_t) \
static_assert(std::is_same<scalar_t, half>::value || \
std::is_same<scalar_t, nv_bfloat16>::value, \
"only float16 and bfloat16 is supported");
template <typename T>
inline std::string str(T x) {
return std::to_string(x);
}
namespace fp8_marlin {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
template <typename scalar_t, // compute dtype, half or nv_float16
const int num_bits, // number of bits used for weights
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 int stages, // number of stages for the async global->shared
// fetch pipeline
const int group_blocks = -1 // number of consecutive 16x16 blocks
// with a separate quantization scale
>
__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
const int4* __restrict__ scales_ptr, // fp16 quantization scales of shape
// (k/groupsize)xn
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
) {}
} // namespace fp8_marlin
torch::Tensor fp8_marlin_gemm(torch::Tensor& a, torch::Tensor& b_q_weight,
torch::Tensor& b_scales, torch::Tensor& workspace,
int64_t num_bits, int64_t size_m, int64_t size_n,
int64_t size_k) {
TORCH_CHECK_NOT_IMPLEMENTED(false,
"marlin_gemm(..) requires CUDA_ARCH >= 8.0");
return torch::empty({1, 1});
}
#else
// m16n8k16 tensor core mma instruction with fp16 inputs and fp32
// output/accumulation.
template <typename scalar_t>
__device__ inline void mma(const typename ScalarType<scalar_t>::FragA& a_frag,
const typename ScalarType<scalar_t>::FragB& frag_b,
typename ScalarType<scalar_t>::FragC& frag_c) {
const uint32_t* a = reinterpret_cast<const uint32_t*>(&a_frag);
const uint32_t* b = reinterpret_cast<const uint32_t*>(&frag_b);
float* c = reinterpret_cast<float*>(&frag_c);
if constexpr (std::is_same<scalar_t, half>::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<scalar_t, nv_bfloat16>::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);
}
}
// Instruction for loading a full 16x16 matrix fragment of operand A from shared
// memory, directly in tensor core layout.
template <typename scalar_t>
__device__ inline void ldsm4(typename ScalarType<scalar_t>::FragA& frag_a,
const void* smem_ptr) {
uint32_t* a = reinterpret_cast<uint32_t*>(&frag_a);
uint32_t smem = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
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));
}
// Fast FP8ToFp16/FP8ToBf16: Efficiently dequantize 8bit fp8_e4m3 values to fp16
// 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 <typename scalar_t>
__device__ inline typename ScalarType<scalar_t>::FragB dequant_8bit(int q) {
STATIC_ASSERT_SCALAR_TYPE_VALID(scalar_t);
}
template <>
__device__ inline typename ScalarType<half>::FragB dequant_8bit<half>(int q) {
// Constants for FP8 (E4M3) and FP16 formats
constexpr int FP8_EXPONENT = 4, FP8_MANTISSA = 3, FP16_EXPONENT = 5;
constexpr int RIGHT_SHIFT = FP16_EXPONENT - FP8_EXPONENT;
// Calculate MASK for extracting mantissa and exponent
constexpr int MASK1 = 0x80000000;
constexpr int MASK2 = MASK1 >> (FP8_EXPONENT + FP8_MANTISSA);
constexpr int MASK3 = MASK2 & 0x7fffffff;
constexpr int MASK = MASK3 | (MASK3 >> 16);
// Final MASK value: 0x7F007F00
// Extract and shift FP8 values to FP16 format
int Out1 = (q & 0x80008000) | ((q & MASK) >> RIGHT_SHIFT);
int Out2 = ((q << 8) & 0x80008000) | (((q << 8) & MASK) >> RIGHT_SHIFT);
// 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
typename ScalarType<half>::FragB frag_b;
// Note: reverse indexing is intentional because weights are permuted
frag_b[1] = __hmul2(*reinterpret_cast<const half2*>(&Out1), bias_reg);
frag_b[0] = __hmul2(*reinterpret_cast<const half2*>(&Out2), bias_reg);
return frag_b;
}
template <>
__device__ inline typename ScalarType<nv_bfloat16>::FragB
dequant_8bit<nv_bfloat16>(int q) {
// Constants for FP8 (E4M3) and BF16 formats
constexpr int FP8_EXPONENT = 4, FP8_MANTISSA = 3, BF16_EXPONENT = 8;
constexpr int RIGHT_SHIFT = BF16_EXPONENT - FP8_EXPONENT;
// Calculate MASK for extracting mantissa and exponent
constexpr int MASK1 = 0x80000000;
constexpr int MASK2 = MASK1 >> (FP8_EXPONENT + FP8_MANTISSA);
constexpr int MASK3 = MASK2 & 0x7fffffff;
constexpr int MASK = MASK3 | (MASK3 >> 16);
// Final MASK value: 0x7F007F00
// Extract and shift FP8 values to BF16 format
int Out1 = (q & 0x80008000) | ((q & MASK) >> RIGHT_SHIFT);
int Out2 = ((q << 8) & 0x80008000) | (((q << 8) & MASK) >> RIGHT_SHIFT);
// 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<const float*>(&BIAS));
// Convert to bfloat162 and apply bias
typename ScalarType<nv_bfloat16>::FragB frag_b;
// Note: reverse indexing is intentional because weights are permuted
frag_b[1] = __hmul2(*reinterpret_cast<const nv_bfloat162*>(&Out1), bias_reg);
frag_b[0] = __hmul2(*reinterpret_cast<const nv_bfloat162*>(&Out2), bias_reg);
return frag_b;
}
// Multiply dequantized values by the corresponding quantization scale; used
// only for grouped quantization.
template <typename scalar_t>
__device__ inline void scale(typename ScalarType<scalar_t>::FragB& frag_b,
typename ScalarType<scalar_t>::FragS& frag_s,
int i) {
using scalar_t2 = typename ScalarType<scalar_t>::scalar_t2;
scalar_t2 s =
ScalarType<scalar_t>::num2num2(reinterpret_cast<scalar_t*>(&frag_s)[i]);
frag_b[0] = __hmul2(frag_b[0], s);
frag_b[1] = __hmul2(frag_b[1], s);
}
// Given 2 floats multiply by 2 scales (halves)
template <typename scalar_t>
__device__ inline void scale_float(float* c,
typename ScalarType<scalar_t>::FragS& s) {
scalar_t* s_ptr = reinterpret_cast<scalar_t*>(&s);
c[0] = __fmul_rn(c[0], ScalarType<scalar_t>::num2float(s_ptr[0]));
c[1] = __fmul_rn(c[1], ScalarType<scalar_t>::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));
}
}
template <typename scalar_t, // compute dtype, half or nv_float16
const int num_bits, // number of bits used for weights
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 int stages, // number of stages for the async global->shared
// fetch pipeline
const int group_blocks = -1 // number of consecutive 16x16 blocks
// with a separate quantization scale
>
__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
const int4* __restrict__ scales_ptr, // fp16 quantization scales of shape
// (k/groupsize)xn
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
) {
// 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<scalar_t>;
using scalar_t2 = typename ScalarType<scalar_t>::scalar_t2;
using FragA = typename ScalarType<scalar_t>::FragA;
using FragB = typename ScalarType<scalar_t>::FragB;
using FragC = typename ScalarType<scalar_t>::FragC;
using FragS = typename ScalarType<scalar_t>::FragS;
constexpr int pack_factor = 32 / num_bits;
// For larger GEMMs we run multiple batchsize 64 versions in parallel for a
// better partitioning with less reductions
int parallel = 1;
if (prob_m > 16 * thread_m_blocks) {
parallel = prob_m / (16 * thread_m_blocks);
prob_m = 16 * thread_m_blocks;
}
int k_tiles = prob_k / 16 / thread_k_blocks;
int n_tiles = prob_n / 16 / thread_n_blocks;
int iters = div_ceil(k_tiles * n_tiles * parallel, gridDim.x);
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
// We can easily implement parallel problem execution by just remapping
// indices and advancing global pointers
if (slice_col_par >= n_tiles) {
A += (slice_col_par / n_tiles) * 16 * thread_m_blocks * prob_k / 8;
C += (slice_col_par / n_tiles) * 16 * thread_m_blocks * prob_n / 8;
locks += (slice_col_par / n_tiles) * n_tiles;
slice_col = slice_col_par % n_tiles;
}
// Compute all information about the current slice which is required for
// synchronization.
auto init_slice = [&]() {
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 * 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 = 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 (slice_col == n_tiles) {
A += 16 * thread_m_blocks * prob_k / 8;
C += 16 * thread_m_blocks * prob_n / 8;
locks += n_tiles;
slice_col = 0;
}
};
init_slice();
// 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 = 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 = num_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 / 8;
constexpr int s_sh_stride = 16 * thread_n_blocks / 8;
// Scale size/strides with act_order
constexpr int tb_k = 16 * thread_k_blocks;
constexpr int g_idx_stage = 0;
// constexpr int act_s_row_stride = 1;
// int act_s_col_stride = act_s_row_stride * num_groups;
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;
// Global A read index of current thread.
int a_gl_rd = a_gl_stride * (threadIdx.x / a_gl_rd_delta_o) +
(threadIdx.x % a_gl_rd_delta_o);
a_gl_rd += a_gl_rd_delta_o * slice_row;
// 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) + (threadIdx.x % 32) / 16;
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;
int b_sh_wr = threadIdx.x * b_thread_vecs;
int b_sh_rd = threadIdx.x * b_thread_vecs;
// For act_order
int slice_k_start = tb_k * slice_row;
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 = s_sh_stride * slice_col + threadIdx.x;
int s_sh_wr = threadIdx.x;
bool s_sh_wr_pred = threadIdx.x < s_sh_stride;
// We scale a `half2` tile in row-major layout for column-wise quantization.
int s_sh_rd =
8 * ((threadIdx.x / 32) % (thread_n_blocks / 4)) + (threadIdx.x % 32) % 4;
// Precompute which thread should not read memory in which iterations; this is
// needed if there are more threads than required for a certain tilesize or
// when the batchsize is not a multiple of 16.
bool a_sh_wr_pred[a_sh_wr_iters];
#pragma unroll
for (int i = 0; i < a_sh_wr_iters; i++)
a_sh_wr_pred[i] = a_sh_wr_delta * i + a_sh_wr < a_sh_stride * prob_m;
// 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;
};
// 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;
extern __shared__ int4 sh[];
// Shared memory storage for global fetch pipelines.
int4* sh_a = sh;
int4* sh_b = sh_a + (stages * a_sh_stage);
int4* sh_g_idx = sh_b + (stages * b_sh_stage);
int4* sh_s = sh_g_idx + (stages * g_idx_stage);
// 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];
// Zero accumulators.
auto zero_accums = [&]() {
#pragma unroll
for (int i = 0; i < thread_m_blocks * 4 * 2 * 4; i++)
reinterpret_cast<float*>(frag_c)[i] = 0;
};
int sh_first_group_id = -1;
int sh_num_groups = -1;
constexpr int sh_max_num_groups = 32;
auto fetch_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 < sh_max_num_groups) {
sh_num_groups = sh_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.
auto fetch_to_shared = [&](int pipe, int a_off, bool pred = true) {
if (pred) {
int4* sh_a_stage = sh_a + a_sh_stage * pipe;
#pragma unroll
for (int i = 0; i < a_sh_wr_iters; i++) {
cp_async4_pred(
&sh_a_stage[a_sh_wr_trans[i]],
&A[a_gl_rd_delta_i * i + a_gl_rd + a_gl_rd_delta_o * a_off],
a_sh_wr_pred[i]);
}
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_ptr[i] += b_gl_rd_delta_o;
}
}
// Insert a fence even when we are winding down the pipeline to ensure that
// waiting is also correct at this point.
cp_async_fence();
};
// 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<stages - 2>();
__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) {
int4* sh_a_stage = sh_a + a_sh_stage * pipe;
#pragma unroll
for (int i = 0; i < thread_m_blocks; i++)
ldsm4<scalar_t>(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<I4*>(
&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) {
is_same_group[pipe] = false;
same_group_id[pipe] = 0;
return;
};
// Execute the actual tensor core matmul of a sub-tile.
auto matmul = [&](int k) {
// 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* frag_b_quant_ptr = reinterpret_cast<int*>(frag_b_quant[k % 2]);
int b_quant_0 = frag_b_quant_ptr[j * 2 + 0];
int b_quant_1 = frag_b_quant_ptr[j * 2 + 1];
frag_b0 = dequant_8bit<scalar_t>(b_quant_0);
frag_b1 = dequant_8bit<scalar_t>(b_quant_1);
#pragma unroll
for (int i = 0; i < thread_m_blocks; i++) {
mma<scalar_t>(frag_a[k % 2][i], frag_b0, frag_c[i][j][0]);
mma<scalar_t>(frag_a[k % 2][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) {
int 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++) {
int red_sh_wr =
red_sh_delta * j + (red_sh_rd - red_sh_stride * i);
if (i < red_off) {
float* c_rd =
reinterpret_cast<float*>(&sh[red_sh_delta * j + red_sh_rd]);
float* c_wr = reinterpret_cast<float*>(&sh[red_sh_wr]);
#pragma unroll
for (int k = 0; k < 4; k++)
reinterpret_cast<FragC*>(frag_c)[4 * 2 * m_block + j][k] +=
c_rd[k] + c_wr[k];
}
sh[red_sh_wr] =
reinterpret_cast<int4*>(&frag_c)[4 * 2 * m_block + j];
}
}
__syncthreads();
}
if (red_idx == 0) {
#pragma unroll
for (int i = 0; i < 4 * 2; i++) {
float* c_rd =
reinterpret_cast<float*>(&sh[red_sh_delta * i + red_sh_rd]);
#pragma unroll
for (int j = 0; j < 4; j++)
reinterpret_cast<FragC*>(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 = [&](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;
if (threadIdx.x < active_threads) {
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 = 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;
int row = (threadIdx.x % 32) / 4;
if (!first) {
// Interestingly, doing direct global accesses here really seems to mess up
// the compiler and lead to slowdowns, hence we also use async-copies even
// though these fetches are not actually asynchronous.
#pragma unroll
for (int i = 0; i < thread_m_blocks * 4; i++) {
cp_async4_pred(
&sh[c_sh_wr + c_sh_wr_delta * i],
&C[c_gl_wr + c_gl_wr_delta_o * (i / 2) +
c_gl_wr_delta_i * (i % 2)],
i < (thread_m_blocks - 1) * 4 || 8 * (i / 2) + row < prob_m);
}
cp_async_fence();
cp_async_wait<0>();
}
#pragma unroll
for (int i = 0; i < thread_m_blocks * 4; i++) {
if (i < (thread_m_blocks - 1) * 4 || 8 * (i / 2) + row < prob_m) {
if (!first) {
int4 c_red = sh[c_sh_wr + i * c_sh_wr_delta];
#pragma unroll
for (int j = 0; j < 2 * 4; j++) {
reinterpret_cast<float*>(
&frag_c)[4 * 2 * 4 * (i / 4) + 4 * j + (i % 4)] +=
Dtype::num2float(reinterpret_cast<scalar_t*>(&c_red)[j]);
}
}
if (!last) {
int4 c;
#pragma unroll
for (int j = 0; j < 2 * 4; j++) {
reinterpret_cast<scalar_t*>(&c)[j] =
Dtype::float2num(reinterpret_cast<float*>(
&frag_c)[4 * 2 * 4 * (i / 4) + 4 * j + (i % 4)]);
}
C[c_gl_wr + c_gl_wr_delta_o * (i / 2) + c_gl_wr_delta_i * (i % 2)] =
c;
}
}
}
}
};
// 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 = [&]() {
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 =
(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));
int c_gl_wr_end = c_gl_stride * prob_m;
// 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) {
scalar_t2 res =
Dtype::nums2num2(Dtype::float2num(c0), Dtype::float2num(c1));
((scalar_t2*)sh)[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++) {
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]);
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]);
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]);
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]);
}
c_sh_wr += 16 * (4 * c_sh_stride);
}
}
__syncthreads();
#pragma unroll
for (int i = 0;
i < div_ceil(16 * thread_m_blocks, threads / (2 * thread_n_blocks));
i++) {
if (c_gl_wr < c_gl_wr_end) {
C[c_gl_wr] = sh[c_sh_rd];
c_gl_wr += c_gl_wr_delta;
c_sh_rd += c_sh_rd_delta;
}
}
};
// Start global fetch and register load pipelines.
auto start_pipes = [&]() {
#pragma unroll
for (int i = 0; i < stages - 1; i++) {
fetch_to_shared(i, i, i < slice_iters);
}
zero_accums();
wait_for_stage();
init_same_group(0);
fetch_to_registers(0, 0);
a_gl_rd += a_gl_rd_delta_o * (stages - 1);
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.
#pragma unroll
for (int pipe = 0; pipe < stages;) {
#pragma unroll
for (int k = 0; k < b_sh_wr_iters; k++) {
fetch_to_registers(k + 1, pipe % stages);
if (k == b_sh_wr_iters - 2) {
fetch_to_shared((pipe + stages - 1) % stages, pipe,
slice_iters >= stages);
pipe++;
wait_for_stage();
init_same_group(pipe % stages);
}
matmul(k);
}
slice_iters--;
if (slice_iters == 0) {
break;
}
}
a_gl_rd += a_gl_rd_delta_o * stages;
slice_k_start += tb_k * stages;
slice_k_start_shared_fetch += tb_k * stages;
// 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 (s_sh_wr_pred) {
cp_async4(&sh_s[s_sh_wr], &scales_ptr[s_gl_rd]);
}
cp_async_fence();
thread_block_reduce();
cp_async_wait<0>();
__syncthreads();
if (threadIdx.x / 32 < thread_n_blocks / 4) {
reinterpret_cast<int4*>(&frag_s)[0] = sh_s[s_sh_rd + 0];
reinterpret_cast<int4*>(&frag_s)[1] = sh_s[s_sh_rd + 4];
}
// 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 (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<scalar_t>(reinterpret_cast<float*>(&frag_c[i][j][0][0]),
frag_s[j / 2][2 * (j % 2) + 0]);
scale_float<scalar_t>(reinterpret_cast<float*>(&frag_c[i][j][0][2]),
frag_s[j / 2][2 * (j % 2) + 0]);
scale_float<scalar_t>(reinterpret_cast<float*>(&frag_c[i][j][1][0]),
frag_s[j / 2][2 * (j % 2) + 1]);
scale_float<scalar_t>(reinterpret_cast<float*>(&frag_c[i][j][1][2]),
frag_s[j / 2][2 * (j % 2) + 1]);
}
}
}
if (slice_count > 1) { // only globally reduce if there is more than one
// block in a slice
barrier_acquire(&locks[slice_col], slice_idx);
global_reduce(slice_idx == 0, last);
barrier_release(&locks[slice_col], last);
}
if (last) // only the last block in a slice actually writes the result
write_result();
slice_row = 0;
slice_col_par++;
slice_col++;
init_slice();
if (slice_iters) {
a_gl_rd = a_gl_stride * (threadIdx.x / a_gl_rd_delta_o) +
(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;
}
// Update slice k/n for scales loading
s_gl_rd = s_sh_stride * slice_col + threadIdx.x;
start_pipes();
}
}
}
}
#define __CALL_IF(NUM_BITS, THREAD_M_BLOCKS, THREAD_N_BLOCKS, \
THREAD_K_BLOCKS, GROUP_BLOCKS, NUM_THREADS) \
else if (num_bits == NUM_BITS && thread_m_blocks == THREAD_M_BLOCKS && \
thread_n_blocks == THREAD_N_BLOCKS && \
thread_k_blocks == THREAD_K_BLOCKS && \
group_blocks == GROUP_BLOCKS && num_threads == NUM_THREADS) { \
cudaFuncSetAttribute( \
Marlin<scalar_t, NUM_BITS, NUM_THREADS, THREAD_M_BLOCKS, \
THREAD_N_BLOCKS, THREAD_K_BLOCKS, pipe_stages, GROUP_BLOCKS>, \
cudaFuncAttributeMaxDynamicSharedMemorySize, max_shared_mem); \
Marlin<scalar_t, NUM_BITS, NUM_THREADS, THREAD_M_BLOCKS, \
THREAD_N_BLOCKS, THREAD_K_BLOCKS, pipe_stages, GROUP_BLOCKS> \
<<<blocks, NUM_THREADS, max_shared_mem, stream>>>( \
A_ptr, B_ptr, C_ptr, s_ptr, num_groups, prob_m, prob_n, prob_k, \
locks); \
}
typedef struct {
int thread_k;
int thread_n;
int num_threads;
} thread_config_t;
typedef struct {
int max_m_blocks;
thread_config_t tb_cfg;
} exec_config_t;
thread_config_t small_batch_thread_configs[] = {
// Ordered by priority
// thread_k, thread_n, num_threads
{128, 128, 256},
{64, 128, 128},
{128, 64, 128},
};
thread_config_t large_batch_thread_configs[] = {
// Ordered by priority
// thread_k, thread_n, num_threads
{64, 256, 256},
{64, 128, 128},
{128, 64, 128},
};
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) {
int tb_n = th_config.thread_n;
// Get max scale groups per thread-block
// Fixed for channelwise
int tb_groups = 1;
int tb_scales = tb_groups * tb_n * 2;
return tb_scales * pipe_stages;
}
bool is_valid_cache_size(thread_config_t const& th_config, int max_m_blocks,
int prob_m, int prob_n, int prob_k, int num_bits,
int scales_cache_size, int max_shared_mem) {
int pack_factor = 32 / num_bits;
// Get B size
int tb_k = th_config.thread_k;
int tb_n = th_config.thread_n;
int b_size = (tb_k * tb_n / pack_factor) * 4;
// Get A size
int m_blocks = div_ceil(prob_m, 16);
int tb_max_m = 16;
while (true) {
if (m_blocks >= max_m_blocks) {
tb_max_m *= max_m_blocks;
break;
}
max_m_blocks--;
if (max_m_blocks == 0) {
TORCH_CHECK(false, "Unexpected m_blocks = ", m_blocks);
}
}
int a_size = (tb_max_m * tb_k) * 2;
float pipe_size = (a_size + b_size) * pipe_stages;
TORCH_CHECK(max_shared_mem / 2 > scales_cache_size); // Sanity
return pipe_size < 0.95f * (max_shared_mem - scales_cache_size);
}
bool is_valid_config(thread_config_t const& th_config, int max_m_blocks,
int prob_m, int prob_n, int prob_k, int num_bits,
int group_size, 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;
}
// Determine cache for scales
int scales_cache_size = get_scales_cache_size(th_config, prob_m, prob_n,
prob_k, num_bits, group_size);
// Check that pipeline fits into cache
if (!is_valid_cache_size(th_config, max_m_blocks, prob_m, prob_n, prob_k,
num_bits, scales_cache_size, max_shared_mem)) {
return false;
}
return true;
}
exec_config_t determine_thread_config(int prob_m, int prob_n, int prob_k,
int num_bits, int group_size,
int max_shared_mem) {
int max_m_blocks = 4;
while (max_m_blocks > 0) {
if (prob_m <= 16) {
for (auto th_config : small_batch_thread_configs) {
if (is_valid_config(th_config, max_m_blocks, prob_m, prob_n, prob_k,
num_bits, group_size, max_shared_mem)) {
return exec_config_t{max_m_blocks, th_config};
}
}
} else {
for (auto th_config : large_batch_thread_configs) {
if (is_valid_config(th_config, max_m_blocks, prob_m, prob_n, prob_k,
num_bits, group_size, max_shared_mem)) {
return exec_config_t{max_m_blocks, th_config};
}
}
}
max_m_blocks--; // Process less M blocks per invocation to reduce cache
// usage
}
return exec_config_t{0, {-1, -1, -1}};
}
#define CALL_IF(NUM_BITS, N_BLOCKS, K_BLOCKS, NUM_THREADS) \
__CALL_IF(NUM_BITS, 1, N_BLOCKS, K_BLOCKS, -1, NUM_THREADS) \
__CALL_IF(NUM_BITS, 2, N_BLOCKS, K_BLOCKS, -1, NUM_THREADS) \
__CALL_IF(NUM_BITS, 3, N_BLOCKS, K_BLOCKS, -1, NUM_THREADS) \
__CALL_IF(NUM_BITS, 4, N_BLOCKS, K_BLOCKS, -1, NUM_THREADS)
template <typename scalar_t>
void marlin_mm_f16i4(const void* A, const void* B, void* C, void* s, int prob_m,
int prob_n, int prob_k, void* workspace, int num_bits,
int num_groups, int group_size, int dev,
cudaStream_t stream, int thread_k, int thread_n, int sms,
int max_par) {
TORCH_CHECK(num_bits == 8, "num_bits must be 8. Got = ", num_bits);
TORCH_CHECK(prob_m > 0 && prob_n > 0 && prob_k > 0, "Invalid MNK = [", prob_m,
", ", prob_n, ", ", prob_k, "]");
int tot_m = prob_m;
int tot_m_blocks = div_ceil(tot_m, 16);
int pad = 16 * tot_m_blocks - tot_m;
if (sms == -1) {
cudaDeviceGetAttribute(&sms, cudaDevAttrMultiProcessorCount, dev);
}
int max_shared_mem = 0;
cudaDeviceGetAttribute(&max_shared_mem,
cudaDevAttrMaxSharedMemoryPerBlockOptin, dev);
TORCH_CHECK(max_shared_mem > 0);
// Set thread config
exec_config_t exec_cfg;
if (thread_k != -1 && thread_n != -1) {
// User-defined config
exec_cfg =
exec_config_t{4, thread_config_t{thread_k, thread_n, default_threads}};
} else {
// Auto config
exec_cfg = determine_thread_config(prob_m, prob_n, prob_k, num_bits,
group_size, max_shared_mem);
}
TORCH_CHECK(
exec_cfg.max_m_blocks > 0 &&
is_valid_config(exec_cfg.tb_cfg, exec_cfg.max_m_blocks, prob_m,
prob_n, prob_k, num_bits, group_size, max_shared_mem),
"Invalid thread config: max_m_blocks = ", exec_cfg.max_m_blocks,
", thread_k = ", exec_cfg.tb_cfg.thread_k,
", thread_n = ", exec_cfg.tb_cfg.thread_n,
", num_threads = ", exec_cfg.tb_cfg.num_threads, " for MKN = [", prob_m,
", ", prob_k, ", ", prob_n, "] and num_bits = ", num_bits,
", group_size = ", group_size, ", max_shared_mem = ", max_shared_mem);
int num_threads = exec_cfg.tb_cfg.num_threads;
thread_k = exec_cfg.tb_cfg.thread_k;
thread_n = exec_cfg.tb_cfg.thread_n;
int thread_k_blocks = thread_k / 16;
int thread_n_blocks = thread_n / 16;
int blocks = sms;
TORCH_CHECK(prob_n % thread_n == 0, "prob_n = ", prob_n,
" is not divisible by thread_n = ", thread_n);
TORCH_CHECK(prob_k % thread_k == 0, "prob_k = ", prob_k,
" is not divisible by thread_k = ", thread_k);
int group_blocks = -1;
const int4* A_ptr = (const int4*)A;
const int4* B_ptr = (const int4*)B;
int4* C_ptr = (int4*)C;
const int4* s_ptr = (const int4*)s;
int* locks = (int*)workspace;
// Main loop
for (int i = 0; i < tot_m_blocks; i += exec_cfg.max_m_blocks) {
int thread_m_blocks = tot_m_blocks - i;
prob_m = tot_m - 16 * i;
int par = 1;
if (thread_m_blocks > exec_cfg.max_m_blocks) {
// Note that parallel > 1 currently only works for inputs without any
// padding
par = (16 * thread_m_blocks - pad) / (16 * exec_cfg.max_m_blocks);
if (par > max_par) par = max_par;
prob_m = (16 * exec_cfg.max_m_blocks) * par;
i += exec_cfg.max_m_blocks * (par - 1);
thread_m_blocks = exec_cfg.max_m_blocks;
}
// Define kernel configurations
if (false) {
}
CALL_IF(8, 32, 2, 256)
CALL_IF(8, 16, 4, 256)
CALL_IF(8, 8, 8, 256)
CALL_IF(8, 8, 4, 128)
CALL_IF(8, 4, 8, 128)
else {
TORCH_CHECK(false, "Unsupported shapes: MNK = [" + str(prob_m) + ", " +
str(prob_n) + ", " + str(prob_k) + "]" +
", num_groups = " + str(num_groups) +
", group_size = " + str(group_size) +
", thread_m_blocks = " + str(thread_m_blocks) +
", thread_n_blocks = " + str(thread_n_blocks) +
", thread_k_blocks = " + str(thread_k_blocks));
}
A_ptr += 16 * thread_m_blocks * (prob_k / 8) * par;
C_ptr += 16 * thread_m_blocks * (prob_n / 8) * par;
}
}
} // namespace fp8_marlin
torch::Tensor fp8_marlin_gemm(torch::Tensor& a, torch::Tensor& b_q_weight,
torch::Tensor& b_scales, torch::Tensor& workspace,
int64_t num_bits, int64_t size_m, int64_t size_n,
int64_t size_k) {
// Verify num_bits
TORCH_CHECK(num_bits == 8, "num_bits must be 8. Got = ", num_bits);
int pack_factor = 32 / num_bits;
// Verify A
TORCH_CHECK(a.size(0) == size_m, "Shape mismatch: a.size(0) = ", a.size(0),
", size_m = ", size_m);
TORCH_CHECK(a.size(1) == size_k, "Shape mismatch: a.size(1) = ", a.size(1),
", size_k = ", size_k);
// Verify B
TORCH_CHECK(size_k % marlin::tile_size == 0, "size_k = ", size_k,
" is not divisible by tile_size = ", marlin::tile_size);
TORCH_CHECK((size_k / marlin::tile_size) == b_q_weight.size(0),
"Shape mismatch: b_q_weight.size(0) = ", b_q_weight.size(0),
", size_k = ", size_k, ", tile_size = ", marlin::tile_size);
TORCH_CHECK(b_q_weight.size(1) % marlin::tile_size == 0,
"b_q_weight.size(1) = ", b_q_weight.size(1),
" is not divisible by tile_size = ", marlin::tile_size);
int actual_size_n = (b_q_weight.size(1) / marlin::tile_size) * pack_factor;
TORCH_CHECK(size_n == actual_size_n, "size_n = ", size_n,
", actual_size_n = ", actual_size_n);
// Verify device and strides
TORCH_CHECK(a.device().is_cuda(), "A is not on GPU");
TORCH_CHECK(a.is_contiguous(), "A is not contiguous");
TORCH_CHECK(b_q_weight.device().is_cuda(), "b_q_weight is not on GPU");
TORCH_CHECK(b_q_weight.is_contiguous(), "b_q_weight is not contiguous");
TORCH_CHECK(b_scales.device().is_cuda(), "b_scales is not on GPU");
TORCH_CHECK(b_scales.is_contiguous(), "b_scales is not contiguous");
// Alloc buffers
const at::cuda::OptionalCUDAGuard device_guard(device_of(a));
auto options = torch::TensorOptions().dtype(a.dtype()).device(a.device());
torch::Tensor c = torch::empty({size_m, size_n}, options);
// thread_k: `k` size of a thread_tile in `weights` (can usually be left as
// auto -1)
int thread_k = -1;
// thread_n: `n` size of a thread_tile in `weights` (can usually be left as
// auto -1)
int thread_n = -1;
// sms: number of SMs to use for the kernel (can usually be left as auto -1)
int sms = -1;
// Detect groupsize and act_order
int num_groups = -1;
int group_size = -1;
int b_rank = b_scales.sizes().size();
TORCH_CHECK(b_rank == 2, "b_scales rank = ", b_rank, " is not 2");
TORCH_CHECK(b_scales.size(1) == size_n, "b_scales dim 1 = ", b_scales.size(1),
" is not size_n = ", size_n);
// Channelwise only for FP8
TORCH_CHECK(b_scales.size(0) == 1)
num_groups = b_scales.size(0);
// Verify workspace size
TORCH_CHECK(size_n % marlin::min_thread_n == 0, "size_n = ", size_n,
", is not divisible by min_thread_n = ", marlin::min_thread_n);
int min_workspace_size = (size_n / marlin::min_thread_n) * marlin::max_par;
TORCH_CHECK(workspace.numel() >= min_workspace_size,
"workspace.numel = ", workspace.numel(),
" is below min_workspace_size = ", min_workspace_size);
int dev = a.get_device();
if (a.scalar_type() == at::ScalarType::Half) {
fp8_marlin::marlin_mm_f16i4<half>(
a.data_ptr<at::Half>(), b_q_weight.data_ptr(), c.data_ptr<at::Half>(),
b_scales.data_ptr<at::Half>(), size_m, size_n, size_k,
workspace.data_ptr(), num_bits, num_groups, group_size, dev,
at::cuda::getCurrentCUDAStream(dev), thread_k, thread_n, sms,
marlin::max_par);
} else if (a.scalar_type() == at::ScalarType::BFloat16) {
fp8_marlin::marlin_mm_f16i4<nv_bfloat16>(
a.data_ptr<at::BFloat16>(), b_q_weight.data_ptr(),
c.data_ptr<at::BFloat16>(), b_scales.data_ptr<at::BFloat16>(), size_m,
size_n, size_k, workspace.data_ptr(), num_bits, num_groups, group_size,
dev, at::cuda::getCurrentCUDAStream(dev), thread_k, thread_n, sms,
marlin::max_par);
} else {
TORCH_CHECK(false, "fp8_marlin_gemm only supports bfloat16 and float16");
}
return c;
}
#endif
TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
m.impl("fp8_marlin_gemm", &fp8_marlin_gemm);
}
\ No newline at end of file
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment