Files
ktransformers/kt-kernel/operators/amx/fp8-moe.hpp
Oql dc6394e501 [fix]: fix moe hpp bug. (#1780)
fix moe hpp init bug.
2026-01-04 19:32:56 +08:00

776 lines
36 KiB
C++

/**
* @Description : FP8 AMX MoE operator for DeepSeek V3.2 native inference
* @Author : oql, Codex and Claude
* @Date : 2025-12-09
* @Version : 1.0.0
* @Copyright (c) 2024 by KVCache.AI, All Rights Reserved.
*
* This file implements FP8 MoE using CRTP pattern, inheriting from moe_base.hpp.
* FP8 weights are stored with 128x128 block-wise scales.
**/
#ifndef CPUINFER_OPERATOR_AMX_FP8_MOE_H
#define CPUINFER_OPERATOR_AMX_FP8_MOE_H
// #define DEBUG_FP8_MOE
#include "la/amx_raw_buffers.hpp"
#include "la/amx_raw_kernels.hpp"
#include "moe_base.hpp"
/**
* @brief FP8 MoE operator using CRTP pattern
* @tparam T Kernel type, defaults to GemmKernel224FP8
*
* This class provides FP8-specific implementations:
* - do_gate_up_gemm, do_down_gemm : FP8 weight -> BF16 conversion mat mul
* - load_weights: Load FP8 weights with 128x128 block scales
*/
template <class T = amx::GemmKernel224FP8>
class AMX_FP8_MOE_TP : public AMX_MOE_BASE<T, AMX_FP8_MOE_TP<T>> {
using Base = AMX_MOE_BASE<T, AMX_FP8_MOE_TP<T>>;
using Base::config_;
using Base::down_ba_;
using Base::down_bb_;
using Base::down_bc_;
using Base::gate_bb_;
using Base::gate_bc_;
using Base::gate_up_ba_;
using Base::m_local_num_;
using Base::tp_part_idx;
using Base::up_bb_;
using Base::up_bc_;
public:
using typename Base::input_t;
using typename Base::output_t;
AMX_FP8_MOE_TP() = default;
AMX_FP8_MOE_TP(GeneralMOEConfig config, int tp_part_idx_ = 0) : Base(config, tp_part_idx_) {
// Initialization now happens in derived_init() which is called by base constructor
}
void derived_init() {
auto& quant_config = config_.quant_config;
if (quant_config.group_size == 0 || quant_config.zero_point) {
throw std::runtime_error("KT-Kernel fp8 MoE only support block-wise FP8");
}
printf("Created AMX_FP8_MOE_TP %d at numa %d\n", tp_part_idx, numa_node_of_cpu(sched_getcpu()));
}
~AMX_FP8_MOE_TP() = default;
// ============================================================================
// CRTP buffer creation - with group_size
// ============================================================================
size_t buffer_a_required_size_impl(size_t m, size_t k) const { return T::BufferA::required_size(m, k); }
size_t buffer_b_required_size_impl(size_t n, size_t k) const {
return T::BufferB::required_size(n, k, config_.quant_config.group_size);
}
size_t buffer_c_required_size_impl(size_t m, size_t n) const { return T::BufferC::required_size(m, n); }
std::shared_ptr<typename T::BufferA> make_buffer_a_impl(size_t m, size_t k, void* data) const {
return std::make_shared<typename T::BufferA>(m, k, data);
}
std::shared_ptr<typename T::BufferB> make_buffer_b_impl(size_t n, size_t k, void* data) const {
return std::make_shared<typename T::BufferB>(n, k, config_.quant_config.group_size, data);
}
std::shared_ptr<typename T::BufferC> make_buffer_c_impl(size_t m, size_t n, void* data) const {
return std::make_shared<typename T::BufferC>(m, n, data);
}
// ============================================================================
// CRTP virtual points - GEMM dispatch
// ============================================================================
void do_gate_up_gemm(bool do_up, int expert_idx, int ith, int nth, int qlen) {
auto& group_size = config_.quant_config.group_size;
int m = m_local_num_[expert_idx];
auto& ba = gate_up_ba_[expert_idx];
auto& bb = do_up ? up_bb_[expert_idx] : gate_bb_[expert_idx];
auto& bc = do_up ? up_bc_[expert_idx] : gate_bc_[expert_idx];
amx::vec_mul_kgroup(m, config_.intermediate_size, config_.hidden_size, group_size, ba, bb, bc, ith, nth);
}
void do_down_gemm(int expert_idx, int ith, int nth, int qlen) {
auto& group_size = config_.quant_config.group_size;
int m = m_local_num_[expert_idx];
amx::vec_mul_kgroup(m, config_.hidden_size, config_.intermediate_size, group_size, down_ba_[expert_idx],
down_bb_[expert_idx], down_bc_[expert_idx], ith, nth);
}
#ifdef DEBUG_FP8_MOE
// Function to dump Buffer B data for debugging FP8 quantization results
inline void dump_buffer_b(const std::string& quantization_type, int expert_idx, const std::string& matrix_type,
typename T::BufferB* buffer) {
auto& quant_config = config_.quant_config;
int& group_size = quant_config.group_size;
printf("[DUMP_BUFFER_B] TP%d %s Expert%d %s:\n", tp_part_idx, quantization_type.c_str(), expert_idx,
matrix_type.c_str());
// Calculate dimensions based on matrix type
int rows, cols;
size_t scale_elem_count;
if (matrix_type == "gate" || matrix_type == "up") {
rows = config_.intermediate_size;
cols = config_.hidden_size;
} else { // down
rows = config_.hidden_size;
cols = config_.intermediate_size;
}
int n_blocks_n = (rows + group_size - 1) / group_size;
int n_blocks_k = (cols + group_size - 1) / group_size;
scale_elem_count = n_blocks_n * n_blocks_k;
// Dump scales (as BF16 converted to float)
printf(" Scales[first 16]: ");
for (int i = 0; i < std::min(16, (int)scale_elem_count); i++) {
printf("%.6f ", buffer->d[i]);
}
printf("\n");
if (scale_elem_count > 16) {
printf(" Scales[last 16]: ");
int start_idx = std::max(0, (int)scale_elem_count - 16);
for (int i = start_idx; i < (int)scale_elem_count; i++) {
printf("%.6f ", buffer->d[i]);
}
printf("\n");
}
// Dump FP8 weights (as hex uint8)
size_t weight_size = (size_t)rows * cols; // FP8 is 1 byte per element
uint8_t* weight_ptr = (uint8_t*)buffer->b;
printf(" FP8 Weights[first 32 bytes]: ");
for (int i = 0; i < std::min(32, (int)weight_size); i++) {
printf("%02x ", weight_ptr[i]);
}
printf("\n");
if (weight_size > 32) {
printf(" FP8 Weights[last 32 bytes]: ");
int start_idx = std::max(32, (int)weight_size - 32);
for (int i = start_idx; i < (int)weight_size; i++) {
printf("%02x ", weight_ptr[i]);
}
printf("\n");
}
printf(" Matrix dimensions: %dx%d (n x k), Scale blocks: %dx%d, Group size: %d, Scale elements: %zu\n", rows, cols,
n_blocks_n, n_blocks_k, group_size, scale_elem_count);
}
#endif
/**
* @brief Load FP8 weights from contiguous memory layout
*
* Loads weights from config_.gate_proj, up_proj, down_proj with scales
* from config_.gate_scale, up_scale, down_scale.
*/
void load_weights() {
auto& quant_config = config_.quant_config;
int& group_size = quant_config.group_size;
const uint64_t* physical_to_logical_map = (const uint64_t*)config_.physical_to_logical_map;
auto pool = config_.pool->get_subpool(tp_part_idx);
if (config_.gate_scale == nullptr) {
throw std::runtime_error("FP8 AVX MOE only support native weight.");
}
// load weight
int nth = T::recommended_nth(config_.intermediate_size);
pool->do_work_stealing_job(
nth * config_.expert_num, nullptr,
[this, nth, physical_to_logical_map, group_size](int task_id) {
uint64_t expert_idx = task_id / nth;
uint64_t logical_expert_id = expert_map(physical_to_logical_map, expert_idx);
int ith = task_id % nth;
// gate part
gate_bb_[expert_idx]->from_mat(
(uint8_t*)config_.gate_proj + (logical_expert_id * config_.intermediate_size * config_.hidden_size),
(float*)config_.gate_scale +
(logical_expert_id * (config_.hidden_size / group_size) * (config_.intermediate_size / group_size)),
ith, nth);
// up part
up_bb_[expert_idx]->from_mat(
(uint8_t*)config_.up_proj + (logical_expert_id * config_.intermediate_size * config_.hidden_size),
(float*)config_.up_scale +
(logical_expert_id * (config_.hidden_size / group_size) * (config_.intermediate_size / group_size)),
ith, nth);
},
nullptr);
nth = T::recommended_nth(config_.hidden_size);
pool->do_work_stealing_job(
nth * config_.expert_num, nullptr,
[this, nth, physical_to_logical_map, group_size](int task_id) {
uint64_t expert_idx = task_id / nth;
uint64_t logical_expert_id = expert_map(physical_to_logical_map, expert_idx);
int ith = task_id % nth;
// down part
down_bb_[expert_idx]->from_mat(
(uint8_t*)config_.down_proj + (logical_expert_id * config_.intermediate_size * config_.hidden_size),
(float*)config_.down_scale +
(logical_expert_id * (config_.hidden_size / group_size) * (config_.intermediate_size / group_size)),
ith, nth);
},
nullptr);
#ifdef DEBUG_FP8_MOE
dump_buffer_b("Native FP8", 0, "gate", gate_bb_[0].get());
dump_buffer_b("Native FP8", 0, "down", down_bb_[0].get());
#endif
}
// Fast 64-byte (512-bit) memcpy using AVX512
static inline void fast_memcpy_64(void* __restrict dst, const void* __restrict src) {
__m512i data = _mm512_loadu_si512(src);
_mm512_storeu_si512(dst, data);
}
// Fast memcpy for arbitrary sizes using AVX512
static inline void fast_memcpy(void* __restrict dst, const void* __restrict src, size_t bytes) {
uint8_t* d = (uint8_t*)dst;
const uint8_t* s = (const uint8_t*)src;
size_t chunks = bytes / 64;
for (size_t i = 0; i < chunks; i++) {
fast_memcpy_64(d, s);
d += 64;
s += 64;
}
bytes -= chunks * 64;
if (bytes > 0) {
std::memcpy(d, s, bytes);
}
}
/**
* @brief Unpack a single N_STEP x K_STEP block from packed BufferB format to n-major format
*
* This is the inverse of the packing done in BufferBFP8Impl::from_mat.
* Optimized with AVX512 gather for efficient non-contiguous reads.
*
* @param src Pointer to packed data (N_STEP * K_STEP bytes in packed layout)
* @param dst Pointer to destination in n-major layout
* @param dst_row_stride Row stride in destination buffer (number of columns in full matrix)
*/
static inline void unpack_nk_block(const uint8_t* src, uint8_t* dst, size_t dst_row_stride) {
// row_map[packed_i] gives the base row for packed index packed_i
static constexpr int row_map[8] = {0, 16, 4, 20, 8, 24, 12, 28};
const uint64_t* src64 = reinterpret_cast<const uint64_t*>(src);
// Gather indices: src64[8*j + packed_i] for j = 0..7
// Offsets in uint64 units: 0, 8, 16, 24, 32, 40, 48, 56 (+ packed_i for each group)
const __m512i gather_offsets = _mm512_set_epi64(56, 48, 40, 32, 24, 16, 8, 0);
// Process each packed group (8 groups of 4 rows each = 32 rows total)
for (int packed_i = 0; packed_i < 8; packed_i++) {
const int base_row = row_map[packed_i];
const uint64_t* base_src = src64 + packed_i;
// Gather 8 values for j=0..7 and j=8..15
__m512i vals_0_7 = _mm512_i64gather_epi64(gather_offsets, base_src, 8);
__m512i vals_8_15 = _mm512_i64gather_epi64(gather_offsets, base_src + 64, 8);
// Extract 4 rows from each set of 8 values
// Row 0: bits 0-15
__m128i row0_lo = _mm512_cvtepi64_epi16(_mm512_and_si512(vals_0_7, _mm512_set1_epi64(0xFFFF)));
__m128i row0_hi = _mm512_cvtepi64_epi16(_mm512_and_si512(vals_8_15, _mm512_set1_epi64(0xFFFF)));
// Row 1: bits 16-31
__m128i row1_lo =
_mm512_cvtepi64_epi16(_mm512_and_si512(_mm512_srli_epi64(vals_0_7, 16), _mm512_set1_epi64(0xFFFF)));
__m128i row1_hi =
_mm512_cvtepi64_epi16(_mm512_and_si512(_mm512_srli_epi64(vals_8_15, 16), _mm512_set1_epi64(0xFFFF)));
// Row 2: bits 32-47
__m128i row2_lo =
_mm512_cvtepi64_epi16(_mm512_and_si512(_mm512_srli_epi64(vals_0_7, 32), _mm512_set1_epi64(0xFFFF)));
__m128i row2_hi =
_mm512_cvtepi64_epi16(_mm512_and_si512(_mm512_srli_epi64(vals_8_15, 32), _mm512_set1_epi64(0xFFFF)));
// Row 3: bits 48-63
__m128i row3_lo = _mm512_cvtepi64_epi16(_mm512_srli_epi64(vals_0_7, 48));
__m128i row3_hi = _mm512_cvtepi64_epi16(_mm512_srli_epi64(vals_8_15, 48));
// Store 32 bytes (16 x uint16) to each row
// Combine two 128-bit values into 256-bit for more efficient stores
uint8_t* row0_dst = dst + (size_t)base_row * dst_row_stride;
uint8_t* row1_dst = dst + (size_t)(base_row + 1) * dst_row_stride;
uint8_t* row2_dst = dst + (size_t)(base_row + 2) * dst_row_stride;
uint8_t* row3_dst = dst + (size_t)(base_row + 3) * dst_row_stride;
// Combine lo and hi into 256-bit and store
__m256i row0_256 = _mm256_set_m128i(row0_hi, row0_lo);
__m256i row1_256 = _mm256_set_m128i(row1_hi, row1_lo);
__m256i row2_256 = _mm256_set_m128i(row2_hi, row2_lo);
__m256i row3_256 = _mm256_set_m128i(row3_hi, row3_lo);
_mm256_storeu_si256((__m256i*)row0_dst, row0_256);
_mm256_storeu_si256((__m256i*)row1_dst, row1_256);
_mm256_storeu_si256((__m256i*)row2_dst, row2_256);
_mm256_storeu_si256((__m256i*)row3_dst, row3_256);
}
}
/**
* @brief Unpack 4 consecutive N_STEP x K_STEP blocks to maximize cache line utilization
*
* Processing 4 blocks together means each row write is 128 bytes = 2 cache lines,
* which greatly improves write efficiency compared to 32 bytes per row.
*
* @param src Array of 4 source pointers (each pointing to a 32x32 packed block)
* @param dst Destination pointer in n-major layout
* @param dst_row_stride Row stride in destination buffer
*/
static inline void unpack_4nk_blocks(const uint8_t* src[4], uint8_t* dst, size_t dst_row_stride) {
static constexpr int row_map[8] = {0, 16, 4, 20, 8, 24, 12, 28};
constexpr int K_STEP = T::K_STEP; // 32
// Reinterpret as uint64 arrays for efficient access
const uint64_t* src0 = reinterpret_cast<const uint64_t*>(src[0]);
const uint64_t* src1 = reinterpret_cast<const uint64_t*>(src[1]);
const uint64_t* src2 = reinterpret_cast<const uint64_t*>(src[2]);
const uint64_t* src3 = reinterpret_cast<const uint64_t*>(src[3]);
// Process all 32 rows, writing 128 bytes (4 x 32) per row
for (int packed_i = 0; packed_i < 8; packed_i++) {
const int base_row = row_map[packed_i];
// Process 4 rows at a time
for (int r = 0; r < 4; r++) {
uint16_t* row_dst = reinterpret_cast<uint16_t*>(dst + (size_t)(base_row + r) * dst_row_stride);
const int shift = r * 16;
// Unroll: process all 4 blocks x 16 columns = 64 uint16 values
// Block 0: columns 0-15
for (int j = 0; j < 16; j++) {
row_dst[j] = static_cast<uint16_t>(src0[8 * j + packed_i] >> shift);
}
// Block 1: columns 16-31
for (int j = 0; j < 16; j++) {
row_dst[16 + j] = static_cast<uint16_t>(src1[8 * j + packed_i] >> shift);
}
// Block 2: columns 32-47
for (int j = 0; j < 16; j++) {
row_dst[32 + j] = static_cast<uint16_t>(src2[8 * j + packed_i] >> shift);
}
// Block 3: columns 48-63
for (int j = 0; j < 16; j++) {
row_dst[48 + j] = static_cast<uint16_t>(src3[8 * j + packed_i] >> shift);
}
}
}
}
/**
* @brief Reconstruct weights for a single expert to the output buffers (no temp buffer version)
*
* Directly unpacks from packed BufferB format to n-major GPU buffers without intermediate storage.
* Optimized version with coarse-grained task splitting for better cache utilization.
*
* Key optimizations:
* - Reduced task count (~40 vs ~350) to minimize scheduling overhead
* - Larger chunks per task for better cache line utilization
* - Process multiple N_STEPs per task for better write locality
*
* @param gpu_tp_count Number of GPU TP parts (1, 2, 4, or 8)
* @param cpu_tp_count Number of CPU TP parts
* @param expert_id Expert index to process
* @param full_config Full configuration (before CPU TP split)
* @param w13_weight_ptrs Pointers to gate+up weight buffers (one per GPU TP)
* @param w13_scale_ptrs Pointers to gate+up scale buffers (one per GPU TP)
* @param w2_weight_ptrs Pointers to down weight buffers (one per GPU TP)
* @param w2_scale_ptrs Pointers to down scale buffers (one per GPU TP)
*/
void write_weights_to_buffer(int gpu_tp_count, [[maybe_unused]] int cpu_tp_count, int expert_id,
const GeneralMOEConfig& full_config, const std::vector<uintptr_t>& w13_weight_ptrs,
const std::vector<uintptr_t>& w13_scale_ptrs,
const std::vector<uintptr_t>& w2_weight_ptrs,
const std::vector<uintptr_t>& w2_scale_ptrs) const {
auto& config = config_;
const int group_size = config.quant_config.group_size;
auto pool = config.pool->get_subpool(tp_part_idx);
constexpr int N_STEP = T::N_STEP;
constexpr int K_STEP = T::K_STEP;
constexpr int N_BLOCK = T::N_BLOCK;
constexpr int K_BLOCK = T::K_BLOCK;
// ========= W13 (gate+up): Shape [intermediate, hidden], split by N only =========
const int cpu_n_w13 = config.intermediate_size;
const int cpu_k_w13 = config.hidden_size;
const int gpu_n_w13 = full_config.intermediate_size / gpu_tp_count;
const int gpu_k_w13 = full_config.hidden_size;
const int global_n_offset_w13 = tp_part_idx * cpu_n_w13;
const size_t gpu_w13_weight_per_mat = (size_t)gpu_n_w13 * gpu_k_w13;
const size_t gpu_w13_scale_per_mat = (size_t)div_up(gpu_n_w13, group_size) * div_up(gpu_k_w13, group_size);
const int cpu_scale_k_blocks_w13 = div_up(cpu_k_w13, group_size);
const int gpu_scale_k_blocks_w13 = div_up(gpu_k_w13, group_size);
// ========= W2 (down): Shape [hidden, intermediate], split by K =========
const int cpu_n_w2 = config.hidden_size;
const int cpu_k_w2 = config.intermediate_size;
const int gpu_n_w2 = full_config.hidden_size;
const int gpu_k_w2 = full_config.intermediate_size / gpu_tp_count;
const int global_k_offset_w2 = tp_part_idx * cpu_k_w2;
const size_t gpu_w2_weight_per_mat = (size_t)gpu_n_w2 * gpu_k_w2;
const size_t gpu_w2_scale_per_mat = (size_t)div_up(gpu_n_w2, group_size) * div_up(gpu_k_w2, group_size);
const int cpu_scale_k_blocks_w2 = div_up(cpu_k_w2, group_size);
const int gpu_scale_k_blocks_w2 = div_up(gpu_k_w2, group_size);
// ========= Scale dimensions =========
const int cpu_scale_n_blocks_w13 = div_up(cpu_n_w13, group_size);
const int gpu_scale_n_blocks_w13 = div_up(gpu_n_w13, group_size);
const int cpu_scale_n_blocks_w2 = div_up(cpu_n_w2, group_size);
// ========= Optimized job layout =========
// Use task count slightly above CPU core count for good work stealing
// For 80-core system, ~100 tasks provides good balance
constexpr int NUM_W13_TASKS = 32; // Per matrix (gate or up), total 64 for w13
constexpr int NUM_W2_TASKS = 32; // For down matrix
constexpr int SCALE_TASKS = 3; // gate_scale, up_scale, down_scale
const int total_tasks = NUM_W13_TASKS * 2 + NUM_W2_TASKS + SCALE_TASKS;
// Calculate N_STEP blocks per task (must be N_STEP aligned for correct BufferB addressing)
const int w13_n_steps = div_up(cpu_n_w13, N_STEP);
const int w13_steps_per_task = div_up(w13_n_steps, NUM_W13_TASKS);
const int w2_n_steps = div_up(cpu_n_w2, N_STEP);
const int w2_steps_per_task = div_up(w2_n_steps, NUM_W2_TASKS);
pool->do_work_stealing_job(
total_tasks, nullptr,
[=, &w13_weight_ptrs, &w13_scale_ptrs, &w2_weight_ptrs, &w2_scale_ptrs, this](int task_id) {
if (task_id < NUM_W13_TASKS * 2) {
// ========= W13 weight task: process chunk of rows x full K =========
const bool is_up = task_id >= NUM_W13_TASKS;
const int chunk_idx = task_id % NUM_W13_TASKS;
const auto& bb = is_up ? up_bb_[expert_id] : gate_bb_[expert_id];
// Calculate row range for this task (N_STEP aligned)
const int step_start = chunk_idx * w13_steps_per_task;
const int step_end = std::min(step_start + w13_steps_per_task, w13_n_steps);
if (step_start >= w13_n_steps) return;
const int chunk_n_start = step_start * N_STEP;
const int chunk_n_end = std::min(step_end * N_STEP, cpu_n_w13);
// Process each N_STEP within this chunk
for (int local_n_start = chunk_n_start; local_n_start < chunk_n_end; local_n_start += N_STEP) {
// Calculate GPU target and offset for each N_STEP (may cross GPU TP boundaries)
const int global_n = global_n_offset_w13 + local_n_start;
const int target_gpu = global_n / gpu_n_w13;
const int n_in_gpu = global_n % gpu_n_w13;
uint8_t* weight_base = (uint8_t*)w13_weight_ptrs[target_gpu];
// Pointer already points to current expert's location, only add offset for up matrix
const size_t expert_weight_off = is_up ? gpu_w13_weight_per_mat : 0;
// Calculate N_BLOCK info for source addressing
const int n_block_idx = local_n_start / N_BLOCK;
const int n_block_begin = n_block_idx * N_BLOCK;
const int n_block_size = std::min(N_BLOCK, cpu_n_w13 - n_block_begin);
const int n_in_block = local_n_start - n_block_begin;
// Process all K in groups of 4 K_STEPs when possible for cache efficiency
for (int k_block_begin = 0; k_block_begin < cpu_k_w13; k_block_begin += K_BLOCK) {
const int k_block_size = std::min(K_BLOCK, cpu_k_w13 - k_block_begin);
// Try to process 4 K_STEPs at once (128 columns = 2 cache lines per row)
int k_begin = 0;
for (; k_begin + 4 * K_STEP <= k_block_size; k_begin += 4 * K_STEP) {
const uint8_t* src_ptrs[4];
for (int i = 0; i < 4; i++) {
src_ptrs[i] = bb->b + (size_t)n_block_begin * cpu_k_w13 + (size_t)k_block_begin * n_block_size +
(size_t)n_in_block * k_block_size + (size_t)(k_begin + i * K_STEP) * N_STEP;
}
uint8_t* dst =
weight_base + expert_weight_off + (size_t)n_in_gpu * gpu_k_w13 + k_block_begin + k_begin;
unpack_4nk_blocks(src_ptrs, dst, gpu_k_w13);
}
// Handle remaining K_STEPs one by one
for (; k_begin < k_block_size; k_begin += K_STEP) {
const uint8_t* src = bb->b + (size_t)n_block_begin * cpu_k_w13 +
(size_t)k_block_begin * n_block_size + (size_t)n_in_block * k_block_size +
(size_t)k_begin * N_STEP;
uint8_t* dst =
weight_base + expert_weight_off + (size_t)n_in_gpu * gpu_k_w13 + k_block_begin + k_begin;
unpack_nk_block(src, dst, gpu_k_w13);
}
}
}
} else if (task_id < NUM_W13_TASKS * 2 + NUM_W2_TASKS) {
// ========= W2 weight task: process chunk of rows x all K slices =========
const int chunk_idx = task_id - NUM_W13_TASKS * 2;
const auto& bb = down_bb_[expert_id];
// Calculate row range for this task (N_STEP aligned)
const int step_start = chunk_idx * w2_steps_per_task;
const int step_end = std::min(step_start + w2_steps_per_task, w2_n_steps);
if (step_start >= w2_n_steps) return;
const int chunk_n_start = step_start * N_STEP;
const int chunk_n_end = std::min(step_end * N_STEP, cpu_n_w2);
// Process each N_STEP within this chunk
for (int local_n_start = chunk_n_start; local_n_start < chunk_n_end; local_n_start += N_STEP) {
// Calculate N_BLOCK info for source addressing
const int n_block_idx = local_n_start / N_BLOCK;
const int n_block_begin = n_block_idx * N_BLOCK;
const int n_block_size = std::min(N_BLOCK, cpu_n_w2 - n_block_begin);
const int n_in_block = local_n_start - n_block_begin;
// Process all K slices (each slice goes to a different GPU TP)
for (int k_slice_start = 0; k_slice_start < cpu_k_w2; k_slice_start += gpu_k_w2) {
const int k_slice_end = std::min(k_slice_start + gpu_k_w2, cpu_k_w2);
const int global_k_start = global_k_offset_w2 + k_slice_start;
const int target_gpu = global_k_start / gpu_k_w2;
const int k_in_gpu_base = global_k_start % gpu_k_w2;
uint8_t* weight_base = (uint8_t*)w2_weight_ptrs[target_gpu];
// Pointer already points to current expert's location
const size_t expert_weight_off = 0;
// Process K within this slice, trying 4 K_STEPs at once when aligned
for (int k_abs = k_slice_start; k_abs < k_slice_end;) {
const int k_block_idx = k_abs / K_BLOCK;
const int k_block_begin = k_block_idx * K_BLOCK;
const int k_block_size = std::min(K_BLOCK, cpu_k_w2 - k_block_begin);
const int k_in_block = k_abs - k_block_begin;
const int k_in_gpu = k_in_gpu_base + (k_abs - k_slice_start);
// Check if we can process 4 K_STEPs at once
const int remaining_in_block = k_block_size - k_in_block;
const int remaining_in_slice = k_slice_end - k_abs;
if (remaining_in_block >= 4 * K_STEP && remaining_in_slice >= 4 * K_STEP) {
const uint8_t* src_ptrs[4];
for (int i = 0; i < 4; i++) {
src_ptrs[i] = bb->b + (size_t)n_block_begin * cpu_k_w2 + (size_t)k_block_begin * n_block_size +
(size_t)n_in_block * k_block_size + (size_t)(k_in_block + i * K_STEP) * N_STEP;
}
uint8_t* dst = weight_base + expert_weight_off + (size_t)local_n_start * gpu_k_w2 + k_in_gpu;
unpack_4nk_blocks(src_ptrs, dst, gpu_k_w2);
k_abs += 4 * K_STEP;
} else {
const uint8_t* src = bb->b + (size_t)n_block_begin * cpu_k_w2 +
(size_t)k_block_begin * n_block_size + (size_t)n_in_block * k_block_size +
(size_t)k_in_block * N_STEP;
uint8_t* dst = weight_base + expert_weight_off + (size_t)local_n_start * gpu_k_w2 + k_in_gpu;
unpack_nk_block(src, dst, gpu_k_w2);
k_abs += K_STEP;
}
}
}
}
} else {
// ========= Scale copy task: simple linear copy with fast_memcpy =========
const int scale_task_id = task_id - NUM_W13_TASKS * 2 - NUM_W2_TASKS;
if (scale_task_id < 2) {
// Gate (0) or Up (1) scale copy
const bool is_up = scale_task_id == 1;
const auto& bb = is_up ? up_bb_[expert_id] : gate_bb_[expert_id];
// W13 scales: copy N blocks corresponding to this CPU TP
// Note: when gpu_tp > cpu_tp, scale blocks may span multiple GPU TPs
const int bn_start_global = global_n_offset_w13 / group_size;
for (int bn = 0; bn < cpu_scale_n_blocks_w13; bn++) {
const int global_bn = bn_start_global + bn;
const int target_gpu = global_bn / gpu_scale_n_blocks_w13;
const int gpu_bn = global_bn % gpu_scale_n_blocks_w13;
float* scale_dst = (float*)w13_scale_ptrs[target_gpu];
// Pointer already points to current expert's location, only add offset for up matrix
const size_t expert_scale_off = is_up ? gpu_w13_scale_per_mat : 0;
fast_memcpy(scale_dst + expert_scale_off + (size_t)gpu_bn * gpu_scale_k_blocks_w13,
bb->d + (size_t)bn * cpu_scale_k_blocks_w13, cpu_scale_k_blocks_w13 * sizeof(float));
}
} else {
// Down scale copy (scale_task_id == 2)
const auto& bb = down_bb_[expert_id];
// W2 scales: K dimension is split, copy to each GPU TP
for (int k_slice_idx = 0; k_slice_idx < div_up(cpu_k_w2, gpu_k_w2); k_slice_idx++) {
const int k_slice_start = k_slice_idx * gpu_k_w2;
const int k_slice_end = std::min(k_slice_start + gpu_k_w2, cpu_k_w2);
const int global_k_start = global_k_offset_w2 + k_slice_start;
const int target_gpu = global_k_start / gpu_k_w2;
const int bk_gpu_base = (global_k_start % gpu_k_w2) / group_size;
float* scale_dst = (float*)w2_scale_ptrs[target_gpu];
// Pointer already points to current expert's location
const size_t expert_scale_off = 0;
const int bk_start = k_slice_start / group_size;
const int bk_end = div_up(k_slice_end, group_size);
const int bk_count = bk_end - bk_start;
for (int bn = 0; bn < cpu_scale_n_blocks_w2; bn++) {
fast_memcpy(scale_dst + expert_scale_off + (size_t)bn * gpu_scale_k_blocks_w2 + bk_gpu_base,
bb->d + (size_t)bn * cpu_scale_k_blocks_w2 + bk_start, bk_count * sizeof(float));
}
}
}
}
},
nullptr);
}
};
template <typename K>
class TP_MOE<AMX_FP8_MOE_TP<K>> : public TP_MOE<AMX_MOE_BASE<K, AMX_FP8_MOE_TP<K>>> {
public:
using Base = TP_MOE<AMX_MOE_BASE<K, AMX_FP8_MOE_TP<K>>>;
using Base::Base;
void load_weights() override {
auto& config = this->config;
auto& tps = this->tps;
auto& tp_count = this->tp_count;
auto pool = config.pool;
const uint64_t* physical_to_logical_map = (const uint64_t*)config.physical_to_logical_map;
const int group_size = config.quant_config.group_size;
if (group_size == 0 || config.quant_config.zero_point) {
throw std::runtime_error("FP8 MoE only supports have group_size, zero_point=false");
}
if (config.gate_projs.empty() && config.gate_proj == nullptr) {
throw std::runtime_error("no weight source");
}
const bool use_per_expert_ptrs = !config.gate_projs.empty();
const size_t full_weight_elems = (size_t)config.intermediate_size * config.hidden_size;
const size_t full_scale_elems =
(size_t)div_up(config.hidden_size, group_size) * div_up(config.intermediate_size, group_size);
pool->dispense_backend()->do_numa_job([&, this](int i) {
auto& tpc = tps[i]->config_;
const size_t tp_weight_elems = (size_t)tpc.intermediate_size * tpc.hidden_size;
const size_t tp_scale_elems =
(size_t)div_up(tpc.intermediate_size, group_size) * div_up(tpc.hidden_size, group_size);
tpc.gate_proj = new uint8_t[tpc.expert_num * tp_weight_elems];
tpc.up_proj = new uint8_t[tpc.expert_num * tp_weight_elems];
tpc.down_proj = new uint8_t[tpc.expert_num * tp_weight_elems];
tpc.gate_scale = new float[tpc.expert_num * tp_scale_elems];
tpc.up_scale = new float[tpc.expert_num * tp_scale_elems];
tpc.down_scale = new float[tpc.expert_num * tp_scale_elems];
const size_t tp_idx = (size_t)i;
const size_t gate_up_weight_src_offset = i * tp_weight_elems;
const size_t gate_up_scale_src_offset = i * tp_scale_elems;
const size_t down_weight_src_col_offset = i * (size_t)tpc.intermediate_size;
const size_t down_scale_src_block_k_offset = down_weight_src_col_offset / (size_t)group_size;
pool->get_subpool(i)->do_work_stealing_job(
tpc.expert_num, nullptr,
[&, &tpc](int expert_id_) {
const size_t expert_id = expert_map(physical_to_logical_map, expert_id_);
uint8_t* gate_dst = (uint8_t*)tpc.gate_proj + expert_id * tp_weight_elems;
uint8_t* up_dst = (uint8_t*)tpc.up_proj + expert_id * tp_weight_elems;
uint8_t* down_dst = (uint8_t*)tpc.down_proj + expert_id * tp_weight_elems;
float* gate_scale_dst = (float*)tpc.gate_scale + expert_id * tp_scale_elems;
float* up_scale_dst = (float*)tpc.up_scale + expert_id * tp_scale_elems;
float* down_scale_dst = (float*)tpc.down_scale + expert_id * tp_scale_elems;
const uint8_t* gate_src;
const uint8_t* up_src;
const uint8_t* down_src;
const float* gate_scale_src;
const float* up_scale_src;
const float* down_scale_src;
if (use_per_expert_ptrs) {
gate_src = (const uint8_t*)config.gate_projs[0][expert_id] + gate_up_weight_src_offset;
up_src = (const uint8_t*)config.up_projs[0][expert_id] + gate_up_weight_src_offset;
down_src = (const uint8_t*)config.down_projs[0][expert_id];
gate_scale_src = (const float*)config.gate_scales[0][expert_id] + gate_up_scale_src_offset;
up_scale_src = (const float*)config.up_scales[0][expert_id] + gate_up_scale_src_offset;
down_scale_src = (const float*)config.down_scales[0][expert_id];
} else {
gate_src = (const uint8_t*)config.gate_proj + expert_id * full_weight_elems + gate_up_weight_src_offset;
up_src = (const uint8_t*)config.up_proj + expert_id * full_weight_elems + gate_up_weight_src_offset;
down_src = (const uint8_t*)config.down_proj + expert_id * full_weight_elems;
gate_scale_src =
(const float*)config.gate_scale + expert_id * full_scale_elems + gate_up_scale_src_offset;
up_scale_src = (const float*)config.up_scale + expert_id * full_scale_elems + gate_up_scale_src_offset;
down_scale_src = (const float*)config.down_scale + expert_id * full_scale_elems;
}
std::memcpy(gate_dst, gate_src, tp_weight_elems);
std::memcpy(up_dst, up_src, tp_weight_elems);
std::memcpy(gate_scale_dst, gate_scale_src, sizeof(float) * tp_scale_elems);
std::memcpy(up_scale_dst, up_scale_src, sizeof(float) * tp_scale_elems);
for (int row = 0; row < config.hidden_size; row++) {
const size_t src_row_offset = (size_t)row * (size_t)config.intermediate_size + down_weight_src_col_offset;
const size_t dst_row_offset = (size_t)row * (size_t)tpc.intermediate_size;
std::memcpy(down_dst + dst_row_offset, down_src + src_row_offset, (size_t)tpc.intermediate_size);
}
const int n_blocks_n = div_up(config.hidden_size, group_size);
const int full_n_blocks_k = div_up(config.intermediate_size, group_size);
const int tp_n_blocks_k = div_up(tpc.intermediate_size, group_size);
for (int bn = 0; bn < n_blocks_n; bn++) {
const float* src = down_scale_src + (size_t)bn * (size_t)full_n_blocks_k + down_scale_src_block_k_offset;
float* dst = down_scale_dst + (size_t)bn * (size_t)tp_n_blocks_k;
std::memcpy(dst, src, sizeof(float) * (size_t)tp_n_blocks_k);
}
},
nullptr);
});
DO_TPS_LOAD_WEIGHTS(pool);
pool->dispense_backend()->do_numa_job([&, this](int i) {
auto& tpc = tps[i]->config_;
delete[] (uint8_t*)tpc.gate_proj;
delete[] (uint8_t*)tpc.up_proj;
delete[] (uint8_t*)tpc.down_proj;
delete[] (float*)tpc.gate_scale;
delete[] (float*)tpc.up_scale;
delete[] (float*)tpc.down_scale;
});
this->weights_loaded = true;
}
void write_weight_scale_to_buffer(int gpu_tp_count, int expert_id, const std::vector<uintptr_t>& w13_weight_ptrs,
const std::vector<uintptr_t>& w13_scale_ptrs,
const std::vector<uintptr_t>& w2_weight_ptrs,
const std::vector<uintptr_t>& w2_scale_ptrs) {
if (this->weights_loaded == false) {
throw std::runtime_error("Not Loaded");
}
if (this->tps.empty()) {
throw std::runtime_error("No TP parts initialized");
}
if ((int)w13_weight_ptrs.size() != gpu_tp_count || (int)w13_scale_ptrs.size() != gpu_tp_count ||
(int)w2_weight_ptrs.size() != gpu_tp_count || (int)w2_scale_ptrs.size() != gpu_tp_count) {
throw std::runtime_error("Pointer arrays size must match gpu_tp_count");
}
this->config.pool->dispense_backend()->do_numa_job([&, this](int i) {
this->tps[i]->write_weights_to_buffer(gpu_tp_count, this->tp_count, expert_id, this->config, w13_weight_ptrs,
w13_scale_ptrs, w2_weight_ptrs, w2_scale_ptrs);
});
}
};
#endif // CPUINFER_OPERATOR_AMX_FP8_MOE_H