biondizzle/nvfp4-megamoe-kernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
cb2ca8591fb87dce58b5cd4e1ff3a4e36963efdf
nvfp4-megamoe-kernel/tests/unit/test_fmha_6warp_multirow.cu

311 lines
13 KiB
Plaintext
Raw Blame History

/**
* Test multi-row FMHA kernel (6-warp, T>1 prefill).
* Compile with -DHD_VAL=64 etc.
*
* Tests:
* 1. Single KV tile, T=1..128 (normalized output + LSE)
* 2. Multi-tile KV via Python merge (s_k=256, 2 segments)
* 3. Multi-head and batched launches
*/
#include <cuda_runtime.h>
#include <cstdio>
#include <cmath>
#include <cstdlib>
#include <cstring>
#ifndef HD_VAL
#define HD_VAL 64
#endif
#include "dsv4/kernels/attention/fmha_common.cuh"
#include "dsv4/kernels/attention/fmha_umma_desc.cuh"
using namespace dsv4::kernels::attention;
static bf16_t f32_to_bf16_host(float f) { uint32_t u; memcpy(&u,&f,4); return (uint16_t)(u>>16); }
static float bf16_to_f32_host(bf16_t h) { uint32_t u=(uint32_t)h<<16; float f; memcpy(&f,&u,4); return f; }
constexpr int HD = HD_VAL;
constexpr int SK = 128;
constexpr int MAX_T = 128;
#include "dsv4/kernels/attention/fmha_6warp_multirow.cuh"
static int compute_smem() {
size_t off = 0;
off += 4; // sTmemBase
off += 128 * sizeof(float); // sRowMax
off += 128 * sizeof(float); // sRowSum
off = (off + 127) & ~(size_t)127;
off += 128 * MMA_K_BF16 * sizeof(bf16_t); // sQ0
off += 128 * MMA_K_BF16 * sizeof(bf16_t); // sK0
off = (off + 127) & ~(size_t)127;
off += 128 * MMA_K_BF16 * sizeof(bf16_t); // sPk
off = (off + 127) & ~(size_t)127;
off += 16 * MMA_K_BF16 * sizeof(bf16_t); // sV
return (int)off;
}
static void reference_attention_multirow(
const bf16_t* q, const bf16_t* k, const bf16_t* v,
float* o_ref, float* lse_ref,
int hd, int T, int s_k, float scale
) {
for (int t = 0; t < T; t++) {
float s[512];
for (int j = 0; j < s_k; j++) {
float dot = 0.0f;
for (int d = 0; d < hd; d++)
dot += bf16_to_f32_host(q[t * hd + d]) * bf16_to_f32_host(k[j * hd + d]);
s[j] = dot * scale;
}
float mx = -INFINITY;
for (int j = 0; j < s_k; j++) mx = fmaxf(mx, s[j]);
float sm = 0.0f;
for (int j = 0; j < s_k; j++) { s[j] = expf(s[j] - mx); sm += s[j]; }
for (int j = 0; j < s_k; j++) s[j] /= sm;
for (int d = 0; d < hd; d++) {
float ov = 0.0f;
for (int j = 0; j < s_k; j++) ov += s[j] * bf16_to_f32_host(v[d * s_k + j]);
o_ref[t * hd + d] = ov;
}
if (lse_ref) lse_ref[t] = logf(sm) + mx;
}
}
static int test_single(int T, int n_h = 1, int batch = 1) {
printf("\n=== T=%d, n_h=%d, batch=%d, HD=%d ===\n", T, n_h, batch, HD);
const float SCALE = 1.0f / sqrtf((float)HD);
int total_heads = batch * n_h;
bf16_t* h_q = (bf16_t*)malloc(total_heads * T * HD * sizeof(bf16_t));
bf16_t* h_k = (bf16_t*)malloc(total_heads * SK * HD * sizeof(bf16_t));
bf16_t* h_v = (bf16_t*)malloc(total_heads * HD * SK * sizeof(bf16_t));
bf16_t* h_o = (bf16_t*)calloc(total_heads * T * HD, sizeof(bf16_t));
float* h_lse = (float*)calloc(total_heads * T, sizeof(float));
srand(42 + T);
for (int i = 0; i < total_heads * T * HD; i++) h_q[i] = f32_to_bf16_host((float)(rand()%100)/100.0f - 0.5f);
for (int i = 0; i < total_heads * SK * HD; i++) h_k[i] = f32_to_bf16_host((float)(rand()%100)/100.0f - 0.5f);
for (int i = 0; i < total_heads * HD * SK; i++) h_v[i] = f32_to_bf16_host((float)(rand()%100)/100.0f - 0.5f);
bf16_t *d_q, *d_k, *d_v, *d_o; float *d_lse;
cudaMalloc(&d_q, total_heads * T * HD * sizeof(bf16_t));
cudaMalloc(&d_k, total_heads * SK * HD * sizeof(bf16_t));
cudaMalloc(&d_v, total_heads * HD * SK * sizeof(bf16_t));
cudaMalloc(&d_o, total_heads * T * HD * sizeof(bf16_t));
cudaMalloc(&d_lse, total_heads * T * sizeof(float));
cudaMemcpy(d_q, h_q, total_heads * T * HD * sizeof(bf16_t), cudaMemcpyHostToDevice);
cudaMemcpy(d_k, h_k, total_heads * SK * HD * sizeof(bf16_t), cudaMemcpyHostToDevice);
cudaMemcpy(d_v, h_v, total_heads * HD * SK * sizeof(bf16_t), cudaMemcpyHostToDevice);
FmhaMultiRowParams params;
params.q = d_q; params.k = d_k; params.v = d_v; params.o = d_o; params.lse = d_lse;
params.s_k = SK; params.T = T; params.scale = SCALE; params.head_dim = HD;
params.q_head_stride = T * HD; params.q_batch_stride = n_h * T * HD;
params.k_head_stride = SK * HD; params.k_batch_stride = n_h * SK * HD;
params.v_head_stride = HD * SK; params.v_batch_stride = n_h * HD * SK;
params.o_head_stride = T * HD; params.o_batch_stride = n_h * T * HD;
params.lse_head_stride = T; params.lse_batch_stride = n_h * T;
int smem = compute_smem();
if (smem > 48 * 1024)
cudaFuncSetAttribute(fmha_6warp_multirow_kernel<HD>, cudaFuncAttributeMaxDynamicSharedMemorySize, smem);
dim3 grid(1, n_h, batch);
fmha_6warp_multirow_kernel<HD><<<grid, 192, smem>>>(params);
cudaError_t err = cudaDeviceSynchronize();
if (err != cudaSuccess) {
printf(" CUDA ERROR: %s\n", cudaGetErrorString(err));
cudaFree(d_q); cudaFree(d_k); cudaFree(d_v); cudaFree(d_o); cudaFree(d_lse);
free(h_q); free(h_k); free(h_v); free(h_o); free(h_lse);
return 0;
}
cudaMemcpy(h_o, d_o, total_heads * T * HD * sizeof(bf16_t), cudaMemcpyDeviceToHost);
cudaMemcpy(h_lse, d_lse, total_heads * T * sizeof(float), cudaMemcpyDeviceToHost);
int failed = 0; float min_cos = 1.0f;
for (int b = 0; b < batch; b++) {
for (int h = 0; h < n_h; h++) {
int idx = b * n_h + h;
float o_ref[MAX_T * 512]; float lse_ref[MAX_T];
reference_attention_multirow(
h_q + idx*T*HD, h_k + idx*SK*HD, h_v + idx*HD*SK,
o_ref, lse_ref, HD, T, SK, SCALE);
for (int t = 0; t < T; t++) {
float cs=0,na=0,nb=0;
for (int d=0;d<HD;d++) {
float a=bf16_to_f32_host(h_o[(idx*T+t)*HD+d]), b2=o_ref[t*HD+d];
if(fabsf(b2)>1e-4f){cs+=a*b2;na+=a*a;nb+=b2*b2;}
}
cs /= (sqrtf(na)*sqrtf(nb)+1e-10f);
if(cs<min_cos) min_cos=cs;
if(cs<0.999f) { printf(" FAIL b=%d h=%d t=%d cos=%.6f\n",b,h,t,cs); failed++; }
// Check LSE
float lse_err = fabsf(h_lse[idx*T+t] - lse_ref[t]);
if(lse_err > 0.01f) { printf(" FAIL LSE b=%d h=%d t=%d kernel=%.6f ref=%.6f err=%.6f\n",b,h,t,h_lse[idx*T+t],lse_ref[t],lse_err); failed++; }
}
}
}
printf(" min_cos=%.8f %s\n", min_cos, failed==0?"PASSED":"FAILED");
cudaFree(d_q); cudaFree(d_k); cudaFree(d_v); cudaFree(d_o); cudaFree(d_lse);
free(h_q); free(h_k); free(h_v); free(h_o); free(h_lse);
return failed == 0;
}
static int test_multitile_merge(int T) {
printf("\n=== MULTI-TILE MERGE T=%d, s_k=256, HD=%d ===\n", T, HD);
constexpr int SK_TOTAL = 256; // 2 KV tiles
constexpr int N_TILES = SK_TOTAL / SK; // 2
const float SCALE = 1.0f / sqrtf((float)HD);
bf16_t* h_q = (bf16_t*)malloc(T * HD * sizeof(bf16_t));
bf16_t* h_k = (bf16_t*)malloc(SK_TOTAL * HD * sizeof(bf16_t));
bf16_t* h_v = (bf16_t*)malloc(HD * SK_TOTAL * sizeof(bf16_t));
srand(42 + T + 2000);
for (int i = 0; i < T * HD; i++) h_q[i] = f32_to_bf16_host((float)(rand()%100)/100.0f - 0.5f);
for (int i = 0; i < SK_TOTAL * HD; i++) h_k[i] = f32_to_bf16_host((float)(rand()%100)/100.0f - 0.5f);
for (int i = 0; i < HD * SK_TOTAL; i++) h_v[i] = f32_to_bf16_host((float)(rand()%100)/100.0f - 0.5f);
bf16_t *d_q, *d_k, *d_v, *d_o; float *d_lse;
cudaMalloc(&d_q, T * HD * sizeof(bf16_t));
cudaMalloc(&d_k, SK * HD * sizeof(bf16_t)); // single tile
cudaMalloc(&d_v, HD * SK * sizeof(bf16_t)); // single tile
cudaMalloc(&d_o, T * HD * sizeof(bf16_t));
cudaMalloc(&d_lse, T * sizeof(float));
cudaMemcpy(d_q, h_q, T * HD * sizeof(bf16_t), cudaMemcpyHostToDevice);
int smem = compute_smem();
if (smem > 48 * 1024)
cudaFuncSetAttribute(fmha_6warp_multirow_kernel<HD>, cudaFuncAttributeMaxDynamicSharedMemorySize, smem);
float* lse_per_tile = (float*)malloc(N_TILES * T * sizeof(float));
float* o_per_tile = (float*)malloc(N_TILES * T * HD * sizeof(float));
// We need separate V buffer per tile since V is (HD, SK_TOTAL) and
// the tile columns are not contiguous in the original array
bf16_t* h_v_tile = (bf16_t*)malloc(HD * SK * sizeof(bf16_t));
for (int tile = 0; tile < N_TILES; tile++) {
// K for this tile: K is (SK_TOTAL, HD), rows are contiguous
cudaMemcpy(d_k, h_k + tile * SK * HD, SK * HD * sizeof(bf16_t), cudaMemcpyHostToDevice);
// V for this tile: V is (HD, SK_TOTAL), tile columns are NOT contiguous
// Copy V[d, tile*SK..(tile+1)*SK-1] into h_v_tile[d, 0..SK-1]
for (int d = 0; d < HD; d++)
memcpy(h_v_tile + d * SK, h_v + d * SK_TOTAL + tile * SK, SK * sizeof(bf16_t));
cudaMemcpy(d_v, h_v_tile, HD * SK * sizeof(bf16_t), cudaMemcpyHostToDevice);
FmhaMultiRowParams params;
params.q = d_q; params.k = d_k; params.v = d_v; params.o = d_o; params.lse = d_lse;
params.s_k = SK; params.T = T; params.scale = SCALE; params.head_dim = HD;
params.q_head_stride = T * HD; params.q_batch_stride = T * HD;
params.k_head_stride = SK * HD; params.k_batch_stride = SK * HD;
params.v_head_stride = HD * SK; params.v_batch_stride = HD * SK;
params.o_head_stride = T * HD; params.o_batch_stride = T * HD;
params.lse_head_stride = T; params.lse_batch_stride = T;
dim3 grid(1, 1, 1);
fmha_6warp_multirow_kernel<HD><<<grid, 192, smem>>>(params);
cudaError_t err = cudaDeviceSynchronize();
if (err != cudaSuccess) {
printf(" CUDA ERROR tile %d: %s\n", tile, cudaGetErrorString(err));
cudaFree(d_q); cudaFree(d_k); cudaFree(d_v); cudaFree(d_o); cudaFree(d_lse);
free(h_q); free(h_k); free(h_v);
free(lse_per_tile); free(o_per_tile);
return 0;
}
bf16_t* h_o_tile = (bf16_t*)malloc(T * HD * sizeof(bf16_t));
cudaMemcpy(h_o_tile, d_o, T * HD * sizeof(bf16_t), cudaMemcpyDeviceToHost);
cudaMemcpy(lse_per_tile + tile * T, d_lse, T * sizeof(float), cudaMemcpyDeviceToHost);
for (int i = 0; i < T * HD; i++)
o_per_tile[tile * T * HD + i] = bf16_to_f32_host(h_o_tile[i]);
free(h_o_tile);
}
// Python KV merge with normalized O + LSE:
// O = Σ exp(lse_i - L) * O_i_norm / Σ exp(lse_i - L)
// where L = max(lse_i) for numerical stability
float* h_o_merged = (float*)calloc(T * HD, sizeof(float));
for (int t = 0; t < T; t++) {
float lse_max = -INFINITY;
for (int tile = 0; tile < N_TILES; tile++)
lse_max = fmaxf(lse_max, lse_per_tile[tile * T + t]);
float sum_w = 0.0f;
for (int tile = 0; tile < N_TILES; tile++)
sum_w += expf(lse_per_tile[tile * T + t] - lse_max);
for (int d = 0; d < HD; d++) {
float ov = 0.0f;
for (int tile = 0; tile < N_TILES; tile++)
ov += expf(lse_per_tile[tile * T + t] - lse_max) * o_per_tile[tile * T * HD + t * HD + d];
h_o_merged[t * HD + d] = ov / sum_w;
}
}
// Compare with full reference
float o_ref[MAX_T * 512];
reference_attention_multirow(h_q, h_k, h_v, o_ref, nullptr, HD, T, SK_TOTAL, SCALE);
int failed = 0; float min_cos = 1.0f;
for (int t = 0; t < T; t++) {
float cs=0,na=0,nb=0;
for (int d=0;d<HD;d++) {
float a=h_o_merged[t*HD+d], b2=o_ref[t*HD+d];
if(fabsf(b2)>1e-4f){cs+=a*b2;na+=a*a;nb+=b2*b2;}
}
cs /= (sqrtf(na)*sqrtf(nb)+1e-10f);
if(cs<min_cos) min_cos=cs;
if(cs<0.999f) { printf(" FAIL merge t=%d cos=%.6f\n",t,cs); failed++; }
}
printf(" min_cos=%.8f %s\n", min_cos, failed==0?"PASSED":"FAILED");
cudaFree(d_q); cudaFree(d_k); cudaFree(d_v); cudaFree(d_o); cudaFree(d_lse);
free(h_q); free(h_k); free(h_v); free(h_o_merged); free(h_v_tile);
free(lse_per_tile); free(o_per_tile);
return failed == 0;
}
int main() {
printf("Multi-row FMHA test (HD=%d)\n", HD);
int ok = 1;
// 1. Single KV tile, normalized output
printf("\n--- Single KV tile tests ---\n");
ok &= test_single(1);
ok &= test_single(2);
ok &= test_single(4);
ok &= test_single(8);
ok &= test_single(16);
ok &= test_single(32);
ok &= test_single(64);
ok &= test_single(128);
// 2. Multi-tile KV merge (s_k=256, 2 segments)
printf("\n--- Multi-tile KV merge tests ---\n");
ok &= test_multitile_merge(1);
ok &= test_multitile_merge(4);
ok &= test_multitile_merge(16);
ok &= test_multitile_merge(32);
ok &= test_multitile_merge(64);
ok &= test_multitile_merge(128);
// 3. Multi-head and batched
printf("\n--- Multi-head and batched tests ---\n");
ok &= test_single(4, 4, 1); // 4 heads, T=4
ok &= test_single(16, 4, 1); // 4 heads, T=16
ok &= test_single(64, 4, 1); // 4 heads, T=64
ok &= test_single(1, 2, 2); // 2 heads, 2 batch, T=1
ok &= test_single(16, 2, 2); // 2 heads, 2 batch, T=16
printf("\n%s\n", ok ? "ALL PASSED" : "SOME FAILED");
return ok ? 0 : 1;
}
Reference in New Issue View Git Blame Copy Permalink