biondizzle/nvfp4-megamoe-kernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: always normalize in kernel, correct KV merge with normalized O + LSE

Browse Source
This commit is contained in:
biondizzle
2026-05-28 23:53:44 +00:00
parent 914f76d30c
commit 2f2259395e
2 changed files with 34 additions and 155 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
dsv4/kernels/attention/fmha_6warp_multirow.cuh
View File

@@ -42,7 +42,7 @@ struct FmhaMultiRowParams {
const bf16_t* __restrict__ k;
const bf16_t* __restrict__ v;
bf16_t* __restrict__ o;
float* __restrict__ lse;
float* __restrict__ lse; // [batch, n_h, T] — per-row LSE for multi-tile KV merge
int s_k, T;
float scale;
int head_dim;
@@ -51,7 +51,6 @@ struct FmhaMultiRowParams {
int v_head_stride, v_batch_stride;
int o_head_stride, o_batch_stride;
int lse_head_stride, lse_batch_stride;
int normalize; // 1 = normalize in kernel, 0 = emit un-normalized O + LSE for multi-tile merge
};
template<int HD, int SK_TILE = 128>
@@ -265,21 +264,21 @@ fmha_6warp_multirow_kernel(FmhaMultiRowParams params) {
__syncthreads();
// ================================================================
// EPILOGUE: TMEM → regs → normalize (optional) → BF16 → GMEM
// EPILOGUE: TMEM → regs → normalize → BF16 → GMEM + LSE output
//
// CRITICAL: TMEM loads (32x32b.x8) are WARP-COLLECTIVE.
// ALL 32 lanes must execute them. The load MUST be outside
// the my_row_active guard. Only the GMEM store is conditional.
//
// When normalize=1 (single KV tile): O_norm = O_unnorm / row_sum
// When normalize=0 (multi-tile merge): emit O_unnorm + LSE for
// Python merge: O = Σ exp(lse_i)·O_i / Σ exp(lse_i)
// Output: normalized O (O_unnorm / row_sum) + per-row LSE.
// LSE = ln(row_sum) + row_max, for multi-tile KV merge:
// O = Σ exp(lse_i - L) * O_i / Σ exp(lse_i - L)
// where L = max(lse_i) for numerical stability.
// ================================================================
const bool do_normalize = (params.normalize != 0);
if (my_warp_active) {
float rm = my_row_active ? sRowMax[my_row] : 0.0f;
float rs = my_row_active ? sRowSum[my_row] : 0.0f;
float inv_rs = (my_row_active && do_normalize) ? (1.0f / rs) : 1.0f;
float inv_rs = my_row_active ? (1.0f / rs) : 0.0f;
// Read O from TMEM: N_NSUB*2 groups of 8 columns
// ALL lanes in the warp must execute the TMEM load (warp-collective)

174
tests/unit/test_fmha_6warp_multirow.cu
View File

@@ -3,10 +3,9 @@
* Compile with -DHD_VAL=64 etc.
*
* Tests:
* 1. Single KV tile, T=1..128 (normalized output)
* 2. Single KV tile, T=1..128 (un-normalized output + LSE)
* 3. Multi-tile KV via Python merge (s_k=256, 2 segments)
* 4. Multi-head and batched launches
* 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>
@@ -75,36 +74,8 @@ static void reference_attention_multirow(
}
}
// Reference that computes un-normalized O + LSE
static void reference_attention_multirow_unnorm(
const bf16_t* q, const bf16_t* k, const bf16_t* v,
float* o_unnorm, 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]; }
// Un-normalized: don't divide by 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_unnorm[t * hd + d] = ov; // un-normalized!
}
if (lse_ref) lse_ref[t] = logf(sm) + mx;
}
}
static int test_normalized(int T, int n_h = 1, int batch = 1) {
printf("\n=== NORMALIZED T=%d, n_h=%d, batch=%d, HD=%d ===\n", T, n_h, batch, HD);
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;
@@ -137,7 +108,6 @@ static int test_normalized(int T, int n_h = 1, int batch = 1) {
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;
params.normalize = 1;
int smem = compute_smem();
if (smem > 48 * 1024)
@@ -184,86 +154,6 @@ static int test_normalized(int T, int n_h = 1, int batch = 1) {
return failed == 0;
}
static int test_unnormalized(int T) {
printf("\n=== UN-NORMALIZED T=%d, HD=%d ===\n", T, HD);
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 * HD * sizeof(bf16_t));
bf16_t* h_v = (bf16_t*)malloc(HD * SK * sizeof(bf16_t));
bf16_t* h_o = (bf16_t*)calloc(T * HD, sizeof(bf16_t));
float* h_lse = (float*)calloc(T, sizeof(float));
srand(42 + T + 1000);
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 * HD; i++) h_k[i] = f32_to_bf16_host((float)(rand()%100)/100.0f - 0.5f);
for (int i = 0; i < 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, T * HD * sizeof(bf16_t));
cudaMalloc(&d_k, SK * HD * sizeof(bf16_t));
cudaMalloc(&d_v, HD * SK * sizeof(bf16_t));
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);
cudaMemcpy(d_k, h_k, SK * HD * sizeof(bf16_t), cudaMemcpyHostToDevice);
cudaMemcpy(d_v, h_v, 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;
params.normalize = 0; // UN-NORMALIZED
int smem = compute_smem();
if (smem > 48 * 1024)
cudaFuncSetAttribute(fmha_6warp_multirow_kernel<HD>, cudaFuncAttributeMaxDynamicSharedMemorySize, smem);
dim3 grid(1, 1, 1);
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, T * HD * sizeof(bf16_t), cudaMemcpyDeviceToHost);
cudaMemcpy(h_lse, d_lse, T * sizeof(float), cudaMemcpyDeviceToHost);
// Verify: O_normalized = O_unnorm / row_sum, where exp(LSE) = row_sum * exp(max)
float o_unnorm_ref[MAX_T * 512]; float lse_ref[MAX_T];
reference_attention_multirow_unnorm(h_q, h_k, h_v, o_unnorm_ref, lse_ref, HD, T, SK, SCALE);
int failed = 0; float min_cos = 1.0f;
for (int t = 0; t < T; t++) {
// Check un-normalized O matches reference
float cs=0,na=0,nb=0;
for (int d=0;d<HD;d++) {
float a=bf16_to_f32_host(h_o[t*HD+d]), b2=o_unnorm_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 unnorm t=%d cos=%.6f\n",t,cs); failed++; }
// Check LSE matches reference
float lse_err = fabsf(h_lse[t] - lse_ref[t]);
if(lse_err > 0.01f) { printf(" FAIL lse t=%d kernel=%.6f ref=%.6f err=%.6f\n",t,h_lse[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
@@ -279,9 +169,6 @@ static int test_multitile_merge(int T) {
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);
// Run kernel per KV tile, get un-normalized O + LSE, merge in Python
float* h_o_merged = (float*)calloc(T * HD, sizeof(float));
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
@@ -310,7 +197,6 @@ static int test_multitile_merge(int T) {
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;
params.normalize = 0; // UN-NORMALIZED for merge
dim3 grid(1, 1, 1);
fmha_6warp_multirow_kernel<HD><<<grid, 192, smem>>>(params);
@@ -318,7 +204,7 @@ static int test_multitile_merge(int T) {
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(h_o_merged);
free(h_q); free(h_k); free(h_v);
free(lse_per_tile); free(o_per_tile);
return 0;
}
@@ -332,7 +218,10 @@ static int test_multitile_merge(int T) {
free(h_o_tile);
}
// Python KV merge: O = Σ exp(lse_i)·O_i / Σ exp(lse_i)
// 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++)
@@ -376,27 +265,18 @@ int main() {
int ok = 1;
// 1. Normalized output (single KV tile)
printf("\n--- Normalized output tests ---\n");
ok &= test_normalized(1);
ok &= test_normalized(2);
ok &= test_normalized(4);
ok &= test_normalized(8);
ok &= test_normalized(16);
ok &= test_normalized(32);
ok &= test_normalized(64);
ok &= test_normalized(128);
// 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. Un-normalized output + LSE
printf("\n--- Un-normalized output + LSE tests ---\n");
ok &= test_unnormalized(1);
ok &= test_unnormalized(4);
ok &= test_unnormalized(16);
ok &= test_unnormalized(32);
ok &= test_unnormalized(64);
ok &= test_unnormalized(128);
// 3. Multi-tile KV merge
// 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);
@@ -405,13 +285,13 @@ int main() {
ok &= test_multitile_merge(64);
ok &= test_multitile_merge(128);
// 4. Multi-head and batched
// 3. Multi-head and batched
printf("\n--- Multi-head and batched tests ---\n");
ok &= test_normalized(4, 4, 1); // 4 heads, T=4
ok &= test_normalized(16, 4, 1); // 4 heads, T=16
ok &= test_normalized(64, 4, 1); // 4 heads, T=64
ok &= test_normalized(1, 2, 2); // 2 heads, 2 batch, T=1
ok &= test_normalized(16, 2, 2); // 2 heads, 2 batch, T=16
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;