vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

flush softmax exp below threshold to 0

Browse source
This commit is contained in:
Johannes Gäßler 2024-04-15 16:05:07 +02:00
parent 6a3b84236d
commit ef9e1593f3
1 changed files with 19 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

22
ggml-cuda/fattn.cu
View file

@ -5,6 +5,7 @@
#define FATTN_KQ_STRIDE 256 #define FATTN_KQ_STRIDE 256
#define HALF_MAX_HALF __float2half(65504.0f/2) // Use neg. of this instead of -INFINITY to initialize KQ max vals to avoid NaN upon subtraction. #define HALF_MAX_HALF __float2half(65504.0f/2) // Use neg. of this instead of -INFINITY to initialize KQ max vals to avoid NaN upon subtraction.
#define SOFTMAX_FTZ_THRESHOLD -20.0f // Softmax exp. of values smaller than this are flushed to zero to avoid NaNs.
template<int D, int parallel_blocks> // D == head size template<int D, int parallel_blocks> // D == head size
__launch_bounds__(((D + WARP_SIZE - 1) / WARP_SIZE)*WARP_SIZE, 1) __launch_bounds__(((D + WARP_SIZE - 1) / WARP_SIZE)*WARP_SIZE, 1)
@ -338,10 +339,16 @@ static __global__ void flash_attn_ext_f16(
#pragma unroll #pragma unroll
for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += WARP_SIZE) { for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += WARP_SIZE) {
const int k = k0 + threadIdx.x; const int k = k0 + threadIdx.x;
KQ_max_new = __hmax2(KQ_max_new, KQ2[j*(kqs_padded/2) + k]); half2 val = KQ2[j*(kqs_padded/2) + k];
val += mask ? mask2[(j*ne11 + k_VKQ_0)/2 + k] : make_half2(0.0f, 0.0f);
KQ_max_new = __hmax2(KQ_max_new, val);
KQ2[j*(kqs_padded/2) + k] = val;
} }
KQ_max_new = __half2half2(warp_reduce_max(__hmax(__low2half(KQ_max_new), __high2half(KQ_max_new)))); KQ_max_new = __half2half2(warp_reduce_max(__hmax(__low2half(KQ_max_new), __high2half(KQ_max_new))));
KQ_max_scale[j0/nwarps] = h2exp(KQ_max[j0/nwarps] - KQ_max_new); const half2 diff = KQ_max[j0/nwarps] - KQ_max_new;
KQ_max_scale[j0/nwarps] = h2exp(diff);
const uint ftz_mask = __hgt2_mask(diff, make_half2(SOFTMAX_FTZ_THRESHOLD, SOFTMAX_FTZ_THRESHOLD));
*((uint *) &KQ_max_scale[j0/nwarps]) &= ftz_mask;
KQ_max[j0/nwarps] = KQ_max_new; KQ_max[j0/nwarps] = KQ_max_new;
half2 KQ_rowsum_add = make_half2(0.0f, 0.0f); half2 KQ_rowsum_add = make_half2(0.0f, 0.0f);
@ -350,8 +357,10 @@ static __global__ void flash_attn_ext_f16(
const int k = k0 + threadIdx.x; const int k = k0 + threadIdx.x;
half2 val = KQ2[j*(kqs_padded/2) + k]; half2 val = KQ2[j*(kqs_padded/2) + k];
val += mask ? mask2[(j*ne11 + k_VKQ_0)/2 + k] : make_half2(0.0f, 0.0f); const half2 diff = val - KQ_max[j0/nwarps];
val = h2exp(val - KQ_max[j0/nwarps]); val = h2exp(diff);
const uint ftz_mask = __hgt2_mask(diff, make_half2(SOFTMAX_FTZ_THRESHOLD, SOFTMAX_FTZ_THRESHOLD));
*((uint *) &val) &= ftz_mask;
KQ_rowsum_add += val; KQ_rowsum_add += val;
KQ2[j*(kqs_padded/2) + k] = val; KQ2[j*(kqs_padded/2) + k] = val;
} }
@ -501,7 +510,10 @@ static __global__ void flash_attn_combine_results(
float VKQ_denominator = 0.0f; float VKQ_denominator = 0.0f;
#pragma unroll #pragma unroll
for (int l = 0; l < parallel_blocks; ++l) { for (int l = 0; l < parallel_blocks; ++l) {
float KQ_max_scale = hexp(__low2half(meta[l]) - kqmax); const half diff = __low2half(meta[l]) - kqmax;
float KQ_max_scale = hexp(diff);
const uint ftz_mask = 0xFFFFFFFF * (diff > __float2half(SOFTMAX_FTZ_THRESHOLD));
*((uint *) &KQ_max_scale) &= ftz_mask;
VKQ_numerator += KQ_max_scale * VKQ_parts[l*gridDim.y*D + blockIdx.y*D + tid]; VKQ_numerator += KQ_max_scale * VKQ_parts[l*gridDim.y*D + blockIdx.y*D + tid];
VKQ_denominator += KQ_max_scale * __high2float(meta[l]); VKQ_denominator += KQ_max_scale * __high2float(meta[l]);