cuda : increase C to 128 for better performance

Author: ggerganov

ggml-cuda.cu

@@ -6495,8 +6495,8 @@ static __global__ void flash_attn_ext_f16(
         half M[Q];
 
         for(int i = 0; i < Q; i++) {
-            S[i] = 0.0f;
-            M[i] = -INFINITY;
+            S[i] = __float2half(0.0f);
+            M[i] = __float2half(-INFINITY);
         }
 
         // assume K and V are same shape
@@ -6579,7 +6579,7 @@ static __global__ void flash_attn_ext_f16(
             }
 
             // used to detect blocks full of -INF
-            half smax = -INFINITY;
+            half smax = __float2half(-INFINITY);
 
             // online softmax
             if (C == 32) {
@@ -6592,8 +6592,8 @@ static __global__ void flash_attn_ext_f16(
                     smax = warp_reduce_max(__hmax(smax, s));
                     M[j] = warp_reduce_max(__hmax(M[j], s));
 
-                    const half ms = __hisinf(m) ? 0.0f : expf(m - M[j]);
-                    const half vs = __hisinf(s) ? 0.0f : expf(s - M[j]);
+                    const half ms = __hisinf(m) ? __float2half(0.0f) : hexp(m - M[j]);
+                    const half vs = __hisinf(s) ? __float2half(0.0f) : hexp(s - M[j]);
 
                     S[j] = S[j]*ms + warp_reduce_sum(vs);
 
@@ -6612,33 +6612,38 @@ static __global__ void flash_attn_ext_f16(
                     for (int64_t p = lane_id; p < C; p += NW) {
                         const half s = ss[j*T + p];
 
-                        smax = warp_reduce_max(__hmax(smax, s));
-                        M[j] = warp_reduce_max(__hmax(M[j], s));
+                        smax = __hmax(smax, s);
+                        M[j] = __hmax(M[j], s);
                     }
 
-                    const half ms = __hisinf(m) ? 0.0f : expf(m - M[j]);
+                    smax = warp_reduce_max(smax);
+                    M[j] = warp_reduce_max(M[j]);
 
-                    S[j] = S[j]*ms;
+                    const half ms = __hisinf(m) ? __float2half(0.0f) : hexp(m - M[j]);
 
                     // create a QxQ diagonal matrix for rescaling the output
                     if (lane_id == j) {
                         ss[j*T + C + j] = ms;
                     }
 
+                    // local sum
+                    half ls = 0.0f;
+
                     for (int64_t p = lane_id; p < C; p += NW) {
                         const half s = ss[j*T + p];
 
-                        const half vs = __hisinf(s) ? 0.0f : expf(s - M[j]);
+                        const half vs = __hisinf(s) ? __float2half(0.0f) : hexp(s - M[j]);
 
-                        S[j] = S[j] + warp_reduce_sum(vs);
+                        ls += vs;
 
                         // the P matrix from the paper (Q rows, C columns)
                         ss[j*T + p] = vs;
                     }
+
+                    S[j] = S[j]*ms + warp_reduce_sum(ls);
                 }
             }
 
-
             // skip -INF blocks
             if (__hisinf(smax)) {
                 continue;
@@ -6669,15 +6674,19 @@ static __global__ void flash_attn_ext_f16(
                 for (int cc = 0; cc < C/16; ++cc) {
                     const half * pv = (const half *) ((const char *) v + ((ic + 16*cc)*nb21 + iv2*nb22 + iv3*nb23));
 
+                    half16x16_b mk[D16];
                     for (int64_t i = 0; i < D16; ++i) {
-                        half16x16_b mk;
-                        nvcuda::wmma::load_matrix_sync(mk, pv + i*16, nb21/sizeof(half));
+                        nvcuda::wmma::load_matrix_sync(mk[i], pv + i*16, nb21/sizeof(half));
+                    }
 
-                        for (int64_t j = 0; j < Q16; ++j) {
-                            half16x16_a mv;
-                            nvcuda::wmma::load_matrix_sync(mv, ss + 16*j*T + 16*cc, T);
+                    half16x16_a mv[Q16];
+                    for (int64_t j = 0; j < Q16; ++j) {
+                        nvcuda::wmma::load_matrix_sync(mv[j], ss + 16*j*T + 16*cc, T);
+                    }
 
-                            nvcuda::wmma::mma_sync(lo[j][i], mv, mk, lo[j][i]);
+                    for (int64_t j = 0; j < Q16; ++j) {
+                        for (int64_t i = 0; i < D16; ++i) {
+                            nvcuda::wmma::mma_sync(lo[j][i], mv[j], mk[i], lo[j][i]);
                         }
                     }
                 }
@@ -6695,8 +6704,8 @@ static __global__ void flash_attn_ext_f16(
 
     // reduce the warps sequentially
     for (int64_t sg = 1; sg < num_warps; ++sg) {
-        half S = 0.0f;
-        half M = -INFINITY;
+        half S = __float2half(0.0f);
+        half M = __float2half(-INFINITY);
 
         __syncthreads();
 
@@ -6722,8 +6731,8 @@ static __global__ void flash_attn_ext_f16(
 
                 M = __hmax(M0, M1);
 
-                const half ms0 = __hisinf(M0) ? 0.0f : expf(M0 - M);
-                const half ms1 = __hisinf(M1) ? 0.0f : expf(M1 - M);
+                const half ms0 = __hisinf(M0) ? __float2half(0.0f) : hexp(M0 - M);
+                const half ms1 = __hisinf(M1) ? __float2half(0.0f) : hexp(M1 - M);
 
                 S = S0*ms0 + S1*ms1;
 
@@ -6770,8 +6779,6 @@ static __global__ void flash_attn_ext_f16(
         }
     }
 
-    // float2 * dst2 = (float2 *) dst;
-
     // final rescale with 1/S and store to global memory
     if (warp_id == 0) {
         for (int64_t j = 0; j < Q && iq1 + j < ne01; ++j) {
@@ -9637,7 +9644,7 @@ static void ggml_cuda_op_soft_max(
 
     const int64_t ne00 = src0->ne[0];
     const int64_t nrows_x = ggml_nrows(src0);
-    const int64_t nrows_y = src1 ? ggml_nrows(src1) : 1;
+    const int64_t nrows_y = src1 ? src0->ne[1] : 1; // note: using number of queries since mask can be padded!
 
     float scale = 1.0f;
     memcpy(&scale, dst->op_params, sizeof(float));
@@ -10932,7 +10939,7 @@ inline void ggml_cuda_flash_attn_ext(const ggml_tensor * Q, const ggml_tensor *
     memcpy(&scale, KQV->op_params, sizeof(float));
 
 #define NQPB 16
-#define NCPW 32
+#define NCPW 128
 
     const int nqpb = NQPB; // queries per block
     const int ncpw = NCPW; // cache values per warp (does not work for other values)

ggml.c

@@ -5089,7 +5089,7 @@ static struct ggml_tensor * ggml_soft_max_impl(
         GGML_ASSERT(ggml_is_contiguous(mask));
         GGML_ASSERT(mask->ne[2] == 1);
         GGML_ASSERT(mask->ne[3] == 1);
-        GGML_ASSERT(ggml_can_repeat_rows(mask, a));
+        GGML_ASSERT(mask->ne[1] >= a->ne[1]);
     }
 
     bool is_node = false;

llama.cpp

@@ -6881,7 +6881,8 @@ static int llama_decode_internal(
     // a heuristic, to avoid attending the full cache if it is not yet utilized
     // after enough generations, the benefit from this heuristic disappears
     // if we start defragmenting the cache, the benefit from this will be more important
-    kv_self.n = std::min((int32_t) cparams.n_ctx, std::max(32, GGML_PAD(llama_kv_cache_cell_max(kv_self), 32)));
+    // note: we pad the n_kv because certain GPU kernels require it (e.g. ggml_flash_attn_ext)
+    kv_self.n = std::min((int32_t) cparams.n_ctx, std::max(128, GGML_PAD(llama_kv_cache_cell_max(kv_self), 128)));
     //kv_self.n = llama_kv_cache_cell_max(kv_self);
 
     //printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self.n, kv_self.used, kv_self.head);

tests/test-backend-ops.cpp

@@ -2210,7 +2210,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     test_cases.emplace_back(new test_leaky_relu());
 
 #if 1
-    for (int hs : { 64,  80, 128, }) {
+    for (int hs : { 128, 64,  80, }) {
         for (int nh : { 32, }) {
             for (int kv : { 512, 1024, 2048, 4096, }) {
                 for (int nb : { 1, 2, 4, 8, 512, 1024, 2048, }) {