ggml : fix FA mask dim 2 and 3 (llama/14505)

* ggml : fix FA mask dim 2 and 3

ggml-ci

* backends : unsupport batched FA in CUDA and Vulkan

ggml-ci

* vulkan : disable FA for mask->ne[2] != 1

Author: ggerganov

ggml/include/ggml.h

@@ -1980,15 +1980,16 @@ extern "C" {
 
 #define GGML_KQ_MASK_PAD 64
 
-    // q:    [n_embd_k, n_batch,     n_head,    ne3]
-    // k:    [n_embd_k, n_kv,        n_head_kv, ne3]
-    // v:    [n_embd_v, n_kv,        n_head_kv, ne3] !! not transposed !!
-    // mask: [n_kv,     n_batch_pad, ne32,      1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
-    // res:  [n_embd_v, n_head,      n_batch,   ne3] !! permuted !!
+    // q:    [n_embd_k, n_batch,     n_head,    ne3 ]
+    // k:    [n_embd_k, n_kv,        n_head_kv, ne3 ]
+    // v:    [n_embd_v, n_kv,        n_head_kv, ne3 ] !! not transposed !!
+    // mask: [n_kv,     n_batch_pad, ne32,      ne33] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
+    // res:  [n_embd_v, n_head,      n_batch,   ne3 ] !! permuted !!
     //
     // broadcast:
     //   n_head % n_head_kv == 0
-    //   ne3    % ne32      == 0
+    //   n_head % ne32      == 0
+    //   ne3    % ne33      == 0
     //
     GGML_API struct ggml_tensor * ggml_flash_attn_ext(
             struct ggml_context * ctx,

ggml/src/ggml-cpu/ops.cpp

@@ -7799,7 +7799,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
             memset(VKQ32, 0, DV*sizeof(float));
         }
 
-        const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1] + (iq3%mask->ne[2])*mask->nb[2]) : NULL;
+        const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1] + (iq2%mask->ne[2])*mask->nb[2] + (iq3%mask->ne[3])*mask->nb[3]) : NULL;
 
         // k indices
         const int ik3 = iq3 / rk3;

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -3390,7 +3390,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 return false;
             }
             // TODO: support broadcast
-            // ref: https://github.com/ggml-org/llama.cpp/pull/14435
+            // note: this was initially implemented in https://github.com/ggml-org/llama.cpp/pull/14500, but
+            //       the interface of ggml_flash_attn_ext() changed in https://github.com/ggml-org/llama.cpp/pull/14505
             if (op->src[0]->ne[3] != 1) {
                 return false;
             }

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -230,8 +230,10 @@ typedef struct {
     uint64_t nb22;
     uint64_t nb23;
     int32_t  ne32;
+    int32_t  ne33;
     uint64_t nb31;
     uint64_t nb32;
+    uint64_t nb33;
     int32_t  ne1;
     int32_t  ne2;
     float    scale;

ggml/src/ggml-metal/ggml-metal.m

@@ -5018,8 +5018,10 @@ static bool ggml_metal_encode_node(
                     /*.nb22          =*/ nb22,
                     /*.nb23          =*/ nb23,
                     /*.ne32          =*/ ne32,
+                    /*.ne33          =*/ ne33,
                     /*.nb31          =*/ nb31,
                     /*.nb32          =*/ nb32,
+                    /*.nb33          =*/ nb33,
                     /*.ne1           =*/ ne1,
                     /*.ne2           =*/ ne2,
                     /*.scale         =*/ scale,

ggml/src/ggml-metal/ggml-metal.metal

@@ -3857,7 +3857,7 @@ kernel void kernel_flash_attn_ext(
                 // load the mask in shared memory
                 #pragma unroll(Q)
                 for (short j = 0; j < Q; ++j) {
-                    device const half * pm = (device const half *) ((device const char *) mask + (iq1 + j)*args.nb31 + (iq3%args.ne32)*args.nb32);
+                    device const half * pm = (device const half *) ((device const char *) mask + (iq1 + j)*args.nb31 + (iq2%args.ne32)*args.nb32 + (iq3%args.ne33)*args.nb33);
 
                     const float m = pm[ic + tiisg];
 
@@ -4343,7 +4343,7 @@ kernel void kernel_flash_attn_ext_vec(
         const bool has_mask = mask != q;
 
         // pointer to the mask
-        device const half * pm = (device const half *) (mask + iq1*args.nb31 + (iq3%args.ne32)*args.nb32);
+        device const half * pm = (device const half *) (mask + iq1*args.nb31 + (iq2%args.ne32)*args.nb32 + (iq3%args.ne33)*args.nb33);
 
         float slope = 1.0f;
 

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -10265,6 +10265,12 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 if (op->src[3] && op->src[3]->type != GGML_TYPE_F16) {
                     return false;
                 }
+                // TODO: support broadcast
+                // note: this was initially implemented in https://github.com/ggml-org/llama.cpp/pull/14449, but
+                //       the interface of ggml_flash_attn_ext() changed in https://github.com/ggml-org/llama.cpp/pull/14505
+                if (op->src[0]->ne[3] != 1 || (op->src[3] && op->src[3]->ne[2] != 1)) {
+                    return false;
+                }
                 // It's straightforward to support different K/V dequant, but would
                 // significantly increase the number of pipelines
                 if (op->src[1]->type != op->src[2]->type) {

ggml/src/ggml.c

@@ -3666,7 +3666,6 @@ static struct ggml_tensor * ggml_soft_max_impl(
     if (mask) {
         GGML_ASSERT(mask->type == GGML_TYPE_F16 || mask->type == GGML_TYPE_F32);
         GGML_ASSERT(ggml_is_contiguous(mask));
-        GGML_ASSERT(ggml_is_3d(mask));
         GGML_ASSERT(mask->ne[0] == a->ne[0]);
         GGML_ASSERT(mask->ne[1] >= a->ne[1]);
         GGML_ASSERT(a->ne[2]%mask->ne[2] == 0);
@@ -4696,12 +4695,12 @@ struct ggml_tensor * ggml_flash_attn_ext(
 
     if (mask) {
         GGML_ASSERT(ggml_is_contiguous(mask));
-        GGML_ASSERT(mask->ne[2] == q->ne[3]);
         GGML_ASSERT(mask->ne[1] >= GGML_PAD(q->ne[1], GGML_KQ_MASK_PAD) &&
                 "the Flash-Attention kernel requires the mask to be padded to GGML_KQ_MASK_PAD and at least n_queries big");
         //GGML_ASSERT(ggml_can_repeat_rows(mask, qk));
 
-        GGML_ASSERT(q->ne[3] % mask->ne[2] == 0);
+        GGML_ASSERT(q->ne[2] % mask->ne[2] == 0);
+        GGML_ASSERT(q->ne[3] % mask->ne[3] == 0);
     }
 
     if (max_bias > 0.0f) {