Allow bf16 kv-cache (#69)

On the CPU I get the exact same PPL with and without FA
using bf16 for kv-cache. But on CUDA the bf16 kv-cache
result is about the same as the fp16 kv-cache CPU result,
so I'm missing some conversion somewhere.

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>

ggml/src/ggml-cuda.cu

@@ -2862,6 +2862,9 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
                 if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
                     return true;
                 }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_BF16) {
+                    return true;
+                }
                 if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q8_0) {
                     return true;
                 }

ggml/src/ggml-cuda/cpy.cu

@@ -16,6 +16,13 @@ static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) {
     *dsti = __float2half(*xi);
 }
 
+static __device__ void cpy_1_f32_bf16(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    nv_bfloat16 * dsti = (nv_bfloat16 *) cdsti;
+
+    *dsti = __float2bfloat16(*xi);
+}
+
 static __device__ void cpy_1_f16_f16(const char * cxi, char * cdsti) {
     const half * xi = (const half *) cxi;
     half * dsti = (half *) cdsti;
@@ -318,6 +325,16 @@ static void ggml_cpy_f32_f16_cuda(
         (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
 
+static void ggml_cpy_f32_bf16_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
+    cpy_f32_f16<cpy_1_f32_bf16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
 static void ggml_cpy_f32_q8_0_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
@@ -432,6 +449,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
         ggml_cpy_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
         ggml_cpy_f32_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
+        ggml_cpy_f32_bf16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
         ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) {
@@ -465,6 +484,8 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
             return (void*) cpy_f32_f16<cpy_1_f32_f32>;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
             return (void*) cpy_f32_f16<cpy_1_f32_f16>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
+            return (void*) cpy_f32_f16<cpy_1_f32_bf16>;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
             return (void*) cpy_f32_q<cpy_blck_f32_q8_0, QK8_0>;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) {

src/llama.cpp

@@ -16816,7 +16816,7 @@ struct llama_context * llama_new_context_with_model(
         params.flash_attn = false;
     }
 
-    if (params.type_v != GGML_TYPE_F16 && !params.flash_attn) {
+    if (params.type_v != GGML_TYPE_F16 && params.type_v != GGML_TYPE_BF16 && !params.flash_attn) {
         LLAMA_LOG_ERROR("%s: V cache quantization requires flash_attn\n", __func__);
         return nullptr;
     }