cuda : add f32 to bf16 copy op (#12806)

This allows BF16 KV-cache on CUDA.

examples/llava/clip.cpp

@@ -380,6 +380,7 @@ struct clip_ctx {
         if (backend_cpu != backend) {
             ggml_backend_free(backend_cpu);
         }
+        clip_image_size_free(load_image_size);
     }
 };
 
@@ -1618,6 +1619,12 @@ struct clip_image_f32 * clip_image_f32_init() {
     return new clip_image_f32();
 }
 
+void clip_image_size_free(struct clip_image_size * load_image_size) {
+    if (load_image_size == nullptr) {
+        return;
+    }
+    delete load_image_size;
+}
 void clip_image_u8_free(struct clip_image_u8  * img) { delete img; }
 void clip_image_f32_free(struct clip_image_f32 * img) { delete img; }
 void clip_image_u8_batch_free(struct clip_image_u8_batch  * batch) {
@@ -2270,6 +2277,9 @@ ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx) {
 }
 
 void clip_free(clip_ctx * ctx) {
+    if (ctx == nullptr) {
+        return;
+    }
     delete ctx;
 }
 

examples/llava/clip.h

@@ -77,6 +77,7 @@ CLIP_API struct clip_image_size * clip_image_size_init();
 CLIP_API struct clip_image_u8  * clip_image_u8_init ();
 CLIP_API struct clip_image_f32 * clip_image_f32_init();
 
+CLIP_API void clip_image_size_free (struct clip_image_size * img_size);
 CLIP_API void clip_image_u8_free (struct clip_image_u8  * img);
 CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
 CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  * batch);

examples/server/server.cpp

@@ -1705,6 +1705,8 @@ private:
 };
 
 struct server_response {
+    bool running = true;
+
     // for keeping track of all tasks waiting for the result
     std::unordered_set<int> waiting_task_ids;
 
@@ -1759,6 +1761,10 @@ struct server_response {
         while (true) {
             std::unique_lock<std::mutex> lock(mutex_results);
             condition_results.wait(lock, [&]{
+                if (!running) {
+                    SRV_DBG("%s : queue result stop\n", __func__);
+                    std::terminate(); // we cannot return here since the caller is HTTP code
+                }
                 return !queue_results.empty();
             });
 
@@ -1789,6 +1795,10 @@ struct server_response {
             }
 
             std::cv_status cr_res = condition_results.wait_for(lock, std::chrono::seconds(timeout));
+            if (!running) {
+                SRV_DBG("%s : queue result stop\n", __func__);
+                std::terminate(); // we cannot return here since the caller is HTTP code
+            }
             if (cr_res == std::cv_status::timeout) {
                 return nullptr;
             }
@@ -1818,6 +1828,12 @@ struct server_response {
             }
         }
     }
+
+    // terminate the waiting loop
+    void terminate() {
+        running = false;
+        condition_results.notify_all();
+    }
 };
 
 struct server_context {
@@ -4491,9 +4507,10 @@ int main(int argc, char ** argv) {
     svr->new_task_queue = [&params] { return new httplib::ThreadPool(params.n_threads_http); };
 
     // clean up function, to be called before exit
-    auto clean_up = [&svr]() {
+    auto clean_up = [&svr, &ctx_server]() {
         SRV_INF("%s: cleaning up before exit...\n", __func__);
         svr->stop();
+        ctx_server.queue_results.terminate();
         llama_backend_free();
     };
 
@@ -4534,7 +4551,7 @@ int main(int argc, char ** argv) {
 
     if (!ctx_server.load_model(params)) {
         clean_up();
-        // t.join(); // FIXME: see below
+        t.join();
         LOG_ERR("%s: exiting due to model loading error\n", __func__);
         return 1;
     }
@@ -4582,7 +4599,7 @@ int main(int argc, char ** argv) {
     ctx_server.queue_tasks.start_loop();
 
     clean_up();
-    // t.join(); // FIXME: http thread may stuck if there is an on-going request. we don't need to care about this for now as the HTTP connection will already be closed at this point, but it's better to fix this
+    t.join();
 
     return 0;
 }

examples/server/tests/unit/test_embedding.py

@@ -49,6 +49,26 @@ def test_embedding_multiple():
         assert len(d['embedding']) > 1
 
 
+def test_embedding_multiple_with_fa():
+    server = ServerPreset.bert_bge_small_with_fa()
+    server.pooling = 'last'
+    server.start()
+    # one of these should trigger the FA branch (i.e. context size % 256 == 0)
+    res = server.make_request("POST", "/v1/embeddings", data={
+        "input": [
+            "a "*253,
+            "b "*254,
+            "c "*255,
+            "d "*256,
+        ],
+    })
+    assert res.status_code == 200
+    assert len(res.body['data']) == 4
+    for d in res.body['data']:
+        assert 'embedding' in d
+        assert len(d['embedding']) > 1
+
+
 @pytest.mark.parametrize(
     "input,is_multi_prompt",
     [

examples/server/tests/utils.py

@@ -323,6 +323,21 @@ class ServerPreset:
         server.server_embeddings = True
         return server
 
+    @staticmethod
+    def bert_bge_small_with_fa() -> ServerProcess:
+        server = ServerProcess()
+        server.model_hf_repo = "ggml-org/models"
+        server.model_hf_file = "bert-bge-small/ggml-model-f16.gguf"
+        server.model_alias = "bert-bge-small"
+        server.n_ctx = 1024
+        server.n_batch = 300
+        server.n_ubatch = 300
+        server.n_slots = 2
+        server.fa = True
+        server.seed = 42
+        server.server_embeddings = True
+        return server
+
     @staticmethod
     def tinyllama_infill() -> ServerProcess:
         server = ServerProcess()

examples/server_embd.py

@@ -15,7 +15,7 @@ async def main():
     model_url = "http://127.0.0.1:6900"
     responses: list[requests.Response] = await asyncio.gather(*[requests_post_async(
         url= f"{model_url}/embedding",
-        json= {"content": str(0)*1024}
+        json= {"content": "a "*1022}
     ) for i in range(n)])
 
     for response in responses:

ggml/src/ggml-cpu/ops.cpp

@@ -6721,8 +6721,8 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     ggml_vec_dot_t    const kq_vec_dot     = ggml_get_type_traits_cpu(k->type)->vec_dot;
     ggml_to_float_t   const v_to_float     = ggml_get_type_traits(v->type)->to_float;
 
-    GGML_ASSERT(q_to_vec_dot && "fattn: unsupported K-type");
-    GGML_ASSERT(v_to_float   && "fattn: unsupported V-type");
+    GGML_ASSERT((                            q_to_vec_dot) && "fattn: unsupported K-type");
+    GGML_ASSERT((v->type == GGML_TYPE_F32 || v_to_float  ) && "fattn: unsupported V-type");
 
     // loop over n_batch and n_head
     for (int ir = ir0; ir < ir1; ++ir) {
@@ -6818,10 +6818,14 @@ static void ggml_compute_forward_flash_attn_ext_f16(
                     vs = expf(s - M);
                 }
 
-                v_to_float(v_data, V32, DV);
-
                 // V += v*expf(s - M)
-                ggml_vec_mad_f32(DV, VKQ32, V32, vs);
+                if (v_to_float) {
+                    v_to_float(v_data, V32, DV);
+                    ggml_vec_mad_f32(DV, VKQ32, V32, vs);
+                } else {
+                    // V is F32
+                    ggml_vec_mad_f32(DV, VKQ32, (const float *) v_data, vs);
+                }
             }
 
             S = S*ms + vs; // scale and increment sum with partial sum

ggml/src/ggml-cuda/cpy.cu

@@ -10,6 +10,13 @@ static __device__ void cpy_1_f32_f32(const char * cxi, char * cdsti) {
     *dsti = *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 = *xi;
+}
+
 static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) {
     const float * xi = (const float *) cxi;
     half * dsti = (half *) cdsti;
@@ -386,6 +393,16 @@ static void ggml_cpy_f32_f32_cuda(
         (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
+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, char ** cdst_indirect, int & graph_cpynode_index) {
+
+    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, cdst_indirect, graph_cpynode_index++);
+}
+
 static void ggml_cpy_f32_f16_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,
@@ -581,6 +598,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
         CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
         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, dest_ptrs_d, graph_cpynode_index);
+    } 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, dest_ptrs_d, graph_cpynode_index);
     } 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, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
@@ -634,6 +653,8 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
         return nullptr;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
         return (void*) cpy_f32_f16<cpy_1_f32_f32>;
+    } 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_F16) {
         return (void*) cpy_f32_f16<cpy_1_f32_f16>;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {

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

@@ -3079,6 +3079,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
                     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_F16) {
                     return true;
                 }

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

@@ -1345,6 +1345,11 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
         case GGML_OP_ARANGE:
             return true;
         case GGML_OP_FLASH_ATTN_EXT:
+            if (op->src[0]->ne[0] == 32) {
+                // head size == 32 (e.g. bert-bge-small)
+                // TODO: not sure if it is worth adding kernels for this size
+                return false;
+            }
             if (op->src[1]->type != op->src[2]->type) {
                 return false;
             }

src/llama-graph.cpp

@@ -1215,6 +1215,15 @@ ggml_tensor * llm_graph_context::build_attn_mha(
             v = ggml_transpose(ctx0, v);
         }
 
+        // this can happen when KV cache is not used (e.g. an embedding model with non-causal attn)
+        if (k->type == GGML_TYPE_F32) {
+            k = ggml_cast(ctx0, k, GGML_TYPE_F16);
+        }
+
+        if (v->type == GGML_TYPE_F32) {
+            v = ggml_cast(ctx0, v, GGML_TYPE_F16);
+        }
+
         cur = ggml_flash_attn_ext(ctx0, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias,
                                   hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f);