diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 0604ccc60..bccda506b 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -1978,7 +1978,7 @@ class Qwen2Model(Model): @Model.register("Qwen2VLForConditionalGeneration") class Qwen2VLModel(Model): - model_arch = gguf.MODEL_ARCH.QWEN2 + model_arch = gguf.MODEL_ARCH.QWEN2VL def set_vocab(self): try: @@ -1986,15 +1986,6 @@ class Qwen2VLModel(Model): except FileNotFoundError: self._set_vocab_gpt2() - # def map_tensor_name(self, name: str, try_suffixes: Sequence[str] = (".weight", ".bias")) -> str: - # new_name = self.tensor_map.get_name(key=name, try_suffixes=try_suffixes) - # if name.startswith("visual."): - # breakpoint() - # return "" - # if new_name is None: - # raise ValueError(f"Can not map tensor {name!r}") - # return new_name - def get_tensors(self) -> Iterator[tuple[str, Tensor]]: for name, data in super().get_tensors(): if name.startswith("visual."): diff --git a/examples/llava/qwen2_vl_surgery.py b/examples/llava/qwen2_vl_surgery.py new file mode 100644 index 000000000..2d5b32fe6 --- /dev/null +++ b/examples/llava/qwen2_vl_surgery.py @@ -0,0 +1,159 @@ +import argparse +import glob +import os +import torch +from safetensors import safe_open +from safetensors.torch import save_file +from typing import Any, ContextManager, cast + +# Function to determine if file is a SafeTensor file +def is_safetensor_file(file_path): + return file_path.endswith('.safetensors') + + +# Unified loading function +def load_model(file_path): + if is_safetensor_file(file_path): + tensors = {} + with cast(ContextManager[Any], safe_open(file_path, framework="pt", device="cpu")) as f: + for key in f.keys(): + tensors[key] = f.get_tensor(key).clone() + # output shape + print(f"{key} : {tensors[key].shape}") + return tensors, 'safetensor' + else: + return torch.load(file_path, map_location=torch.device('cpu')), 'pytorch' + + +# Unified saving function +def save_model(model, file_path, file_type): + if file_type == 'safetensor': + # safe_save(model, file_path) + save_file(model, file_path) + else: + torch.save(model, file_path) + + +# Adapted function to clean vision tower from checkpoint +def clean_vision_tower_from_checkpoint(checkpoint_path): + checkpoint, file_type = load_model(checkpoint_path) + # file_type = 'pytorch' + model_path = os.path.dirname(checkpoint_path) + print(f"Searching for vision tower tensors in {checkpoint_path}") + clip_tensors = [k for k, v in checkpoint.items() if (k.startswith("model.vision_tower") or k.startswith("vit."))] + + if len(clip_tensors) > 0: + print(f"Found {len(clip_tensors)} tensors to extract from {checkpoint_path}") + # Adapted for file type + clip_path = os.path.join(model_path, "llava.clip") + + if os.path.exists(clip_path): + print(f"Loading existing llava.clip from {clip_path}") + existing_clip, _ = load_model(clip_path) + else: + print(f"Creating new llava.clip at {clip_path}") + existing_clip = {} + # Update existing_clip with new tensors, avoid duplicates + for name in clip_tensors: + simple_name = name[name.index('vision_model.'):] if 'vision_model.' in name else name + print(f"Adding {simple_name} to llava.clip") + if simple_name not in existing_clip: + existing_clip[simple_name] = checkpoint[name] + + # Save the updated clip tensors back to llava.clip + save_model(existing_clip, clip_path, 'pytorch') + + # Remove the tensors from the original checkpoint + for name in clip_tensors: + del checkpoint[name] + + checkpoint_path = checkpoint_path + return True + return False + +def find_relevant_checkpoints(checkpoint_paths, newline_criteria, projector): + newline_checkpoint_path = None + projector_checkpoint_path = None + + for path in checkpoint_paths: + checkpoint, _ = load_model(path) + if newline_criteria(checkpoint) and newline_checkpoint_path is None: + newline_checkpoint_path = path + if projector(checkpoint): + projector_checkpoint_path = path + + return newline_checkpoint_path, projector_checkpoint_path + +def newline_criteria(checkpoint): + return any(k.startswith("model.image_newline") for k in checkpoint.keys()) + +def proj_criteria(checkpoint): + return any(k.startswith("model.mm_projector") or k.startswith("vision_proj.") for k in checkpoint.keys()) + + +# Command-line interface setup +ap = argparse.ArgumentParser() +ap.add_argument("-m", "--model", required=True, help="Path to LLaVA v1.5+ model") +ap.add_argument("-C", "--clean-vision-tower", action="store_true", help="Remove any vision tower from the model files") +args = ap.parse_args() + +if args.clean_vision_tower: + # Generalized to handle both PyTorch and SafeTensors models + model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True) + # checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and path.startswith('pytorch')) or (path.endswith('.safetensors') and path.startswith('model'))] + checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])] + for projector_checkpoint_path in checkpoint_paths: + print(f"Cleaning {projector_checkpoint_path}") + if not clean_vision_tower_from_checkpoint(projector_checkpoint_path): + print(f"No vision tower found in {projector_checkpoint_path}") + # we break once none is found, so far all models append them at the end + # break + print("Done! All vision tower tensors are removed from the model files and stored in llava.clip file.") + +# Now we look for the projector in the last checkpoint +model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True) +checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])] +# last_checkpoint_path = checkpoint_paths[0] +# first_checkpoint_path = checkpoint_paths[-1] +newline_checkpoint_path, projector_checkpoint_path = find_relevant_checkpoints(checkpoint_paths, newline_criteria, proj_criteria) + +print(f"Taking projector from {projector_checkpoint_path}") +first_mm_tensors = [] +first_checkpoint = None +if newline_checkpoint_path is not None: + print(f"Taking newline from {newline_checkpoint_path}") + first_checkpoint, file_type = load_model(newline_checkpoint_path) + first_mm_tensors = [k for k, v in first_checkpoint.items() if k.startswith("model.image_newline")] + +# Load the checkpoint +mm_tensors = [] +last_checkpoint = None +if projector_checkpoint_path is not None: + last_checkpoint, file_type = load_model(projector_checkpoint_path) + mm_tensors = [k for k, v in last_checkpoint.items() if k.startswith("model.mm_projector") or k.startswith("vision_proj.")] + +if len(mm_tensors) == 0: + if last_checkpoint is not None: + for k, v in last_checkpoint.items(): + print(k) + print(f"Found {len(mm_tensors)} tensors to extract out of {len(last_checkpoint) if last_checkpoint is not None else 0} tensors.") + print("No tensors found. Is this a LLaVA model?") + exit() + +print(f"Found {len(mm_tensors)} tensors to extract.") +print(f"Found additional {len(first_mm_tensors)} tensors to extract.") +# projector = {name: checkpoint.[name].float() for name in mm_tensors} +projector = {} +for name in mm_tensors: + assert last_checkpoint is not None + projector[name] = last_checkpoint[name].float() +for name in first_mm_tensors: + assert first_checkpoint is not None + projector[name] = first_checkpoint[name].float() + +if len(projector) > 0: + save_model(projector, f"{args.model}/llava.projector", 'pytorch') + +print("Done!") +print(f"Now you can convert {args.model} to a a regular LLaMA GGUF file.") +print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.") diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index 65cb92c44..07d66af20 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -1445,6 +1445,21 @@ extern "C" { float beta_fast, float beta_slow); + GGML_API struct ggml_tensor * ggml_mrope_ext( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c, + int n_dims, + int mode, + int n_ctx_orig, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow); + // in-place, returns view(a) GGML_API struct ggml_tensor * ggml_rope_ext_inplace( struct ggml_context * ctx, diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index 1a9a7efaf..c5926c10f 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -3515,15 +3515,23 @@ static struct ggml_tensor * ggml_rope_impl( GGML_ASSERT(c->ne[0] >= n_dims / 2); } + bool is_node = false; + int sections[3] = {0, 0, 0}; + + if (a->grad) { + is_node = true; + } + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - int32_t params[11] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig }; + int32_t params[14] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig }; memcpy(params + 5, &freq_base, sizeof(float)); memcpy(params + 6, &freq_scale, sizeof(float)); memcpy(params + 7, &ext_factor, sizeof(float)); memcpy(params + 8, &attn_factor, sizeof(float)); memcpy(params + 9, &beta_fast, sizeof(float)); memcpy(params + 10, &beta_slow, sizeof(float)); + memcpy(params + 11, §ions, sizeof(int) * 3); ggml_set_op_params(result, params, sizeof(params)); result->op = GGML_OP_ROPE; @@ -3545,6 +3553,61 @@ struct ggml_tensor * ggml_rope( ); } +struct ggml_tensor * ggml_mrope_ext( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c, + int n_dims, + int mode, + int n_ctx_orig, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow) { + + int sections[3] = {16, 24, 24}; // TODO: move this into gguf model file. + + GGML_ASSERT((mode & 1) == 0 && "mode & 1 == 1 is no longer supported"); + + GGML_ASSERT(ggml_is_vector(b)); + GGML_ASSERT(b->type == GGML_TYPE_I32); + GGML_ASSERT(a->ne[2] * 3 == b->ne[0]); + + if (c) { + GGML_ASSERT(c->type == GGML_TYPE_F32); + GGML_ASSERT(c->ne[0] >= n_dims / 2); + } + + bool is_node = false; + + if (a->grad) { + is_node = true; + } + + struct ggml_tensor * result = ggml_dup_tensor(ctx, a); + + int32_t params[11 + 3] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig }; + memcpy(params + 5, &freq_base, sizeof(float)); + memcpy(params + 6, &freq_scale, sizeof(float)); + memcpy(params + 7, &ext_factor, sizeof(float)); + memcpy(params + 8, &attn_factor, sizeof(float)); + memcpy(params + 9, &beta_fast, sizeof(float)); + memcpy(params + 10, &beta_slow, sizeof(float)); + memcpy(params + 11, §ions, sizeof(int) * 3); + ggml_set_op_params(result, params, sizeof(params)); + + result->op = GGML_OP_ROPE; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src[0] = a; + result->src[1] = b; + result->src[2] = c; + + return result; +} + struct ggml_tensor * ggml_rope_inplace( struct ggml_context * ctx, struct ggml_tensor * a, @@ -4967,6 +5030,9626 @@ struct ggml_tensor * ggml_opt_step_adamw( //////////////////////////////////////////////////////////////////////////////// +// ggml_compute_forward_dup + +static void ggml_compute_forward_dup_same_cont( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); + GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); + GGML_ASSERT(src0->type == dst->type); + + const size_t nb0 = ggml_type_size(src0->type); + + const int ith = params->ith; // thread index + const int nth = params->nth; // number of threads + + // parallelize by elements + const int ne = ggml_nelements(dst); + const int dr = (ne + nth - 1) / nth; + const int ie0 = dr * ith; + const int ie1 = MIN(ie0 + dr, ne); + + if (ie0 < ie1) { + memcpy( + ((char *) dst->data + ie0*nb0), + ((char *) src0->data + ie0*nb0), + (ie1 - ie0) * nb0); + } +} + +static void ggml_compute_forward_dup_f16( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); + + GGML_TENSOR_UNARY_OP_LOCALS + + const int ith = params->ith; // thread index + const int nth = params->nth; // number of threads + + // parallelize by rows + const int nr = ne01; + // number of rows per thread + const int dr = (nr + nth - 1) / nth; + // row range for this thread + const int ir0 = dr * ith; + const int ir1 = MIN(ir0 + dr, nr); + + if (src0->type == dst->type && + ne00 == ne0 && + nb00 == ggml_type_size(src0->type) && nb0 == ggml_type_size(dst->type)) { + // copy by rows + const size_t rs = ne00*nb00; + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = ir0; i01 < ir1; i01++) { + memcpy( + ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3), + ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03), + rs); + } + } + } + return; + } + + // TODO: add more special-case implementations for tensor shapes/strides that can benefit from memcpy + + if (ggml_is_contiguous(dst)) { + if (nb00 == sizeof(ggml_fp16_t)) { + if (dst->type == GGML_TYPE_F16) { + size_t id = 0; + const size_t rs = ne00 * nb00; + char * dst_ptr = (char *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03; + memcpy(dst_ptr + id, src0_ptr, rs); + id += rs; + } + id += rs * (ne01 - ir1); + } + } + } else if (dst->type == GGML_TYPE_F32) { + size_t id = 0; + float * dst_ptr = (float *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + for (int i00 = 0; i00 < ne00; i00++) { + dst_ptr[id] = GGML_FP16_TO_FP32(src0_ptr[i00]); + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else if (type_traits[dst->type].from_float) { + ggml_from_float_t const quantize_row_q = type_traits[dst->type].from_float; + float * src0_f32 = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith; + + size_t id = 0; + size_t rs = nb0 * (ne00 / ggml_blck_size(dst->type)); + char * dst_ptr = (char *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + + for (int i00 = 0; i00 < ne00; i00++) { + src0_f32[i00] = GGML_FP16_TO_FP32(src0_ptr[i00]); + } + + quantize_row_q(src0_f32, dst_ptr + id, ne00); + id += rs; + } + id += rs * (ne01 - ir1); + } + } + } else { + GGML_ABORT("fatal error"); // TODO: implement + } + } else { + //printf("%s: this is not optimal - fix me\n", __func__); + + if (dst->type == GGML_TYPE_F32) { + size_t id = 0; + float * dst_ptr = (float *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = GGML_FP16_TO_FP32(*src0_ptr); + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else if (dst->type == GGML_TYPE_F16) { + size_t id = 0; + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = *src0_ptr; + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else { + GGML_ABORT("fatal error"); // TODO: implement + } + } + return; + } + + // dst counters + int64_t i10 = 0; + int64_t i11 = 0; + int64_t i12 = 0; + int64_t i13 = 0; + + if (dst->type == GGML_TYPE_F16) { + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + memcpy(dst_ptr, src0_ptr, sizeof(ggml_fp16_t)); + + if (++i10 == ne00) { + i10 = 0; + if (++i11 == ne01) { + i11 = 0; + if (++i12 == ne02) { + i12 = 0; + if (++i13 == ne03) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + } else if (dst->type == GGML_TYPE_F32) { + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + *(float *) dst_ptr = GGML_FP16_TO_FP32(*(const ggml_fp16_t *) src0_ptr); + + if (++i10 == ne0) { + i10 = 0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + } else { + GGML_ABORT("fatal error"); // TODO: implement + } +} + +static void ggml_compute_forward_dup_bf16( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); + + GGML_TENSOR_UNARY_OP_LOCALS + + const int ith = params->ith; // thread index + const int nth = params->nth; // number of threads + + // parallelize by rows + const int nr = ne01; + // number of rows per thread + const int dr = (nr + nth - 1) / nth; + // row range for this thread + const int ir0 = dr * ith; + const int ir1 = MIN(ir0 + dr, nr); + + if (src0->type == dst->type && + ne00 == ne0 && + nb00 == ggml_type_size(src0->type) && nb0 == ggml_type_size(dst->type)) { + // copy by rows + const size_t rs = ne00*nb00; + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = ir0; i01 < ir1; i01++) { + memcpy( + ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3), + ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03), + rs); + } + } + } + return; + } + + // TODO: add more special-case implementations for tensor shapes/strides that can benefit from memcpy + + if (ggml_is_contiguous(dst)) { + if (nb00 == sizeof(ggml_bf16_t)) { + if (dst->type == GGML_TYPE_BF16) { + size_t id = 0; + const size_t rs = ne00 * nb00; + char * dst_ptr = (char *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03; + memcpy(dst_ptr + id, src0_ptr, rs); + id += rs; + } + id += rs * (ne01 - ir1); + } + } + } else if (dst->type == GGML_TYPE_F16) { + size_t id = 0; + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + for (int i00 = 0; i00 < ne00; i00++) { + dst_ptr[id] = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(src0_ptr[i00])); + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else if (dst->type == GGML_TYPE_F32) { + size_t id = 0; + float * dst_ptr = (float *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + for (int i00 = 0; i00 < ne00; i00++) { + dst_ptr[id] = GGML_BF16_TO_FP32(src0_ptr[i00]); + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else if (type_traits[dst->type].from_float) { + ggml_from_float_t const quantize_row_q = type_traits[dst->type].from_float; + float * src0_f32 = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith; + + size_t id = 0; + size_t rs = nb0 * (ne00 / ggml_blck_size(dst->type)); + char * dst_ptr = (char *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + + for (int i00 = 0; i00 < ne00; i00++) { + src0_f32[i00] = GGML_BF16_TO_FP32(src0_ptr[i00]); + } + + quantize_row_q(src0_f32, dst_ptr + id, ne00); + id += rs; + } + id += rs * (ne01 - ir1); + } + } + } else { + GGML_ABORT("fatal error"); // TODO: implement + } + } else { + //printf("%s: this is not optimal - fix me\n", __func__); + + if (dst->type == GGML_TYPE_F32) { + size_t id = 0; + float * dst_ptr = (float *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = GGML_BF16_TO_FP32(*src0_ptr); + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else if (dst->type == GGML_TYPE_BF16) { + size_t id = 0; + ggml_bf16_t * dst_ptr = (ggml_bf16_t *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = *src0_ptr; + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else if (dst->type == GGML_TYPE_F16) { + size_t id = 0; + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(*src0_ptr)); + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else { + GGML_ABORT("fatal error"); // TODO: implement + } + } + return; + } + + // dst counters + int64_t i10 = 0; + int64_t i11 = 0; + int64_t i12 = 0; + int64_t i13 = 0; + + if (dst->type == GGML_TYPE_BF16) { + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + memcpy(dst_ptr, src0_ptr, sizeof(ggml_bf16_t)); + + if (++i10 == ne00) { + i10 = 0; + if (++i11 == ne01) { + i11 = 0; + if (++i12 == ne02) { + i12 = 0; + if (++i13 == ne03) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + } else if (dst->type == GGML_TYPE_F16) { + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + *(ggml_fp16_t *) dst_ptr = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(*(const ggml_bf16_t *) src0_ptr)); + + if (++i10 == ne0) { + i10 = 0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + } else if (dst->type == GGML_TYPE_F32) { + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + *(float *) dst_ptr = GGML_BF16_TO_FP32(*(const ggml_bf16_t *) src0_ptr); + + if (++i10 == ne0) { + i10 = 0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + } else { + GGML_ABORT("fatal error"); // TODO: implement + } +} + +static void ggml_compute_forward_dup_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); + + GGML_TENSOR_UNARY_OP_LOCALS + + const int ith = params->ith; // thread index + const int nth = params->nth; // number of threads + + // parallelize by rows + const int nr = ne01; + // number of rows per thread + const int dr = (nr + nth - 1) / nth; + // row range for this thread + const int ir0 = dr * ith; + const int ir1 = MIN(ir0 + dr, nr); + + if (src0->type == dst->type && + ne00 == ne0 && + nb00 == ggml_type_size(src0->type) && nb0 == ggml_type_size(dst->type)) { + // copy by rows + const size_t rs = ne00*nb00; + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = ir0; i01 < ir1; i01++) { + memcpy( + ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3), + ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03), + rs); + } + } + } + return; + } + + if (ggml_is_contiguous(dst)) { + // TODO: simplify + if (nb00 == sizeof(float)) { + if (dst->type == GGML_TYPE_F32) { + size_t id = 0; + const size_t rs = ne00 * nb00; + char * dst_ptr = (char *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03; + memcpy(dst_ptr + id, src0_ptr, rs); + id += rs; + } + id += rs * (ne01 - ir1); + } + } + } else if (type_traits[dst->type].from_float) { + ggml_from_float_t const quantize_row_q = type_traits[dst->type].from_float; + + size_t id = 0; + size_t rs = nb0 * (ne00 / ggml_blck_size(dst->type)); + char * dst_ptr = (char *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const float * src0_ptr = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + quantize_row_q(src0_ptr, dst_ptr + id, ne00); + id += rs; + } + id += rs * (ne01 - ir1); + } + } + } else { + GGML_ABORT("fatal error"); // TODO: implement + } + } else { + //printf("%s: this is not optimal - fix me\n", __func__); + + if (dst->type == GGML_TYPE_F32) { + size_t id = 0; + float * dst_ptr = (float *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = *src0_ptr; + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else if (dst->type == GGML_TYPE_F16) { + size_t id = 0; + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = GGML_FP32_TO_FP16(*src0_ptr); + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else if (dst->type == GGML_TYPE_BF16) { + size_t id = 0; + ggml_bf16_t * dst_ptr = (ggml_bf16_t *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = GGML_FP32_TO_BF16(*src0_ptr); + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else { + GGML_ABORT("fatal error"); // TODO: implement + } + } + + return; + } + + // dst counters + + int64_t i10 = 0; + int64_t i11 = 0; + int64_t i12 = 0; + int64_t i13 = 0; + + if (dst->type == GGML_TYPE_F32) { + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + memcpy(dst_ptr, src0_ptr, sizeof(float)); + + if (++i10 == ne0) { + i10 = 0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + } else if (dst->type == GGML_TYPE_F16) { + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + *(ggml_fp16_t *) dst_ptr = GGML_FP32_TO_FP16(*(const float *) src0_ptr); + + if (++i10 == ne0) { + i10 = 0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + } else if (dst->type == GGML_TYPE_BF16) { + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + *(ggml_bf16_t *) dst_ptr = GGML_FP32_TO_BF16(*(const float *) src0_ptr); + + if (++i10 == ne0) { + i10 = 0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + } else { + GGML_ABORT("fatal error"); // TODO: implement + } +} + +// A simplified version of ggml_compute_forward_dup that doesn't do float upcasting, and just plain old memcpy. +static void ggml_compute_forward_dup_bytes( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); + GGML_ASSERT(src0->type == dst->type); + + GGML_TENSOR_UNARY_OP_LOCALS; + + if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst)) { + ggml_compute_forward_dup_same_cont(params, dst); + return; + } + + const size_t type_size = ggml_type_size(src0->type); + const int ith = params->ith; // thread index + const int nth = params->nth; // number of threads + + + // parallelize by rows + const int nr = ne01; + // number of rows per thread + const int dr = (nr + nth - 1) / nth; + // row range for this thread + const int ir0 = dr * ith; + const int ir1 = MIN(ir0 + dr, nr); + + if (src0->type == dst->type && + ne00 == ne0 && + nb00 == type_size && nb0 == type_size) { + // copy by rows + const size_t rs = ne00 * type_size; + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = ir0; i01 < ir1; i01++) { + memcpy( + ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3), + ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03), + rs); + } + } + } + return; + } + + if (ggml_is_contiguous(dst)) { + size_t id = 0; + char * dst_ptr = (char *) dst->data; + const size_t rs = ne00 * type_size; + + if (nb00 == type_size) { + // src0 is contigous on first dimension, copy by rows + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int64_t i01 = ir0; i01 < ir1; i01++) { + const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03; + memcpy(dst_ptr + id, src0_ptr, rs); + id += rs; + } + id += rs * (ne01 - ir1); + } + } + } else { + //printf("%s: this is not optimal - fix me\n", __func__); + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = (char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03; + memcpy(dst_ptr + id, src0_ptr, type_size); + + id += type_size; + } + } + id += rs * (ne01 - ir1); + } + } + } + + return; + } + + // dst counters + + int64_t i10 = 0; + int64_t i11 = 0; + int64_t i12 = 0; + int64_t i13 = 0; + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + memcpy(dst_ptr, src0_ptr, type_size); + + if (++i10 == ne0) { + i10 = 0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } +} + +static void ggml_compute_forward_dup( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (src0->type == dst->type) { + ggml_compute_forward_dup_bytes(params, dst); + return; + } + + switch (src0->type) { + case GGML_TYPE_F16: + { + ggml_compute_forward_dup_f16(params, dst); + } break; + case GGML_TYPE_BF16: + { + ggml_compute_forward_dup_bf16(params, dst); + } break; + case GGML_TYPE_F32: + { + ggml_compute_forward_dup_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_add + +static void ggml_compute_forward_add_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_BINARY_OP_LOCALS + + GGML_ASSERT( nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + if (nb10 == sizeof(float)) { + for (int ir = ir0; ir < ir1; ++ir) { + // src1 is broadcastable across src0 and dst in i1, i2, i3 + const int64_t i03 = ir/(ne02*ne01); + const int64_t i02 = (ir - i03*ne02*ne01)/ne01; + const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); + + const int64_t i13 = i03 % ne13; + const int64_t i12 = i02 % ne12; + const int64_t i11 = i01 % ne11; + const int64_t nr0 = ne00 / ne10; + + float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); + float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); + float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11); + + for (int64_t r = 0; r < nr0; ++r) { +#ifdef GGML_USE_ACCELERATE + vDSP_vadd(src0_ptr + r*ne10, 1, src1_ptr, 1, dst_ptr + r*ne10, 1, ne10); +#else + ggml_vec_add_f32(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr); +#endif + } + } + } else { + // src1 is not contiguous + for (int ir = ir0; ir < ir1; ++ir) { + // src1 is broadcastable across src0 and dst in i1, i2, i3 + const int64_t i03 = ir/(ne02*ne01); + const int64_t i02 = (ir - i03*ne02*ne01)/ne01; + const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); + + const int64_t i13 = i03 % ne13; + const int64_t i12 = i02 % ne12; + const int64_t i11 = i01 % ne11; + + float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); + float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); + + for (int64_t i0 = 0; i0 < ne0; ++i0) { + const int64_t i10 = i0 % ne10; + float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10); + + dst_ptr[i0] = src0_ptr[i0] + *src1_ptr; + } + } + } +} + +static void ggml_compute_forward_add_f16_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_BINARY_OP_LOCALS + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + if (dst->type == GGML_TYPE_F32) { + GGML_ASSERT( nb0 == sizeof(float)); + } + else { + GGML_ASSERT(dst->type == GGML_TYPE_F16); + GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); + } + + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + if (nb10 == sizeof(float)) { + if (dst->type == GGML_TYPE_F16) { + for (int ir = ir0; ir < ir1; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); + + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + src1_ptr[i]); + } + } + } else { + for (int ir = ir0; ir < ir1; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + float * dst_ptr = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); + + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP16_TO_FP32(src0_ptr[i]) + src1_ptr[i]; + } + } + } + } + else { + // src1 is not contiguous + GGML_ABORT("fatal error"); + } +} + +static void ggml_compute_forward_add_bf16_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_BINARY_OP_LOCALS + + GGML_ASSERT(src0->type == GGML_TYPE_BF16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + if (dst->type == GGML_TYPE_F32) { + GGML_ASSERT( nb0 == sizeof(float)); + } + else { + GGML_ASSERT(dst->type == GGML_TYPE_BF16); + GGML_ASSERT( nb0 == sizeof(ggml_bf16_t)); + } + + GGML_ASSERT(nb00 == sizeof(ggml_bf16_t)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + if (nb10 == sizeof(float)) { + if (dst->type == GGML_TYPE_BF16) { + for (int ir = ir0; ir < ir1; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + ggml_bf16_t * dst_ptr = (ggml_bf16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); + ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); + + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_BF16(GGML_BF16_TO_FP32(src0_ptr[i]) + src1_ptr[i]); + } + } + } else { + for (int ir = ir0; ir < ir1; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + float * dst_ptr = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); + ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); + + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_BF16_TO_FP32(src0_ptr[i]) + src1_ptr[i]; + } + } + } + } + else { + // src1 is not contiguous + GGML_ABORT("fatal error"); + } +} + +static void ggml_compute_forward_add_f16_f16( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_BINARY_OP_LOCALS + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F16); + + GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + if (nb10 == sizeof(ggml_fp16_t)) { + for (int ir = ir0; ir < ir1; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + ggml_fp16_t * src1_ptr = (ggml_fp16_t *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); + + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + GGML_FP16_TO_FP32(src1_ptr[i])); + } + } + } + else { + // src1 is not contiguous + GGML_ABORT("fatal error"); + } +} + +static void ggml_compute_forward_add_bf16_bf16( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_BINARY_OP_LOCALS + + GGML_ASSERT(src0->type == GGML_TYPE_BF16); + GGML_ASSERT(src1->type == GGML_TYPE_BF16); + GGML_ASSERT(dst->type == GGML_TYPE_BF16); + + GGML_ASSERT( nb0 == sizeof(ggml_bf16_t)); + GGML_ASSERT(nb00 == sizeof(ggml_bf16_t)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + if (nb10 == sizeof(ggml_bf16_t)) { + for (int ir = ir0; ir < ir1; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + ggml_bf16_t * dst_ptr = (ggml_bf16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); + ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + ggml_bf16_t * src1_ptr = (ggml_bf16_t *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); + + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_BF16(GGML_BF16_TO_FP32(src0_ptr[i]) + GGML_BF16_TO_FP32(src1_ptr[i])); + } + } + } + else { + // src1 is not contiguous + GGML_ABORT("fatal error"); + } +} + +static void ggml_compute_forward_add_q_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_BINARY_OP_LOCALS + + const int ith = params->ith; + const int nth = params->nth; + + const enum ggml_type type = src0->type; + const enum ggml_type dtype = dst->type; + ggml_to_float_t const dequantize_row_q = type_traits[type].to_float; + ggml_from_float_t const quantize_row_q = type_traits[dtype].from_float; + + // we don't support permuted src0 or src1 + GGML_ASSERT(nb00 == ggml_type_size(type)); + GGML_ASSERT(nb10 == sizeof(float)); + + // dst cannot be transposed or permuted + GGML_ASSERT(nb0 <= nb1); + GGML_ASSERT(nb1 <= nb2); + GGML_ASSERT(nb2 <= nb3); + + GGML_ASSERT(ggml_is_quantized(src0->type)); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + float * wdata = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith; + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 indices + const int i03 = ir/(ne02*ne01); + const int i02 = (ir - i03*ne02*ne01)/ne01; + const int i01 = (ir - i03*ne02*ne01 - i02*ne01); + + // src1 and dst are same shape as src0 => same indices + const int i13 = i03; + const int i12 = i02; + const int i11 = i01; + + const int i3 = i03; + const int i2 = i02; + const int i1 = i01; + + void * src0_row = (void *) ((char *) src0->data + (i01*nb01 + i02*nb02 + i03*nb03)); + float * src1_row = (float *)((char *) src1->data + (i11*nb11 + i12*nb12 + i13*nb13)); + void * dst_row = (void *) ((char *) dst->data + ( i1*nb1 + i2*nb2 + i3*nb3)); + + assert(ne00 % 32 == 0); + + // unquantize row from src0 to temp buffer + dequantize_row_q(src0_row, wdata, ne00); + // add src1 + ggml_vec_acc_f32(ne00, wdata, src1_row); + // quantize row to dst + if (quantize_row_q != NULL) { + quantize_row_q(wdata, dst_row, ne00); + } else { + memcpy(dst_row, wdata, ne0*nb0); + } + } +} + +static void ggml_compute_forward_add( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + if (src1->type == GGML_TYPE_F32) { + ggml_compute_forward_add_f32(params, dst); + } + else { + GGML_ABORT("fatal error"); + } + } break; + case GGML_TYPE_F16: + { + if (src1->type == GGML_TYPE_F16) { + ggml_compute_forward_add_f16_f16(params, dst); + } + else if (src1->type == GGML_TYPE_F32) { + ggml_compute_forward_add_f16_f32(params, dst); + } + else { + GGML_ABORT("fatal error"); + } + } break; + case GGML_TYPE_BF16: + { + if (src1->type == GGML_TYPE_BF16) { + ggml_compute_forward_add_bf16_bf16(params, dst); + } + else if (src1->type == GGML_TYPE_F32) { + ggml_compute_forward_add_bf16_f32(params, dst); + } + else { + GGML_ABORT("fatal error"); + } + } break; + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q2_K: + case GGML_TYPE_Q3_K: + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + case GGML_TYPE_Q6_K: + case GGML_TYPE_TQ1_0: + case GGML_TYPE_TQ2_0: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ4_NL: + case GGML_TYPE_IQ4_XS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ2_S: + case GGML_TYPE_Q4_0_4_4: + case GGML_TYPE_Q4_0_4_8: + case GGML_TYPE_Q4_0_8_8: + { + ggml_compute_forward_add_q_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_add1 + +static void ggml_compute_forward_add1_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_scalar(src1)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_UNARY_OP_LOCALS + + GGML_ASSERT( nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + +#ifdef GGML_USE_ACCELERATE + UNUSED(ggml_vec_add1_f32); + + vDSP_vadd( + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1, + (float *) ((char *) src1->data), 0, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), 1, + ne0); +#else + ggml_vec_add1_f32(ne0, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), + *(float *) src1->data); +#endif + } +} + +static void ggml_compute_forward_add1_f16_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_scalar(src1)); + + // scalar to add + const float v = *(float *) src1->data; + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_UNARY_OP_LOCALS + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F16); + + GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + v); + } + } +} + +static void ggml_compute_forward_add1_f16_f16( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_scalar(src1)); + + // scalar to add + const float v = GGML_FP16_TO_FP32(*(ggml_fp16_t *) src1->data); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_UNARY_OP_LOCALS + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F16); + + GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + v); + } + } +} + +static void ggml_compute_forward_add1_q_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_scalar(src1)); + + // scalar to add + const float v = *(float *) src1->data; + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_UNARY_OP_LOCALS + + const enum ggml_type type = src0->type; + ggml_to_float_t const dequantize_row_q = type_traits[type].to_float; + ggml_from_float_t const quantize_row_q = type_traits[type].from_float; + + // we don't support permuted src0 + GGML_ASSERT(nb00 == ggml_type_size(type)); + + // dst cannot be transposed or permuted + GGML_ASSERT(nb0 <= nb1); + GGML_ASSERT(nb1 <= nb2); + GGML_ASSERT(nb2 <= nb3); + + GGML_ASSERT(ggml_is_quantized(src0->type)); + GGML_ASSERT(dst->type == src0->type); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + float * wdata = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32) * ith; + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + void * src0_row = (void *) ((char *) src0->data + (i1*nb01 + i2*nb02 + i3*nb03)); + void * dst_row = (void *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb0 )); + + assert(ne0 % 32 == 0); + + // unquantize row from src0 to temp buffer + dequantize_row_q(src0_row, wdata, ne0); + // add src1 + ggml_vec_acc1_f32(ne0, wdata, v); + // quantize row to dst + quantize_row_q(wdata, dst_row, ne0); + } +} + +static void ggml_compute_forward_add1_bf16_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_scalar(src1)); + + // scalar to add + const float v = *(float *) src1->data; + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_UNARY_OP_LOCALS + + GGML_ASSERT(src0->type == GGML_TYPE_BF16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_BF16); + + GGML_ASSERT( nb0 == sizeof(ggml_bf16_t)); + GGML_ASSERT(nb00 == sizeof(ggml_bf16_t)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + ggml_bf16_t * dst_ptr = (ggml_bf16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); + ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_BF16(GGML_BF16_TO_FP32(src0_ptr[i]) + v); + } + } +} + +static void ggml_compute_forward_add1_bf16_bf16( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_scalar(src1)); + + // scalar to add + const float v = GGML_BF16_TO_FP32(*(ggml_bf16_t *) src1->data); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_UNARY_OP_LOCALS + + GGML_ASSERT(src0->type == GGML_TYPE_BF16); + GGML_ASSERT(src1->type == GGML_TYPE_BF16); + GGML_ASSERT(dst->type == GGML_TYPE_BF16); + + GGML_ASSERT( nb0 == sizeof(ggml_bf16_t)); + GGML_ASSERT(nb00 == sizeof(ggml_bf16_t)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + ggml_bf16_t * dst_ptr = (ggml_bf16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); + ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_BF16(GGML_BF16_TO_FP32(src0_ptr[i]) + v); + } + } +} + +static void ggml_compute_forward_add1( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_add1_f32(params, dst); + } break; + case GGML_TYPE_F16: + { + if (src1->type == GGML_TYPE_F16) { + ggml_compute_forward_add1_f16_f16(params, dst); + } + else if (src1->type == GGML_TYPE_F32) { + ggml_compute_forward_add1_f16_f32(params, dst); + } + else { + GGML_ABORT("fatal error"); + } + } break; + case GGML_TYPE_BF16: + { + if (src1->type == GGML_TYPE_BF16) { + ggml_compute_forward_add1_bf16_bf16(params, dst); + } + else if (src1->type == GGML_TYPE_F32) { + ggml_compute_forward_add1_bf16_f32(params, dst); + } + else { + GGML_ABORT("fatal error"); + } + } break; + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q8_1: + case GGML_TYPE_Q2_K: + case GGML_TYPE_Q3_K: + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + case GGML_TYPE_Q6_K: + case GGML_TYPE_TQ1_0: + case GGML_TYPE_TQ2_0: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ4_NL: + case GGML_TYPE_IQ4_XS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ2_S: + case GGML_TYPE_Q4_0_4_4: + case GGML_TYPE_Q4_0_4_8: + case GGML_TYPE_Q4_0_8_8: + { + ggml_compute_forward_add1_q_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_acc + +static void ggml_compute_forward_acc_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); + + // view src0 and dst with these strides and data offset inbytes during acc + // nb0 is implicitly element_size because src0 and dst are contiguous + size_t nb1 = ((int32_t *) dst->op_params)[0]; + size_t nb2 = ((int32_t *) dst->op_params)[1]; + size_t nb3 = ((int32_t *) dst->op_params)[2]; + size_t offset = ((int32_t *) dst->op_params)[3]; + bool inplace = (bool) ((int32_t *) dst->op_params)[4]; + + if (!inplace) { + if (params->ith == 0) { + // memcpy needs to be synchronized across threads to avoid race conditions. + // => do it in INIT phase + memcpy( + ((char *) dst->data), + ((char *) src0->data), + ggml_nbytes(dst)); + } + ggml_barrier(params->threadpool); + } + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src1); + const int nc = src1->ne[0]; + + GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) + GGML_TENSOR_LOCALS(size_t, nb1, src1, nb) + + // src0 and dst as viewed during acc + const size_t nb0 = ggml_element_size(src0); + + const size_t nb00 = nb0; + const size_t nb01 = nb1; + const size_t nb02 = nb2; + const size_t nb03 = nb3; + + GGML_ASSERT(offset + (ne10 == 0 ? 0 : ne10-1)*nb0 + (ne11 == 0 ? 0 : ne11-1)*nb1 + (ne12 == 0 ? 0 : ne12-1)*nb2 + (ne13 == 0 ? 0 : ne13-1)*nb3 < ggml_nbytes(dst)); + GGML_ASSERT(offset + (ne10 == 0 ? 0 : ne10-1)*nb00 + (ne11 == 0 ? 0 : ne11-1)*nb01 + (ne12 == 0 ? 0 : ne12-1)*nb02 + (ne13 == 0 ? 0 : ne13-1)*nb03 < ggml_nbytes(src0)); + + GGML_ASSERT(nb10 == sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are viewed with shape of src1 and offset + // => same indices + const int i3 = ir/(ne12*ne11); + const int i2 = (ir - i3*ne12*ne11)/ne11; + const int i1 = (ir - i3*ne12*ne11 - i2*ne11); + +#ifdef GGML_USE_ACCELERATE + vDSP_vadd( + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + offset), 1, + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + offset), 1, nc); +#else + ggml_vec_add_f32(nc, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + offset), + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + offset), + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); +#endif + } +} + +static void ggml_compute_forward_acc( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_acc_f32(params, dst); + } break; + case GGML_TYPE_F16: + case GGML_TYPE_BF16: + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q8_1: + case GGML_TYPE_Q2_K: + case GGML_TYPE_Q3_K: + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + case GGML_TYPE_Q6_K: + case GGML_TYPE_TQ1_0: + case GGML_TYPE_TQ2_0: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ4_NL: + case GGML_TYPE_IQ4_XS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ2_S: + case GGML_TYPE_Q4_0_4_4: + case GGML_TYPE_Q4_0_4_8: + case GGML_TYPE_Q4_0_8_8: + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_sub + +static void ggml_compute_forward_sub_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + assert(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_BINARY_OP_LOCALS + + GGML_ASSERT( nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + if (nb10 == sizeof(float)) { + for (int ir = ir0; ir < ir1; ++ir) { + // src1 is broadcastable across src0 and dst in i1, i2, i3 + const int64_t i03 = ir/(ne02*ne01); + const int64_t i02 = (ir - i03*ne02*ne01)/ne01; + const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); + + const int64_t i13 = i03 % ne13; + const int64_t i12 = i02 % ne12; + const int64_t i11 = i01 % ne11; + const int64_t nr0 = ne00 / ne10; + + float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); + float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); + float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11); + + for (int64_t r = 0; r < nr0; ++r) { +#ifdef GGML_USE_ACCELERATE + vDSP_vsub(src1_ptr, 1, src0_ptr + r*ne10, 1, dst_ptr + r*ne10, 1, ne10); +#else + ggml_vec_sub_f32(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr); +#endif + } + } + } else { + // src1 is not contiguous + for (int ir = ir0; ir < ir1; ++ir) { + // src1 is broadcastable across src0 and dst in i1, i2, i3 + const int64_t i03 = ir/(ne02*ne01); + const int64_t i02 = (ir - i03*ne02*ne01)/ne01; + const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); + + const int64_t i13 = i03 % ne13; + const int64_t i12 = i02 % ne12; + const int64_t i11 = i01 % ne11; + + float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); + float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); + + for (int64_t i0 = 0; i0 < ne0; ++i0) { + const int64_t i10 = i0 % ne10; + float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10); + + dst_ptr[i0] = src0_ptr[i0] - *src1_ptr; + } + } + } +} + +static void ggml_compute_forward_sub( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_sub_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_mul + +static void ggml_compute_forward_mul_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst)); + + const int ith = params->ith; + const int nth = params->nth; + + const int64_t nr = ggml_nrows(src0); + + GGML_TENSOR_BINARY_OP_LOCALS + + GGML_ASSERT( nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); + + if (nb10 == sizeof(float)) { + for (int64_t ir = ith; ir < nr; ir += nth) { + // src0 and dst are same shape => same indices + const int64_t i03 = ir/(ne02*ne01); + const int64_t i02 = (ir - i03*ne02*ne01)/ne01; + const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); + + const int64_t i13 = i03 % ne13; + const int64_t i12 = i02 % ne12; + const int64_t i11 = i01 % ne11; + const int64_t nr0 = ne00 / ne10; + + float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); + float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); + float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11); + + for (int64_t r = 0 ; r < nr0; ++r) { +#ifdef GGML_USE_ACCELERATE + UNUSED(ggml_vec_mul_f32); + + vDSP_vmul(src0_ptr + r*ne10, 1, src1_ptr, 1, dst_ptr + r*ne10, 1, ne10); +#else + ggml_vec_mul_f32(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr); +#endif + } + } + } else { + // src1 is not contiguous + for (int64_t ir = ith; ir < nr; ir += nth) { + // src0 and dst are same shape => same indices + // src1 is broadcastable across src0 and dst in i1, i2, i3 + const int64_t i03 = ir/(ne02*ne01); + const int64_t i02 = (ir - i03*ne02*ne01)/ne01; + const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); + + const int64_t i13 = i03 % ne13; + const int64_t i12 = i02 % ne12; + const int64_t i11 = i01 % ne11; + + float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); + float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); + + for (int64_t i0 = 0; i0 < ne00; ++i0) { + const int64_t i10 = i0 % ne10; + float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10); + + dst_ptr[i0] = src0_ptr[i0] * (*src1_ptr); + } + } + } +} + +static void ggml_compute_forward_mul( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(src1->type == GGML_TYPE_F32 && "only f32 src1 supported for now"); + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_mul_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_div + +static void ggml_compute_forward_div_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst)); + + const int ith = params->ith; + const int nth = params->nth; + + const int64_t nr = ggml_nrows(src0); + + GGML_TENSOR_BINARY_OP_LOCALS + + GGML_ASSERT( nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); + + if (nb10 == sizeof(float)) { + for (int64_t ir = ith; ir < nr; ir += nth) { + // src0 and dst are same shape => same indices + const int64_t i03 = ir/(ne02*ne01); + const int64_t i02 = (ir - i03*ne02*ne01)/ne01; + const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); + + const int64_t i13 = i03 % ne13; + const int64_t i12 = i02 % ne12; + const int64_t i11 = i01 % ne11; + const int64_t nr0 = ne00 / ne10; + + float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); + float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); + float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11); + + for (int64_t r = 0; r < nr0; ++r) { +#ifdef GGML_USE_ACCELERATE + UNUSED(ggml_vec_div_f32); + + vDSP_vdiv(src1_ptr, 1, src0_ptr + r*ne10, 1, dst_ptr + r*ne10, 1, ne10); +#else + ggml_vec_div_f32(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr); +#endif + } + } + } else { + // src1 is not contiguous + for (int64_t ir = ith; ir < nr; ir += nth) { + // src0 and dst are same shape => same indices + // src1 is broadcastable across src0 and dst in i1, i2, i3 + const int64_t i03 = ir/(ne02*ne01); + const int64_t i02 = (ir - i03*ne02*ne01)/ne01; + const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); + + const int64_t i13 = i03 % ne13; + const int64_t i12 = i02 % ne12; + const int64_t i11 = i01 % ne11; + + float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); + float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); + + for (int64_t i0 = 0; i0 < ne00; ++i0) { + const int64_t i10 = i0 % ne10; + float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10); + + dst_ptr[i0] = src0_ptr[i0] / (*src1_ptr); + } + } + } +} + +static void ggml_compute_forward_div( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_div_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_sqr + +static void ggml_compute_forward_sqr_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + assert( dst->nb[0] == sizeof(float)); + assert(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < n; i++) { + ggml_vec_sqr_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_sqr( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_sqr_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_sqrt + +static void ggml_compute_forward_sqrt_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + assert( dst->nb[0] == sizeof(float)); + assert(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < n; i++) { + ggml_vec_sqrt_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_sqrt( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_sqrt_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_log + +static void ggml_compute_forward_log_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + GGML_ASSERT( dst->nb[0] == sizeof(float)); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < n; i++) { + ggml_vec_log_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_log( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_log_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_sin + +static void ggml_compute_forward_sin_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + GGML_ASSERT( dst->nb[0] == sizeof(float)); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < n; i++) { + ggml_vec_sin_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_sin( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_sin_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_cos + +static void ggml_compute_forward_cos_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + GGML_ASSERT( dst->nb[0] == sizeof(float)); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < n; i++) { + ggml_vec_cos_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_cos( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_cos_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_sum + +static void ggml_compute_forward_sum_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_scalar(dst)); + assert(src0->nb[0] == sizeof(float)); + + GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) + GGML_TENSOR_LOCALS(size_t, nb0, src0, nb) + + ggml_float sum = 0; + ggml_float row_sum = 0; + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = 0; i01 < ne01; i01++) { + ggml_vec_sum_f32_ggf(ne00, + &row_sum, + (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03)); + sum += row_sum; + } + } + } + ((float *) dst->data)[0] = sum; +} + +static void ggml_compute_forward_sum_f16( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_scalar(dst)); + + assert(src0->nb[0] == sizeof(ggml_fp16_t)); + + GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) + GGML_TENSOR_LOCALS(size_t, nb0, src0, nb) + + float sum = 0; + float row_sum = 0; + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = 0; i01 < ne01; i01++) { + ggml_vec_sum_f16_ggf(ne00, + &row_sum, + (ggml_fp16_t *) ((char *) src0->data + i01 * nb01 + i02 * nb02 + i03 * nb03)); + sum += row_sum; + } + } + } + ((ggml_fp16_t *) dst->data)[0] = GGML_FP32_TO_FP16(sum); +} + +static void ggml_compute_forward_sum_bf16( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_scalar(dst)); + + assert(src0->nb[0] == sizeof(ggml_bf16_t)); + + GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) + GGML_TENSOR_LOCALS(size_t, nb0, src0, nb) + + float sum = 0; + float row_sum = 0; + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = 0; i01 < ne01; i01++) { + ggml_vec_sum_bf16_ggf(ne00, + &row_sum, + (ggml_bf16_t *) ((char *) src0->data + i01 * nb01 + i02 * nb02 + i03 * nb03)); + sum += row_sum; + } + } + } + ((ggml_bf16_t *) dst->data)[0] = GGML_FP32_TO_BF16(sum); +} + +static void ggml_compute_forward_sum( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_sum_f32(params, dst); + } break; + case GGML_TYPE_F16: + { + ggml_compute_forward_sum_f16(params, dst); + } break; + case GGML_TYPE_BF16: + { + ggml_compute_forward_sum_bf16(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_sum_rows + +static void ggml_compute_forward_sum_rows_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + GGML_ASSERT(dst->nb[0] == sizeof(float)); + + GGML_TENSOR_UNARY_OP_LOCALS + + GGML_ASSERT(ne0 == 1); + GGML_ASSERT(ne1 == ne01); + GGML_ASSERT(ne2 == ne02); + GGML_ASSERT(ne3 == ne03); + + for (int64_t i3 = 0; i3 < ne03; i3++) { + for (int64_t i2 = 0; i2 < ne02; i2++) { + for (int64_t i1 = 0; i1 < ne01; i1++) { + float * src_row = (float *) ((char *) src0->data + i1*nb01 + i2*nb02 + i3*nb03); + float * dst_row = (float *) ((char *) dst->data + i1*nb1 + i2*nb2 + i3*nb3); + float row_sum = 0; + ggml_vec_sum_f32(ne00, &row_sum, src_row); + dst_row[0] = row_sum; + } + } + } +} + +static void ggml_compute_forward_sum_rows( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_sum_rows_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_mean + +static void ggml_compute_forward_mean_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(src0->nb[0] == sizeof(float)); + + GGML_TENSOR_UNARY_OP_LOCALS + + assert(ne0 == 1); + assert(ne1 == ne01); + assert(ne2 == ne02); + assert(ne3 == ne03); + + UNUSED(ne0); + UNUSED(ne1); + UNUSED(ne2); + UNUSED(ne3); + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = 0; i01 < ne01; i01++) { + ggml_vec_sum_f32(ne00, + (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3), + (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03)); + + *(float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3) /= (float) ne00; + } + } + } +} + +static void ggml_compute_forward_mean( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_mean_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_argmax + +static void ggml_compute_forward_argmax_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(src0->nb[0] == sizeof(float)); + assert(dst->nb[0] == sizeof(float)); + + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + + const size_t nb01 = src0->nb[1]; + const size_t nb0 = dst->nb[0]; + + for (int64_t i1 = 0; i1 < ne01; i1++) { + float * src = (float *) ((char *) src0->data + i1*nb01); + int32_t * dst_ = (int32_t *) ((char *) dst->data + i1*nb0); + int v = 0; + ggml_vec_argmax_f32(ne00, &v, src); + dst_[0] = v; + } +} + +static void ggml_compute_forward_argmax( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_argmax_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_repeat + +static void ggml_compute_forward_repeat_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + GGML_ASSERT(ggml_can_repeat(src0, dst)); + + GGML_TENSOR_UNARY_OP_LOCALS + + // guaranteed to be an integer due to the check in ggml_can_repeat + const int nr0 = (int)(ne0/ne00); + const int nr1 = (int)(ne1/ne01); + const int nr2 = (int)(ne2/ne02); + const int nr3 = (int)(ne3/ne03); + + // TODO: support for transposed / permuted tensors + GGML_ASSERT(nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); + + // TODO: maybe this is not optimal? + for (int i3 = 0; i3 < nr3; i3++) { + for (int k3 = 0; k3 < ne03; k3++) { + for (int i2 = 0; i2 < nr2; i2++) { + for (int k2 = 0; k2 < ne02; k2++) { + for (int i1 = 0; i1 < nr1; i1++) { + for (int k1 = 0; k1 < ne01; k1++) { + for (int i0 = 0; i0 < nr0; i0++) { + ggml_vec_cpy_f32(ne00, + (float *) ((char *) dst->data + (i3*ne03 + k3)*nb3 + (i2*ne02 + k2)*nb2 + (i1*ne01 + k1)*nb1 + (i0*ne00)*nb0), + (float *) ((char *) src0->data + ( k3)*nb03 + ( k2)*nb02 + ( k1)*nb01)); + } + } + } + } + } + } + } +} + +static void ggml_compute_forward_repeat_f16( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + GGML_ASSERT(ggml_can_repeat(src0, dst)); + + GGML_TENSOR_UNARY_OP_LOCALS + + // guaranteed to be an integer due to the check in ggml_can_repeat + const int nr0 = (int)(ne0/ne00); + const int nr1 = (int)(ne1/ne01); + const int nr2 = (int)(ne2/ne02); + const int nr3 = (int)(ne3/ne03); + + // TODO: support for transposed / permuted tensors + GGML_ASSERT(nb0 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + + // TODO: maybe this is not optimal? + for (int i3 = 0; i3 < nr3; i3++) { + for (int k3 = 0; k3 < ne03; k3++) { + for (int i2 = 0; i2 < nr2; i2++) { + for (int k2 = 0; k2 < ne02; k2++) { + for (int i1 = 0; i1 < nr1; i1++) { + for (int k1 = 0; k1 < ne01; k1++) { + for (int i0 = 0; i0 < nr0; i0++) { + ggml_fp16_t * y = (ggml_fp16_t *) ((char *) dst->data + (i3*ne03 + k3)*nb3 + (i2*ne02 + k2)*nb2 + (i1*ne01 + k1)*nb1 + (i0*ne00)*nb0); + ggml_fp16_t * x = (ggml_fp16_t *) ((char *) src0->data + ( k3)*nb03 + ( k2)*nb02 + ( k1)*nb01); + // ggml_vec_cpy_f16(ne00, y, x) + for (int i = 0; i < ne00; ++i) { + y[i] = x[i]; + } + } + } + } + } + } + } + } +} + +static void ggml_compute_forward_repeat( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F16: + case GGML_TYPE_BF16: + case GGML_TYPE_I16: + { + ggml_compute_forward_repeat_f16(params, dst); + } break; + case GGML_TYPE_F32: + case GGML_TYPE_I32: + { + ggml_compute_forward_repeat_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_repeat_back + +static void ggml_compute_forward_repeat_back_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + GGML_ASSERT(ggml_can_repeat(dst, src0)); + + GGML_TENSOR_UNARY_OP_LOCALS + + // guaranteed to be an integer due to the check in ggml_can_repeat + const int nr0 = (int)(ne00/ne0); + const int nr1 = (int)(ne01/ne1); + const int nr2 = (int)(ne02/ne2); + const int nr3 = (int)(ne03/ne3); + + // TODO: support for transposed / permuted tensors + GGML_ASSERT(nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); + + if (ggml_is_contiguous(dst)) { + ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0); + } else { + for (int k3 = 0; k3 < ne3; k3++) { + for (int k2 = 0; k2 < ne2; k2++) { + for (int k1 = 0; k1 < ne1; k1++) { + ggml_vec_set_f32(ne0, + (float *) ((char *) dst->data + k1*nb1 + k2*nb2 + k3*nb3), + 0); + } + } + } + } + + // TODO: maybe this is not optimal? + for (int i3 = 0; i3 < nr3; i3++) { + for (int k3 = 0; k3 < ne3; k3++) { + for (int i2 = 0; i2 < nr2; i2++) { + for (int k2 = 0; k2 < ne2; k2++) { + for (int i1 = 0; i1 < nr1; i1++) { + for (int k1 = 0; k1 < ne1; k1++) { + for (int i0 = 0; i0 < nr0; i0++) { + ggml_vec_acc_f32(ne0, + (float *) ((char *) dst->data + ( k3)*nb3 + ( k2)*nb2 + ( k1)*nb1), + (float *) ((char *) src0->data + (i3*ne3 + k3)*nb03 + (i2*ne2 + k2)*nb02 + (i1*ne1 + k1)*nb01 + (i0*ne0)*nb00)); + } + } + } + } + } + } + } +} + +static void ggml_compute_forward_repeat_back( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_repeat_back_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_concat + +static void ggml_compute_forward_concat_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int32_t dim = ggml_get_op_params_i32(dst, 0); + + GGML_ASSERT(dim >= 0 && dim < 4); + + int64_t o[4] = {0, 0, 0, 0}; + o[dim] = src0->ne[dim]; + + const float * x; + + // TODO: smarter multi-theading + for (int i3 = 0; i3 < ne3; i3++) { + for (int i2 = ith; i2 < ne2; i2 += nth) { + for (int i1 = 0; i1 < ne1; i1++) { + for (int i0 = 0; i0 < ne0; i0++) { + if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) { + x = (const float *) ((const char *)src0->data + (i0 )*nb00 + (i1 )*nb01 + (i2 )*nb02 + (i3 )*nb03); + } else { + x = (const float *) ((const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13); + } + + float * y = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3); + + *y = *x; + } + } + } + } +} + +static void ggml_compute_forward_concat( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + case GGML_TYPE_I32: + { + ggml_compute_forward_concat_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_abs + +static void ggml_compute_forward_abs_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + for (int i = 0; i < n; i++) { + ggml_vec_abs_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_abs( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_abs_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_sgn + +static void ggml_compute_forward_sgn_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + for (int i = 0; i < n; i++) { + ggml_vec_sgn_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_sgn( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_sgn_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_neg + +static void ggml_compute_forward_neg_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + for (int i = 0; i < n; i++) { + ggml_vec_neg_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_neg( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_neg_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_step + +static void ggml_compute_forward_step_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + for (int i = 0; i < n; i++) { + ggml_vec_step_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_step( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_step_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_tanh + +static void ggml_compute_forward_tanh_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + for (int i = 0; i < n; i++) { + ggml_vec_tanh_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_tanh( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_tanh_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_elu + +static void ggml_compute_forward_elu_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + for (int i = 0; i < n; i++) { + ggml_vec_elu_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_elu( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_elu_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_relu + +static void ggml_compute_forward_relu_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + for (int i = 0; i < n; i++) { + ggml_vec_relu_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_relu( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_relu_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_sigmoid + +static void ggml_compute_forward_sigmoid_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + for (int i = 0; i < n; i++) { + ggml_vec_sigmoid_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_sigmoid( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_sigmoid_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_gelu + +static void ggml_compute_forward_gelu_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int i1 = ir0; i1 < ir1; i1++) { + ggml_vec_gelu_f32(nc, + (float *) ((char *) dst->data + i1*( dst->nb[1])), + (float *) ((char *) src0->data + i1*(src0->nb[1]))); + +#ifndef NDEBUG + for (int k = 0; k < nc; k++) { + const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k]; + UNUSED(x); + assert(!isnan(x)); + assert(!isinf(x)); + } +#endif + } +} + +static void ggml_compute_forward_gelu( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_gelu_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_gelu_quick + +static void ggml_compute_forward_gelu_quick_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int i1 = ir0; i1 < ir1; i1++) { + ggml_vec_gelu_quick_f32(nc, + (float *) ((char *) dst->data + i1*( dst->nb[1])), + (float *) ((char *) src0->data + i1*(src0->nb[1]))); + +#ifndef NDEBUG + for (int k = 0; k < nc; k++) { + const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k]; + UNUSED(x); + assert(!isnan(x)); + assert(!isinf(x)); + } +#endif + } +} + +static void ggml_compute_forward_gelu_quick( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_gelu_quick_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_silu + +static void ggml_compute_forward_silu_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int i1 = ir0; i1 < ir1; i1++) { + ggml_vec_silu_f32(nc, + (float *) ((char *) dst->data + i1*( dst->nb[1])), + (float *) ((char *) src0->data + i1*(src0->nb[1]))); + +#ifndef NDEBUG + for (int k = 0; k < nc; k++) { + const float x = ((float *) ((char *) dst->data + i1*(dst->nb[1])))[k]; + UNUSED(x); + assert(!isnan(x)); + assert(!isinf(x)); + } +#endif + } +} + +static void ggml_compute_forward_silu( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_silu_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} +// ggml_compute_forward_leaky_relu + +static void ggml_compute_forward_leaky_relu_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + float negative_slope; + memcpy(&negative_slope, dst->op_params, sizeof(float)); + + assert(dst->nb[0] == sizeof(float)); + assert(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < n; i++) { + ggml_vec_leaky_relu_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1])), negative_slope); + } +} + +static void ggml_compute_forward_leaky_relu( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_leaky_relu_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_silu_back + +static void ggml_compute_forward_silu_back_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * grad = dst->src[1]; + + assert(ggml_is_contiguous_1(grad)); + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + assert(ggml_are_same_shape(src0, grad)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int i1 = ir0; i1 < ir1; i1++) { + ggml_vec_silu_backward_f32(nc, + (float *) ((char *) dst->data + i1*( dst->nb[1])), + (float *) ((char *) src0->data + i1*(src0->nb[1])), + (float *) ((char *) grad->data + i1*(grad->nb[1]))); + +#ifndef NDEBUG + for (int k = 0; k < nc; k++) { + const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k]; + UNUSED(x); + assert(!isnan(x)); + assert(!isinf(x)); + } +#endif + } +} + +static void ggml_compute_forward_silu_back( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_silu_back_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + + +static void ggml_compute_forward_hardswish_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + for (int i = 0; i < n; i++) { + ggml_vec_hardswish_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} +static void ggml_compute_forward_hardswish( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_hardswish_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +static void ggml_compute_forward_hardsigmoid_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + for (int i = 0; i < n; i++) { + ggml_vec_hardsigmoid_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_hardsigmoid( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_hardsigmoid_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +static void ggml_compute_forward_exp_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + for (int i = 0; i < n; i++) { + ggml_vec_exp_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_exp( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_exp_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + + +// ggml_compute_forward_norm + +static void ggml_compute_forward_norm_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_UNARY_OP_LOCALS + + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); + + GGML_ASSERT(eps > 0.0f); + + // TODO: optimize + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = ith; i01 < ne01; i01 += nth) { + const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + + ggml_float sum = 0.0; + for (int64_t i00 = 0; i00 < ne00; i00++) { + sum += (ggml_float)x[i00]; + } + + float mean = sum/ne00; + + float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3); + + ggml_float sum2 = 0.0; + for (int64_t i00 = 0; i00 < ne00; i00++) { + float v = x[i00] - mean; + y[i00] = v; + sum2 += (ggml_float)(v*v); + } + + float variance = sum2/ne00; + const float scale = 1.0f/sqrtf(variance + eps); + + ggml_vec_scale_f32(ne00, y, scale); + } + } + } +} + +static void ggml_compute_forward_norm( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_norm_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_group_rms_norm + +static void ggml_compute_forward_rms_norm_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_UNARY_OP_LOCALS + + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); + + GGML_ASSERT(eps > 0.0f); + + // TODO: optimize + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = ith; i01 < ne01; i01 += nth) { + const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + + ggml_float sum = 0.0; + for (int64_t i00 = 0; i00 < ne00; i00++) { + sum += (ggml_float)(x[i00] * x[i00]); + } + + const float mean = sum/ne00; + + float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3); + + memcpy(y, x, ne00 * sizeof(float)); + // for (int i00 = 0; i00 < ne00; i00++) { + // y[i00] = x[i00]; + // } + + const float scale = 1.0f/sqrtf(mean + eps); + + ggml_vec_scale_f32(ne00, y, scale); + } + } + } +} + +static void ggml_compute_forward_rms_norm( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_rms_norm_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +static void ggml_compute_forward_rms_norm_back_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst) && ggml_are_same_shape(src0, src1)); + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_BINARY_OP_LOCALS + + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); + + // TODO: optimize + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = ith; i01 < ne01; i01 += nth) { + // src1 is same shape as src0 => same indices + const int64_t i11 = i01; + const int64_t i12 = i02; + const int64_t i13 = i03; + + const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + const float * dz = (float *) ((char *) src1->data + i11*nb11 + i12*nb12 + i13*nb13); + + ggml_float sum_xx = 0.0; + ggml_float sum_xdz = 0.0; + + for (int64_t i00 = 0; i00 < ne00; i00++) { + sum_xx += (ggml_float)(x[i00] * x[i00]); + sum_xdz += (ggml_float)(x[i00] * dz[i00]); + } + + //const float mean = (float)(sum_xx)/ne00; + const float mean_eps = (float)(sum_xx)/ne00 + eps; + const float sum_eps = (float)(sum_xx) + eps*ne00; + //const float mean_xdz = (float)(sum_xdz)/ne00; + // we could cache rms from forward pass to improve performance. + // to do this implement ggml_rms and compose ggml_rms_norm using ggml_rms. + //const float rms = sqrtf(mean_eps); + const float rrms = 1.0f / sqrtf(mean_eps); + //const float scale = -rrms/(ne00 * mean_eps); // -1/(n*rms**3) + + { + // z = rms_norm(x) + // + // rms_norm(src0) = + // scale( + // src0, + // div( + // 1, + // sqrt( + // add( + // scale( + // sum( + // sqr( + // src0)), + // (1.0/N)), + // eps)))); + + // postorder: + // ## op args grad + // 00 param src0 grad[#00] + // 01 const 1 + // 02 sqr (#00) grad[#02] + // 03 sum (#02) grad[#03] + // 04 const 1/N + // 05 scale (#03, #04) grad[#05] + // 06 const eps + // 07 add (#05, #06) grad[#07] + // 08 sqrt (#07) grad[#08] + // 09 div (#01,#08) grad[#09] + // 10 scale (#00,#09) grad[#10] + // + // backward pass, given grad[#10] + // #10: scale + // grad[#00] += scale(grad[#10],#09) + // grad[#09] += sum(mul(grad[#10],#00)) + // #09: div + // grad[#08] += neg(mul(grad[#09], div(#09,#08))) + // #08: sqrt + // grad[#07] += mul(grad[#08], div(0.5, #08)) + // #07: add + // grad[#05] += grad[#07] + // #05: scale + // grad[#03] += scale(grad[#05],#04) + // #03: sum + // grad[#02] += repeat(grad[#03], #02) + // #02: + // grad[#00] += scale(mul(#00, grad[#02]), 2.0) + // + // substitute and simplify: + // grad[#00] = scale(grad(#10), #09) + scale(mul(#00, grad[#02]), 2.0) + // grad[#02] = repeat(grad[#03], #02) + // grad[#02] = repeat(scale(grad[#05],#04), #02) + // grad[#02] = repeat(scale(grad[#07],#04), #02) + // grad[#02] = repeat(scale(mul(grad[#08], div(0.5, #08)),#04), #02) + // grad[#02] = repeat(scale(mul(neg(mul(grad[#09], div(#09,#08))), div(0.5, #08)),#04), #02) + // grad[#02] = repeat(scale(mul(neg(mul(sum(mul(grad[#10],#00)), div(#09,#08))), div(0.5, #08)),#04), #02) + // grad[#02] = repeat(-(sum(mul(grad[#10],#00)) * div(#09,#08) * div(0.5, #08) * (1/N)), #02) + // grad[#02] = repeat(-(sum(mul(grad[#10],#00)) * div(div(#01,#08),#08) * div(0.5, #08) * (1/N)), #02) + // grad[#02] = repeat(-(sum(mul(grad[#10],#00)) * div(1,#08*#08) * div(0.5, #08) * (1/N)), #02) + // grad[#02] = repeat(-(sum(mul(grad[#10],#00)) * div(1,#07) * div(0.5, #08) * (1/N)), #02) + // grad[#00] = scale(grad(#10), #09) + scale(mul(#00, grad[#02]), 2.0) + // grad[#00] = scale(grad(#10), #09) + scale(mul(#00, repeat(-(sum(mul(grad[#10],#00)) * div(1,#07) * div(0.5, #08) * (1/N)), #02)), 2.0) + // grad[#00] = scale(grad(#10), #09) + scale(scale(#00, -(sum(mul(grad[#10],#00)) * div(1,#07) * div(0.5, #08) * (1/N))), 2.0) + // grad[#00] = scale(grad(#10), #09) + scale(#00, -(sum(mul(grad[#10],#00)) * div(1,#07) * div(1,#08) * (1/N))) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(1,#07*#08) * (-1/N)) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(1,#07*#08) * (-1/N)) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(1,mean_eps*rms) * (-1/N)) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(-1,rms*N*mean_eps)) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(-1,rms*N*(sum_xx/N+eps))) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(-1,rms*N*sum_xx+rms*N*eps)) + // grad[#00] = scale(dz, rrms) + scale(x, sum(mul(dz,x)) * div(-1,rms*N*mean_eps)) + // grad[#00] = scale(dz, rrms) + scale(x, sum_xdz * div(-1,rms*N*mean_eps)) + // a = b*c + d*e + // a = b*c*f/f + d*e*f/f + // a = (b*c*f + d*e*f)*(1/f) + // a = (b*c*(1/c) + d*e*(1/c))*(1/(1/c)) + // a = (b + d*e/c)*c + // b = dz, c = rrms, d = x, e = sum_xdz * div(-1,rms*N*mean_eps) + // a = (dz + x*sum_xdz * div(-1,rms*N*mean_eps)/rrms)*rrms + // a = (dz + x*sum_xdz * div(-1,rms*N*mean_eps)*rms)*rrms + // a = (dz + x*sum_xdz * div(-rms,rms*N*mean_eps))*rrms + // a = (dz + x*sum_xdz * div(-1,N*mean_eps))*rrms + // a = (dz + x*div(-sum_xdz,N*mean_eps))*rrms + // a = (dz + x*div(-mean_xdz,mean_eps))*rrms + // grad[#00] = scale(dz + scale(x, div(-mean_xdz,mean_eps)),rrms) + // grad[#00] = scale(dz + scale(x, -mean_xdz/mean_eps),rrms) + // dx = scale(dz + scale(x, -mean_xdz/mean_eps),rrms) + } + // dx = scale(dz + scale(x, -mean_xdz/mean_eps),rrms) + // post-order: + // dx := x + // dx := scale(dx,-mean_xdz/mean_eps) + // dx := add(dx, dz) + // dx := scale(dx, rrms) + float * dx = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3); + + ggml_vec_cpy_f32 (ne00, dx, x); + // ggml_vec_scale_f32(ne00, dx, -mean_xdz/mean_eps); + ggml_vec_scale_f32(ne00, dx, (float)(-sum_xdz)/sum_eps); + ggml_vec_acc_f32 (ne00, dx, dz); + ggml_vec_scale_f32(ne00, dx, rrms); + } + } + } +} + +static void ggml_compute_forward_rms_norm_back( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_rms_norm_back_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_group_norm + +static void ggml_compute_forward_group_norm_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_UNARY_OP_LOCALS + + // TODO: optimize + + float eps; + memcpy(&eps, dst->op_params + 1, sizeof(float)); + + int n_channels = src0->ne[2]; + int n_groups = dst->op_params[0]; + int n_channels_per_group = (n_channels + n_groups - 1) / n_groups; + for (int i = ith; i < n_groups; i += nth) { + int start = i * n_channels_per_group; + int end = start + n_channels_per_group; + if (end > n_channels) { + end = n_channels; + } + int step = end - start; + + for (int64_t i03 = 0; i03 < ne03; i03++) { + ggml_float sum = 0.0; + for (int64_t i02 = start; i02 < end; i02++) { + for (int64_t i01 = 0; i01 < ne01; i01++) { + const float * x = (float *)((char *) src0->data + i01 * nb01 + i02 * nb02 + i03 * nb03); + + ggml_float sumr = 0.0; + for (int64_t i00 = 0; i00 < ne00; i00++) { + sumr += (ggml_float)x[i00]; + } + sum += sumr; + } + } + const float mean = sum / (ne00 * ne01 * step); + + ggml_float sum2 = 0.0; + for (int64_t i02 = start; i02 < end; i02++) { + for (int64_t i01 = 0; i01 < ne01; i01++) { + const float * x = (float *)((char *) src0->data + i01 * nb01 + i02 * nb02 + i03 * nb03); + + float * y = (float *)((char *) dst->data + i01 * nb1 + i02 * nb2 + i03 * nb3); + + ggml_float sumr = 0.0; + for (int64_t i00 = 0; i00 < ne00; i00++) { + float v = x[i00] - mean; + y[i00] = v; + sumr += (ggml_float)(v * v); + } + sum2 += sumr; + } + } + const float variance = sum2 / (ne00 * ne01 * step); + const float scale = 1.0f / sqrtf(variance + eps); + + for (int64_t i02 = start; i02 < end; i02++) { + for (int64_t i01 = 0; i01 < ne01; i01++) { + float * y = (float *)((char *) dst->data + i01 * nb1 + i02 * nb2 + i03 * nb3); + ggml_vec_scale_f32(ne00, y, scale); + } + } + } + } +} + +static void ggml_compute_forward_group_norm( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_group_norm_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_mul_mat + +static void ggml_compute_forward_mul_mat_one_chunk( + const struct ggml_compute_params * params, + struct ggml_tensor * dst, + const int64_t num_rows_per_vec_dot, + const int64_t ir0_start, + const int64_t ir0_end, + const int64_t ir1_start, + const int64_t ir1_end) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_TENSOR_BINARY_OP_LOCALS + + const enum ggml_type type = src0->type; + + const bool src1_cont = ggml_is_contiguous(src1); + + ggml_vec_dot_t const vec_dot = type_traits[type].vec_dot; + enum ggml_type const vec_dot_type = type_traits[type].vec_dot_type; + + // broadcast factors + const int64_t r2 = ne12 / ne02; + const int64_t r3 = ne13 / ne03; + + //printf("ir0_start = %6lld, ir0_end = %6lld, ir1_start = %6lld, ir1_end = %6lld\n", ir0_start, ir0_end, ir1_start, ir1_end); + + // threads with no work simply yield (not sure if it helps) + if (ir0_start >= ir0_end || ir1_start >= ir1_end) { + return; + } + + const void * wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata; + const size_t row_size = ggml_row_size(vec_dot_type, ne10); + + assert(ne12 % ne02 == 0); + assert(ne13 % ne03 == 0); + + // block-tiling attempt + const int64_t blck_0 = 16; + const int64_t blck_1 = 16; + + const size_t src1_col_stride = src1_cont || src1->type != vec_dot_type ? row_size : nb11; + + // attempt to reduce false-sharing (does not seem to make a difference) + // 16 * 2, accounting for mmla kernels + float tmp[32]; + + for (int64_t iir1 = ir1_start; iir1 < ir1_end; iir1 += blck_1) { + for (int64_t iir0 = ir0_start; iir0 < ir0_end; iir0 += blck_0) { + for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir1_end; ir1 += num_rows_per_vec_dot) { + const int64_t i13 = (ir1 / (ne12 * ne1)); + const int64_t i12 = (ir1 - i13 * ne12 * ne1) / ne1; + const int64_t i11 = (ir1 - i13 * ne12 * ne1 - i12 * ne1); + + // broadcast src0 into src1 + const int64_t i03 = i13 / r3; + const int64_t i02 = i12 / r2; + + const int64_t i1 = i11; + const int64_t i2 = i12; + const int64_t i3 = i13; + + const char * src0_row = (const char*)src0->data + (0 + i02 * nb02 + i03 * nb03); + + // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides + // if it is, then we have either copied the data to params->wdata and made it contiguous or we are using + // the original src1 data pointer, so we should index using the indices directly + // TODO: this is a bit of a hack, we should probably have a better way to handle this + const char * src1_col = (const char*)wdata + + (src1_cont || src1->type != vec_dot_type + ? (i11 + i12 * ne11 + i13 * ne12 * ne11) * row_size + : (i11 * nb11 + i12 * nb12 + i13 * nb13)); + float * dst_col = (float*)((char*)dst->data + (i1 * nb1 + i2 * nb2 + i3 * nb3)); + + //for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir0_end; ++ir0) { + // vec_dot(ne00, &dst_col[ir0], src0_row + ir0*nb01, src1_col); + //} + + for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir0_end; ir0 += num_rows_per_vec_dot) { + vec_dot(ne00, &tmp[ir0 - iir0], (num_rows_per_vec_dot > 1 ? 16 : 0), src0_row + ir0 * nb01, (num_rows_per_vec_dot > 1 ? nb01 : 0), src1_col, (num_rows_per_vec_dot > 1 ? src1_col_stride : 0), num_rows_per_vec_dot); + } + + for (int cn = 0; cn < num_rows_per_vec_dot; ++cn) { + memcpy(&dst_col[iir0 + cn * nb1 / nb0], tmp + (cn * 16), (MIN(iir0 + blck_0, ir0_end) - iir0) * sizeof(float)); + } + } + } + } +} + +static void ggml_compute_forward_mul_mat( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int ith = params->ith; + const int nth = params->nth; + + const enum ggml_type type = src0->type; + + enum ggml_type const vec_dot_type = type_traits[type].vec_dot_type; + ggml_from_float_t const from_float = type_traits[vec_dot_type].from_float; + ggml_from_float_to_mat_t const from_float_to_mat = type_traits[vec_dot_type].from_float_to_mat; + int64_t const vec_dot_num_rows = type_traits[type].nrows; + int64_t const matmul_num_cols = type_traits[type].ncols; + int64_t const blck_size_interleave = type_traits[type].blck_size_interleave; + ggml_gemv_t const gemv = type_traits[type].gemv; + ggml_gemm_t const gemm = type_traits[type].gemm; + + GGML_ASSERT(ne0 == ne01); + GGML_ASSERT(ne1 == ne11); + GGML_ASSERT(ne2 == ne12); + GGML_ASSERT(ne3 == ne13); + + // we don't support permuted src0 or src1 + GGML_ASSERT(nb00 == ggml_type_size(type)); + GGML_ASSERT(nb10 == ggml_type_size(src1->type)); + + // dst cannot be transposed or permuted + GGML_ASSERT(nb0 == sizeof(float)); + GGML_ASSERT(nb0 <= nb1); + GGML_ASSERT(nb1 <= nb2); + GGML_ASSERT(nb2 <= nb3); + + // nb01 >= nb00 - src0 is not transposed + // compute by src0 rows + +#if GGML_USE_LLAMAFILE + // broadcast factors + const int64_t r2 = ne12 / ne02; + const int64_t r3 = ne13 / ne03; + + const bool src1_cont = ggml_is_contiguous(src1); + + if (src1_cont) { + for (int64_t i13 = 0; i13 < ne13; i13++) + for (int64_t i12 = 0; i12 < ne12; i12++) + if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type), + (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03, + nb01/ggml_type_size(src0->type), + (const char *)src1->data + i12*nb12 + i13*nb13, + nb11/ggml_type_size(src1->type), + (char *)dst->data + i12*nb2 + i13*nb3, + nb1/ggml_type_size(dst->type), + ith, nth, + src0->type, + src1->type, + dst->type)) + goto UseGgmlGemm1; + return; + } +UseGgmlGemm1:; +#endif + + if (src1->type != vec_dot_type) { + char * wdata = params->wdata; + + const size_t nbw1 = ggml_row_size(vec_dot_type, ne10); + const size_t nbw2 = nbw1*ne11; + const size_t nbw3 = nbw2*ne12; + + assert(params->wsize >= ne13*nbw3); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + for (int64_t i13 = 0; i13 < ne13; ++i13) { + for (int64_t i12 = 0; i12 < ne12; ++i12) { + int64_t i11_processed = 0; + if ((ggml_n_dims(src1) == 2) && from_float_to_mat && gemm) { + for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) { + from_float_to_mat((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), + (void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1), + 4, ne10, blck_size_interleave); + } + i11_processed = ne11 - ne11 % 4; + } + for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) { + from_float((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), + (void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1), + ne10); + } + } + } + } + + if (ith == 0) { + // Every thread starts at ith, so the first unprocessed chunk is nth. This save a bit of coordination right at the start. + atomic_store_explicit(¶ms->threadpool->current_chunk, nth, memory_order_relaxed); + } + + ggml_barrier(params->threadpool); + +#if GGML_USE_LLAMAFILE + if (src1->type != vec_dot_type) { + const void* wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata; + const size_t row_size = ggml_row_size(vec_dot_type, ne10); + + for (int64_t i13 = 0; i13 < ne13; i13++) + for (int64_t i12 = 0; i12 < ne12; i12++) + if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type), + (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03, + nb01/ggml_type_size(src0->type), + (const char *)wdata + (i12*ne11 + i13*ne12*ne11)*row_size, + row_size/ggml_type_size(vec_dot_type), + (char *)dst->data + i12*nb2 + i13*nb3, + nb1/ggml_type_size(dst->type), + ith, nth, + src0->type, + vec_dot_type, + dst->type)) + goto UseGgmlGemm2; + return; + } +UseGgmlGemm2:; +#endif + + // This is the size of the first dimension of the result, so we can iterate that way. (see the ASSERT above, these are the same numbers) + const int64_t nr0 = ne0; + + // This is the size of the rest of the dimensions of the result + const int64_t nr1 = ne1 * ne2 * ne3; + + // dot kernels can handle 1 row and col at a time, but mmla kernels can process 2 rows and cols + int64_t num_rows_per_vec_dot = vec_dot_num_rows; + // TODO: currently the mmla kernels support only even numbered rows/cols. + // this check can be removed once they are extended to support odd numbered rows/cols too + if ((nr0 % 2 != 0) || (ne11 % 2 != 0)) { + num_rows_per_vec_dot = 1; + } + + // Now select a reasonable chunk size. + int chunk_size = 16; + + // We need to step up the size if it's small + if (nr0 == 1 || nr1 == 1) { + chunk_size = 64; + } + + // distribute the work across the inner or outer loop based on which one is larger + // The number of chunks in the 0/1 dim. + // CEIL(nr0/chunk_size) + int64_t nchunk0 = (nr0 + chunk_size - 1) / chunk_size; + int64_t nchunk1 = (nr1 + chunk_size - 1) / chunk_size; + + // If the chunking is poor for the number of threads on this setup, scrap the whole plan. Re-chunk it by thread. + // Also, chunking by thread was measured to have perform better on NUMA systems. See https://github.com/ggerganov/llama.cpp/pull/6915 + // In theory, chunking should be just as useful on NUMA and non NUMA systems, but testing disagreed with that. + if (nchunk0 * nchunk1 < nth * 4 || ggml_is_numa()) { + // distribute the thread work across the inner or outer loop based on which one is larger + nchunk0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows + nchunk1 = nr0 > nr1 ? 1 : nth; // parallelize by src1 rows + } + + // The number of elements in each chunk + const int64_t dr0 = (nr0 + nchunk0 - 1) / nchunk0; + const int64_t dr1 = (nr1 + nchunk1 - 1) / nchunk1; + + if ((ggml_n_dims(src0) == 2) && gemv) { + const void * src1_wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata; + const size_t src1_col_stride = ggml_is_contiguous(src1) || src1->type != vec_dot_type ? ggml_row_size(vec_dot_type, ne10) : nb11; + int64_t src0_start = (ith * ne01) / nth; + int64_t src0_end = ((ith + 1) * ne01) / nth; + src0_start = (src0_start % matmul_num_cols) ? src0_start + matmul_num_cols - (src0_start % matmul_num_cols): src0_start; + src0_end = (src0_end % matmul_num_cols) ? src0_end + matmul_num_cols - (src0_end % matmul_num_cols): src0_end; + if (src0_start >= src0_end) return; + + // If there are more than three rows in src1, use gemm; otherwise, use gemv. + if (gemm && (ne11 > 3)) { + gemm(ne00, (float *)((char *) dst->data) + src0_start, ne01, (const char *) src0->data + src0_start * nb01, + (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start); + } + for (int iter = gemm ? ne11 - ne11 % 4 : 0; iter < ne11; iter++) { + gemv(ne00, (float *)((char *) dst->data + (iter * nb1)) + src0_start, ne01, + (const char *) src0->data + src0_start * nb01, (const char *) src1_wdata + (src1_col_stride * iter), 1, + src0_end - src0_start); + } + return; + } + + // The first chunk comes from our thread_id, the rest will get auto-assigned. + int current_chunk = ith; + + while (current_chunk < nchunk0 * nchunk1) { + const int64_t ith0 = current_chunk % nchunk0; + const int64_t ith1 = current_chunk / nchunk0; + + const int64_t ir0_start = dr0 * ith0; + const int64_t ir0_end = MIN(ir0_start + dr0, nr0); + + const int64_t ir1_start = dr1 * ith1; + const int64_t ir1_end = MIN(ir1_start + dr1, nr1); + + ggml_compute_forward_mul_mat_one_chunk(params, dst, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end); + + if (nth >= nchunk0 * nchunk1) { + break; + } + + current_chunk = atomic_fetch_add_explicit(¶ms->threadpool->current_chunk, 1, memory_order_relaxed); + } +} + +// ggml_compute_forward_mul_mat_id + +static void ggml_compute_forward_mul_mat_id( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + const struct ggml_tensor * ids = dst->src[2]; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int ith = params->ith; + const int nth = params->nth; + + const enum ggml_type type = src0->type; + + const bool src1_cont = ggml_is_contiguous(src1); + + ggml_vec_dot_t const vec_dot = type_traits[type].vec_dot; + enum ggml_type const vec_dot_type = type_traits[type].vec_dot_type; + ggml_from_float_t const from_float = type_traits[vec_dot_type].from_float; + int64_t const matmul_num_cols = type_traits[type].ncols; + ggml_gemv_t const gemv = type_traits[type].gemv; + + // we don't support permuted src0 or src1 + GGML_ASSERT(nb00 == ggml_type_size(type)); + GGML_ASSERT(nb10 == ggml_type_size(src1->type)); + + // dst cannot be transposed or permuted + GGML_ASSERT(nb0 == sizeof(float)); + GGML_ASSERT(nb0 <= nb1); + GGML_ASSERT(nb1 <= nb2); + GGML_ASSERT(nb2 <= nb3); + + // row groups + const int n_ids = ids->ne[0]; // n_expert_used + const int n_as = ne02; // n_expert + + char * wdata_src1_end = (src1->type == vec_dot_type) ? + (char *) params->wdata : + (char *) params->wdata + GGML_PAD(ggml_row_size(vec_dot_type, ggml_nelements(src1)), sizeof(int64_t)); + + struct mmid_row_mapping { + int32_t i1; + int32_t i2; + }; + + int64_t * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as] + struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *)(matrix_row_counts + n_as); // [n_as][ne11] + + if (src1->type != vec_dot_type) { + char * wdata = params->wdata; + + const size_t nbw1 = ggml_row_size(vec_dot_type, ne10); + const size_t nbw2 = nbw1*ne11; + const size_t nbw3 = nbw2*ne12; + + assert(params->wsize >= ne13*nbw3); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + for (int64_t i13 = 0; i13 < ne13; ++i13) { + for (int64_t i12 = 0; i12 < ne12; ++i12) { + for (int64_t i11 = ith; i11 < ne11; i11 += nth) { + from_float((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), + (void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1), + ne10); + } + } + } + } + +#define MMID_MATRIX_ROW(row_id, i1) matrix_rows[(row_id)*ne12 + (i1)] + + if (ith == 0) { + // initialize matrix_row_counts + memset(matrix_row_counts, 0, n_as*sizeof(int64_t)); + + // group rows by src0 matrix + for (int64_t iid1 = 0; iid1 < ids->ne[1]; ++iid1) { + for (int id = 0; id < n_ids; ++id) { + const int32_t i02 = *(const int32_t *) ((const char *) ids->data + iid1*ids->nb[1] + id*ids->nb[0]); + + assert(i02 >= 0 && i02 < n_as); + + MMID_MATRIX_ROW(i02, matrix_row_counts[i02]) = (struct mmid_row_mapping) {id, iid1}; + matrix_row_counts[i02] += 1; + } + } + } + + ggml_barrier(params->threadpool); + + // compute each matrix multiplication in sequence + for (int cur_a = 0; cur_a < n_as; ++cur_a) { + const int64_t cne1 = matrix_row_counts[cur_a]; + + if (cne1 == 0) { + continue; + } + + const char * src0_cur = (const char *) src0->data + cur_a*nb02; + + const void * wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata; + const size_t row_size = ggml_row_size(vec_dot_type, ne10); + + const int64_t nr0 = ne01; // src0 rows + const int64_t nr1 = cne1; // src1 rows + + if (((ggml_n_dims(src0) - 1) == 2) && gemv) { + int64_t src0_cur_start = (ith * ne01) / nth; + int64_t src0_cur_end = ((ith + 1) * ne01) / nth; + src0_cur_start = (src0_cur_start % matmul_num_cols) ? src0_cur_start + matmul_num_cols - (src0_cur_start % matmul_num_cols): src0_cur_start; + src0_cur_end = (src0_cur_end % matmul_num_cols) ? src0_cur_end + matmul_num_cols - (src0_cur_end % matmul_num_cols): src0_cur_end; + if (src0_cur_start >= src0_cur_end) return; + + for (int ir1 = 0; ir1 < nr1; ir1++) { + struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, ir1); + const int id = row_mapping.i1; // selected expert index + + const int64_t i11 = id % ne11; + const int64_t i12 = row_mapping.i2; // row index in src1 + + const int64_t i1 = id; // selected expert index + const int64_t i2 = i12; // row + + const char * src1_col = (const char *) wdata + + (src1_cont || src1->type != vec_dot_type + ? (i11 + i12 * ne11) * row_size + : (i11 * nb11 + i12 * nb12)); + + gemv(ne00, (float *)((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start, ne01, + (const char *) src0_cur + src0_cur_start * nb01, src1_col, 1, src0_cur_end - src0_cur_start); + } + continue; + } + + // distribute the thread work across the inner or outer loop based on which one is larger + + const int64_t nth0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows + const int64_t nth1 = nr0 > nr1 ? 1 : nth; // parallelize by src1 rows + + const int64_t ith0 = ith % nth0; + const int64_t ith1 = ith / nth0; + + const int64_t dr0 = (nr0 + nth0 - 1)/nth0; + const int64_t dr1 = (nr1 + nth1 - 1)/nth1; + + const int64_t ir010 = dr0*ith0; + const int64_t ir011 = MIN(ir010 + dr0, nr0); + + const int64_t ir110 = dr1*ith1; + const int64_t ir111 = MIN(ir110 + dr1, nr1); + + // threads with no work simply yield (not sure if it helps) + //if (ir010 >= ir011 || ir110 >= ir111) { + // sched_yield(); + // continue; + //} + + // block-tiling attempt + const int64_t blck_0 = 16; + const int64_t blck_1 = 16; + + // attempt to reduce false-sharing (does not seem to make a difference) + float tmp[16]; + + for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) { + for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) { + for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) { + const int64_t _i12 = ir1; // logical row index for this expert + + struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, _i12); + const int id = row_mapping.i1; // selected expert index + + const int64_t i11 = id % ne11; + const int64_t i12 = row_mapping.i2; // row index in src1 + + const int64_t i1 = id; // selected expert index + const int64_t i2 = i12; // row + + // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides + // if it is, then we have either copied the data to params->wdata and made it contiguous or we are using + // the original src1 data pointer, so we should index using the indices directly + // TODO: this is a bit of a hack, we should probably have a better way to handle this + const char * src1_col = (const char *) wdata + + (src1_cont || src1->type != vec_dot_type + ? (i11 + i12*ne11)*row_size + : (i11*nb11 + i12*nb12)); + + float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2)); + + //for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) { + // vec_dot(ne00, &dst_col[ir0], src0_row + ir0*nb01, src1_col); + //} + + for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) { + vec_dot(ne00, &tmp[ir0 - iir0], 0, src0_cur + ir0*nb01, 0, src1_col, 0, 1); + } + + memcpy(&dst_col[iir0], tmp, (MIN(iir0 + blck_0, ir011) - iir0)*sizeof(float)); + } + } + } + } + +#undef MMID_MATRIX_ROW +} + +// ggml_compute_forward_out_prod + +static void ggml_compute_forward_out_prod_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int ith = params->ith; + const int nth = params->nth; + + GGML_ASSERT(ne0 == ne00); + GGML_ASSERT(ne1 == ne10); + GGML_ASSERT(ne2 == ne02); + GGML_ASSERT(ne02 == ne12); + GGML_ASSERT(ne3 == ne13); + GGML_ASSERT(ne03 == ne13); + + // we don't support permuted src0 or src1 + GGML_ASSERT(nb00 == sizeof(float)); + + // dst cannot be transposed or permuted + GGML_ASSERT(nb0 == sizeof(float)); + // GGML_ASSERT(nb0 <= nb1); + // GGML_ASSERT(nb1 <= nb2); + // GGML_ASSERT(nb2 <= nb3); + + // nb01 >= nb00 - src0 is not transposed + // compute by src0 rows + + if (ith == 0) { + ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0); + } + ggml_barrier(params->threadpool); + + // dst[:,:,:,:] = 0 + // for i2,i3: + // for i1: + // for i01: + // for i0: + // dst[i0,i1,i2,i3] += src0[i0,i01,i2,i3] * src1[i1,i01,i2,i3] + + // parallelize by last three dimensions + + // total rows in dst + const int64_t nr = ne1*ne2*ne3; + + // rows per thread + const int64_t dr = (nr + nth - 1)/nth; + + // row range for this thread + const int64_t ir0 = dr*ith; + const int64_t ir1 = MIN(ir0 + dr, nr); + + // block-tiling attempt + const int64_t blck_0 = MAX(GGML_VEC_MAD_UNROLL, 32); + const int64_t blck_1 = 16; + + for (int64_t bir = ir0; bir < ir1; bir += blck_1) { + const int64_t bir1 = MIN(bir + blck_1, ir1); + for (int64_t bi01 = 0; bi01 < ne01; bi01 += blck_0) { + const int64_t bne01 = MIN(bi01 + blck_0, ne01); + for (int64_t ir = bir; ir < bir1; ++ir) { + // dst indices + const int64_t i3 = ir/(ne2*ne1); + const int64_t i2 = (ir - i3*ne2*ne1)/ne1; + const int64_t i1 = (ir - i3*ne2*ne1 - i2*ne1); + + const int64_t i02 = i2; + const int64_t i03 = i3; + + //const int64_t i10 = i1; + const int64_t i12 = i2; + const int64_t i13 = i3; + +#if GGML_VEC_MAD_UNROLL > 2 + const int64_t bne01_unroll = bne01 - (bne01 % GGML_VEC_MAD_UNROLL); + for (int64_t i01 = bi01; i01 < bne01_unroll; i01 += GGML_VEC_MAD_UNROLL) { + const int64_t i11 = i01; + + float * s0 = (float *) ((char *) src0->data + ( i01*nb01 + i02*nb02 + i03*nb03)); + float * s1 = (float *) ((char *) src1->data + (i1*nb10 + i11*nb11 + i12*nb12 + i13*nb13)); + float * d = (float *) ((char *) dst->data + ( i1*nb1 + i2*nb2 + i3*nb3)); + + ggml_vec_mad_f32_unroll(ne0, nb01, nb11, d, s0, s1); + } + for (int64_t i01 = bne01_unroll; i01 < bne01; ++i01) { + const int64_t i11 = i01; + + float * s0 = (float *) ((char *) src0->data + ( i01*nb01 + i02*nb02 + i03*nb03)); + float * s1 = (float *) ((char *) src1->data + (i1*nb10 + i11*nb11 + i12*nb12 + i13*nb13)); + float * d = (float *) ((char *) dst->data + ( i1*nb1 + i2*nb2 + i3*nb3)); + + ggml_vec_mad_f32(ne0, d, s0, *s1); + } +#else + for (int64_t i01 = bi01; i01 < bne01; ++i01) { + const int64_t i11 = i01; + + float * s0 = (float *) ((char *) src0->data + ( i01*nb01 + i02*nb02 + i03*nb03)); + float * s1 = (float *) ((char *) src1->data + (i1*nb10 + i11*nb11 + i12*nb12 + i13*nb13)); + float * d = (float *) ((char *) dst->data + ( i1*nb1 + i2*nb2 + i3*nb3)); + + ggml_vec_mad_f32(ne0, d, s0, *s1); + } +#endif + } + } + } +} + +static void ggml_compute_forward_out_prod_q_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_TENSOR_BINARY_OP_LOCALS; + + const int ith = params->ith; + const int nth = params->nth; + + const enum ggml_type type = src0->type; + ggml_to_float_t const dequantize_row_q = type_traits[type].to_float; + + GGML_ASSERT(ne02 == ne12); + GGML_ASSERT(ne03 == ne13); + GGML_ASSERT(ne2 == ne12); + GGML_ASSERT(ne3 == ne13); + + // we don't support permuted src0 dim0 + GGML_ASSERT(nb00 == ggml_type_size(type)); + + // dst dim0 cannot be transposed or permuted + GGML_ASSERT(nb0 == sizeof(float)); + // GGML_ASSERT(nb0 <= nb1); + // GGML_ASSERT(nb1 <= nb2); + // GGML_ASSERT(nb2 <= nb3); + + GGML_ASSERT(ne0 == ne00); + GGML_ASSERT(ne1 == ne10); + GGML_ASSERT(ne2 == ne02); + GGML_ASSERT(ne3 == ne03); + + // nb01 >= nb00 - src0 is not transposed + // compute by src0 rows + + if (ith == 0) { + ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0); + } + ggml_barrier(params->threadpool); + + // parallelize by last three dimensions + + // total rows in dst + const int64_t nr = ne1*ne2*ne3; + + // rows per thread + const int64_t dr = (nr + nth - 1)/nth; + + // row range for this thread + const int64_t ir0 = dr*ith; + const int64_t ir1 = MIN(ir0 + dr, nr); + + // dst[:,:,:,:] = 0 + // for i2,i3: + // for i1: + // for i01: + // for i0: + // dst[i0,i1,i2,i3] += src0[i0,i01,i2,i3] * src1[i1,i01,i2,i3] + + float * wdata = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32) * ith; + + for (int64_t ir = ir0; ir < ir1; ++ir) { + // dst indices + const int64_t i3 = ir/(ne2*ne1); + const int64_t i2 = (ir - i3*ne2*ne1)/ne1; + const int64_t i1 = (ir - i3*ne2*ne1 - i2*ne1); + + const int64_t i02 = i2; + const int64_t i03 = i3; + + //const int64_t i10 = i1; + const int64_t i12 = i2; + const int64_t i13 = i3; + + for (int64_t i01 = 0; i01 < ne01; ++i01) { + const int64_t i11 = i01; + + float * s0 = (float *) ((char *) src0->data + ( i01*nb01 + i02*nb02 + i03*nb03)); + float * s1 = (float *) ((char *) src1->data + (i1*nb10 + i11*nb11 + i12*nb12 + i13*nb13)); + float * d = (float *) ((char *) dst->data + ( i1*nb1 + i2*nb2 + i3*nb3)); + + dequantize_row_q(s0, wdata, ne0); + ggml_vec_mad_f32(ne0, d, wdata, *s1); + } + } +} + +static void ggml_compute_forward_out_prod( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q2_K: + case GGML_TYPE_Q3_K: + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + case GGML_TYPE_Q6_K: + case GGML_TYPE_TQ1_0: + case GGML_TYPE_TQ2_0: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ4_NL: + case GGML_TYPE_IQ4_XS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ2_S: + case GGML_TYPE_Q4_0_4_4: + case GGML_TYPE_Q4_0_4_8: + case GGML_TYPE_Q4_0_8_8: + { + ggml_compute_forward_out_prod_q_f32(params, dst); + } break; + case GGML_TYPE_F16: + { + GGML_ABORT("fatal error"); // todo + // ggml_compute_forward_out_prod_f16_f32(params, dst); + } + case GGML_TYPE_F32: + { + ggml_compute_forward_out_prod_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_scale + +static void ggml_compute_forward_scale_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(ggml_is_contiguous(dst)); + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + + // scale factor + float v; + memcpy(&v, dst->op_params, sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + const size_t nb01 = src0->nb[1]; + + const size_t nb1 = dst->nb[1]; + + for (int i1 = ir0; i1 < ir1; i1++) { + if (dst->data != src0->data) { + // src0 is same shape as dst => same indices + memcpy((char *)dst->data + i1*nb1, (char *)src0->data + i1*nb01, nc * sizeof(float)); + } + ggml_vec_scale_f32(nc, (float *) ((char *) dst->data + i1*nb1), v); + } +} + +static void ggml_compute_forward_scale( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_scale_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_set + +static void ggml_compute_forward_set_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); + + // view src0 and dst with these strides and data offset inbytes during set + // nb0 is implicitly element_size because src0 and dst are contiguous + size_t nb1 = ((int32_t *) dst->op_params)[0]; + size_t nb2 = ((int32_t *) dst->op_params)[1]; + size_t nb3 = ((int32_t *) dst->op_params)[2]; + size_t offset = ((int32_t *) dst->op_params)[3]; + bool inplace = (bool) ((int32_t *) dst->op_params)[4]; + + if (!inplace) { + if (params->ith == 0) { + // memcpy needs to be synchronized across threads to avoid race conditions. + // => do it in INIT phase + memcpy( + ((char *) dst->data), + ((char *) src0->data), + ggml_nbytes(dst)); + } + ggml_barrier(params->threadpool); + } + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src1); + const int nc = src1->ne[0]; + + GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) + GGML_TENSOR_LOCALS(size_t, nb1, src1, nb) + + // src0 and dst as viewed during set + const size_t nb0 = ggml_element_size(src0); + + const int im0 = (ne10 == 0 ? 0 : ne10-1); + const int im1 = (ne11 == 0 ? 0 : ne11-1); + const int im2 = (ne12 == 0 ? 0 : ne12-1); + const int im3 = (ne13 == 0 ? 0 : ne13-1); + + GGML_ASSERT(offset + im0*nb0 + im1*nb1 + im2*nb2 + im3*nb3 <= ggml_nbytes(dst)); + + GGML_ASSERT(nb10 == sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are viewed with shape of src1 and offset + // => same indices + const int i3 = ir/(ne12*ne11); + const int i2 = (ir - i3*ne12*ne11)/ne11; + const int i1 = (ir - i3*ne12*ne11 - i2*ne11); + + ggml_vec_cpy_f32(nc, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + offset), + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); + } +} + +static void ggml_compute_forward_set( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_set_f32(params, dst); + } break; + case GGML_TYPE_F16: + case GGML_TYPE_BF16: + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q8_1: + case GGML_TYPE_Q2_K: + case GGML_TYPE_Q3_K: + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + case GGML_TYPE_Q6_K: + case GGML_TYPE_TQ1_0: + case GGML_TYPE_TQ2_0: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ4_NL: + case GGML_TYPE_IQ4_XS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ2_S: + case GGML_TYPE_Q4_0_4_4: + case GGML_TYPE_Q4_0_4_8: + case GGML_TYPE_Q4_0_8_8: + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_cpy + +static void ggml_compute_forward_cpy( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + ggml_compute_forward_dup(params, dst); +} + +// ggml_compute_forward_cont + +static void ggml_compute_forward_cont( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + ggml_compute_forward_dup(params, dst); +} + +// ggml_compute_forward_reshape + +static void ggml_compute_forward_reshape( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + // NOP + UNUSED(params); + UNUSED(dst); +} + +// ggml_compute_forward_view + +static void ggml_compute_forward_view( + const struct ggml_compute_params * params, + const struct ggml_tensor * dst) { + // NOP + UNUSED(params); + UNUSED(dst); +} + +// ggml_compute_forward_permute + +static void ggml_compute_forward_permute( + const struct ggml_compute_params * params, + const struct ggml_tensor * dst) { + // NOP + UNUSED(params); + UNUSED(dst); +} + +// ggml_compute_forward_transpose + +static void ggml_compute_forward_transpose( + const struct ggml_compute_params * params, + const struct ggml_tensor * dst) { + // NOP + UNUSED(params); + UNUSED(dst); +} + +// ggml_compute_forward_get_rows + +static void ggml_compute_forward_get_rows_q( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int64_t nc = ne00; + const int64_t nr = ggml_nelements(src1); + + const enum ggml_type type = src0->type; + ggml_to_float_t const dequantize_row_q = type_traits[type].to_float; + + assert(ne0 == nc); + assert(ne02 == ne11); + assert(nb00 == ggml_type_size(type)); + assert(ggml_nrows(dst) == nr); + + const int ith = params->ith; + const int nth = params->nth; + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int64_t i = ir0; i < ir1; ++i) { + const int64_t i12 = i/(ne11*ne10); + const int64_t i11 = (i - i12*ne11*ne10)/ne10; + const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10); + const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12); + + GGML_ASSERT(i01 >= 0 && i01 < ne01); + + dequantize_row_q( + (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03), + (float *) ((char *) dst->data + i10*nb1 + i11*nb2 + i12*nb3), nc); + } +} + +static void ggml_compute_forward_get_rows_f16( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int64_t nc = ne00; + const int64_t nr = ggml_nelements(src1); + + assert(ne0 == nc); + assert(ne02 == ne11); + assert(nb00 == sizeof(ggml_fp16_t)); + assert(ggml_nrows(dst) == nr); + + const int ith = params->ith; + const int nth = params->nth; + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int64_t i = ir0; i < ir1; ++i) { + const int64_t i12 = i/(ne11*ne10); + const int64_t i11 = (i - i12*ne11*ne10)/ne10; + const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10); + const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12); + + GGML_ASSERT(i01 >= 0 && i01 < ne01); + + ggml_fp16_to_fp32_row( + (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03), + (float *) ((char *) dst->data + i10*nb1 + i11*nb2 + i12*nb3), nc); + } +} + +static void ggml_compute_forward_get_rows_bf16( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int64_t nc = ne00; + const int64_t nr = ggml_nelements(src1); + + assert(ne0 == nc); + assert(ne02 == ne11); + assert(nb00 == sizeof(ggml_bf16_t)); + assert(ggml_nrows(dst) == nr); + + const int ith = params->ith; + const int nth = params->nth; + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int64_t i = ir0; i < ir1; ++i) { + const int64_t i12 = i/(ne11*ne10); + const int64_t i11 = (i - i12*ne11*ne10)/ne10; + const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10); + const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12); + + GGML_ASSERT(i01 >= 0 && i01 < ne01); + + ggml_bf16_to_fp32_row( + (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03), + (float *) ((char *) dst->data + i10*nb1 + i11*nb2 + i12*nb3), nc); + } +} + +static void ggml_compute_forward_get_rows_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int64_t nc = ne00; + const int64_t nr = ggml_nelements(src1); + + assert(ne0 == nc); + assert(ne02 == ne11); + assert(nb00 == sizeof(float)); + assert(ggml_nrows(dst) == nr); + + const int ith = params->ith; + const int nth = params->nth; + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int64_t i = ir0; i < ir1; ++i) { + const int64_t i12 = i/(ne11*ne10); + const int64_t i11 = (i - i12*ne11*ne10)/ne10; + const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10); + const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12); + + GGML_ASSERT(i01 >= 0 && i01 < ne01); + + ggml_vec_cpy_f32(nc, + (float *) ((char *) dst->data + i10*nb1 + i11*nb2 + i12*nb3), + (float *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03)); + } +} + +static void ggml_compute_forward_get_rows( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q8_1: + case GGML_TYPE_Q2_K: + case GGML_TYPE_Q3_K: + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + case GGML_TYPE_Q6_K: + case GGML_TYPE_TQ1_0: + case GGML_TYPE_TQ2_0: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ4_NL: + case GGML_TYPE_IQ4_XS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ2_S: + case GGML_TYPE_Q4_0_4_4: + case GGML_TYPE_Q4_0_4_8: + case GGML_TYPE_Q4_0_8_8: + { + ggml_compute_forward_get_rows_q(params, dst); + } break; + case GGML_TYPE_F16: + { + ggml_compute_forward_get_rows_f16(params, dst); + } break; + case GGML_TYPE_BF16: + { + ggml_compute_forward_get_rows_bf16(params, dst); + } break; + case GGML_TYPE_F32: + case GGML_TYPE_I32: + { + ggml_compute_forward_get_rows_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } + + //static bool first = true; + //printf("ne0 = %d, ne1 = %d, ne2 = %d\n", dst->ne[0], dst->ne[1], dst->ne[2]); + //if (first) { + // first = false; + //} else { + // for (int k = 0; k < dst->ne[1]; ++k) { + // for (int j = 0; j < dst->ne[0]/16; ++j) { + // for (int i = 0; i < 16; ++i) { + // printf("%8.4f ", ((float *) dst->data)[k*dst->ne[0] + j*16 + i]); + // } + // printf("\n"); + // } + // printf("\n"); + // } + // printf("\n"); + // exit(0); + //} +} + +// ggml_compute_forward_get_rows_back + +static void ggml_compute_forward_get_rows_back_f32_f16( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + if (params->ith != 0) { + return; + } + + GGML_ASSERT(ggml_is_contiguous(dst)); + + // ggml_compute_forward_dup_same_cont(params, opt0, dst); + + memset(dst->data, 0, ggml_nbytes(dst)); + + const int nc = src0->ne[0]; + const int nr = ggml_nelements(src1); + + GGML_ASSERT( dst->ne[0] == nc); + GGML_ASSERT(src0->nb[0] == sizeof(ggml_fp16_t)); + + for (int i = 0; i < nr; ++i) { + const int r = ((int32_t *) src1->data)[i]; + + for (int j = 0; j < nc; ++j) { + ggml_fp16_t v = ((ggml_fp16_t *) ((char *) src0->data + i*src0->nb[1]))[j]; + ((float *) ((char *) dst->data + r*dst->nb[1]))[j] += GGML_FP16_TO_FP32(v); + } + } +} + +static void ggml_compute_forward_get_rows_back_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + if (params->ith != 0) { + return; + } + + GGML_ASSERT(ggml_is_contiguous(dst)); + + // ggml_compute_forward_dup_same_cont(params, opt0, dst); + + memset(dst->data, 0, ggml_nbytes(dst)); + + const int nc = src0->ne[0]; + const int nr = ggml_nelements(src1); + + GGML_ASSERT( dst->ne[0] == nc); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < nr; ++i) { + const int r = ((int32_t *) src1->data)[i]; + + ggml_vec_add_f32(nc, + (float *) ((char *) dst->data + r*dst->nb[1]), + (float *) ((char *) dst->data + r*dst->nb[1]), + (float *) ((char *) src0->data + i*src0->nb[1])); + } +} + +static void ggml_compute_forward_get_rows_back( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F16: + { + ggml_compute_forward_get_rows_back_f32_f16(params, dst); + } break; + case GGML_TYPE_F32: + { + ggml_compute_forward_get_rows_back_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } + + //static bool first = true; + //printf("ne0 = %d, ne1 = %d, ne2 = %d\n", dst->ne[0], dst->ne[1], dst->ne[2]); + //if (first) { + // first = false; + //} else { + // for (int k = 0; k < dst->ne[1]; ++k) { + // for (int j = 0; j < dst->ne[0]/16; ++j) { + // for (int i = 0; i < 16; ++i) { + // printf("%8.4f ", ((float *) dst->data)[k*dst->ne[0] + j*16 + i]); + // } + // printf("\n"); + // } + // printf("\n"); + // } + // printf("\n"); + // exit(0); + //} +} + +// ggml_compute_forward_diag + +static void ggml_compute_forward_diag_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + // TODO: handle transposed/permuted matrices + + GGML_TENSOR_UNARY_OP_LOCALS + + GGML_ASSERT(ne00 == ne0); + GGML_ASSERT(ne00 == ne1); + GGML_ASSERT(ne01 == 1); + GGML_ASSERT(ne02 == ne2); + GGML_ASSERT(ne03 == ne3); + + GGML_ASSERT(nb00 == sizeof(float)); + GGML_ASSERT(nb0 == sizeof(float)); + + for (int i3 = 0; i3 < ne3; i3++) { + for (int i2 = 0; i2 < ne2; i2++) { + for (int i1 = 0; i1 < ne1; i1++) { + float * d = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); + float * s = (float *)((char *) src0->data + i3*nb03 + i2*nb02); + for (int i0 = 0; i0 < i1; i0++) { + d[i0] = 0; + } + d[i1] = s[i1]; + for (int i0 = i1+1; i0 < ne0; i0++) { + d[i0] = 0; + } + } + } + } +} + +static void ggml_compute_forward_diag( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_diag_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_diag_mask_inf + +static void ggml_compute_forward_diag_mask_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst, + const float value) { + + const struct ggml_tensor * src0 = dst->src[0]; + + const int ith = params->ith; + const int nth = params->nth; + + const int n_past = ((int32_t *) dst->op_params)[0]; + const bool inplace = src0->data == dst->data; + + GGML_ASSERT(n_past >= 0); + + if (!inplace) { + if (ith == 0) { + // memcpy needs to be synchronized across threads to avoid race conditions. + // => do it in INIT phase + GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); + GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); + memcpy( + ((char *) dst->data), + ((char *) src0->data), + ggml_nbytes(dst)); + } + ggml_barrier(params->threadpool); + } + + // TODO: handle transposed/permuted matrices + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + const int nr = src0->ne[1]; + const int nz = n/nr; + + GGML_ASSERT( dst->nb[0] == sizeof(float)); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + for (int k = 0; k < nz; k++) { + for (int j = ith; j < nr; j += nth) { + for (int i = n_past; i < nc; i++) { + if (i > n_past + j) { + *(float *)((char *) dst->data + k*dst->nb[2] + j*dst->nb[1] + i*dst->nb[0]) = value; + } + } + } + } +} + +static void ggml_compute_forward_diag_mask_inf( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_diag_mask_f32(params, dst, -INFINITY); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +static void ggml_compute_forward_diag_mask_zero( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_diag_mask_f32(params, dst, 0); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_soft_max + +static void ggml_compute_forward_soft_max_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + assert(ggml_is_contiguous(dst)); + assert(ggml_are_same_shape(src0, dst)); + + float scale = 1.0f; + float max_bias = 0.0f; + + memcpy(&scale, (float *) dst->op_params + 0, sizeof(float)); + memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float)); + + // TODO: handle transposed/permuted matrices + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_UNARY_OP_LOCALS + + //const int64_t ne11 = src1 ? src1->ne[1] : 1; + + // TODO: is this supposed to be ceil instead of floor? + // https://huggingface.co/mosaicml/mpt-7b/blob/main/attention.py#L370 + const uint32_t n_head = ne02; + const uint32_t n_head_log2 = 1u << (uint32_t) floor(log2(n_head)); + + const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); + const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); + + const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + float * wp = (float *) params->wdata + (nc + CACHE_LINE_SIZE_F32) * ith; + + const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16); + + for (int i1 = ir0; i1 < ir1; i1++) { + // ALiBi + const uint32_t h = (i1/ne01)%ne02; // head + const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f; + + float * sp = (float *)((char *) src0->data + i1*src0->nb[1]); + float * dp = (float *)((char *) dst->data + i1*dst->nb[1]); + + // broadcast the mask across rows + ggml_fp16_t * mp_f16 = src1 ? (ggml_fp16_t *)((char *) src1->data) + (i1%ne01)*ne00 : NULL; + float * mp_f32 = src1 ? (float *)((char *) src1->data) + (i1%ne01)*ne00 : NULL; + + ggml_vec_cpy_f32 (nc, wp, sp); + ggml_vec_scale_f32(nc, wp, scale); + if (mp_f32) { + if (use_f16) { + for (int i = 0; i < nc; ++i) { + wp[i] += slope*GGML_FP16_TO_FP32(mp_f16[i]); + } + } else { + for (int i = 0; i < nc; ++i) { + wp[i] += slope*mp_f32[i]; + } + } + } + +#ifndef NDEBUG + for (int i = 0; i < nc; ++i) { + //printf("p[%d] = %f\n", i, p[i]); + assert(!isnan(wp[i])); + } +#endif + + float max = -INFINITY; + ggml_vec_max_f32(nc, &max, wp); + + ggml_float sum = ggml_vec_soft_max_f32(nc, dp, wp, max); + assert(sum > 0.0); + + sum = 1.0/sum; + ggml_vec_scale_f32(nc, dp, sum); + +#ifndef NDEBUG + for (int i = 0; i < nc; ++i) { + assert(!isnan(dp[i])); + assert(!isinf(dp[i])); + } +#endif + } +} + +static void ggml_compute_forward_soft_max( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_soft_max_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + + +// ggml_compute_forward_soft_max_back + +static void ggml_compute_forward_soft_max_back_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(ggml_is_contiguous(src1)); + GGML_ASSERT(ggml_is_contiguous(dst)); + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_are_same_shape(src1, dst)); + + // TODO: handle transposed/permuted matrices + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int i1 = ir0; i1 < ir1; i1++) { + float *dy = (float *)((char *) src0->data + i1*src0->nb[1]); + float *y = (float *)((char *) src1->data + i1*src1->nb[1]); + float *dx = (float *)((char *) dst->data + i1*dst->nb[1]); + +#ifndef NDEBUG + for (int i = 0; i < nc; ++i) { + //printf("p[%d] = %f\n", i, p[i]); + assert(!isnan(dy[i])); + assert(!isnan(y[i])); + } +#endif + // Jii = yi - yi*yi + // Jij = -yi*yj + // J = diag(y)-y.T*y + // dx = J * dy + // dxk = sum_i(Jki * dyi) + // dxk = sum_i(-yk*yi * dyi) - (-yk*yk)*dyk + (yk - yk*yk)*dyk + // dxk = sum_i(-yk*yi * dyi) + yk*yk*dyk + yk*dyk - yk*yk*dyk + // dxk = sum_i(-yk*yi * dyi) + yk*dyk + // dxk = -yk * sum_i(yi * dyi) + yk*dyk + // dxk = -yk * dot(y, dy) + yk*dyk + // dxk = yk * (- dot(y, dy) + dyk) + // dxk = yk * (dyk - dot(y, dy)) + // + // post-order: + // dot_y_dy := dot(y, dy) + // dx := dy + // dx := dx - dot_y_dy + // dx := dx * y + + // linear runtime, no additional memory + float dot_y_dy = 0; + ggml_vec_dot_f32 (nc, &dot_y_dy, 0, y, 0, dy, 0, 1); + ggml_vec_cpy_f32 (nc, dx, dy); + ggml_vec_acc1_f32(nc, dx, -dot_y_dy); + ggml_vec_mul_f32 (nc, dx, dx, y); + +#ifndef NDEBUG + for (int i = 0; i < nc; ++i) { + assert(!isnan(dx[i])); + assert(!isinf(dx[i])); + } +#endif + } +} + +static void ggml_compute_forward_soft_max_back( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_soft_max_back_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_clamp + +static void ggml_compute_forward_clamp_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + float min; + float max; + memcpy(&min, (float *) dst->op_params + 0, sizeof(float)); + memcpy(&max, (float *) dst->op_params + 1, sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + + GGML_ASSERT( nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); + + for (int j = ith; j < n; j += nth) { + float * dst_ptr = (float *) ((char *) dst->data + j*nb1); + float * src0_ptr = (float *) ((char *) src0->data + j*nb01); + + for (int i = 0; i < nc; i++) { + dst_ptr[i] = MAX(MIN(src0_ptr[i], max), min); + } + } +} + +static void ggml_compute_forward_clamp( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_clamp_f32(params, dst); + } break; + case GGML_TYPE_F16: + case GGML_TYPE_BF16: + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q8_1: + case GGML_TYPE_Q2_K: + case GGML_TYPE_Q3_K: + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + case GGML_TYPE_Q6_K: + case GGML_TYPE_TQ1_0: + case GGML_TYPE_TQ2_0: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ4_NL: + case GGML_TYPE_IQ4_XS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ2_S: + case GGML_TYPE_Q8_K: + case GGML_TYPE_Q4_0_4_4: + case GGML_TYPE_Q4_0_4_8: + case GGML_TYPE_Q4_0_8_8: + case GGML_TYPE_I8: + case GGML_TYPE_I16: + case GGML_TYPE_I32: + case GGML_TYPE_I64: + case GGML_TYPE_F64: + case GGML_TYPE_COUNT: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_rope + +static float rope_yarn_ramp(const float low, const float high, const int i0) { + const float y = (i0 / 2 - low) / MAX(0.001f, high - low); + return 1 - MIN(1, MAX(0, y)); +} + +// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn +// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng. +static void rope_yarn( + float theta_extrap, float freq_scale, float corr_dims[2], int64_t i0, float ext_factor, float mscale, + float * cos_theta, float * sin_theta) { + // Get n-d rotational scaling corrected for extrapolation + float theta_interp = freq_scale * theta_extrap; + float theta = theta_interp; + if (ext_factor != 0.0f) { + float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor; + theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix; + + // Get n-d magnitude scaling corrected for interpolation + mscale *= 1.0f + 0.1f * logf(1.0f / freq_scale); + } + *cos_theta = cosf(theta) * mscale; + *sin_theta = sinf(theta) * mscale; +} + +// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get +// `corr_dim(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))` +static float ggml_rope_yarn_corr_dim(int n_dims, int n_ctx_orig, float n_rot, float base) { + return n_dims * logf(n_ctx_orig / (n_rot * 2 * (float)M_PI)) / (2 * logf(base)); +} + +static void ggml_rope_cache_init( + float theta_base, float freq_scale, const float * freq_factors, float corr_dims[2], int64_t ne0, float ext_factor, float mscale, + float * cache, float sin_sign, float theta_scale) { + // ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py + float theta = theta_base; + for (int64_t i0 = 0; i0 < ne0; i0 += 2) { + const float ff = freq_factors ? freq_factors[i0/2] : 1.0f; + rope_yarn( + theta/ff, freq_scale, corr_dims, i0, ext_factor, mscale, &cache[i0 + 0], &cache[i0 + 1] + ); + cache[i0 + 1] *= sin_sign; + + theta *= theta_scale; + } +} + +static void ggml_mrope_cache_init( + float theta_base_t, float theta_base_h, float theta_base_w, int sections[3], + float freq_scale, const float * freq_factors, float corr_dims[2], int64_t ne0, float ext_factor, float mscale, + float * cache, float sin_sign, float theta_scale) { + // ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py + float theta_t = theta_base_t; + float theta_h = theta_base_h; + float theta_w = theta_base_w; + int sect_dims = sections[0] + sections[1] + sections[2]; + + for (int64_t i0 = 0; i0 < ne0; i0 += 2) { + const float ff = freq_factors ? freq_factors[i0/2] : 1.0f; + + float theta = theta_base_t; + int sector = i0 % sect_dims; + if (sector > sections[1] && sector >= sections[0]) + theta = theta_base_h; + else if (sector >= sections[1]) + theta = theta_base_w; + + rope_yarn( + theta/ff, freq_scale, corr_dims, i0, ext_factor, mscale, &cache[i0 + 0], &cache[i0 + 1] + ); + cache[i0 + 1] *= sin_sign; + + theta *= theta_scale; + } +} + +GGML_CALL void ggml_rope_yarn_corr_dims( + int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2] +) { + // start and end correction dims + float start = floorf(ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_fast, freq_base)); + float end = ceilf(ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_slow, freq_base)); + dims[0] = MAX(0, start); + dims[1] = MIN(n_dims - 1, end); +} + +static void ggml_compute_forward_rope_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst, + const bool forward) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + const struct ggml_tensor * src2 = dst->src[2]; + + float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; + int sections[3]; + + //const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_dims = ((int32_t *) dst->op_params)[1]; + const int mode = ((int32_t *) dst->op_params)[2]; + //const int n_ctx = ((int32_t *) dst->op_params)[3]; + const int n_ctx_orig = ((int32_t *) dst->op_params)[4]; + + memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); + memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); + memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); + memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); + memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); + memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); + memcpy(§ions, (int32_t *) dst->op_params + 11, sizeof(int) * 3); + + GGML_TENSOR_UNARY_OP_LOCALS + + //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3); + //printf("n_past = %d, ne2 = %d\n", n_past, ne2); + + GGML_ASSERT(nb00 == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(dst); + + GGML_ASSERT(n_dims <= ne0); + GGML_ASSERT(n_dims % 2 == 0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + // row index used to determine which thread to use + int ir = 0; + + const float theta_scale = powf(freq_base, -2.0f/n_dims); + + float corr_dims[2]; + ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims); + + const bool is_neox = mode & GGML_ROPE_TYPE_NEOX; + const bool is_mrope = sections[0] > 0 || sections[1] > 0 || sections[2] > 0; + + const float * freq_factors = NULL; + if (src2 != NULL) { + GGML_ASSERT(src2->type == GGML_TYPE_F32); + GGML_ASSERT(src2->ne[0] >= n_dims / 2); + freq_factors = (const float *) src2->data; + } + + // backward process uses inverse rotation by cos and sin. + // cos and sin build a rotation matrix, where the inverse is the transpose. + // this essentially just switches the sign of sin. + const float sin_sign = forward ? 1.0f : -1.0f; + + const int32_t * pos = (const int32_t *) src1->data; + + for (int64_t i3 = 0; i3 < ne3; i3++) { // batch + for (int64_t i2 = 0; i2 < ne2; i2++) { // seq-len + + float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith; + if (!is_mrope) { + const int64_t p = pos[i2]; + ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale); + } + else { + const int64_t p_t = pos[i2]; + const int64_t p_h = pos[i2 + ne2]; + const int64_t p_w = pos[i2 + ne2 * 2]; + ggml_mrope_cache_init( + p_t, p_h, p_w, sections, + freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale); + } + + for (int64_t i1 = 0; i1 < ne1; i1++) { // attn-heads + if (ir++ < ir0) continue; + if (ir > ir1) break; + + if (!is_neox) { + for (int64_t i0 = 0; i0 < n_dims; i0 += 2) { + const float cos_theta = cache[i0 + 0]; + const float sin_theta = cache[i0 + 1]; + + const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + const float x0 = src[0]; + const float x1 = src[1]; + + dst_data[0] = x0*cos_theta - x1*sin_theta; + dst_data[1] = x0*sin_theta + x1*cos_theta; + } + } else { + for (int64_t i0 = 0; i0 < n_dims; i0 += 2) { + const int64_t ic = i0/2; + + const float cos_theta = cache[i0 + 0]; + const float sin_theta = cache[i0 + 1]; + + const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); + float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); + + const float x0 = src[0]; + const float x1 = src[n_dims/2]; + + dst_data[0] = x0*cos_theta - x1*sin_theta; + dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta; + } + } + + for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) { + const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + dst_data[0] = src[0]; + dst_data[1] = src[1]; + } + } + } + } +} + +// TODO: deduplicate f16/f32 code +static void ggml_compute_forward_rope_f16( + const struct ggml_compute_params * params, + struct ggml_tensor * dst, + const bool forward) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + const struct ggml_tensor * src2 = dst->src[2]; + + float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; + + //const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_dims = ((int32_t *) dst->op_params)[1]; + const int mode = ((int32_t *) dst->op_params)[2]; + //const int n_ctx = ((int32_t *) dst->op_params)[3]; + const int n_ctx_orig = ((int32_t *) dst->op_params)[4]; + memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); + memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); + memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); + memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); + memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); + memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); + + GGML_TENSOR_UNARY_OP_LOCALS + + //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3); + //printf("n_past = %d, ne2 = %d\n", n_past, ne2); + + GGML_ASSERT(nb0 == sizeof(ggml_fp16_t)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(dst); + + GGML_ASSERT(n_dims <= ne0); + GGML_ASSERT(n_dims % 2 == 0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + // row index used to determine which thread to use + int ir = 0; + + const float theta_scale = powf(freq_base, -2.0f/n_dims); + + float corr_dims[2]; + ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims); + + const bool is_neox = mode & GGML_ROPE_TYPE_NEOX; + + const float * freq_factors = NULL; + if (src2 != NULL) { + GGML_ASSERT(src2->type == GGML_TYPE_F32); + GGML_ASSERT(src2->ne[0] >= n_dims / 2); + freq_factors = (const float *) src2->data; + } + + // backward process uses inverse rotation by cos and sin. + // cos and sin build a rotation matrix, where the inverse is the transpose. + // this essentially just switches the sign of sin. + const float sin_sign = forward ? 1.0f : -1.0f; + + const int32_t * pos = (const int32_t *) src1->data; + + for (int64_t i3 = 0; i3 < ne3; i3++) { + for (int64_t i2 = 0; i2 < ne2; i2++) { + const int64_t p = pos[i2]; + + float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith; + ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale); + + for (int64_t i1 = 0; i1 < ne1; i1++) { + if (ir++ < ir0) continue; + if (ir > ir1) break; + + if (!is_neox) { + for (int64_t i0 = 0; i0 < n_dims; i0 += 2) { + const float cos_theta = cache[i0 + 0]; + const float sin_theta = cache[i0 + 1]; + + const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + const float x0 = GGML_FP16_TO_FP32(src[0]); + const float x1 = GGML_FP16_TO_FP32(src[1]); + + dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); + dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); + } + } else { + for (int64_t i0 = 0; i0 < n_dims; i0 += 2) { + const int64_t ic = i0/2; + + const float cos_theta = cache[i0 + 0]; + const float sin_theta = cache[i0 + 1]; + + const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); + ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); + + const float x0 = GGML_FP16_TO_FP32(src[0]); + const float x1 = GGML_FP16_TO_FP32(src[n_dims/2]); + + dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); + dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); + } + } + + for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) { + const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + dst_data[0] = src[0]; + dst_data[1] = src[1]; + } + } + } + } +} + +static void ggml_compute_forward_rope( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F16: + { + ggml_compute_forward_rope_f16(params, dst, true); + } break; + case GGML_TYPE_F32: + { + ggml_compute_forward_rope_f32(params, dst, true); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_rope_back + +static void ggml_compute_forward_rope_back( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F16: + { + ggml_compute_forward_rope_f16(params, dst, false); + } break; + case GGML_TYPE_F32: + { + ggml_compute_forward_rope_f32(params, dst, false); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_conv_transpose_1d + +static void ggml_compute_forward_conv_transpose_1d_f16_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + GGML_TENSOR_BINARY_OP_LOCALS + + const int ith = params->ith; + const int nth = params->nth; + + const int nk = ne00*ne01*ne02; + + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb10 == sizeof(float)); + + if (ith == 0) { + memset(params->wdata, 0, params->wsize); + + // permute kernel data (src0) from (K x Cout x Cin) to (Cin x K x Cout) + { + ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; + + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = 0; i01 < ne01; i01++) { + const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i02*nb02 + i01*nb01); + ggml_fp16_t * dst_data = wdata + i01*ne00*ne02; + for (int64_t i00 = 0; i00 < ne00; i00++) { + dst_data[i00*ne02 + i02] = src[i00]; + } + } + } + } + + // permute source data (src1) from (L x Cin) to (Cin x L) + { + ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + nk; + ggml_fp16_t * dst_data = wdata; + + for (int64_t i11 = 0; i11 < ne11; i11++) { + const float * const src = (float *)((char *) src1->data + i11*nb11); + for (int64_t i10 = 0; i10 < ne10; i10++) { + dst_data[i10*ne11 + i11] = GGML_FP32_TO_FP16(src[i10]); + } + } + } + + // need to zero dst since we are accumulating into it + memset(dst->data, 0, ggml_nbytes(dst)); + } + ggml_barrier(params->threadpool); + + const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; + + // total rows in dst + const int nr = ne1; + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; + ggml_fp16_t * const wdata_src = wdata + nk; + + for (int i1 = ir0; i1 < ir1; i1++) { + float * dst_data = (float *)((char *) dst->data + i1*nb1); + ggml_fp16_t * wdata_kernel = wdata + i1*ne02*ne00; + for (int i10 = 0; i10 < ne10; i10++) { + const int i1n = i10*ne11; + for (int i00 = 0; i00 < ne00; i00++) { + float v = 0; + ggml_vec_dot_f16(ne02, &v, 0, + (ggml_fp16_t *) wdata_src + i1n, 0, + (ggml_fp16_t *) wdata_kernel + i00*ne02, 0, 1); + dst_data[i10*s0 + i00] += v; + } + } + } +} + +static void ggml_compute_forward_conv_transpose_1d_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + GGML_TENSOR_BINARY_OP_LOCALS + + const int ith = params->ith; + const int nth = params->nth; + + const int nk = ne00*ne01*ne02; + + GGML_ASSERT(nb00 == sizeof(float)); + GGML_ASSERT(nb10 == sizeof(float)); + + if (ith == 0) { + memset(params->wdata, 0, params->wsize); + + // prepare kernel data (src0) from (K x Cout x Cin) to (Cin x K x Cout) + { + float * const wdata = (float *) params->wdata + 0; + + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = 0; i01 < ne01; i01++) { + const float * const src = (float *)((char *) src0->data + i02*nb02 + i01*nb01); + float * dst_data = wdata + i01*ne00*ne02; + for (int64_t i00 = 0; i00 < ne00; i00++) { + dst_data[i00*ne02 + i02] = src[i00]; + } + } + } + } + + // prepare source data (src1) + { + float * const wdata = (float *) params->wdata + nk; + float * dst_data = wdata; + + for (int64_t i11 = 0; i11 < ne11; i11++) { + const float * const src = (float *)((char *) src1->data + i11*nb11); + for (int64_t i10 = 0; i10 < ne10; i10++) { + dst_data[i10*ne11 + i11] = src[i10]; + } + } + } + + // need to zero dst since we are accumulating into it + memset(dst->data, 0, ggml_nbytes(dst)); + } + ggml_barrier(params->threadpool); + + const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; + + // total rows in dst + const int nr = ne1; + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + float * const wdata = (float *) params->wdata + 0; + float * const wdata_src = wdata + nk; + + for (int i1 = ir0; i1 < ir1; i1++) { + float * dst_data = (float *)((char *) dst->data + i1*nb1); + float * wdata_kernel = wdata + i1*ne02*ne00; + for (int i10 = 0; i10 < ne10; i10++) { + const int i1n = i10*ne11; + for (int i00 = 0; i00 < ne00; i00++) { + float v = 0; + ggml_vec_dot_f32(ne02, &v, 0, + wdata_src + i1n, 0, + wdata_kernel + i00*ne02, 0, 1); + dst_data[i10*s0 + i00] += v; + } + } + } +} + +static void ggml_compute_forward_conv_transpose_1d( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F16: + { + ggml_compute_forward_conv_transpose_1d_f16_f32(params, dst); + } break; + case GGML_TYPE_F32: + { + ggml_compute_forward_conv_transpose_1d_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_im2col_f32 +// src0: kernel [OC, IC, KH, KW] +// src1: image [N, IC, IH, IW] +// dst: result [N, OH, OW, IC*KH*KW] +static void ggml_compute_forward_im2col_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + GGML_TENSOR_BINARY_OP_LOCALS; + + const int32_t s0 = ((const int32_t *)(dst->op_params))[0]; + const int32_t s1 = ((const int32_t *)(dst->op_params))[1]; + const int32_t p0 = ((const int32_t *)(dst->op_params))[2]; + const int32_t p1 = ((const int32_t *)(dst->op_params))[3]; + const int32_t d0 = ((const int32_t *)(dst->op_params))[4]; + const int32_t d1 = ((const int32_t *)(dst->op_params))[5]; + const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1; + + const int ith = params->ith; + const int nth = params->nth; + + const int64_t N = is_2D ? ne13 : ne12; + const int64_t IC = is_2D ? ne12 : ne11; + const int64_t IH = is_2D ? ne11 : 1; + const int64_t IW = ne10; + + const int64_t KH = is_2D ? ne01 : 1; + const int64_t KW = ne00; + + const int64_t OH = is_2D ? ne2 : 1; + const int64_t OW = ne1; + + int ofs0 = is_2D ? nb13 : nb12; + int ofs1 = is_2D ? nb12 : nb11; + + GGML_ASSERT(nb10 == sizeof(float)); + + // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW] + { + float * const wdata = (float *) dst->data; + + for (int64_t in = 0; in < N; in++) { + for (int64_t ioh = 0; ioh < OH; ioh++) { // 1 + for (int64_t iow = 0; iow < OW; iow++) { + for (int64_t iic = ith; iic < IC; iic += nth) { + + // micro kernel + float * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW] + const float * const src_data = (float *)((char *) src1->data + in*ofs0 + iic*ofs1); // [IH, IW] + + for (int64_t ikh = 0; ikh < KH; ikh++) { // 1 + for (int64_t ikw = 0; ikw < KW; ikw++) { + const int64_t iiw = iow*s0 + ikw*d0 - p0; + const int64_t iih = ioh*s1 + ikh*d1 - p1; + + if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { + dst_data[iic*(KH*KW) + ikh*KW + ikw] = 0; + } else { + dst_data[iic*(KH*KW) + ikh*KW + ikw] = (src_data[iih*IW + iiw]); + } + } + } + } + } + } + } + } +} + + +// ggml_compute_forward_im2col_f16 +// src0: kernel [OC, IC, KH, KW] +// src1: image [N, IC, IH, IW] +// dst: result [N, OH, OW, IC*KH*KW] +static void ggml_compute_forward_im2col_f16( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F16); + + GGML_TENSOR_BINARY_OP_LOCALS; + + const int32_t s0 = ((const int32_t *)(dst->op_params))[0]; + const int32_t s1 = ((const int32_t *)(dst->op_params))[1]; + const int32_t p0 = ((const int32_t *)(dst->op_params))[2]; + const int32_t p1 = ((const int32_t *)(dst->op_params))[3]; + const int32_t d0 = ((const int32_t *)(dst->op_params))[4]; + const int32_t d1 = ((const int32_t *)(dst->op_params))[5]; + const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1; + + const int ith = params->ith; + const int nth = params->nth; + + const int64_t N = is_2D ? ne13 : ne12; + const int64_t IC = is_2D ? ne12 : ne11; + const int64_t IH = is_2D ? ne11 : 1; + const int64_t IW = ne10; + + const int64_t KH = is_2D ? ne01 : 1; + const int64_t KW = ne00; + + const int64_t OH = is_2D ? ne2 : 1; + const int64_t OW = ne1; + + int ofs0 = is_2D ? nb13 : nb12; + int ofs1 = is_2D ? nb12 : nb11; + + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb10 == sizeof(float)); + + // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW] + { + ggml_fp16_t * const wdata = (ggml_fp16_t *) dst->data; + + for (int64_t in = 0; in < N; in++) { + for (int64_t ioh = 0; ioh < OH; ioh++) { // 1 + for (int64_t iow = 0; iow < OW; iow++) { + for (int64_t iic = ith; iic < IC; iic += nth) { + + // micro kernel + ggml_fp16_t * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW] + const float * const src_data = (float *)((char *) src1->data + in*ofs0 + iic*ofs1); // [IH, IW] + + for (int64_t ikh = 0; ikh < KH; ikh++) { // 1 + for (int64_t ikw = 0; ikw < KW; ikw++) { + const int64_t iiw = iow*s0 + ikw*d0 - p0; + const int64_t iih = ioh*s1 + ikh*d1 - p1; + + if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { + dst_data[iic*(KH*KW) + ikh*KW + ikw] = 0; + } else { + dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_FP32_TO_FP16(src_data[iih*IW + iiw]); + } + } + } + } + } + } + } + } +} + +static void ggml_compute_forward_im2col( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + switch (dst->type) { + case GGML_TYPE_F16: + { + ggml_compute_forward_im2col_f16(params, dst); + } break; + case GGML_TYPE_F32: + { + ggml_compute_forward_im2col_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_im2col_back_f32 + +static void ggml_compute_forward_im2col_back_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + GGML_TENSOR_BINARY_OP_LOCALS; + + const int32_t s0 = ((const int32_t *)(dst->op_params))[0]; + const int32_t s1 = ((const int32_t *)(dst->op_params))[1]; + const int32_t p0 = ((const int32_t *)(dst->op_params))[2]; + const int32_t p1 = ((const int32_t *)(dst->op_params))[3]; + const int32_t d0 = ((const int32_t *)(dst->op_params))[4]; + const int32_t d1 = ((const int32_t *)(dst->op_params))[5]; + const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1; + + const int ith = params->ith; + const int nth = params->nth; + + const int64_t N = is_2D ? ne3 : ne2; + const int64_t IC = is_2D ? ne2 : ne1; + const int64_t IH = is_2D ? ne1 : 1; + const int64_t IW = ne0; + + const int64_t KH = is_2D ? ne01 : 1; + const int64_t KW = ne00; + + const int64_t OH = is_2D ? ne12 : 1; + const int64_t OW = ne11; + + int ofs0 = is_2D ? nb3 : nb2; + int ofs1 = is_2D ? nb2 : nb1; + + GGML_ASSERT(nb0 == sizeof(float)); + + // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW] + { + float * const wdata = (float *) dst->data; + + for (int64_t in = 0; in < N; in++) { + for (int64_t iic = ith; iic < IC; iic += nth) { + for (int64_t iih = 0; iih < IH; iih++) { + for (int64_t iiw = 0; iiw < IW; iiw++) { + + // micro kernel + float grad = 0.0f; + for (int64_t ikh = 0; ikh < KH; ikh++) { + for (int64_t ikw = 0; ikw < KW; ikw++) { + // For s0 > 1 some values were skipped over in the forward pass. + // These values have tmpw % s0 != 0 and need to be skipped in the backwards pass as well. + const int64_t tmpw = (iiw + p0 - ikw*d0); + if (tmpw % s0 != 0) { + continue; + } + const int64_t iow = tmpw / s0; + + // Equivalent logic as above except for s1. + int64_t ioh; + if (is_2D) { + const int64_t tmph = iih + p1 - ikh*d1; + + if (tmph % s1 != 0) { + continue; + } + + ioh = tmph / s1; + } else { + ioh = 0; + } + + if (iow < 0 || iow >= OW || ioh < 0 || ioh >= OH) { + continue; + } + + const float * const src_data = (const float *) src1->data + + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW] + grad += src_data[iic*(KH*KW) + ikh*KW + ikw]; + } + } + float * dst_data = (float *)((char *) wdata + (in*ofs0 + iic*ofs1)); // [IH, IW] + dst_data[iih*IW + iiw] = grad; + } + } + } + } + } +} + +// ggml_compute_forward_conv_transpose_2d + +static void ggml_compute_forward_conv_transpose_2d( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + GGML_TENSOR_BINARY_OP_LOCALS + + const int ith = params->ith; + const int nth = params->nth; + + const int nk = ne00*ne01*ne02*ne03; + + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb10 == sizeof(float)); + + if (ith == 0) { + memset(params->wdata, 0, params->wsize); + + // permute kernel data (src0) from (Kw x Kh x Cout x Cin) to (Cin x Kw x Kh x Cout) + { + ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i03*nb03 + i02*nb02); + ggml_fp16_t * dst_data = wdata + i02*ne01*ne00*ne03; + for (int64_t i01 = 0; i01 < ne01; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + dst_data[i01*ne00*ne03 + i00*ne03 + i03] = src[i01 * ne00 + i00]; + } + } + } + } + } + + // permute source data (src1) from (Sw x Sh x Cin) to (Cin x Sw x Sh) + { + ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + nk; + for (int i12 = 0; i12 < ne12; i12++) { + for (int i11 = 0; i11 < ne11; i11++) { + const float * const src = (float *)((char *) src1->data + i12*nb12 + i11*nb11); + ggml_fp16_t * dst_data = wdata + i11*ne10*ne12; + for (int i10 = 0; i10 < ne10; i10++) { + dst_data[i10*ne12 + i12] = GGML_FP32_TO_FP16(src[i10]); + } + } + } + } + + memset(dst->data, 0, ggml_nbytes(dst)); + } + ggml_barrier(params->threadpool); + + const int32_t stride = ggml_get_op_params_i32(dst, 0); + + // total patches in dst + const int np = ne2; + + // patches per thread + const int dp = (np + nth - 1)/nth; + + // patch range for this thread + const int ip0 = dp*ith; + const int ip1 = MIN(ip0 + dp, np); + + ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; + ggml_fp16_t * const wdata_src = wdata + nk; + + for (int i2 = ip0; i2 < ip1; i2++) { // Cout + float * dst_data = (float *)((char *) dst->data + i2*nb2); + ggml_fp16_t * wdata_kernel = wdata + i2*ne01*ne00*ne03; + for (int i11 = 0; i11 < ne11; i11++) { + for (int i10 = 0; i10 < ne10; i10++) { + const int i1n = i11*ne10*ne12 + i10*ne12; + for (int i01 = 0; i01 < ne01; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + float v = 0; + ggml_vec_dot_f16(ne03, &v, 0, + wdata_src + i1n, 0, + wdata_kernel + i01*ne00*ne03 + i00*ne03, 0, 1); + dst_data[(i11*stride + i01)*ne0 + i10*stride + i00] += v; + } + } + } + } + } +} + +// ggml_compute_forward_pool_1d_sk_p0 + +static void ggml_compute_forward_pool_1d_sk_p0( + const struct ggml_compute_params * params, + const enum ggml_op_pool op, + const int k, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src = dst->src[0]; + + assert(src->type == GGML_TYPE_F32 || src->type == GGML_TYPE_F16); + + if (params->ith != 0) { + return; + } + + const char * cdata = (const char *)src->data; + const char * const data_end = cdata + ggml_nbytes(src); + float * drow = (float *)dst->data; + + const int64_t rs = dst->ne[0]; + + while (cdata < data_end) { + const void * srow = (const void *)cdata; + int j = 0; + for (int64_t i = 0; i < rs; ++i) { + switch (op) { + case GGML_OP_POOL_AVG: drow[i] = 0; break; + case GGML_OP_POOL_MAX: drow[i] = -FLT_MAX; break; + case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error"); + } + for (int ki = 0; ki < k; ++ki) { + const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]); + switch (op) { + case GGML_OP_POOL_AVG: drow[i] += srow_j; break; + case GGML_OP_POOL_MAX: if (srow_j > drow[i]) drow[i] = srow_j; break; + case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error"); + } + ++j; + } + switch (op) { + case GGML_OP_POOL_AVG: drow[i] /= k; break; + case GGML_OP_POOL_MAX: break; + case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error"); + } + } + + cdata += src->nb[1]; + drow += rs; + } +} + +// ggml_compute_forward_pool_1d + +static void ggml_compute_forward_pool_1d( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const int32_t * opts = (const int32_t *)dst->op_params; + enum ggml_op_pool op = opts[0]; + const int k0 = opts[1]; + const int s0 = opts[2]; + const int p0 = opts[3]; + GGML_ASSERT(p0 == 0); // padding not supported + GGML_ASSERT(k0 == s0); // only s = k supported + + ggml_compute_forward_pool_1d_sk_p0(params, op, k0, dst); +} + +// ggml_compute_forward_pool_2d + +static void ggml_compute_forward_pool_2d( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src = dst->src[0]; + + assert(src->type == GGML_TYPE_F32 || src->type == GGML_TYPE_F16); + + if (params->ith != 0) { + return; + } + + const int32_t * opts = (const int32_t *)dst->op_params; + enum ggml_op_pool op = opts[0]; + const int k0 = opts[1]; + const int k1 = opts[2]; + const int s0 = opts[3]; + const int s1 = opts[4]; + const int p0 = opts[5]; + const int p1 = opts[6]; + const char * cdata = (const char*)src->data; + const char * const data_end = cdata + ggml_nbytes(src); + + const int64_t px = dst->ne[0]; + const int64_t py = dst->ne[1]; + const int64_t pa = px * py; + + float * dplane = (float *)dst->data; + + const int ka = k0 * k1; + const int offset0 = -p0; + const int offset1 = -p1; + + while (cdata < data_end) { + for (int oy = 0; oy < py; ++oy) { + float * const drow = dplane + oy * px; + for (int ox = 0; ox < px; ++ox) { + float * const out = drow + ox; + switch (op) { + case GGML_OP_POOL_AVG: *out = 0; break; + case GGML_OP_POOL_MAX: *out = -FLT_MAX; break; + case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error"); + } + + const int ix = offset0 + ox * s0; + const int iy = offset1 + oy * s1; + + for (int ky = 0; ky < k1; ++ky) { + if (iy + ky < 0 || iy + ky >= src->ne[1]) continue; + const void * srow = (const void *)(cdata + src->nb[1] * (iy + ky)); + for (int kx = 0; kx < k0; ++kx) { + int j = ix + kx; + if (j < 0 || j >= src->ne[0]) continue; + const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]); + switch (op) { + case GGML_OP_POOL_AVG: *out += srow_j; break; + case GGML_OP_POOL_MAX: if (srow_j > *out) *out = srow_j; break; + case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error"); + } + } + } + switch (op) { + case GGML_OP_POOL_AVG: *out /= ka; break; + case GGML_OP_POOL_MAX: break; + case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error"); + } + } + } + + cdata += src->nb[2]; + dplane += pa; + } +} + +// ggml_compute_forward_pool_2d_back + +static void ggml_compute_forward_pool_2d_back( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src = dst->src[0]; + const struct ggml_tensor * dstf = dst->src[1]; // forward tensor of dst + + assert(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); + + if (params->ith != 0) { + return; + } + + const int32_t * opts = (const int32_t *)dst->op_params; + enum ggml_op_pool op = opts[0]; + const int k0 = opts[1]; + const int k1 = opts[2]; + const int s0 = opts[3]; + const int s1 = opts[4]; + const int p0 = opts[5]; + const int p1 = opts[6]; + + char * cdata = (char *) dst->data; + const char * cdataf = (const char *) dstf->data; + const char * const data_end = cdata + ggml_nbytes(dst); + + GGML_ASSERT(params->ith == 0); + memset(cdata, 0, ggml_nbytes(dst)); + + const int64_t px = src->ne[0]; + const int64_t py = src->ne[1]; + const int64_t pa = px * py; + + const float * splane = (const float *) src->data; + + const int ka = k0 * k1; + const int offset0 = -p0; + const int offset1 = -p1; + + while (cdata < data_end) { + for (int oy = 0; oy < py; ++oy) { + const float * const srow = splane + oy * px; + for (int ox = 0; ox < px; ++ox) { + const float grad0 = srow[ox]; + + const int ix = offset0 + ox * s0; + const int iy = offset1 + oy * s1; + + if (op == GGML_OP_POOL_MAX) { + float maxval = -FLT_MAX; + int kxmax = -1; + int kymax = -1; + + for (int ky = 0; ky < k1; ++ky) { + if (iy + ky < 0 || iy + ky >= dst->ne[1]) { + continue; + } + const void * drowf = (const void *)(cdataf + dst->nb[1] * (iy + ky)); + for (int kx = 0; kx < k0; ++kx) { + int j = ix + kx; + if (j < 0 || j >= dst->ne[0]) { + continue; + } + + const float val = dst->type == GGML_TYPE_F32 ? + ((const float *) drowf)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t *) drowf)[j]); + if (val <= maxval) { + continue; + } + + maxval = val; + kxmax = kx; + kymax = ky; + } + } + + if (kxmax == -1 || kymax == -1) { + continue; + } + + void * drow = (void *)(cdata + dst->nb[1] * (iy + kymax)); + const int j = ix + kxmax; + if (dst->type == GGML_TYPE_F32) { + ((float *) drow)[j] += grad0; + } else { + ((ggml_fp16_t *) drow)[j] = GGML_FP32_TO_FP16(grad0 + GGML_FP16_TO_FP32(((const ggml_fp16_t *) drow)[j])); + } + } else if (op == GGML_OP_POOL_AVG) { + const float grad = grad0 / ka; + + for (int ky = 0; ky < k1; ++ky) { + if (iy + ky < 0 || iy + ky >= dst->ne[1]) { + continue; + } + void * drow = (void *)(cdata + dst->nb[1] * (iy + ky)); + for (int kx = 0; kx < k0; ++kx) { + int j = ix + kx; + if (j < 0 || j >= dst->ne[0]) { + continue; + } + + if (dst->type == GGML_TYPE_F32) { + ((float *) drow)[j] += grad; + } else { + ((ggml_fp16_t *) drow)[j] += GGML_FP32_TO_FP16(grad); + } + } + } + } else { + GGML_ASSERT(false); + } + } + } + + cdata += dst->nb[2]; + cdataf += dst->nb[2]; + splane += pa; + } +} + +// ggml_compute_forward_upscale + +static void ggml_compute_forward_upscale_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_UNARY_OP_LOCALS + + const float sf0 = (float)ne0/src0->ne[0]; + const float sf1 = (float)ne1/src0->ne[1]; + const float sf2 = (float)ne2/src0->ne[2]; + const float sf3 = (float)ne3/src0->ne[3]; + + // TODO: optimize + + for (int64_t i3 = 0; i3 < ne3; i3++) { + const int64_t i03 = i3 / sf3; + for (int64_t i2 = ith; i2 < ne2; i2 += nth) { + const int64_t i02 = i2 / sf2; + for (int64_t i1 = 0; i1 < ne1; i1++) { + const int64_t i01 = i1 / sf1; + for (int64_t i0 = 0; i0 < ne0; i0++) { + const int64_t i00 = i0 / sf0; + + const float * x = (float *)((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + float * y = (float *)((char *) dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3); + + *y = *x; + } + } + } + } +} + +static void ggml_compute_forward_upscale( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_upscale_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + + +// ggml_compute_forward_pad + +static void ggml_compute_forward_pad_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + GGML_ASSERT( dst->nb[0] == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_UNARY_OP_LOCALS + + float * dst_ptr = (float *) dst->data; + + // TODO: optimize + + for (int64_t i2 = 0; i2 < ne2; ++i2) { + for (int64_t i1 = ith; i1 < ne1; i1 += nth) { + for (int64_t i0 = 0; i0 < ne0; ++i0) { + for (int64_t i3 = 0; i3 < ne3; ++i3) { + const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0; + + const float * src_ptr = (const float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + + if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) { + dst_ptr[dst_idx] = *src_ptr; + } else { + dst_ptr[dst_idx] = 0; + } + } + } + } + } +} + +static void ggml_compute_forward_pad( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_pad_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + + +// ggml_compute_forward_arange + +static void ggml_compute_forward_arange_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + GGML_ASSERT(dst->nb[0] == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + const float start = ggml_get_op_params_f32(dst, 0); + const float stop = ggml_get_op_params_f32(dst, 1); + const float step = ggml_get_op_params_f32(dst, 2); + + const int64_t steps = (int64_t) ceilf((stop - start) / step); + + GGML_ASSERT(ggml_nelements(dst) == steps); + + for (int64_t i = ith; i < steps; i+= nth) { + float value = start + step * i; + ((float *)dst->data)[i] = value; + } +} + +static void ggml_compute_forward_arange( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + switch (dst->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_arange_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +static void ggml_compute_forward_timestep_embedding_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_UNARY_OP_LOCALS + + const int dim = ggml_get_op_params_i32(dst, 0); + const int max_period = ggml_get_op_params_i32(dst, 1); + + int half = dim / 2; + + for (int64_t i = 0; i < ne00; i++) { + float * embed_data = (float *)((char *) dst->data + i*nb1); + for (int64_t j = ith; j < half; j += nth) { + float timestep = ((float *)src0->data)[i]; + float freq = (float)expf(-logf(max_period) * j / half); + float arg = timestep * freq; + embed_data[j] = cosf(arg); + embed_data[j + half] = sinf(arg); + } + if (dim % 2 != 0 && ith == 0) { + embed_data[dim] = 0.f; + } + } +} + +static void ggml_compute_forward_timestep_embedding( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_timestep_embedding_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_argsort + +static void ggml_compute_forward_argsort_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + GGML_TENSOR_UNARY_OP_LOCALS + + GGML_ASSERT(nb0 == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + const int64_t nr = ggml_nrows(src0); + + enum ggml_sort_order order = (enum ggml_sort_order) ggml_get_op_params_i32(dst, 0); + + for (int64_t i = ith; i < nr; i += nth) { + int32_t * dst_data = (int32_t *)((char *) dst->data + i*nb1); + const float * src_data = (float *)((char *) src0->data + i*nb01); + + for (int64_t j = 0; j < ne0; j++) { + dst_data[j] = j; + } + + // C doesn't have a functional sort, so we do a bubble sort instead + for (int64_t j = 0; j < ne0; j++) { + for (int64_t k = j + 1; k < ne0; k++) { + if ((order == GGML_SORT_ORDER_ASC && src_data[dst_data[j]] > src_data[dst_data[k]]) || + (order == GGML_SORT_ORDER_DESC && src_data[dst_data[j]] < src_data[dst_data[k]])) { + int32_t tmp = dst_data[j]; + dst_data[j] = dst_data[k]; + dst_data[k] = tmp; + } + } + } + } +} + +static void ggml_compute_forward_argsort( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_argsort_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_flash_attn_ext + +static void ggml_compute_forward_flash_attn_ext_f16( + const struct ggml_compute_params * params, + const struct ggml_tensor * q, + const struct ggml_tensor * k, + const struct ggml_tensor * v, + const struct ggml_tensor * mask, + struct ggml_tensor * dst) { + + GGML_TENSOR_LOCALS(int64_t, neq, q, ne) + GGML_TENSOR_LOCALS(size_t, nbq, q, nb) + GGML_TENSOR_LOCALS(int64_t, nek, k, ne) + GGML_TENSOR_LOCALS(size_t, nbk, k, nb) + GGML_TENSOR_LOCALS(int64_t, nev, v, ne) + GGML_TENSOR_LOCALS(size_t, nbv, v, nb) + GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) + GGML_TENSOR_LOCALS(size_t, nb, dst, nb) + + const int ith = params->ith; + const int nth = params->nth; + + const int64_t D = neq0; + const int64_t N = neq1; + + GGML_ASSERT(ne0 == D); + GGML_ASSERT(ne2 == N); + + // input tensor rows must be contiguous + GGML_ASSERT(nbq0 == ggml_type_size(q->type)); + GGML_ASSERT(nbk0 == ggml_type_size(k->type)); + GGML_ASSERT(nbv0 == ggml_type_size(v->type)); + + GGML_ASSERT(neq0 == D); + GGML_ASSERT(nek0 == D); + GGML_ASSERT(nev0 == D); + + GGML_ASSERT(neq1 == N); + GGML_ASSERT(nev0 == D); + + // dst cannot be transposed or permuted + GGML_ASSERT(nb0 == sizeof(float)); + GGML_ASSERT(nb0 <= nb1); + GGML_ASSERT(nb1 <= nb2); + GGML_ASSERT(nb2 <= nb3); + + // broadcast factors + const int64_t rk2 = neq2/nek2; + const int64_t rk3 = neq3/nek3; + + const int64_t rv2 = neq2/nev2; + const int64_t rv3 = neq3/nev3; + + // parallelize by q rows using ggml_vec_dot_f32 + + // total rows in q + const int nr = neq1*neq2*neq3; + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + float scale = 1.0f; + float max_bias = 0.0f; + float logit_softcap = 0.0f; + + memcpy(&scale, (float *) dst->op_params + 0, sizeof(float)); + memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float)); + memcpy(&logit_softcap, (float *) dst->op_params + 2, sizeof(float)); + + if (logit_softcap != 0) { + scale /= logit_softcap; + } + + const uint32_t n_head = neq2; + const uint32_t n_head_log2 = 1u << (uint32_t) floor(log2(n_head)); + + const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); + const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); + + enum ggml_type const k_vec_dot_type = type_traits[k->type].vec_dot_type; + ggml_from_float_t const q_to_vec_dot = type_traits[k_vec_dot_type].from_float; + ggml_vec_dot_t const kq_vec_dot = type_traits[k->type].vec_dot; + ggml_to_float_t const v_to_float = type_traits[v->type].to_float; + + // loop over n_batch and n_head + for (int ir = ir0; ir < ir1; ++ir) { + // q indices + const int iq3 = ir/(neq2*neq1); + const int iq2 = (ir - iq3*neq2*neq1)/neq1; + const int iq1 = (ir - iq3*neq2*neq1 - iq2*neq1); + + const uint32_t h = iq2; // head index + const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f; + + float S = 0.0f; // sum + float M = -INFINITY; // maximum KQ value + + float * VKQ32 = (float *) params->wdata + ith*(3*D + CACHE_LINE_SIZE_F32); // FP32 VKQ accumulator + float * V32 = (VKQ32 + 1*D); // (temporary) FP32 V buffer + ggml_fp16_t * VKQ16 = (ggml_fp16_t *) (VKQ32 + 1*D); // (temporary) FP16 VKQ accumulator + ggml_fp16_t * Q_q = (ggml_fp16_t *) (VKQ32 + 2*D); // (temporary) buffer for Q converted to quantized/FP16 + + if (v->type == GGML_TYPE_F16) { + memset(VKQ16, 0, D*sizeof(ggml_fp16_t)); + } else { + memset(VKQ32, 0, D*sizeof(float)); + } + + const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1]) : NULL; + + // k indices + const int ik3 = iq3 / rk3; + const int ik2 = iq2 / rk2; + + // v indices + const int iv3 = iq3 / rv3; + const int iv2 = iq2 / rv2; + + const float * pq = (const float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)); + q_to_vec_dot(pq, Q_q, D); + + // online softmax / attention + // loop over n_kv and n_head_kv + // ref: https://arxiv.org/pdf/2112.05682.pdf + for (int64_t ic = 0; ic < nek1; ++ic) { + const float mv = mp ? slope*GGML_FP16_TO_FP32(mp[ic]) : 0.0f; + if (mv == -INFINITY) { + continue; + } + + float s; // KQ value + + const char * k_data = (const char *) k->data + ( ic*nbk1 + ik2*nbk2 + ik3*nbk3); + kq_vec_dot(D, &s, 0, k_data, 0, Q_q, 0, 1); + + s = s*scale; // scale KQ value + + if (logit_softcap != 0.0f) { + s = logit_softcap*tanhf(s); + } + + s += mv; // apply mask + + const float Mold = M; + + float ms = 1.0f; // upon new higher max val, scale VKQ and KQ sum with this value + float vs = 1.0f; // post-softmax KQ value, expf(s - M) + + const char * v_data = ((const char *) v->data + (ic*nbv1 + iv2*nbv2 + iv3*nbv3)); + + if (v->type == GGML_TYPE_F16) { + if (s > M) { + // s is new maximum, ms < 1.0f, vs == expf(s - s) == 1.0f + M = s; + ms = expf(Mold - M); + + // V = V*expf(Mold - M) + ggml_vec_scale_f16(D, VKQ16, ms); + } else { + // no new maximum, ms == 1.0f, vs != 1.0f + vs = expf(s - M); + } + + // V += v*expf(s - M) + ggml_vec_mad_f16(D, VKQ16, (const ggml_fp16_t *) v_data, vs); + } else { + if (s > M) { + // s is new maximum, ms < 1.0f, vs == expf(s - s) == 1.0f + M = s; + ms = expf(Mold - M); + + // V = V*expf(Mold - M) + ggml_vec_scale_f32(D, VKQ32, ms); + } else { + // no new maximum, ms == 1.0f, vs != 1.0f + vs = expf(s - M); + } + + v_to_float(v_data, V32, D); + + // V += v*expf(s - M) + ggml_vec_mad_f32(D, VKQ32, V32, vs); + } + + S = S*ms + vs; // scale and increment sum with partial sum + } + + if (v->type == GGML_TYPE_F16) { + for (int64_t d = 0; d < D; ++d) { + VKQ32[d] = GGML_FP16_TO_FP32(VKQ16[d]); + } + } + + // V /= S + const float S_inv = 1.0f/S; + ggml_vec_scale_f32(D, VKQ32, S_inv); + + // dst indices + const int i1 = iq1; + const int i2 = iq2; + const int i3 = iq3; + + // original + //memcpy((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3), V, nev0*sizeof(float)); + + // permute(0, 2, 1, 3) + memcpy((char *) dst->data + (i3*ne2*ne1 + i2 + i1*ne1)*nb1, VKQ32, nb1); + } +} + +static void ggml_compute_forward_flash_attn_ext( + const struct ggml_compute_params * params, + const struct ggml_tensor * q, + const struct ggml_tensor * k, + const struct ggml_tensor * v, + const struct ggml_tensor * mask, + struct ggml_tensor * dst) { + switch (dst->op_params[3]) { + case GGML_PREC_DEFAULT: + case GGML_PREC_F32: + { + // uses F32 accumulators + ggml_compute_forward_flash_attn_ext_f16(params, q, k, v, mask, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_flash_attn_back + +static void ggml_compute_forward_flash_attn_back_f32( + const struct ggml_compute_params * params, + const bool masked, + struct ggml_tensor * dst) { + + const struct ggml_tensor * q = dst->src[0]; + const struct ggml_tensor * k = dst->src[1]; + const struct ggml_tensor * v = dst->src[2]; + const struct ggml_tensor * d = dst->src[3]; + + GGML_TENSOR_LOCALS(int64_t, neq, q, ne) + GGML_TENSOR_LOCALS(size_t, nbq, q, nb) + GGML_TENSOR_LOCALS(int64_t, nek, k, ne) + GGML_TENSOR_LOCALS(size_t, nbk, k, nb) + GGML_TENSOR_LOCALS(int64_t, nev, v, ne) + GGML_TENSOR_LOCALS(size_t, nbv, v, nb) + GGML_TENSOR_LOCALS(int64_t, ned, d, ne) + GGML_TENSOR_LOCALS(size_t, nbd, d, nb) + GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) + GGML_TENSOR_LOCALS(size_t, nb, dst, nb) + + const int ith = params->ith; + const int nth = params->nth; + + const int64_t D = neq0; + const int64_t N = neq1; + const int64_t P = nek1 - N; + const int64_t M = P + N; + + const int Mup = ggml_up(M, GGML_SOFT_MAX_UNROLL); + const int mxDM = MAX(D, Mup); + + // GGML_ASSERT(ne0 == D); + // GGML_ASSERT(ne1 == N); + GGML_ASSERT(P >= 0); + + GGML_ASSERT(nbq0 == sizeof(float)); + GGML_ASSERT(nbk0 == sizeof(float)); + GGML_ASSERT(nbv0 == sizeof(float)); + + GGML_ASSERT(neq0 == D); + GGML_ASSERT(nek0 == D); + GGML_ASSERT(nev1 == D); + GGML_ASSERT(ned0 == D); + + GGML_ASSERT(neq1 == N); + GGML_ASSERT(nek1 == N + P); + GGML_ASSERT(nev1 == D); + GGML_ASSERT(ned1 == N); + + // dst cannot be transposed or permuted + GGML_ASSERT(nb0 == sizeof(float)); + GGML_ASSERT(nb0 <= nb1); + GGML_ASSERT(nb1 <= nb2); + GGML_ASSERT(nb2 <= nb3); + + if (ith == 0) { + memset(dst->data, 0, nb0*ne0*ne1*ne2*ne3); + } + ggml_barrier(params->threadpool); + + const int64_t elem_q = ggml_nelements(q); + const int64_t elem_k = ggml_nelements(k); + + enum ggml_type result_type = dst->type; + GGML_ASSERT(ggml_blck_size(result_type) == 1); + const size_t tsize = ggml_type_size(result_type); + + const size_t offs_q = 0; + const size_t offs_k = offs_q + GGML_PAD(elem_q * tsize, GGML_MEM_ALIGN); + const size_t offs_v = offs_k + GGML_PAD(elem_k * tsize, GGML_MEM_ALIGN); + + void * grad_q = (char *) dst->data; + void * grad_k = (char *) dst->data + offs_k; + void * grad_v = (char *) dst->data + offs_v; + + const size_t nbgq1 = nb0*neq0; + const size_t nbgq2 = nb0*neq0*neq1; + const size_t nbgq3 = nb0*neq0*neq1*neq2; + + const size_t nbgk1 = nb0*nek0; + const size_t nbgk2 = nb0*nek0*nek1; + const size_t nbgk3 = nb0*nek0*nek1*neq2; + + const size_t nbgv1 = nb0*nev0; + const size_t nbgv2 = nb0*nev0*nev1; + const size_t nbgv3 = nb0*nev0*nev1*neq2; + + // parallelize by k rows using ggml_vec_dot_f32 + + // total rows in k + const int nr = nek2*nek3; + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + const float scale = 1.0f/sqrtf(D); + + //printf("P=%d N=%d D=%d ir0=%d ir1=%d scale = %f\n", P, N, D, ir0, ir1, scale); + + // how often k2 (and v2) is repeated in q2 + int nrep = neq2/nek2; + + for (int ir = ir0; ir < ir1; ++ir) { + // q indices + const int ik3 = ir/(nek2); + const int ik2 = ir - ik3*nek2; + + const int iq3 = ik3; + const int id3 = ik3; + const int iv3 = ik3; + const int iv2 = ik2; + + for (int irep = 0; irep < nrep; ++irep) { + const int iq2 = ik2 + irep*nek2; + const int id2 = iq2; + + // (ik2 + irep*nek2) % nek2 == ik2 + for (int iq1 = 0; iq1 < neq1; ++iq1) { + const int id1 = iq1; + + // not sure about CACHE_LINE_SIZE_F32.. + // - maybe it must not be multiplied by 2 and excluded from .. in SM 1*(..) offset? + float * S = (float *) params->wdata + ith*2*(mxDM + CACHE_LINE_SIZE_F32) + 0*(mxDM+CACHE_LINE_SIZE_F32); + float * SM = (float *) params->wdata + ith*2*(mxDM + CACHE_LINE_SIZE_F32) + 1*(mxDM+CACHE_LINE_SIZE_F32); + + for (int i = M; i < Mup; ++i) { + S[i] = -INFINITY; + } + + const int64_t masked_begin = masked ? (P + iq1 + 1) : M; + for (int64_t ic = 0; ic < masked_begin; ++ic) { + // k indices + const int ik1 = ic; + + // S indices + const int i1 = ik1; + + ggml_vec_dot_f32(neq0, + S + i1, 0, + (float *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)), 0, + (float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)), 0, 1); + } + + // scale + ggml_vec_scale_f32(masked_begin, S, scale); + + for (int64_t i = masked_begin; i < M; i++) { + S[i] = -INFINITY; + } + + // softmax + // exclude known -INF S[..] values from max and loop + // dont forget to set their SM values to zero + { + float max = -INFINITY; + ggml_vec_max_f32(masked_begin, &max, S); + + ggml_float sum = 0.0; + { +#ifdef GGML_SOFT_MAX_ACCELERATE + max = -max; + vDSP_vsadd(SM, 1, &max, SM, 1, Mup); + vvexpf(SM, SM, &Mup); + ggml_vec_sum_f32(Mup, &sum, SM); +#else + sum = ggml_vec_soft_max_f32(Mup, SM, S, max); +#endif + } + + assert(sum > 0.0); + + sum = 1.0/sum; + ggml_vec_scale_f32(masked_begin, SM, sum); + + } + + // step-by-step explanation + { + // forward-process shape grads from backward process + // parallel_for ik2,ik3: + // for irep: + // iq2 = ik2 + irep*nek2 + // k[:D,:M,:,:] [D,M,:,:] grad[k][:D,:M,ik2,ik3] += grad[kcur] + // q[:D,:N,:,:] [D,N,:,:] grad[q][:D,iq1,iq2,iq3] += grad[qcur] + // v[:M,:D,:,:] [M,D,:,:] grad[v][:M,:D,iv2,iv3] += grad[vcur] + // for iq1: + // kcur = k[:D,:M,ik2,ik3] [D,M,1,1] grad[kcur] = grad[S1].T @ qcur + // qcur = q[:D,iq1,iq2,iq3] [D,1,1,1] grad[qcur] = grad[S1] @ kcur + // vcur = v[:M,:D,iv2,iv3] [M,D,1,1] grad[vcur] = grad[S5].T @ S4 + // S0 = -Inf [D,1,1,1] + // ~S1[i] = dot(kcur[:D,i], qcur) + // S1 = qcur @ kcur.T [M,1,1,1] grad[S1] = grad[S2] * scale + // S2 = S1 * scale [M,1,1,1] grad[S2] = diag_mask_zero(grad[S3], P) + // S3 = diag_mask_inf(S2, P) [M,1,1,1] grad[S3] = S4 * (grad[S4] - dot(S4, grad[S4])) + // S4 = softmax(S3) [M,1,1,1] grad[S4] = grad[S5] @ vcur + // ~S5[i] = dot(vcur[:,i], S4) + // S5 = S4 @ vcur.T [D,1,1,1] grad[S5] = d[:D,id1,id2,id3] + // ~dst[i,iq1,iq2,iq3] = S5[i] ^ + // dst[:D,iq1,iq2,iq3] = S5 | grad[dst[:D,iq1,iq2,iq3]] = d[:D,id1,id2,id3] + // dst backward-/ grad[dst] = d + // + // output gradients with their dependencies: + // + // grad[kcur] = grad[S1].T @ qcur + // grad[S1] = diag_mask_zero(grad[S3], P) * scale + // grad[S3] = S4 * (grad[S4] - dot(S4, grad[S4])) + // grad[S4] = grad[S5] @ vcur + // grad[S4] = d[:D,id1,id2,id3] @ vcur + // grad[qcur] = grad[S1] @ kcur + // grad[vcur] = grad[S5].T @ S4 + // grad[vcur] = d[:D,id1,id2,id3].T @ S4 + // + // in post-order: + // + // S1 = qcur @ kcur.T + // S2 = S1 * scale + // S3 = diag_mask_inf(S2, P) + // S4 = softmax(S3) + // grad[S4] = d[:D,id1,id2,id3] @ vcur + // grad[S3] = S4 * (grad[S4] - dot(S4, grad[S4])) + // grad[S1] = diag_mask_zero(grad[S3], P) * scale + // grad[qcur] = grad[S1] @ kcur + // grad[kcur] = grad[S1].T @ qcur + // grad[vcur] = d[:D,id1,id2,id3].T @ S4 + // + // using less variables (SM=S4): + // + // S = diag_mask_inf(qcur @ kcur.T * scale, P) + // SM = softmax(S) + // S = d[:D,iq1,iq2,iq3] @ vcur + // dot_SM_gradSM = dot(SM, S) + // S = SM * (S - dot(SM, S)) + // S = diag_mask_zero(S, P) * scale + // + // grad[q][:D,iq1,iq2,iq3] += S @ kcur + // grad[k][:D,:M,ik2,ik3] += S.T @ qcur + // grad[v][:M,:D,iv2,iv3] += d[:D,id1,id2,id3].T @ SM + } + + // S = gradSM = d[:D,id1,id2,id3] @ vcur[:,:,iv2,iv3] + // S = d[:D,id1,id2,id3] @ vcur[:,:,iv2,iv3] + // for ic: + // S[:M] += vcur[:M,ic,iv2,iv3] * d[ic,id1,id2,id3] + // exclude known future zero S[..] values from operation + ggml_vec_set_f32(masked_begin, S, 0); + for (int64_t ic = 0; ic < D; ++ic) { + ggml_vec_mad_f32(masked_begin, + S, + (float *) ((char *) v->data + ( ic*nbv1 + iv2*nbv2 + iv3*nbv3)), + *(float *) ((char *) d->data + (ic*nbd0 + id1*nbd1 + id2*nbd2 + id3*nbd3))); + } + + // S = SM * (S - dot(SM, S)) + float dot_SM_gradSM = 0; + ggml_vec_dot_f32 (masked_begin, &dot_SM_gradSM, 0, SM, 0, S, 0, 1); + ggml_vec_acc1_f32(M, S, -dot_SM_gradSM); + ggml_vec_mul_f32 (masked_begin, S, S, SM); + + // S = diag_mask_zero(S, P) * scale + // already done by above ggml_vec_set_f32 + + // exclude known zero S[..] values from operation + ggml_vec_scale_f32(masked_begin, S, scale); + + // S shape [M,1] + // SM shape [M,1] + // kcur shape [D,M] + // qcur shape [D,1] + // vcur shape [M,D] + + // grad[q][:D,iq1,iq2,iq3] += S @ kcur + // grad[q][:D,iq1,iq2,iq3] += shape[M,1] @ shape[D,M] + // for ic: + // grad[q][:D,iq1,iq2,iq3] += S[ic] * kcur[:D,ic,ik2,ik3] + // exclude known zero S[..] values from loop + for (int64_t ic = 0; ic < masked_begin; ++ic) { + ggml_vec_mad_f32(D, + (float *) ((char *) grad_q + (iq1*nbgq1 + iq2*nbgq2 + iq3*nbgq3)), + (float *) ((char *) k->data + (ic*nbk1 + ik2*nbk2 + ik3*nbk3)), + S[ic]); + } + + // grad[k][:D,:M,iq2,iq3] += S.T @ qcur + // for ic: + // grad[k][:D,ic,iq2,iq3] += S.T[0,ic] * qcur[:D,0] + // grad[k][:D,ic,iq2,iq3] += S[ic] * qcur[:D,0] + // exclude known zero S[..] values from loop + for (int64_t ic = 0; ic < masked_begin; ++ic) { + ggml_vec_mad_f32(D, + (float *) ((char *) grad_k + (ic*nbgk1 + ik2*nbgk2 + ik3*nbgk3)), + (float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)), + S[ic]); + } + + // grad[v][:M,:D,iv2,iv3] += d[:D,id1,id2,id3].T @ SM + // for ic: + // grad[v][:M,ic,iv2,iv3] += d[:D,id1,id2,id3].T[0,ic] * SM[:M] + // grad[v][:M,ic,iv2,iv3] += d[ic,id1,id2,id3] * SM[:M] + // exclude known zero SM[..] values from mad + for (int64_t ic = 0; ic < D; ++ic) { + ggml_vec_mad_f32(masked_begin, + (float *) ((char *) grad_v + ( ic*nbgv1 + iv2*nbgv2 + iv3*nbgv3)), + SM, + *(float *) ((char *) d->data + (ic*nbd0 + id1*nbd1 + id2*nbd2 + id3*nbd3))); + } + } + } + } +} + +static void ggml_compute_forward_flash_attn_back( + const struct ggml_compute_params * params, + const bool masked, + struct ggml_tensor * dst) { + + const struct ggml_tensor * q = dst->src[0]; + + switch (q->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_flash_attn_back_f32(params, masked, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_ssm_conv + +static void ggml_compute_forward_ssm_conv_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + const struct ggml_tensor * src0 = dst->src[0]; // conv_x + const struct ggml_tensor * src1 = dst->src[1]; // conv1d.weight + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src1->ne[0]; // d_conv + const int ncs = src0->ne[0]; // d_conv - 1 + n_t + const int nr = src0->ne[1]; // d_inner + const int n_t = dst->ne[1]; // tokens per sequence + const int n_s = dst->ne[2]; // number of sequences in the batch + + GGML_ASSERT( dst->ne[0] == nr); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + GGML_ASSERT(src1->nb[0] == sizeof(float)); + GGML_ASSERT(src0->nb[1] == src0->ne[0]*sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + const int ir = ir1 - ir0; + + for (int i3 = 0; i3 < n_s; ++i3) { + for (int i2 = 0; i2 < n_t; ++i2) { + // {d_conv - 1 + n_t, d_inner, n_seqs} + // sliding window + const float * s = (const float *) ((const char *) src0->data + ir0*(src0->nb[1]) + i2*(src0->nb[0]) + i3*(src0->nb[2])); // {d_conv, d_inner, n_s} + const float * c = (const float *) ((const char *) src1->data + ir0*(src1->nb[1])); // {d_conv, d_inner} + float * x = (float *) ((char *) dst->data + ir0*(dst->nb[0]) + i2*(dst->nb[1]) + i3*(dst->nb[2])); // {d_inner, n_t, n_s} + + // TODO: transpose the output for smaller strides for big batches? + // d_inner + for (int i1 = 0; i1 < ir; ++i1) { + // rowwise dot product + // NOTE: not using ggml_vec_dot_f32, because its sum is in double precision + float sumf = 0.0f; + + // d_conv + for (int i0 = 0; i0 < nc; ++i0) { + sumf += s[i0 + i1*ncs] * c[i0 + i1*nc]; + } + x[i1] = sumf; + } + } + } +} + +static void ggml_compute_forward_ssm_conv( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + switch (dst->src[0]->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_ssm_conv_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_ssm_scan + +static void ggml_compute_forward_ssm_scan_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + const struct ggml_tensor * src0 = dst->src[0]; // s + const struct ggml_tensor * src1 = dst->src[1]; // x + const struct ggml_tensor * src2 = dst->src[2]; // dt + const struct ggml_tensor * src3 = dst->src[3]; // A + const struct ggml_tensor * src4 = dst->src[4]; // B + const struct ggml_tensor * src5 = dst->src[5]; // C + + const int ith = params->ith; + const int nth = params->nth; + + const int64_t nc = src0->ne[0]; // d_state + const int64_t nr = src0->ne[1]; // d_inner + const int64_t n_t = src1->ne[1]; // number of tokens per sequence + const int64_t n_s = src0->ne[2]; // number of sequences in the batch + + GGML_ASSERT(ggml_nelements(src1) + ggml_nelements(src0) == ggml_nelements(dst)); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + GGML_ASSERT(src1->nb[0] == sizeof(float)); + GGML_ASSERT(src2->nb[0] == sizeof(float)); + GGML_ASSERT(src3->nb[0] == sizeof(float)); + GGML_ASSERT(src4->nb[0] == sizeof(float)); + GGML_ASSERT(src5->nb[0] == sizeof(float)); + // required for the dot product between s and C + GGML_ASSERT(src0->nb[1] == src0->ne[0]*sizeof(float)); + // required for per-sequence offsets for states + GGML_ASSERT(src0->nb[2] == src0->ne[0]*src0->ne[1]*sizeof(float)); + // required to get correct offset for state destination (i.e. src1->nb[3]) + GGML_ASSERT(src1->nb[3] == src1->ne[0]*src1->ne[1]*src1->ne[2]*sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + const int ir = ir1 - ir0; + + for (int i3 = 0; i3 < n_s; ++i3) { + for (int i2 = 0; i2 < n_t; ++i2) { + const float * s0 = (const float *) ((const char *) src0->data + ir0*(src0->nb[1]) + i3*(src0->nb[2])); // {d_state, d_inner, n_s} + const float * x = (const float *) ((const char *) src1->data + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s} + const float * dt = (const float *) ((const char *) src2->data + ir0*(src2->nb[0]) + i2*(src2->nb[1]) + i3*(src2->nb[2])); // {d_inner, n_t, n_s} + const float * A = (const float *) ((const char *) src3->data + ir0*(src3->nb[1])); // {d_state, d_inner} + const float * B = (const float *) ((const char *) src4->data + i2*(src4->nb[1]) + i3*(src4->nb[2])); // {d_state, n_t, n_s} + const float * C = (const float *) ((const char *) src5->data + i2*(src5->nb[1]) + i3*(src5->nb[2])); // {d_state, n_t, n_s} + float * y = ( float *) (( char *) dst->data + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s} + float * s = ( float *) (( char *) dst->data + ir0*(src0->nb[1]) + i3*(src0->nb[2]) + src1->nb[3]); // {d_state, d_inner, n_s} + + // use the output as the source for the next token-wise iterations + if (i2 > 0) { s0 = s; } + + // d_inner + for (int i1 = 0; i1 < ir; ++i1) { + // ref: https://github.com/state-spaces/mamba/blob/34076d664838588a3c97727b263478ab9f621a07/mamba_ssm/ops/triton/selective_state_update.py#L78 + float dt_soft_plus = dt[i1] <= 20.0f ? log1pf(expf(dt[i1])) : dt[i1]; + float x_dt = x[i1] * dt_soft_plus; + float sumf = 0.0f; + // d_state + for (int i0 = 0; i0 < nc; ++i0) { + int i = i0 + i1*nc; + // state = prev_state * dA + dB * x + float state = (s0[i] * expf(dt_soft_plus * A[i])) + (B[i0] * x_dt); + // y = rowwise_dotprod(state, C) + sumf += state * C[i0]; + s[i] = state; + } + y[i1] = sumf; + } + } + } +} + +static void ggml_compute_forward_ssm_scan( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + switch (dst->src[0]->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_ssm_scan_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_win_part + +static void ggml_compute_forward_win_part_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + UNUSED(params); + + const struct ggml_tensor * src0 = dst->src[0]; + + GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) + GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) + + const int32_t nep0 = ((const int32_t *)(dst->op_params))[0]; + const int32_t nep1 = ((const int32_t *)(dst->op_params))[1]; + const int32_t w = ((const int32_t *)(dst->op_params))[2]; + + assert(ne00 == ne0); + assert(ne3 == nep0*nep1); + + // TODO: optimize / multi-thread + for (int py = 0; py < nep1; ++py) { + for (int px = 0; px < nep0; ++px) { + const int64_t i3 = py*nep0 + px; + for (int64_t i2 = 0; i2 < ne2; ++i2) { + for (int64_t i1 = 0; i1 < ne1; ++i1) { + for (int64_t i0 = 0; i0 < ne0; ++i0) { + const int64_t i02 = py*w + i2; + const int64_t i01 = px*w + i1; + const int64_t i00 = i0; + + const int64_t i = i3*ne2*ne1*ne0 + i2*ne1*ne0 + i1*ne0 + i0; + const int64_t j = i02*ne01*ne00 + i01*ne00 + i00; + + if (py*w + i2 >= ne02 || px*w + i1 >= ne01) { + ((float *) dst->data)[i] = 0.0f; + } else { + ((float *) dst->data)[i] = ((float *) src0->data)[j]; + } + } + } + } + } + } +} + +static void ggml_compute_forward_win_part( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_win_part_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_win_unpart + +static void ggml_compute_forward_win_unpart_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + UNUSED(params); + + const struct ggml_tensor * src0 = dst->src[0]; + + GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) + GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) + + const int32_t w = ((const int32_t *)(dst->op_params))[0]; + + // padding + const int px = (w - ne1%w)%w; + //const int py = (w - ne2%w)%w; + + const int npx = (px + ne1)/w; + //const int npy = (py + ne2)/w; + + assert(ne0 == ne00); + + // TODO: optimize / multi-thread + for (int64_t i2 = 0; i2 < ne2; ++i2) { + for (int64_t i1 = 0; i1 < ne1; ++i1) { + for (int64_t i0 = 0; i0 < ne0; ++i0) { + const int ip2 = i2/w; + const int ip1 = i1/w; + + const int64_t i02 = i2%w; + const int64_t i01 = i1%w; + const int64_t i00 = i0; + + const int64_t i = (ip2*npx + ip1)*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00 + i00; + const int64_t j = i2*ne1*ne0 + i1*ne0 + i0; + + ((float *) dst->data)[j] = ((float *) src0->data)[i]; + } + } + } +} + +static void ggml_compute_forward_win_unpart( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_win_unpart_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +//gmml_compute_forward_unary + +static void ggml_compute_forward_unary( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const enum ggml_unary_op op = ggml_get_unary_op(dst); + + switch (op) { + case GGML_UNARY_OP_ABS: + { + ggml_compute_forward_abs(params, dst); + } break; + case GGML_UNARY_OP_SGN: + { + ggml_compute_forward_sgn(params, dst); + } break; + case GGML_UNARY_OP_NEG: + { + ggml_compute_forward_neg(params, dst); + } break; + case GGML_UNARY_OP_STEP: + { + ggml_compute_forward_step(params, dst); + } break; + case GGML_UNARY_OP_TANH: + { + ggml_compute_forward_tanh(params, dst); + } break; + case GGML_UNARY_OP_ELU: + { + ggml_compute_forward_elu(params, dst); + } break; + case GGML_UNARY_OP_RELU: + { + ggml_compute_forward_relu(params, dst); + } break; + case GGML_UNARY_OP_SIGMOID: + { + ggml_compute_forward_sigmoid(params, dst); + } break; + case GGML_UNARY_OP_GELU: + { + ggml_compute_forward_gelu(params, dst); + } break; + case GGML_UNARY_OP_GELU_QUICK: + { + ggml_compute_forward_gelu_quick(params, dst); + } break; + case GGML_UNARY_OP_SILU: + { + ggml_compute_forward_silu(params, dst); + } break; + case GGML_UNARY_OP_HARDSWISH: + { + ggml_compute_forward_hardswish(params, dst); + } break; + case GGML_UNARY_OP_HARDSIGMOID: + { + ggml_compute_forward_hardsigmoid(params, dst); + } break; + case GGML_UNARY_OP_EXP: + { + ggml_compute_forward_exp(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_get_rel_pos + +static void ggml_compute_forward_get_rel_pos_f16( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + UNUSED(params); + + const struct ggml_tensor * src0 = dst->src[0]; + + // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/image_encoder.py#L292-L322 + + GGML_TENSOR_UNARY_OP_LOCALS + + const int64_t w = ne1; + + ggml_fp16_t * src0_data = (ggml_fp16_t *) src0->data; + ggml_fp16_t * dst_data = (ggml_fp16_t *) dst->data; + + for (int64_t i2 = 0; i2 < ne2; ++i2) { + for (int64_t i1 = 0; i1 < ne1; ++i1) { + const int64_t pos = (w - i1 - 1) + i2; + for (int64_t i0 = 0; i0 < ne0; ++i0) { + dst_data[i2*ne1*ne0 + i1*ne0 + i0] = src0_data[pos*ne00 + i0]; + } + } + } +} + +static void ggml_compute_forward_get_rel_pos( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F16: + case GGML_TYPE_BF16: + { + ggml_compute_forward_get_rel_pos_f16(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_add_rel_pos + +static void ggml_compute_forward_add_rel_pos_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + const struct ggml_tensor * src2 = dst->src[2]; + + const bool inplace = (bool) ((int32_t *) dst->op_params)[0]; + if (!inplace) { + if (params->ith == 0) { + memcpy((char *) dst->data, (char *) src0->data, ggml_nbytes(dst)); + } + ggml_barrier(params->threadpool); + } + // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/image_encoder.py#L357-L359 + + float * src1_data = (float *) src1->data; + float * src2_data = (float *) src2->data; + float * dst_data = (float *) dst->data; + + const int64_t ne10 = src1->ne[0]; + const int64_t ne11 = src1->ne[1]; + const int64_t ne12 = src1->ne[2]; + const int64_t ne13 = src1->ne[3]; + + const int ith = params->ith; + const int nth = params->nth; + + // total patches in dst + const int np = ne13; + + // patches per thread + const int dp = (np + nth - 1)/nth; + + // patch range for this thread + const int ip0 = dp*ith; + const int ip1 = MIN(ip0 + dp, np); + + for (int64_t i13 = ip0; i13 < ip1; ++i13) { + for (int64_t i12 = 0; i12 < ne12; ++i12) { + for (int64_t i11 = 0; i11 < ne11; ++i11) { + const int64_t jp1 = i13*ne12*ne11*ne10 + i12*ne11*ne10 + i11*ne10; + for (int64_t i10 = 0; i10 < ne10; ++i10) { + const int64_t jp0 = jp1 + i10; + const float src1_e = src1_data[jp0]; + const float src2_e = src2_data[jp0]; + + const int64_t jdh = jp0 * ne10; + const int64_t jdw = jdh - (ne10 - 1) * i10; + + for (int64_t j = 0; j < ne10; ++j) { + dst_data[jdh + j ] += src2_e; + dst_data[jdw + j*ne10] += src1_e; + } + } + } + } + } +} + +static void ggml_compute_forward_add_rel_pos( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_add_rel_pos_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_rwkv_wkv + +static void ggml_compute_forward_rwkv_wkv_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + const size_t T = dst->src[1]->ne[3]; + const size_t C = dst->ne[0]; + const size_t H = dst->src[1]->ne[2]; + const size_t n_seqs = dst->src[5]->ne[1]; + + float * dst_data = (float *) dst->data; + float * state = ((float *) dst->data) + C * T; + + if (params->ith != 0) { + return; + } + + memset(dst_data, 0, T * C * sizeof(float)); + + float * k = (float *) dst->src[0]->data; + float * v = (float *) dst->src[1]->data; + float * r = (float *) dst->src[2]->data; + float * time_faaaa = (float *) dst->src[3]->data; + float * time_decay = (float *) dst->src[4]->data; + + size_t t_stride = H * (C / H); + + size_t h_stride = C / H; + size_t h_stride_2d = (C / H) * (C / H); + + // basically fused operations: + // dst = r @ (time_faaaa * (k @ v) + state), + // state = time_decay * state + (k @ v), + // recursive through each token + for (size_t t = 0; t < T; t++) { + size_t t_offset = t * t_stride; + size_t state_offset = (C / H) * C * (t / (T / n_seqs)); + float * state_cur = state + state_offset; + float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[5]->data + state_offset; + + for (size_t h = 0; h < H; h++) { + size_t h_offset = h * h_stride; + size_t t_h_offset = t_offset + h_offset; + size_t h_2d_offset = h * h_stride_2d; + + for (size_t i = 0; i < C / H; i++) { + size_t t_h_i_offset = t_h_offset + i; + size_t h_i_offset = h_offset + i; + size_t h_2d_i_offset = h_2d_offset + i * h_stride; + + float k_val = k[t_h_i_offset]; + float r_val = r[t_h_i_offset]; + float time_faaaa_val = time_faaaa[h_i_offset]; + // RWKV v6: different time_decay for each token. + float time_decay_val = time_decay[t_h_i_offset]; + + for (size_t j = 0; j < C / H; j ++) { + size_t t_h_j_offset = t_h_offset + j; + size_t h_2d_i_j_offset = h_2d_i_offset + j; + + float v_val = v[t_h_j_offset]; + float kv_val = v_val * k_val; + float prev_state_val = state_prev[h_2d_i_j_offset]; + float temp_val = kv_val * time_faaaa_val + prev_state_val; + dst_data[t_h_j_offset] += temp_val * r_val; + state_cur[h_2d_i_j_offset] = prev_state_val * time_decay_val + kv_val; + } + } + } + } +} + +static void ggml_compute_forward_rwkv_wkv( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_rwkv_wkv_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_map_unary + +static void ggml_compute_forward_map_unary_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst, + const ggml_unary_op_f32_t fun) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + for (int i = 0; i < n; i++) { + fun(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_map_unary( + const struct ggml_compute_params * params, + struct ggml_tensor * dst, + const ggml_unary_op_f32_t fun) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_map_unary_f32(params, dst, fun); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_map_binary + +static void ggml_compute_forward_map_binary_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst, + const ggml_binary_op_f32_t fun) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(src1)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + for (int i = 0; i < n; i++) { + fun(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1])), + (float *) ((char *) src1->data + i*(src1->nb[1]))); + } +} + +static void ggml_compute_forward_map_binary( + const struct ggml_compute_params * params, + struct ggml_tensor * dst, + const ggml_binary_op_f32_t fun) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_map_binary_f32(params, dst, fun); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_map_custom1 + +static void ggml_compute_forward_map_custom1_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst, + const ggml_custom1_op_f32_t fun) { + + const struct ggml_tensor * a = dst->src[0]; + + if (params->ith != 0) { + return; + } + + fun(dst, a); +} + +// ggml_compute_forward_map_custom2 + +static void ggml_compute_forward_map_custom2_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst, + const ggml_custom2_op_f32_t fun) { + + const struct ggml_tensor * a = dst->src[0]; + const struct ggml_tensor * b = dst->src[1]; + + if (params->ith != 0) { + return; + } + + fun(dst, a, b); +} + +// ggml_compute_forward_map_custom3 + +static void ggml_compute_forward_map_custom3_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst, + const ggml_custom3_op_f32_t fun) { + + const struct ggml_tensor * a = dst->src[0]; + const struct ggml_tensor * b = dst->src[1]; + const struct ggml_tensor * c = dst->src[1]; + + if (params->ith != 0) { + return; + } + + fun(dst, a, b, c); +} + +// ggml_compute_forward_map_custom1 + +static void ggml_compute_forward_map_custom1( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * a = dst->src[0]; + + struct ggml_map_custom1_op_params p; + memcpy(&p, dst->op_params, sizeof(p)); + + p.fun(dst, a, params->ith, params->nth, p.userdata); +} + +// ggml_compute_forward_map_custom2 + +static void ggml_compute_forward_map_custom2( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * a = dst->src[0]; + const struct ggml_tensor * b = dst->src[1]; + + struct ggml_map_custom2_op_params p; + memcpy(&p, dst->op_params, sizeof(p)); + + p.fun(dst, a, b, params->ith, params->nth, p.userdata); +} + +// ggml_compute_forward_map_custom3 + +static void ggml_compute_forward_map_custom3( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * a = dst->src[0]; + const struct ggml_tensor * b = dst->src[1]; + const struct ggml_tensor * c = dst->src[2]; + + struct ggml_map_custom3_op_params p; + memcpy(&p, dst->op_params, sizeof(p)); + + p.fun(dst, a, b, c, params->ith, params->nth, p.userdata); +} + +// ggml_compute_forward_cross_entropy_loss + +static void ggml_compute_forward_cross_entropy_loss_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(ggml_is_contiguous(src1)); + GGML_ASSERT(ggml_is_scalar(dst)); + GGML_ASSERT(ggml_are_same_shape(src0, src1)); + + const int ith = params->ith; + const int nth = params->nth; + + float * sums = (float *) params->wdata; + + // TODO: handle transposed/permuted matrices + const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + + GGML_ASSERT(params->wsize >= sizeof(float) * (nth + nth * nc)); + + if (ith == 0) { + memset(sums, 0, sizeof(float) * (nth + nth * nc)); + } + ggml_barrier(params->threadpool); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int i1 = ir0; i1 < ir1; i1++) { + float * s0 = (float *)((char *) src0->data + i1*src0->nb[1]); + float * s1 = (float *)((char *) src1->data + i1*src1->nb[1]); + float * st = ((float *) params->wdata) + nth + ith*nc; + +#ifndef NDEBUG + for (int i = 0; i < nc; ++i) { + //printf("p[%d] = %f\n", i, p[i]); + assert(!isnan(s0[i])); + assert(!isnan(s1[i])); + } +#endif + + float max = -INFINITY; + ggml_vec_max_f32(nc, &max, s0); + ggml_float sum = ggml_vec_log_soft_max_f32(nc, st, s0, max); + assert(sum >= 0.0); + + ggml_vec_add1_f32(nc, st, st, -sum); + ggml_vec_mul_f32(nc, st, st, s1); + + float st_sum = 0.0f; + ggml_vec_sum_f32(nc, &st_sum, st); + sums[ith] += st_sum; + +#ifndef NDEBUG + for (int i = 0; i < nc; ++i) { + assert(!isnan(st[i])); + assert(!isinf(st[i])); + } +#endif + } + ggml_barrier(params->threadpool); + + if (ith == 0) { + float * dp = (float *) dst->data; + ggml_vec_sum_f32(nth, dp, sums); + dp[0] *= -1.0f / (float) nr; + } +} + +static void ggml_compute_forward_cross_entropy_loss( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_cross_entropy_loss_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_cross_entropy_loss_back + +static void ggml_compute_forward_cross_entropy_loss_back_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + const struct ggml_tensor * opt0 = dst->src[2]; + + GGML_ASSERT(ggml_is_contiguous(dst)); + GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(ggml_is_contiguous(src1)); + GGML_ASSERT(ggml_is_contiguous(opt0)); + GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); + + const int64_t ith = params->ith; + const int64_t nth = params->nth; + + // TODO: handle transposed/permuted matrices + const int64_t nc = src0->ne[0]; + const int64_t nr = ggml_nrows(src0); + + // rows per thread + const int64_t dr = (nr + nth - 1)/nth; + + // row range for this thread + const int64_t ir0 = dr*ith; + const int64_t ir1 = MIN(ir0 + dr, nr); + + const float d_by_nr = ((const float *) opt0->data)[0] / (float) nr; + + for (int64_t i1 = ir0; i1 < ir1; i1++) { + float * ds0 = (float *)((char *) dst->data + i1*dst->nb[1]); + float * s0 = (float *)((char *) src0->data + i1*src0->nb[1]); + float * s1 = (float *)((char *) src1->data + i1*src1->nb[1]); + +#ifndef NDEBUG + for (int i = 0; i < nc; ++i) { + //printf("p[%d] = %f\n", i, p[i]); + assert(!isnan(s0[i])); + assert(!isnan(s1[i])); + } +#endif + + // soft_max + float max = -INFINITY; + ggml_vec_max_f32(nc, &max, s0); + ggml_float sum = ggml_vec_soft_max_f32(nc, ds0, s0, max); + assert(sum > 0.0); + ggml_vec_scale_f32(nc, ds0, 1.0/sum); + + // grad(src0) = (softmax(src0) - src1) * grad(cross_entropy_loss(src0, src1)) / nr + ggml_vec_sub_f32(nc, ds0, ds0, s1); + ggml_vec_scale_f32(nc, ds0, d_by_nr); + +#ifndef NDEBUG + for (int i = 0; i < nc; ++i) { + assert(!isnan(ds0[i])); + assert(!isinf(ds0[i])); + } +#endif + } +} + +static void ggml_compute_forward_cross_entropy_loss_back( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_cross_entropy_loss_back_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +static void ggml_compute_forward_opt_step_adamw_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src0_grad = dst->src[1]; + const struct ggml_tensor * src0_grad_m = dst->src[2]; + const struct ggml_tensor * src0_grad_v = dst->src[3]; + GGML_ASSERT(ggml_are_same_shape(src0, src0_grad)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_UNARY_OP_LOCALS + GGML_ASSERT(nb00 == sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + /* const float gnorm = 1.0f; */ + int64_t iter; memcpy(&iter, &dst->op_params[0], sizeof(int64_t)); + const float alpha = ggml_get_op_params_f32(dst, 2); + const float beta1 = ggml_get_op_params_f32(dst, 3); + const float beta2 = ggml_get_op_params_f32(dst, 4); + const float eps = ggml_get_op_params_f32(dst, 5); + const float wd = ggml_get_op_params_f32(dst, 6); + + const float beta1h = alpha/(1.0f - powf(beta1, iter)); + const float beta2h = 1.0f/(1.0f - powf(beta2, iter)); + + for (int ir = ir0; ir < ir1; ++ir) { + const int64_t i03 = ir/(ne02*ne01); + const int64_t i02 = (ir - i03*ne02*ne01)/ne01; + const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); + + const size_t offset = i03*nb03 + i02*nb02 + i01*nb01; + + float * w = (float *) ((char *) src0->data + offset); // weight + const float * g = (const float *) ((const char *) src0_grad->data + offset); // grad + float * m = (float *) ((char *) src0_grad_m->data + offset); + float * v = (float *) ((char *) src0_grad_v->data + offset); + + for (int i00 = 0; i00 < ne00; ++i00) { + m[i00] = m[i00]*beta1 + g[i00]*(1.0f - beta1); + v[i00] = v[i00]*beta2 + g[i00]*g[i00]*(1.0f - beta2); + + const float mh = m[i00]*beta1h; + const float vh = sqrtf(v[i00]*beta2h) + eps; + + // The weight decay is applied independently of the Adam momenta m and v. + // This is NOT equivalent to l2 regularization that adds w[i00]*w[i00] to the loss. + // See: https://arxiv.org/pdf/1711.05101v3.pdf + w[i00] = w[i00]*(1.0f - alpha*wd) - mh/vh; + } + } + + ggml_barrier(params->threadpool); + if (ith != 0) { + return; + } + + iter++; + memcpy(&dst->op_params[0], &iter, sizeof(int64_t)); +} + +static void ggml_compute_forward_opt_step_adamw( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_opt_step_adamw_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} +///////////////////////////////// + +static void ggml_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) { + GGML_ASSERT(params); + + if (tensor->op == GGML_OP_NONE || ggml_is_empty(tensor)) { + return; + } + + switch (tensor->op) { + case GGML_OP_DUP: + { + ggml_compute_forward_dup(params, tensor); + } break; + case GGML_OP_ADD: + { + ggml_compute_forward_add(params, tensor); + } break; + case GGML_OP_ADD1: + { + ggml_compute_forward_add1(params, tensor); + } break; + case GGML_OP_ACC: + { + ggml_compute_forward_acc(params, tensor); + } break; + case GGML_OP_SUB: + { + ggml_compute_forward_sub(params, tensor); + } break; + case GGML_OP_MUL: + { + ggml_compute_forward_mul(params, tensor); + } break; + case GGML_OP_DIV: + { + ggml_compute_forward_div(params, tensor); + } break; + case GGML_OP_SQR: + { + ggml_compute_forward_sqr(params, tensor); + } break; + case GGML_OP_SQRT: + { + ggml_compute_forward_sqrt(params, tensor); + } break; + case GGML_OP_LOG: + { + ggml_compute_forward_log(params, tensor); + } break; + case GGML_OP_SIN: + { + ggml_compute_forward_sin(params, tensor); + } break; + case GGML_OP_COS: + { + ggml_compute_forward_cos(params, tensor); + } break; + case GGML_OP_SUM: + { + ggml_compute_forward_sum(params, tensor); + } break; + case GGML_OP_SUM_ROWS: + { + ggml_compute_forward_sum_rows(params, tensor); + } break; + case GGML_OP_MEAN: + { + ggml_compute_forward_mean(params, tensor); + } break; + case GGML_OP_ARGMAX: + { + ggml_compute_forward_argmax(params, tensor); + } break; + case GGML_OP_REPEAT: + { + ggml_compute_forward_repeat(params, tensor); + } break; + case GGML_OP_REPEAT_BACK: + { + ggml_compute_forward_repeat_back(params, tensor); + } break; + case GGML_OP_CONCAT: + { + ggml_compute_forward_concat(params, tensor); + } break; + case GGML_OP_SILU_BACK: + { + ggml_compute_forward_silu_back(params, tensor); + } break; + case GGML_OP_NORM: + { + ggml_compute_forward_norm(params, tensor); + } break; + case GGML_OP_RMS_NORM: + { + ggml_compute_forward_rms_norm(params, tensor); + } break; + case GGML_OP_RMS_NORM_BACK: + { + ggml_compute_forward_rms_norm_back(params, tensor); + } break; + case GGML_OP_GROUP_NORM: + { + ggml_compute_forward_group_norm(params, tensor); + } break; + case GGML_OP_MUL_MAT: + { + ggml_compute_forward_mul_mat(params, tensor); + } break; + case GGML_OP_MUL_MAT_ID: + { + ggml_compute_forward_mul_mat_id(params, tensor); + } break; + case GGML_OP_OUT_PROD: + { + ggml_compute_forward_out_prod(params, tensor); + } break; + case GGML_OP_SCALE: + { + ggml_compute_forward_scale(params, tensor); + } break; + case GGML_OP_SET: + { + ggml_compute_forward_set(params, tensor); + } break; + case GGML_OP_CPY: + { + ggml_compute_forward_cpy(params, tensor); + } break; + case GGML_OP_CONT: + { + ggml_compute_forward_cont(params, tensor); + } break; + case GGML_OP_RESHAPE: + { + ggml_compute_forward_reshape(params, tensor); + } break; + case GGML_OP_VIEW: + { + ggml_compute_forward_view(params, tensor); + } break; + case GGML_OP_PERMUTE: + { + ggml_compute_forward_permute(params, tensor); + } break; + case GGML_OP_TRANSPOSE: + { + ggml_compute_forward_transpose(params, tensor); + } break; + case GGML_OP_GET_ROWS: + { + ggml_compute_forward_get_rows(params, tensor); + } break; + case GGML_OP_GET_ROWS_BACK: + { + ggml_compute_forward_get_rows_back(params, tensor); + } break; + case GGML_OP_DIAG: + { + ggml_compute_forward_diag(params, tensor); + } break; + case GGML_OP_DIAG_MASK_INF: + { + ggml_compute_forward_diag_mask_inf(params, tensor); + } break; + case GGML_OP_DIAG_MASK_ZERO: + { + ggml_compute_forward_diag_mask_zero(params, tensor); + } break; + case GGML_OP_SOFT_MAX: + { + ggml_compute_forward_soft_max(params, tensor); + } break; + case GGML_OP_SOFT_MAX_BACK: + { + ggml_compute_forward_soft_max_back(params, tensor); + } break; + case GGML_OP_ROPE: + { + ggml_compute_forward_rope(params, tensor); + } break; + case GGML_OP_ROPE_BACK: + { + ggml_compute_forward_rope_back(params, tensor); + } break; + case GGML_OP_CLAMP: + { + ggml_compute_forward_clamp(params, tensor); + } break; + case GGML_OP_CONV_TRANSPOSE_1D: + { + ggml_compute_forward_conv_transpose_1d(params, tensor); + } break; + case GGML_OP_IM2COL: + { + ggml_compute_forward_im2col(params, tensor); + } break; + case GGML_OP_IM2COL_BACK: + { + ggml_compute_forward_im2col_back_f32(params, tensor); + } break; + case GGML_OP_CONV_TRANSPOSE_2D: + { + ggml_compute_forward_conv_transpose_2d(params, tensor); + } break; + case GGML_OP_POOL_1D: + { + ggml_compute_forward_pool_1d(params, tensor); + } break; + case GGML_OP_POOL_2D: + { + ggml_compute_forward_pool_2d(params, tensor); + } break; + case GGML_OP_POOL_2D_BACK: + { + ggml_compute_forward_pool_2d_back(params, tensor); + } break; + case GGML_OP_UPSCALE: + { + ggml_compute_forward_upscale(params, tensor); + } break; + case GGML_OP_PAD: + { + ggml_compute_forward_pad(params, tensor); + } break; + case GGML_OP_ARANGE: + { + ggml_compute_forward_arange(params, tensor); + } break; + case GGML_OP_TIMESTEP_EMBEDDING: + { + ggml_compute_forward_timestep_embedding(params, tensor); + } break; + case GGML_OP_ARGSORT: + { + ggml_compute_forward_argsort(params, tensor); + } break; + case GGML_OP_LEAKY_RELU: + { + ggml_compute_forward_leaky_relu(params, tensor); + } break; + case GGML_OP_FLASH_ATTN_EXT: + { + ggml_compute_forward_flash_attn_ext(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor->src[3], tensor); + } break; + case GGML_OP_FLASH_ATTN_BACK: + { + int32_t t = ggml_get_op_params_i32(tensor, 0); + GGML_ASSERT(t == 0 || t == 1); + bool masked = t != 0; + ggml_compute_forward_flash_attn_back(params, masked, tensor); + } break; + case GGML_OP_SSM_CONV: + { + ggml_compute_forward_ssm_conv(params, tensor); + } break; + case GGML_OP_SSM_SCAN: + { + ggml_compute_forward_ssm_scan(params, tensor); + } break; + case GGML_OP_WIN_PART: + { + ggml_compute_forward_win_part(params, tensor); + } break; + case GGML_OP_WIN_UNPART: + { + ggml_compute_forward_win_unpart(params, tensor); + } break; + case GGML_OP_UNARY: + { + ggml_compute_forward_unary(params, tensor); + } break; + case GGML_OP_GET_REL_POS: + { + ggml_compute_forward_get_rel_pos(params, tensor); + } break; + case GGML_OP_ADD_REL_POS: + { + ggml_compute_forward_add_rel_pos(params, tensor); + } break; + case GGML_OP_RWKV_WKV: + { + ggml_compute_forward_rwkv_wkv(params, tensor); + } break; + case GGML_OP_MAP_UNARY: + { + ggml_unary_op_f32_t fun; + memcpy(&fun, tensor->op_params, sizeof(fun)); + ggml_compute_forward_map_unary(params, tensor, fun); + } + break; + case GGML_OP_MAP_BINARY: + { + ggml_binary_op_f32_t fun; + memcpy(&fun, tensor->op_params, sizeof(fun)); + ggml_compute_forward_map_binary(params, tensor, fun); + } + break; + case GGML_OP_MAP_CUSTOM1_F32: + { + ggml_custom1_op_f32_t fun; + memcpy(&fun, tensor->op_params, sizeof(fun)); + ggml_compute_forward_map_custom1_f32(params, tensor, fun); + } + break; + case GGML_OP_MAP_CUSTOM2_F32: + { + ggml_custom2_op_f32_t fun; + memcpy(&fun, tensor->op_params, sizeof(fun)); + ggml_compute_forward_map_custom2_f32(params, tensor, fun); + } + break; + case GGML_OP_MAP_CUSTOM3_F32: + { + ggml_custom3_op_f32_t fun; + memcpy(&fun, tensor->op_params, sizeof(fun)); + ggml_compute_forward_map_custom3_f32(params, tensor, fun); + } + break; + case GGML_OP_MAP_CUSTOM1: + { + ggml_compute_forward_map_custom1(params, tensor); + } + break; + case GGML_OP_MAP_CUSTOM2: + { + ggml_compute_forward_map_custom2(params, tensor); + } + break; + case GGML_OP_MAP_CUSTOM3: + { + ggml_compute_forward_map_custom3(params, tensor); + } + break; + case GGML_OP_CROSS_ENTROPY_LOSS: + { + ggml_compute_forward_cross_entropy_loss(params, tensor); + } + break; + case GGML_OP_CROSS_ENTROPY_LOSS_BACK: + { + ggml_compute_forward_cross_entropy_loss_back(params, tensor); + } + break; + case GGML_OP_OPT_STEP_ADAMW: + { + ggml_compute_forward_opt_step_adamw(params, tensor); + } + break; + case GGML_OP_NONE: + { + // nop + } break; + case GGML_OP_COUNT: + { + GGML_ABORT("fatal error"); + } + } +} + +//////////////////////////////////////////////////////////////////////////////// + struct ggml_hash_set ggml_hash_set_new(size_t size) { size = ggml_hash_size(size); struct ggml_hash_set result; diff --git a/src/llama.cpp b/src/llama.cpp index 22b951ba2..995d6c8d6 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -163,6 +163,7 @@ enum llm_arch { LLM_ARCH_QWEN, LLM_ARCH_QWEN2, LLM_ARCH_QWEN2MOE, + LLM_ARCH_QWEN2VL, LLM_ARCH_PHI2, LLM_ARCH_PHI3, LLM_ARCH_PLAMO, @@ -217,6 +218,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_QWEN, "qwen" }, { LLM_ARCH_QWEN2, "qwen2" }, { LLM_ARCH_QWEN2MOE, "qwen2moe" }, + { LLM_ARCH_QWEN2VL, "qwen2vl" }, { LLM_ARCH_PHI2, "phi2" }, { LLM_ARCH_PHI3, "phi3" }, { LLM_ARCH_PLAMO, "plamo" }, @@ -898,6 +900,23 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_QWEN2VL, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, { LLM_ARCH_QWEN2MOE, { @@ -3329,6 +3348,8 @@ struct llama_context { struct ggml_tensor * inp_tokens; // I32 [n_batch] struct ggml_tensor * inp_embd; // F32 [n_embd, n_batch] struct ggml_tensor * inp_pos; // I32 [n_batch] + struct ggml_tensor * inp_pos_w; // I32 [n_batch] second-dimension of m-rope position index + struct ggml_tensor * inp_pos_h; // I32 [n_batch] third-dimension of m-rope position index struct ggml_tensor * inp_out_ids; // I32 [n_outputs] struct ggml_tensor * inp_KQ_mask; // F32 [kv_size, n_batch] struct ggml_tensor * inp_KQ_mask_swa; // F32 [kv_size, n_batch] @@ -5686,6 +5707,7 @@ static void llm_load_hparams( } } break; case LLM_ARCH_QWEN2: + case LLM_ARCH_QWEN2VL: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); switch (hparams.n_layer) { @@ -8096,6 +8118,7 @@ static bool llm_load_tensors( } } break; case LLM_ARCH_QWEN2: + case LLM_ARCH_QWEN2VL: { model.tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -12484,6 +12507,123 @@ struct llm_build_context { return gf; } + + struct ggml_cgraph * build_qwen2vl() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); + + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); + + // inp_pos - contains the positions + // struct ggml_tensor * inp_pos = build_inp_pos(); + lctx.inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens * 3); + cb(lctx.inp_pos, "inp_pos", -1); + ggml_set_input(lctx.inp_pos); + struct ggml_tensor * inp_pos = lctx.inp_pos; + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + + Qcur = ggml_mrope_ext( + ctx0, + ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_mrope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = llm_build_kv(ctx0, lctx, kv_self, gf, + model.layers[il].wo, model.layers[il].bo, + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, lctx, cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cur = lctx.cvec.apply_to(ctx0, cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, NULL, + LLM_NORM_RMS, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } struct ggml_cgraph * build_qwen2moe() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); @@ -16732,6 +16872,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_qwen2(); } break; + case LLM_ARCH_QWEN2VL: + { + result = llm.build_qwen2vl(); + } break; case LLM_ARCH_QWEN2MOE: { result = llm.build_qwen2moe(); @@ -20088,6 +20232,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_BITNET: case LLM_ARCH_QWEN: case LLM_ARCH_QWEN2: + case LLM_ARCH_QWEN2VL: case LLM_ARCH_QWEN2MOE: case LLM_ARCH_OLMO2: case LLM_ARCH_OLMOE: