diff --git a/common/common.cpp b/common/common.cpp index 91057041b..015764f5f 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -886,7 +886,8 @@ std::string fs_get_cache_directory() { if (getenv("LLAMA_CACHE")) { cache_directory = std::getenv("LLAMA_CACHE"); } else { -#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || defined(__OpenBSD__) +#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || \ + defined(__OpenBSD__) || defined(__NetBSD__) if (std::getenv("XDG_CACHE_HOME")) { cache_directory = std::getenv("XDG_CACHE_HOME"); } else if (std::getenv("HOME")) { @@ -1230,7 +1231,7 @@ common_init_result_ptr common_init_from_params(common_params & params) { return res; } - int err = llama_apply_adapter_cvec( + int err = llama_set_adapter_cvec( lctx, cvec.data.data(), cvec.data.size(), @@ -1332,12 +1333,15 @@ std::string get_model_endpoint() { } void common_set_adapter_lora(struct llama_context * ctx, std::vector & lora) { - llama_clear_adapter_lora(ctx); - for (auto & la : lora) { - if (la.scale != 0.0f) { - llama_set_adapter_lora(ctx, la.ptr, la.scale); - } + std::vector loras; + std::vector scales; + + for (auto & la: lora) { + loras.push_back(la.ptr); + scales.push_back(la.scale); } + + llama_set_adapters_lora(ctx, loras.data(), loras.size(), scales.data()); } struct llama_model_params common_model_params_to_llama(common_params & params) { diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 825080b58..7f341a58d 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -570,6 +570,7 @@ class ModelBase: self.match_model_tensor_name(new_name, key, bid) for key in ( gguf.MODEL_TENSOR.FFN_GATE_INP, + gguf.MODEL_TENSOR.FFN_GATE_INP_SHEXP, gguf.MODEL_TENSOR.POS_EMBD, gguf.MODEL_TENSOR.TOKEN_TYPES, gguf.MODEL_TENSOR.SSM_CONV1D, @@ -4073,6 +4074,87 @@ class InternVisionModel(MmprojModel): yield from super().modify_tensors(data_torch, name, bid) +@ModelBase.register( + "NemotronH_Nano_VL_V2", + "RADIOModel", +) +class NemotronNanoV2VLModel(MmprojModel): + # ViT-Huge architecture parameters for RADIO v2.5-h + _vit_hidden_size = 1280 + _vit_intermediate_size = 5120 + _vit_num_layers = 32 + _vit_num_heads = 16 + + def get_vision_config(self) -> dict[str, Any] | None: + # RADIO config doesn't have standard ViT parameters, so they need to be constructed manually + vision_config = self.global_config.get("vision_config") + if vision_config is None: + return None + # Add ViT-H parameters + vision_config = { + **vision_config, + "hidden_size": self._vit_hidden_size, + "intermediate_size": self._vit_intermediate_size, + "num_hidden_layers": self._vit_num_layers, + "num_attention_heads": self._vit_num_heads, + "image_size": self.global_config.get("force_image_size", 512), + } + return vision_config + + def set_gguf_parameters(self): + if "image_mean" not in self.preprocessor_config: + self.preprocessor_config["image_mean"] = [0.485, 0.456, 0.406] + if "image_std" not in self.preprocessor_config: + self.preprocessor_config["image_std"] = [0.229, 0.224, 0.225] + + super().set_gguf_parameters() + hparams = self.global_config + self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.NEMOTRON_V2_VL) + self.gguf_writer.add_vision_attention_layernorm_eps(1e-6) + self.gguf_writer.add_vision_use_gelu(True) + downsample_ratio = hparams.get("downsample_ratio", 0.5) + self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / downsample_ratio)) + + def tensor_force_quant(self, name, new_name, bid, n_dims): + if ".position_embd." in new_name or "pos_embed" in new_name: + return gguf.GGMLQuantizationType.F32 + return super().tensor_force_quant(name, new_name, bid, n_dims) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + if "input_conditioner" in name: + return + + # RADIO's pos_embed doesn't have .weight suffix, but clip.cpp expects it + if "patch_generator.pos_embed" in name: + if not name.endswith(".weight"): + name += ".weight" + # Downsample position embeddings for fixed 512x512 image size + import torch.nn.functional as F + n_embd = self.hparams["hidden_size"] + image_size = self.global_config.get("force_image_size", 512) + patch_size = self.hparams["patch_size"] + target_patches_per_side = image_size // patch_size # 32 + max_patches_per_side = int((data_torch.shape[1]) ** 0.5) # 128 + if target_patches_per_side != max_patches_per_side: + # Reshape to grid, interpolate, flatten back + data_torch = data_torch.reshape(1, max_patches_per_side, max_patches_per_side, n_embd) + data_torch = data_torch.permute(0, 3, 1, 2).float() # [1, n_embd, 128, 128] + data_torch = F.interpolate(data_torch, size=(target_patches_per_side, target_patches_per_side), + mode='bilinear', align_corners=True) + data_torch = data_torch.permute(0, 2, 3, 1) # [1, 32, 32, n_embd] + data_torch = data_torch.reshape(1, target_patches_per_side * target_patches_per_side, n_embd) + + # Reshape linear patch embedding to conv2d format for ggml_conv_2d + # From [n_embd, patch_size*patch_size*3] to [n_embd, 3, patch_size, patch_size] + if "patch_generator.embedder" in name: + patch_size = self.hparams["patch_size"] + n_embd = self.hparams["hidden_size"] + data_torch = data_torch.reshape(n_embd, 3, patch_size, patch_size) + + if name.startswith("vision_model.radio_model.model.") or name.startswith("mlp1."): + yield from super().modify_tensors(data_torch, name, bid) + + @ModelBase.register("WavTokenizerDec") class WavTokenizerDecModel(TextModel): model_arch = gguf.MODEL_ARCH.WAVTOKENIZER_DEC @@ -7054,6 +7136,8 @@ class Mamba2Model(TextModel): if hparams is None: with open(dir_model / "config.json", "r", encoding="utf-8") as f: hparams = json.load(f) + if "llm_config" in hparams: + hparams["text_config"] = hparams["llm_config"] super().__init__(dir_model, *args, hparams=hparams, **kwargs) self.d_model = self.find_hparam(["hidden_size", "d_model", "dim"]) self.d_inner = self.find_hparam(["mamba_d_ssm", "intermediate_size", "d_inner"], optional=True) or 2 * self.d_model @@ -9541,6 +9625,14 @@ class NemotronHModel(GraniteHybridModel): self.gguf_writer.add_add_bos_token(True) def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + # Skip vision model and projector tensors for VLM models (handled by mmproj) (e.g., Nemotron Nano 12B v2 VL) + if name.startswith(("vision_model.", "mlp1.")): + return + + # Strip language_model. prefix for VLM models (e.g., Nemotron Nano 12B v2 VL) + if name.startswith("language_model."): + name = name[len("language_model."):] + if self.is_moe and bid is not None: if name.endswith("mixer.gate.e_score_correction_bias"): new_name = name.replace("e_score_correction_bias", "e_score_correction.bias") diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c index 55f46d3c5..9584692fc 100644 --- a/ggml/src/ggml-cpu/ggml-cpu.c +++ b/ggml/src/ggml-cpu/ggml-cpu.c @@ -3796,7 +3796,11 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { /*.use_ref =*/ cplan->use_ref, }; - GGML_PRINT_DEBUG("thread #%d compute-start cplan %p last-graph %d \n", state->ith, cplan, state->last_graph); +#ifdef GGML_USE_OPENMP + GGML_PRINT_DEBUG("thread #%d compute-start cplan %p\n", state->ith, (const void *)cplan); +#else + GGML_PRINT_DEBUG("thread #%d compute-start cplan %p last-graph %d\n", state->ith, (const void *)cplan, state->last_graph); +#endif for (int node_n = 0; node_n < cgraph->n_nodes && atomic_load_explicit(&tp->abort, memory_order_relaxed) != node_n; node_n++) { struct ggml_tensor * node = cgraph->nodes[node_n]; @@ -3823,7 +3827,11 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { } } - GGML_PRINT_DEBUG("thread #%d compute-done cplan %p last-graph %d \n", state->ith, cplan, state->last_graph); +#ifdef GGML_USE_OPENMP + GGML_PRINT_DEBUG("thread #%d compute-done cplan %p\n", state->ith, (const void *)cplan); +#else + GGML_PRINT_DEBUG("thread #%d compute-done cplan %p last-graph %d\n", state->ith, (const void *)cplan, state->last_graph); +#endif ggml_barrier(state->threadpool); diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index 6d21cb5b1..c611a7766 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -2884,6 +2884,7 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) { const std::string ffn_moe_down_bias_prefix = "ffn_moe_down_biased"; const std::string nemotron_h_block_out_prefix = "nemotron_h_block_out"; const std::string mamba2_y_add_d_prefix = "mamba2_y_add_d"; + const std::string delta_net_prefix = "dnet_add"; for (int i = 0; i < cgraph->n_nodes; i++) { ggml_tensor * node = cgraph->nodes[i]; @@ -2914,7 +2915,8 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) { strncmp(node->name, ffn_moe_up_bias_prefix.c_str(), ffn_moe_up_bias_prefix.size()) != 0 && strncmp(node->name, ffn_moe_down_bias_prefix.c_str(), ffn_moe_down_bias_prefix.size()) != 0 && strncmp(node->name, nemotron_h_block_out_prefix.c_str(), nemotron_h_block_out_prefix.size()) != 0 && - strncmp(node->name, mamba2_y_add_d_prefix.c_str(), mamba2_y_add_d_prefix.size()) != 0) { + strncmp(node->name, mamba2_y_add_d_prefix.c_str(), mamba2_y_add_d_prefix.size()) != 0 && + strncmp(node->name, delta_net_prefix.c_str(), delta_net_prefix.size()) != 0) { // disable CUDA graphs for batch size > 1 for now while excluding the matrix-matrix addition as part of Gemma3n's `project_per_layer_input` operation // by means of matching node names. See // https://github.com/ggml-org/llama.cpp/blob/f9a31eea06a859e34cecb88b4d020c7f03d86cc4/src/llama-model.cpp#L10199-L10241 and @@ -4561,6 +4563,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g case GGML_UNARY_OP_CEIL: case GGML_UNARY_OP_ROUND: case GGML_UNARY_OP_TRUNC: + // TODO: should become: + //return ggml_is_contiguous_rows(op->src[0]); return ggml_is_contiguous(op->src[0]); default: return false; diff --git a/ggml/src/ggml-metal/ggml-metal-common.cpp b/ggml/src/ggml-metal/ggml-metal-common.cpp index 87e137868..2eb9820bf 100644 --- a/ggml/src/ggml-metal/ggml-metal-common.cpp +++ b/ggml/src/ggml-metal/ggml-metal-common.cpp @@ -273,6 +273,7 @@ static std::vector ggml_metal_graph_optimize_reorder(const std::vectorsrc[0]) && ggml_is_contiguous_rows(op->src[1]) && op->src[0]->type == GGML_TYPE_F32; case GGML_OP_ACC: + return ggml_is_contiguous_rows(op->src[0]) && ggml_is_contiguous_rows(op->src[1]) && op->src[0]->type == GGML_TYPE_F32; case GGML_OP_REPEAT: case GGML_OP_CONV_TRANSPOSE_1D: return true; diff --git a/ggml/src/ggml-metal/ggml-metal-ops.cpp b/ggml/src/ggml-metal/ggml-metal-ops.cpp index c04e9fc7f..3d5db0b79 100644 --- a/ggml/src/ggml-metal/ggml-metal-ops.cpp +++ b/ggml/src/ggml-metal/ggml-metal-ops.cpp @@ -620,8 +620,8 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) { GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32); GGML_ASSERT(op->type == GGML_TYPE_F32); - GGML_ASSERT(ggml_is_contiguous(op->src[0])); - GGML_ASSERT(ggml_is_contiguous(op->src[1])); + GGML_ASSERT(ggml_is_contiguous_rows(op->src[0])); + GGML_ASSERT(ggml_is_contiguous_rows(op->src[1])); const size_t pnb1 = ((const int32_t *) op->op_params)[0]; const size_t pnb2 = ((const int32_t *) op->op_params)[1]; @@ -671,10 +671,10 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) { } ggml_metal_kargs_bin args = { - /*.ne00 =*/ ne00, - /*.ne01 =*/ ne01, - /*.ne02 =*/ ne02, - /*.ne03 =*/ ne03, + /*.ne00 =*/ ne10, + /*.ne01 =*/ ne11, + /*.ne02 =*/ ne12, + /*.ne03 =*/ ne13, /*.nb00 =*/ nb00, /*.nb01 =*/ pnb1, /*.nb02 =*/ pnb2, @@ -687,10 +687,10 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) { /*.nb11 =*/ nb11, /*.nb12 =*/ nb12, /*.nb13 =*/ nb13, - /*.ne0 =*/ ne0, - /*.ne1 =*/ ne1, - /*.ne2 =*/ ne2, - /*.ne3 =*/ ne3, + /*.ne0 =*/ ne10, + /*.ne1 =*/ ne11, + /*.ne2 =*/ ne12, + /*.ne3 =*/ ne13, /*.nb0 =*/ nb0, /*.nb1 =*/ pnb1, /*.nb2 =*/ pnb2, @@ -707,7 +707,13 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) { ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op->src[1]), 2); ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op), 3); - const int nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne00); + const int nth_max = MIN(256, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)); + + int nth = 1; + + while (2*nth < args.ne0 && nth < nth_max) { + nth *= 2; + } ggml_metal_encoder_dispatch_threadgroups(enc, ne11, ne12, ne13, nth, 1, 1); diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index 37d07ce72..81e3240b2 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -108,6 +108,7 @@ static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; } #define VK_VENDOR_ID_APPLE 0x106b #define VK_VENDOR_ID_INTEL 0x8086 #define VK_VENDOR_ID_NVIDIA 0x10de +#define VK_VENDOR_ID_QUALCOMM 0x5143 #define VK_DEVICE_DESCRIPTOR_POOL_SIZE 256 @@ -703,6 +704,7 @@ struct vk_device_struct { vk_pipeline pipeline_get_rows[GGML_TYPE_COUNT]; vk_pipeline pipeline_get_rows_f32[GGML_TYPE_COUNT]; vk_pipeline pipeline_acc_f32; + vk_pipeline pipeline_set_f32; // [src0 0=fp32,1=fp16][src1 0=fp32,1=fp16][dst 0=fp32,1=fp16] vk_pipeline pipeline_add[2][2][2]; @@ -4096,7 +4098,7 @@ static void ggml_vk_load_shaders(vk_device& device) { } ggml_vk_create_pipeline(device, device->pipeline_rms_norm_back_f32, "rms_norm_back_f32", rms_norm_back_f32_len, rms_norm_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_l2_norm_f32, "l2_norm_f32", l2_norm_f32_len, l2_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_l2_norm_f32, "l2_norm_f32", l2_norm_f32_len, l2_norm_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {1, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f32, "cpy_f32_f32", cpy_f32_f32_len, cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f16, "cpy_f32_f16", cpy_f32_f16_len, cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); @@ -4197,7 +4199,8 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_add_id_f32, "add_id_f32", add_id_f32_len, add_id_f32_data, "main", 4, sizeof(vk_op_add_id_push_constants), {1, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_acc_f32, "acc_f32", acc_f32_len, acc_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_acc_f32, "acc_f32", acc_f32_len, acc_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0, 1}, 1); + ggml_vk_create_pipeline(device, device->pipeline_set_f32, "set_f32", acc_f32_len, acc_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0, 0}, 1); ggml_vk_create_pipeline(device, device->pipeline_concat_f32, "concat_f32", concat_f32_len, concat_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_concat_f16, "concat_f16", concat_f16_len, concat_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1); @@ -5671,6 +5674,10 @@ static void ggml_vk_instance_init() { driver_priorities[vk::DriverId::eMesaNvk] = 2; #endif break; + case VK_VENDOR_ID_QUALCOMM: + driver_priorities[vk::DriverId::eQualcommProprietary] = 1; + driver_priorities[vk::DriverId::eMesaTurnip] = 2; + break; } driver_priorities[vk::DriverId::eMesaDozen] = 100; @@ -8855,6 +8862,12 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_acc_f32; } return nullptr; + case GGML_OP_SET: + if (src0->type == src1->type && src0->type == dst->type && + (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_I32)) { + return ctx->device->pipeline_set_f32; + } + return nullptr; case GGML_OP_ADD: case GGML_OP_SUB: case GGML_OP_MUL: @@ -9846,7 +9859,7 @@ static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const int nb3 = dst->op_params[2] / src0_type_size; // 4 bytes of float32 int offset = dst->op_params[3] / src0_type_size; // offset in bytes - ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_ACC, { + ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, dst->op, { (uint32_t)ggml_nelements(src0), (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)nb1, (uint32_t)nb2, (uint32_t)nb3, (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size, @@ -10664,8 +10677,10 @@ static void ggml_vk_rms_norm_back(ggml_backend_vk_context * ctx, vk_context& sub } static void ggml_vk_l2_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { - float * op_params = (float *)dst->op_params; - ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_L2_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f, 0.0f, 0.0f }); + const float * op_params = (const float *)dst->op_params; + vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst); + p.param1 = op_params[0]; + ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_L2_NORM, std::move(p)); } static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) { @@ -12540,6 +12555,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr break; case GGML_OP_ACC: + case GGML_OP_SET: ggml_vk_acc(ctx, compute_ctx, src0, src1, node); break; @@ -14936,8 +14952,10 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm return true; case GGML_OP_NORM: case GGML_OP_GROUP_NORM: - case GGML_OP_L2_NORM: return ggml_is_contiguous(op->src[0]); + case GGML_OP_L2_NORM: + return ggml_is_contiguous_rows(op->src[0]) && + op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32; case GGML_OP_ADD: case GGML_OP_SUB: case GGML_OP_MUL: @@ -15000,7 +15018,10 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm } return op->src[0]->type == GGML_TYPE_F32; case GGML_OP_ACC: - return op->src[0]->type == GGML_TYPE_F32; + return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32; + case GGML_OP_SET: + return op->src[0]->type == op->src[1]->type && op->src[0]->type == op->type && + (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_I32); case GGML_OP_CONCAT: return ggml_type_size(op->src[0]->type) == ggml_type_size(GGML_TYPE_F32); case GGML_OP_ADD1: @@ -15651,6 +15672,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph * tensor_clone = ggml_add(ggml_ctx, src_clone[0], src_clone[1]); } else if (tensor->op == GGML_OP_ACC) { tensor_clone = ggml_acc(ggml_ctx, src_clone[0], src_clone[1], tensor->op_params[0], tensor->op_params[1], tensor->op_params[2], tensor->op_params[3]); + } else if (tensor->op == GGML_OP_SET) { + tensor_clone = ggml_set(ggml_ctx, src_clone[0], src_clone[1], tensor->op_params[0], tensor->op_params[1], tensor->op_params[2], tensor->op_params[3]); } else if (tensor->op == GGML_OP_NORM) { tensor_clone = ggml_norm(ggml_ctx, src_clone[0], *(float *)tensor->op_params); } else if (tensor->op == GGML_OP_GROUP_NORM) { diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp b/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp index 3d61168b5..6ba3d1d89 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp @@ -3,6 +3,9 @@ #include "types.glsl" #include "generic_binary_head.glsl" +// false for SET, true for ACC +layout(constant_id = 1) const bool ACC = true; + layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; void main() { @@ -23,7 +26,11 @@ void main() { uint i00, i01, i02, i03; if (i0 < p.ne10 && i1 < p.ne11 && i2 < p.ne12 && i3 < p.ne13) { - data_d[get_doffset() + idx] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + idx]) + FLOAT_TYPE(data_b[get_boffset() + src1_idx(i0, i1, i2, i3)])); + if (ACC) { + data_d[get_doffset() + idx] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + idx]) + FLOAT_TYPE(data_b[get_boffset() + src1_idx(i0, i1, i2, i3)])); + } else { + data_d[get_doffset() + idx] = D_TYPE(FLOAT_TYPE(data_b[get_boffset() + src1_idx(i0, i1, i2, i3)])); + } } else { data_d[get_doffset() + idx] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + idx])); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp index 83ef2f879..7d0a1de0d 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp @@ -1,6 +1,6 @@ #version 450 -#include "generic_head.glsl" +#include "generic_unary_head.glsl" #include "types.glsl" #extension GL_EXT_control_flow_attributes : enable @@ -8,19 +8,22 @@ layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in; -layout (binding = 0) readonly buffer X {A_TYPE data_a[];}; -layout (binding = 1) writeonly buffer D {D_TYPE data_d[];}; - shared FLOAT_TYPE sum[BLOCK_SIZE]; void main() { const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x; const uint tid = gl_LocalInvocationID.x; + const uint i3 = row / (p.ne11 * p.ne12); + const uint i3_offset = i3 * p.ne12 * p.ne11; + const uint i2 = (row - i3_offset) / p.ne11; + const uint i2_offset = i2 * p.ne11; + const uint i1 = row - i3_offset - i2_offset; + sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp - [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) { - const FLOAT_TYPE xi = FLOAT_TYPE(data_a[row*p.KX + col]); + [[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) { + const FLOAT_TYPE xi = FLOAT_TYPE(data_a[i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0]); sum[tid] += xi * xi; } @@ -35,7 +38,7 @@ void main() { const FLOAT_TYPE scale = inversesqrt(max(sum[0], FLOAT_TYPE(p.param1))); - [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) { - data_d[row*p.KX + col] = D_TYPE(scale * FLOAT_TYPE(data_a[row*p.KX + col])); + [[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) { + data_d[i3*p.nb13 + i2*p.nb12 + i1*p.nb11 + i0] = D_TYPE(scale * FLOAT_TYPE(data_a[i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0])); } } diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index eb8770af0..ea0fde920 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -3830,6 +3830,7 @@ class VisionProjectorType: MUSIC_FLAMINGO = "musicflamingo" # audio GLM4V = "glm4v" YOUTUVL = "youtuvl" + NEMOTRON_V2_VL = "nemotron_v2_vl" # Items here are (block size, type size) diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index c1538b3ff..b5c6e4bc5 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -1346,6 +1346,7 @@ class TensorNameMap: "model.vision_tower.embeddings.cls_token", # Intern-S1 "vision_model.class_embedding", # llama 4 "model.vision.patch_embedding.cls_embedding", # cogvlm + "vision_model.radio_model.model.patch_generator.cls_token.token", # Nemotron Nano v2 VL ), MODEL_TENSOR.V_ENC_EMBD_PATCH: ( @@ -1360,6 +1361,7 @@ class TensorNameMap: "vision_tower.patch_embed.proj", # kimi-vl "model.vision.patch_embedding.proj", # cogvlm "siglip2.vision_model.embeddings.patch_embedding", + "vision_model.radio_model.model.patch_generator.embedder", # Nemotron Nano v2 VL ), MODEL_TENSOR.V_ENC_EMBD_NORM: ( @@ -1376,12 +1378,14 @@ class TensorNameMap: "visual.pos_embed", # qwen3vl "model.vision.patch_embedding.position_embedding", # cogvlm "visual.embeddings.position_embedding", # glm4v + "vision_model.radio_model.model.patch_generator.pos_embed", # Nemotron Nano v2 VL ), MODEL_TENSOR.V_ENC_ATTN_QKV: ( "visual.blocks.{bid}.attn.qkv", # qwen3vl "model.vision.transformer.layers.{bid}.attention.query_key_value", # cogvlm - "vision_tower.encoder.blocks.{bid}.wqkv" # Kimi-K2.5 + "vision_tower.encoder.blocks.{bid}.wqkv", # Kimi-K2.5 + "vision_model.radio_model.model.blocks.{bid}.attn.qkv", # Nemotron Nano v2 VL ), MODEL_TENSOR.V_ENC_ATTN_Q: ( @@ -1446,6 +1450,7 @@ class TensorNameMap: "vision_tower.encoder.blocks.{bid}.norm0", # kimi-vl (norm0/norm1) "model.vision.transformer.layers.{bid}.input_layernorm", # cogvlm "siglip2.vision_model.encoder.layers.{bid}.layer_norm1", + "vision_model.radio_model.model.blocks.{bid}.norm1", # Nemotron Nano v2 VL ), MODEL_TENSOR.V_ENC_ATTN_O: ( @@ -1462,6 +1467,7 @@ class TensorNameMap: "vision_tower.encoder.blocks.{bid}.wo", # kimi-vl "model.vision.transformer.layers.{bid}.attention.dense", # cogvlm "siglip2.vision_model.encoder.layers.{bid}.self_attn.out_proj", # youtuvl + "vision_model.radio_model.model.blocks.{bid}.attn.proj", # Nemotron Nano v2 VL ), MODEL_TENSOR.V_ENC_POST_ATTN_NORM: ( @@ -1477,6 +1483,7 @@ class TensorNameMap: "vision_tower.encoder.blocks.{bid}.norm1", # kimi-vl (norm0/norm1) "model.vision.transformer.layers.{bid}.post_attention_layernorm", # cogvlm "siglip2.vision_model.encoder.layers.{bid}.layer_norm2", + "vision_model.radio_model.model.blocks.{bid}.norm2", # Nemotron Nano v2 VL ), MODEL_TENSOR.V_ENC_FFN_UP: ( @@ -1493,6 +1500,7 @@ class TensorNameMap: "vision_tower.encoder.blocks.{bid}.mlp.fc0", # kimi-vl (fc0/fc1) "model.vision.transformer.layers.{bid}.mlp.fc1", # cogvlm "siglip2.vision_model.encoder.layers.{bid}.mlp.fc1", + "vision_model.radio_model.model.blocks.{bid}.mlp.fc1", # Nemotron Nano v2 VL ), MODEL_TENSOR.V_ENC_FFN_GATE: ( @@ -1515,6 +1523,7 @@ class TensorNameMap: "vision_tower.encoder.blocks.{bid}.mlp.fc1", # kimi-vl (fc0/fc1) "model.vision.transformer.layers.{bid}.mlp.fc2", # cogvlm "siglip2.vision_model.encoder.layers.{bid}.mlp.fc2", + "vision_model.radio_model.model.blocks.{bid}.mlp.fc2", # Nemotron Nano v2 VL ), MODEL_TENSOR.V_LAYER_SCALE_1: ( diff --git a/gpttype_adapter.cpp b/gpttype_adapter.cpp index 48bdd2879..490d887c5 100644 --- a/gpttype_adapter.cpp +++ b/gpttype_adapter.cpp @@ -2681,6 +2681,8 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in } llama_attach_threadpool(llama_ctx_v4, threadpool1, threadpool2); + std::vector loras; + std::vector lorascales; if (lora_filename != "") { printf("\nAttempting to apply LORA adapter: %s\n", lora_filename.c_str()); @@ -2689,7 +2691,10 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__); return ModelLoadResult::FAIL; } - llama_set_adapter_lora(llama_ctx_v4, adapter, inputs.lora_multiplier); + + loras.push_back(adapter); + lorascales.push_back(inputs.lora_multiplier); + llama_set_adapters_lora(llama_ctx_v4, loras.data(), loras.size(), lorascales.data()); } if(mmproj_filename != "" && file_format==FileFormat::GGUF_GENERIC) diff --git a/include/llama.h b/include/llama.h index e2036ec5c..2667f685c 100644 --- a/include/llama.h +++ b/include/llama.h @@ -659,21 +659,12 @@ extern "C" { // The following functions operate on a llama_context, hence the naming: llama_verb_... - // Add a loaded LoRA adapter to given context - // This will not modify model's weight - LLAMA_API int32_t llama_set_adapter_lora( + // Set LoRa adapters on the context. Will only modify if the adapters currently in context are different. + LLAMA_API int32_t llama_set_adapters_lora( struct llama_context * ctx, - struct llama_adapter_lora * adapter, - float scale); - - // Remove a specific LoRA adapter from given context - // Return -1 if the adapter is not present in the context - LLAMA_API int32_t llama_rm_adapter_lora( - struct llama_context * ctx, - struct llama_adapter_lora * adapter); - - // Remove all LoRA adapters from given context - LLAMA_API void llama_clear_adapter_lora(struct llama_context * ctx); + struct llama_adapter_lora ** adapters, + size_t n_adapters, + float * scales); // Apply a loaded control vector to a llama_context, or if data is NULL, clear // the currently loaded vector. @@ -681,7 +672,7 @@ extern "C" { // to an n_embd x n_layers buffer starting from layer 1. // il_start and il_end are the layer range the vector should apply to (both inclusive) // See llama_control_vector_load in common to load a control vector. - LLAMA_API int32_t llama_apply_adapter_cvec( + LLAMA_API int32_t llama_set_adapter_cvec( struct llama_context * ctx, const float * data, size_t len, diff --git a/src/llama-context.cpp b/src/llama-context.cpp index bd75b0870..e683fd537 100644 --- a/src/llama-context.cpp +++ b/src/llama-context.cpp @@ -1067,51 +1067,43 @@ bool llama_context::set_sampler(llama_seq_id seq_id, llama_sampler * sampler) { return true; } -void llama_context::set_adapter_lora( - llama_adapter_lora * adapter, - float scale) { - LLAMA_LOG_DEBUG("%s: adapter = %p, scale = %f\n", __func__, (void *) adapter, scale); +void llama_context::set_adapters_lora(llama_adapter_lora ** adapters, size_t n_adapters, float * scales) { + LLAMA_LOG_DEBUG("%s: adapters = %p\n", __func__, (void *) adapters); - if (auto it = loras.find(adapter); it != loras.end()) { - if (it->second == scale) { - return; - } - } - - loras[adapter] = scale; - - sched_need_reserve = true; -} - -bool llama_context::rm_adapter_lora( - llama_adapter_lora * adapter) { - LLAMA_LOG_DEBUG("%s: adapter = %p\n", __func__, (void *) adapter); - - auto it = loras.find(adapter); - if (it != loras.end()) { - loras.erase(it); - - sched_need_reserve = true; - - return true; - } - - return false; -} - -void llama_context::clear_adapter_lora() { - LLAMA_LOG_DEBUG("%s: call\n", __func__); - - if (loras.empty()) { + if (adapters_lora_are_same(adapters, n_adapters, scales)) { return; } loras.clear(); + for (size_t i = 0; i < n_adapters; i ++) { + if (scales[i] != 0.0f) { + loras[adapters[i]] = scales[i]; + } + } + sched_need_reserve = true; } -bool llama_context::apply_adapter_cvec( +bool llama_context::adapters_lora_are_same(llama_adapter_lora ** adapters, size_t n_adapters, float * scales) { + LLAMA_LOG_DEBUG("%s: adapters = %p\n", __func__, (void *) adapters); + + if (n_adapters != loras.size()) { + return false; + } + + for (size_t i = 0; i < n_adapters; i ++) { + auto it = loras.find(adapters[i]); + + if (it == loras.end() || it->second != scales[i]) { + return false; + } + } + + return true; +} + +bool llama_context::set_adapter_cvec( const float * data, size_t len, int32_t n_embd, @@ -3219,35 +3211,28 @@ uint32_t llama_get_sampled_probs_count_ith(llama_context * ctx, int32_t i) { // llama adapter API -int32_t llama_set_adapter_lora( +int32_t llama_set_adapters_lora( llama_context * ctx, - llama_adapter_lora * adapter, - float scale) { - ctx->set_adapter_lora(adapter, scale); + llama_adapter_lora ** adapters, + size_t n_adapters, + float * scales) { + if (adapters == nullptr || scales == nullptr) { + GGML_ASSERT(n_adapters == 0 && "invalid llama_set_adapters_lora call"); + } + + ctx->set_adapters_lora(adapters, n_adapters, scales); return 0; } -int32_t llama_rm_adapter_lora( - llama_context * ctx, - llama_adapter_lora * adapter) { - bool res = ctx->rm_adapter_lora(adapter); - - return res ? 0 : -1; -} - -void llama_clear_adapter_lora(llama_context * ctx) { - ctx->clear_adapter_lora(); -} - -int32_t llama_apply_adapter_cvec( +int32_t llama_set_adapter_cvec( llama_context * ctx, - const float * data, - size_t len, - int32_t n_embd, - int32_t il_start, - int32_t il_end) { - bool res = ctx->apply_adapter_cvec(data, len, n_embd, il_start, il_end); + const float * data, + size_t len, + int32_t n_embd, + int32_t il_start, + int32_t il_end) { + bool res = ctx->set_adapter_cvec(data, len, n_embd, il_start, il_end); return res ? 0 : -1; } diff --git a/src/llama-context.h b/src/llama-context.h index d99511757..37117ba7b 100644 --- a/src/llama-context.h +++ b/src/llama-context.h @@ -105,16 +105,11 @@ struct llama_context { void set_causal_attn(bool value); void set_warmup(bool value); - void set_adapter_lora( - llama_adapter_lora * adapter, - float scale); + void set_adapters_lora(llama_adapter_lora ** adapters, size_t n_adapters, float * scales); - bool rm_adapter_lora( - llama_adapter_lora * adapter); + bool adapters_lora_are_same(llama_adapter_lora ** adapters, size_t n_adapters, float * scales); - void clear_adapter_lora(); - - bool apply_adapter_cvec( + bool set_adapter_cvec( const float * data, size_t len, int32_t n_embd, diff --git a/src/llama-mmap.cpp b/src/llama-mmap.cpp index b5582849f..1325492cc 100644 --- a/src/llama-mmap.cpp +++ b/src/llama-mmap.cpp @@ -506,6 +506,8 @@ struct llama_mmap::impl { } } #elif defined(_WIN32) + HANDLE hMapping = nullptr; + impl(struct llama_file * file, size_t prefetch, bool numa) { GGML_UNUSED(numa); @@ -513,7 +515,7 @@ struct llama_mmap::impl { HANDLE hFile = (HANDLE) _get_osfhandle(file->file_id()); - HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL); + hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL); if (hMapping == NULL) { DWORD error = GetLastError(); @@ -522,9 +524,9 @@ struct llama_mmap::impl { addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0); DWORD error = GetLastError(); - CloseHandle(hMapping); if (addr == NULL) { + CloseHandle(hMapping); throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str())); } @@ -560,9 +562,17 @@ printf("\nPrefetchVirtualMemory skipped in compatibility mode.\n"); } ~impl() { - if (!UnmapViewOfFile(addr)) { - LLAMA_LOG_WARN("warning: UnmapViewOfFile failed: %s\n", - llama_format_win_err(GetLastError()).c_str()); + if (hMapping) { + if (addr) { + if (!UnmapViewOfFile(addr)) { + LLAMA_LOG_WARN("warning: UnmapViewOfFile failed: %s\n", + llama_format_win_err(GetLastError()).c_str()); + } + } + if (!CloseHandle(hMapping)) { + LLAMA_LOG_WARN("warning: CloseHandle failed: %s\n", + llama_format_win_err(GetLastError()).c_str()); + } } } #else diff --git a/src/models/models.h b/src/models/models.h index 3c66d3253..ec6f80e52 100644 --- a/src/models/models.h +++ b/src/models/models.h @@ -489,9 +489,6 @@ private: ggml_tensor * build_layer_attn_linear( llm_graph_input_rs * inp, ggml_tensor * cur, - ggml_tensor * causal_mask, - ggml_tensor * identity, - ggml_tensor * diag_mask, int il); ggml_tensor * build_layer_ffn( @@ -506,9 +503,6 @@ private: ggml_tensor * g, ggml_tensor * beta, ggml_tensor * state, - ggml_tensor * causal_mask, - ggml_tensor * identity, - ggml_tensor * diag_mask, int il); // returns pair of output and new state diff --git a/src/models/qwen3next.cpp b/src/models/qwen3next.cpp index 99b1a76a4..aea8b2951 100644 --- a/src/models/qwen3next.cpp +++ b/src/models/qwen3next.cpp @@ -16,17 +16,6 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr ggml_tensor * inp_pos = build_inp_pos(); ggml_tensor * inp_out_ids = build_inp_out_ids(); - ggml_tensor * causal_mask = - ggml_tri(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, CHUNK_SIZE, CHUNK_SIZE), 1.0f), - GGML_TRI_TYPE_LOWER); - - ggml_tensor * identity = ggml_diag(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, CHUNK_SIZE), 1.0f)); - ggml_tensor * diag_mask = ggml_add(ctx0, causal_mask, identity); - - ggml_build_forward_expand(gf, causal_mask); - ggml_build_forward_expand(gf, identity); - ggml_build_forward_expand(gf, diag_mask); - for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -36,7 +25,7 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr // Determine layer type and build appropriate attention mechanism if (hparams.is_recurrent(il)) { // Linear attention layer (gated delta net) - cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, diag_mask, il); + cur = build_layer_attn_linear(inp->get_recr(), cur, il); } else { // Full attention layer cur = build_layer_attn(inp->get_attn(), cur, inp_pos, il); @@ -99,11 +88,8 @@ std::pair llm_build_qwen3next::build_delta_net_chu ggml_tensor * k, ggml_tensor * v, ggml_tensor * g, - ggml_tensor * beta, - ggml_tensor * state, - ggml_tensor * causal_mask, - ggml_tensor * identity, - ggml_tensor * diag_mask, + ggml_tensor * b, + ggml_tensor * s, int il) { const int64_t S_k = q->ne[0]; const int64_t H_k = q->ne[1]; @@ -113,134 +99,123 @@ std::pair llm_build_qwen3next::build_delta_net_chu const int64_t S_v = v->ne[0]; const int64_t H_v = v->ne[1]; - GGML_ASSERT(v->ne[2] == n_tokens); - GGML_ASSERT(k->ne[2] == n_tokens); - GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs); - GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs); - GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs); + GGML_ASSERT(S_k == S_v); + GGML_ASSERT(H_v % H_k == 0); GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs); GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs); + GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs); - GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case + GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs); + GGML_ASSERT(b->ne[0] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs); + GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs); - const float eps_norm = hparams.f_norm_rms_eps; - - q = ggml_l2_norm(ctx0, q, eps_norm); - k = ggml_l2_norm(ctx0, k, eps_norm); - - const float scale = 1.0f / sqrtf(S_v); + const float scale = 1.0f / sqrtf(S_k); q = ggml_scale(ctx0, q, scale); - beta = ggml_sigmoid(ctx0, beta); - cb(q, "q_in", il); cb(k, "k_in", il); cb(v, "v_in", il); - cb(beta, "beta_in", il); + cb(b, "b_in", il); cb(g, "g_in", il); - q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); - k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); - v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); - g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs); + q = ggml_permute(ctx0, q, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs] + k = ggml_permute(ctx0, k, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs] + v = ggml_permute(ctx0, v, 0, 2, 1, 3); // [S_v, n_tokens, H_v, n_seqs] + g = ggml_permute(ctx0, g, 2, 1, 3, 0); // [ 1, n_tokens, H_v, n_seqs] + b = ggml_permute(ctx0, b, 2, 0, 1, 3); // [ 1, n_tokens, H_v, n_seqs] - beta = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3)); - state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs); + const int CS = CHUNK_SIZE; - cb(q, "q_perm", il); - cb(k, "k_perm", il); - cb(v, "v_perm", il); - cb(beta, "beta_perm", il); - cb(g, "g_perm", il); - cb(state, "state_in", il); - - GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs); - GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs); - GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs); - GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs); - - // Do padding - const int64_t chunk_size = CHUNK_SIZE; - - const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size; - const int64_t n_chunks = (n_tokens + pad) / chunk_size; + const int pad = (CS - n_tokens % CS) % CS; + const int n_chunks = (n_tokens + pad) / CS; q = ggml_pad(ctx0, q, 0, pad, 0, 0); k = ggml_pad(ctx0, k, 0, pad, 0, 0); v = ggml_pad(ctx0, v, 0, pad, 0, 0); - g = ggml_pad(ctx0, g, pad, 0, 0, 0); - beta = ggml_pad(ctx0, beta, 0, pad, 0, 0); + g = ggml_pad(ctx0, g, 0, pad, 0, 0); + b = ggml_pad(ctx0, b, 0, pad, 0, 0); - cb(q, "q_pad", il); - cb(k, "k_pad", il); - cb(v, "v_pad", il); - cb(beta, "beta_pad", il); - cb(g, "g_pad", il); + ggml_tensor * v_b = ggml_mul(ctx0, v, b); + ggml_tensor * k_b = ggml_mul(ctx0, k, b); - ggml_tensor * v_beta = ggml_mul(ctx0, v, beta); - ggml_tensor * k_beta = ggml_mul(ctx0, k, beta); + cb(v_b, "v_b", il); + cb(k_b, "k_b", il); - cb(v_beta, "v_beta", il); - cb(k_beta, "k_beta", il); + q = ggml_reshape_4d(ctx0, q, S_k, CS, n_chunks, H_k * n_seqs); + k = ggml_reshape_4d(ctx0, k, S_k, CS, n_chunks, H_k * n_seqs); + k_b = ggml_reshape_4d(ctx0, k_b, S_k, CS, n_chunks, H_v * n_seqs); + v = ggml_reshape_4d(ctx0, v, S_v, CS, n_chunks, H_v * n_seqs); + v_b = ggml_reshape_4d(ctx0, v_b, S_v, CS, n_chunks, H_v * n_seqs); - q = ggml_reshape_4d(ctx0, q, S_k, chunk_size, n_chunks, H_k * n_seqs); - k = ggml_reshape_4d(ctx0, k, S_k, chunk_size, n_chunks, H_k * n_seqs); - k_beta = ggml_reshape_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs); - v = ggml_reshape_4d(ctx0, v, S_v, chunk_size, n_chunks, H_v * n_seqs); - v_beta = ggml_reshape_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs); + g = ggml_reshape_4d(ctx0, g, CS, 1, n_chunks, H_v * n_seqs); + b = ggml_reshape_4d(ctx0, b, 1, CS, n_chunks, H_v * n_seqs); - g = ggml_reshape_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs); - beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs); + // [CS, 1, n_chunks, H_v * n_seqs] + ggml_tensor * g_cs = ggml_cumsum(ctx0, g); + cb(g_cs, "g_cs", il); - ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g); - cb(g_cumsum, "g_cumsum", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs) + ggml_tensor * g_cs_i = g_cs; + ggml_tensor * g_cs_j = ggml_reshape_4d(ctx0, g_cs, 1, CS, n_chunks, H_v * n_seqs); - ggml_tensor * gcs_i = g_cumsum; // ggml_reshape_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs); - ggml_tensor * gcs_j = ggml_reshape_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs); + g_cs_j = ggml_repeat_4d(ctx0, g_cs_j, CS, CS, n_chunks, H_v * n_seqs); - ggml_tensor * gcs_j_broadcast = - ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs); - - ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i); - cb(decay_mask, "decay_mask", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) - - decay_mask = ggml_mul(ctx0, decay_mask, diag_mask); + // [CS, CS, n_chunks, H_v * n_seqs] + ggml_tensor * decay_mask; + decay_mask = ggml_sub(ctx0, g_cs_j, g_cs_i); + decay_mask = ggml_tri(ctx0, decay_mask, GGML_TRI_TYPE_LOWER_DIAG); decay_mask = ggml_exp(ctx0, decay_mask); - decay_mask = ggml_mul(ctx0, decay_mask, diag_mask); + cb(decay_mask, "decay_mask", il); - ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta); + // [CS, CS, n_chunks, H_k * n_seqs] + ggml_tensor * kb; + kb = ggml_mul_mat(ctx0, k, k_b); + kb = ggml_mul (ctx0, kb, decay_mask); - ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask); - ggml_tensor * attn = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask)); - cb(attn, "attn_pre_solve", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + // [CS, CS, n_chunks, H_k * n_seqs] + ggml_tensor * attn; + attn = ggml_tri(ctx0, kb, GGML_TRI_TYPE_LOWER); - ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask); - ggml_tensor * lhs = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower); + ggml_tensor * identity; + identity = ggml_view_1d(ctx0, attn, CS, 0); + identity = ggml_fill (ctx0, identity, 1.0f); + identity = ggml_diag (ctx0, identity); - ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false); - attn = ggml_mul(ctx0, lin_solve, causal_mask); - attn = ggml_add(ctx0, attn, identity); - cb(attn, "attn_solved", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + ggml_tensor * lhs = ggml_add(ctx0, attn, identity); + cb(lhs, "dnet_add_ch_lhs", il); - v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn); + attn = ggml_neg(ctx0, attn); - ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum)); - ggml_tensor * gexp = ggml_exp(ctx0, g_cumsum_t); + ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false); + attn = ggml_add(ctx0, lin_solve, identity); + cb(attn, "dnet_add_ch_attn_solved", il); // [CS, CS, n_chunks, H_k * n_seqs] - ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp); - cb(kbeta_gexp, "kbeta_gexp", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs) + // [S_v, CS, n_chunks, H_v * n_seqs] + v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_b)), attn); - ggml_tensor * k_cumdecay = - ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp))))); - cb(k_cumdecay, "k_cumdecay", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + // [CS, 1, n_chunks, H_v * n_seqs] + ggml_tensor * g_exp = ggml_exp(ctx0, g_cs); - ggml_tensor * attn_kq = ggml_mul_mat(ctx0, k, q); - attn_kq = ggml_mul(ctx0, attn_kq, decay_mask); - attn_kq = ggml_mul(ctx0, attn_kq, diag_mask); - cb(attn_kq, "attn_kq", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + k_b = ggml_cont(ctx0, ggml_transpose(ctx0, k_b)); + // [CS, S_k, n_chunks, H_k * n_seqs] + ggml_tensor * kbg = ggml_mul(ctx0, k_b, g_exp); + cb(kbg, "k_beta_g_exp", il); + + // [S_k, CS, n_chunks, H_k * n_seqs] + ggml_tensor * k_cd = ggml_mul_mat(ctx0, kbg, attn); + cb(k_cd, "k_cumdecay", il); + + // [S_k, CS, n_chunks, H_k * n_seqs] + ggml_tensor * g_exp_t = ggml_transpose(ctx0, g_exp); + ggml_tensor * q_g_exp = ggml_mul(ctx0, q, g_exp_t); + + // [CS, CS, n_chunks, H_k * n_seqs] + ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + kq = ggml_mul(ctx0, kq, decay_mask); + kq = ggml_tri(ctx0, kq, GGML_TRI_TYPE_LOWER_DIAG); + cb(kq, "kq", il); // vectorized calculation of key_gdiff // improved from the chunked version: @@ -250,109 +225,98 @@ std::pair llm_build_qwen3next::build_delta_net_chu // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew - // get last element in g_cumsum along chunk_size dimension (ne0) + // get last element in g_cumsum along CS dimension (ne0) // example: [[x, y, z, ..., last], ...] -> [[last], ...] - ggml_tensor * g_last = ggml_view_4d(ctx0, g_cumsum, 1, 1, g_cumsum->ne[2], g_cumsum->ne[3], - g_cumsum->nb[1], g_cumsum->nb[2], g_cumsum->nb[3], - (g_cumsum->ne[0] - 1) * ggml_element_size(g_cumsum)); + // [1, 1, n_chunks, H_v * n_seqs] + ggml_tensor * g_last = ggml_view_4d(ctx0, g_cs, 1, 1, g_cs->ne[2], g_cs->ne[3], + g_cs->nb[1], + g_cs->nb[2], + g_cs->nb[3], + ggml_row_size(g_cs->type, g_cs->ne[0] - 1)); + cb(g_last, "g_last", il); + + // TODO: remove this cont when CUDA supports non-cont unary ops g_last = ggml_cont(ctx0, g_last); - cb(g_last, "g_last", il); // shape: (1, 1, n_chunks, H_v * n_seqs) + // [1, 1, n_chunks, H_v * n_seqs] ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last); - cb(g_last_exp, "g_last_exp", il); // shape: (1, 1, n_chunks, H_v * n_seqs) + cb(g_last_exp, "g_last_exp", il); - ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum, g_last)); - cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs) + // [CS, 1, n_chunks, H_v * n_seqs] + ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cs, g_last)); + cb(g_diff, "g_diff", il); - ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff); - ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp, - 1, chunk_size, n_chunks, g_diff_exp->ne[3]); + ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff); + ggml_tensor * g_diff_exp_t = ggml_transpose(ctx0, g_diff_exp); - ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t); - cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs) + // [S_k, CS, n_chunks, H_v * n_seqs] + ggml_tensor * kg = ggml_mul(ctx0, k, g_diff_exp_t); + cb(kg, "key_gdiff", il); - ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff)); - cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs) + // [CS, S_k, n_chunks, H_v * n_seqs] + ggml_tensor * kg_t = ggml_cont(ctx0, ggml_transpose(ctx0, kg)); + cb(kg_t, "key_gdiff_t", il); + ggml_tensor * s_t = ggml_transpose(ctx0, s); + s_t = ggml_cont_4d(ctx0, s_t, S_v, S_v, 1, H_v * n_seqs); + cb(s_t, "dnet_add_ch_state", il); - // state to be updated per chunk - ggml_tensor * new_state = state; // ggml_dup(ctx0, state); - cb(new_state, "new_state", il); // shape: (S_v, S_v, H_v, n_seqs) - - // shape after loop of chunks: (S_v, chunk_size, n_chunks, H_v * n_seqs) - ggml_tensor * core_attn_out = nullptr; + // [CS, S_v, n_chunks, H_v * n_seqs] + ggml_tensor * v_t = ggml_cont(ctx0, ggml_transpose(ctx0, v)); for (int64_t chunk = 0; chunk < n_chunks; chunk++) { - // shape: (S_k, chunk_size, 1, H_k * n_seqs) - ggml_tensor * q_chunk = get_slice_2d(ctx0, q, chunk); // (no cont), next op: ggml_mul + ggml_tensor * ch_k_cd = get_slice_2d(ctx0, k_cd, chunk); // [S_k, CS, 1, H_k * n_seqs] + ggml_tensor * ch_v_t = get_slice_2d(ctx0, v_t, chunk); // [ CS, S_v, 1, H_v * n_seqs] + ggml_tensor * ch_kq = get_slice_2d(ctx0, kq, chunk); // [ CS, CS, 1, H_k * n_seqs] + ggml_tensor * ch_q_g_exp = get_slice_2d(ctx0, q_g_exp, chunk); // [S_k, CS, 1, H_k * n_seqs] + ggml_tensor * ch_kg_t = get_slice_2d(ctx0, kg_t, chunk); // [ CS, S_k, 1, H_v * n_seqs] - // shape: (S_v, chunk_size, 1, H_v * n_seqs) - ggml_tensor * v_chunk = get_slice_2d(ctx0, v, chunk); // (no cont), next op: ggml_repeat + // [CS, S_v, 1, H_v * n_seqs] + ggml_tensor * v_t_p = ggml_mul_mat(ctx0, ch_k_cd, s_t); + cb(v_t_p, "v_prime", il); - // shape: (chunk_size, 1, n_chunks, H_v * n_seqs) - ggml_tensor * gexp_chunk = get_slice_2d(ctx0, gexp, chunk); // (no cont), next op: ggml_mul + // [CS, S_v, 1, H_v * n_seqs] + ggml_tensor * v_t_new = ggml_sub(ctx0, ch_v_t, v_t_p); + cb(v_t_new, "v_t_new", il); - // shape: (chunk_size, 1, H_v * n_seqs) - ggml_tensor * k_cumdecay_chunk = get_slice_2d(ctx0, k_cumdecay, chunk); // (no cont), next op: ggml_mul_mat + // [S_v, CS, 1, H_v * n_seqs] + ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_t_new, ch_kq); + cb(v_attn, "v_attn", il); - // attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0) - // replaced by precomputed attn_kq - ggml_tensor * attn_chunk = get_slice_2d(ctx0, attn_kq, chunk); - cb(attn_chunk, "attn_chunk", il); + // [S_v, CS, 1, H_v * n_seqs] + ggml_tensor * attn_inter = ggml_mul_mat(ctx0, s_t, ch_q_g_exp); + cb(attn_inter, "attn_inter", il); - ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs); + // [S_v, CS, 1, H_v * n_seqs] + ggml_tensor * o_ch = ggml_add(ctx0, attn_inter, v_attn); + cb(o_ch, "dnet_add_ch_attn_out", il); - // v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state - ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk); - cb(v_prime, "v_prime_chunk", il); // shape: (S_v, 1, H_v * n_seqs) - - // v_new = v_i - v_prime - ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v_chunk, v_prime), v_prime); - ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new)); - cb(v_new, "v_new_chunk", il); - - // attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state - ggml_tensor * q_g_exp = ggml_mul(ctx0, q_chunk, gexp_chunk); - ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp); - cb(attn_inter, "attn_inter_chunk", il); - - // core_attn_out[:, :, i] = attn_inter + attn @ v_new - ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn_chunk); - cb(v_attn, "v_attn_chunk", il); - - ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn); - cb(core_attn_out_chunk, "core_attn_out_chunk", il); // shape: (S_v, chunk_size, 1, H_v * n_seqs) - - core_attn_out = core_attn_out == nullptr - ? core_attn_out_chunk - : ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2); + v = ggml_set_inplace(ctx0, v, o_ch, v->nb[1], v->nb[2], v->nb[3], chunk * v->nb[2]); // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new - ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk); - //ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why? - ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t); + // TODO: head broadcast might not work here - probably will need a transpose + ggml_tensor * kgv = ggml_mul_mat(ctx0, ch_kg_t, v_t_new); // [S_k, S_v, 1, H_k * n_seqs] // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew - ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk)); - new_state = ggml_add(ctx0, - ggml_mul(ctx0, new_state, ggml_reshape_4d(ctx0, gexp_last_chunk, gexp_last_chunk->ne[0], gexp_last_chunk->ne[1], H_v, n_seqs)), - ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs)); + ggml_tensor * ch_g_last_exp = get_slice_2d(ctx0, g_last_exp, chunk); + s_t = ggml_mul(ctx0, s_t, ch_g_last_exp); + s_t = ggml_add(ctx0, s_t, kgv); + cb(s_t, "dnet_add_ch_state", il); } + s_t = ggml_reshape_4d(ctx0, s_t, S_v, S_v, H_v, n_seqs); + // truncate padded tokens - ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out, + ggml_tensor * o = ggml_view_4d(ctx0, v, S_v, n_tokens, H_v, n_seqs, - ggml_row_size(core_attn_out->type, S_v), - ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks), - ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks * H_v), 0); - output_tokens = ggml_cont(ctx0, output_tokens); - cb(output_tokens, "output_tokens", il); + ggml_row_size(v->type, S_v), + ggml_row_size(v->type, S_v * CS * n_chunks), + ggml_row_size(v->type, S_v * CS * n_chunks * H_v), 0); - // permute back to (S_v, H_v, n_tokens, n_seqs) - output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3); - output_tokens = ggml_cont(ctx0, output_tokens); + o = ggml_permute (ctx0, o, 0, 2, 1, 3); // [S_v, H_v, n_tokens, n_seqs] + s = ggml_transpose(ctx0, s_t); // [S_v, S_v, H_v, n_seqs] - return {output_tokens, new_state}; + return {o, s}; } std::pair llm_build_qwen3next::build_delta_net_autoregressive( @@ -360,8 +324,8 @@ std::pair llm_build_qwen3next::build_delta_net_aut ggml_tensor * k, ggml_tensor * v, ggml_tensor * g, - ggml_tensor * beta, - ggml_tensor * state, + ggml_tensor * b, // beta + ggml_tensor * s, // state int il) { const int64_t S_k = q->ne[0]; const int64_t H_k = q->ne[1]; @@ -371,75 +335,72 @@ std::pair llm_build_qwen3next::build_delta_net_aut const int64_t S_v = v->ne[0]; const int64_t H_v = v->ne[1]; - GGML_ASSERT(n_tokens == 1); // This function is optimized for single token processing - GGML_ASSERT(v->ne[2] == n_tokens); - GGML_ASSERT(k->ne[2] == n_tokens); - GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs); - GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs); - GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs); + GGML_ASSERT(n_tokens == 1); + + GGML_ASSERT(S_k == S_v); + GGML_ASSERT(H_v % H_k == 0); GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs); GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs); + GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs); - GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case + GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs); + GGML_ASSERT(b->ne[0] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs); + GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs); - const float eps_norm = hparams.f_norm_rms_eps; + const float scale = 1.0f / sqrtf(S_k); - q = ggml_l2_norm(ctx0, q, eps_norm); - k = ggml_l2_norm(ctx0, k, eps_norm); + q = ggml_scale(ctx0, q, scale); - const float scale = 1.0f / sqrtf(S_v); - - q = ggml_scale(ctx0, q, scale); - beta = ggml_sigmoid(ctx0, beta); + q = ggml_permute(ctx0, q, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs] + k = ggml_permute(ctx0, k, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs] + v = ggml_permute(ctx0, v, 0, 2, 1, 3); // [S_v, n_tokens, H_v, n_seqs] cb(q, "q_in", il); cb(k, "k_in", il); cb(v, "v_in", il); - cb(beta, "beta_in", il); + cb(b, "b_in", il); cb(g, "g_in", il); - state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs); + g = ggml_reshape_4d(ctx0, g, 1, 1, H_v, n_seqs); + b = ggml_reshape_4d(ctx0, b, 1, 1, H_v, n_seqs); - ggml_tensor * g_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, g), 1, 1, H_k, n_seqs); - ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs); + // [S_v, S_v, H_v, n_seqs] + g = ggml_exp(ctx0, g); + s = ggml_mul(ctx0, s, g); - // Apply exponential to g_t - g_t = ggml_exp(ctx0, g_t); + ggml_tensor * s_t = ggml_cont(ctx0, ggml_transpose(ctx0, s)); - // Apply the gated delta rule for the single timestep - // last_recurrent_state = last_recurrent_state * g_t - state = ggml_mul(ctx0, state, g_t); + // [1, S_v, H_v, n_seqs] + ggml_tensor * sk; + sk = ggml_mul (ctx0, s_t, k); + sk = ggml_sum_rows(ctx0, sk); - // kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2) - ggml_tensor * k_t_unsqueezed = ggml_reshape_4d(ctx0, k, 1, S_v, H_v, n_seqs); - ggml_tensor * kv_mem = ggml_mul(ctx0, state, k_t_unsqueezed); - // we need to sum over dim=-2, so we transpose, sum, then transpose again - kv_mem = ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kv_mem)))); + // [S_v, 1, H_v, n_seqs] + ggml_tensor * d; + d = ggml_sub(ctx0, v, ggml_transpose(ctx0, sk)); + d = ggml_mul(ctx0, d, b); - // v_t = v.unsqueeze(2) (we insert the singleton dimension after n_seqs and H_v) - ggml_tensor * v_t = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs); - // delta = (v_t - kv_mem) * beta_t - ggml_tensor * v_diff = ggml_sub(ctx0, v_t, kv_mem); // both should be [S_v, 1, H_v, n_seqs] - ggml_tensor * delta = ggml_mul(ctx0, v_diff, beta_t); + // [1, S_v, H_v, n_seqs] + ggml_tensor * d_t; + d_t = ggml_transpose(ctx0, d); - // last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta - ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta); - state = ggml_add(ctx0, state, k_t_delta); + // [S_v, S_v, H_v, n_seqs] + ggml_tensor * kd; + k = ggml_repeat(ctx0, k, s); + kd = ggml_mul (ctx0, k, d_t); - // Compute the attention output - // core_attn_out = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2) - ggml_tensor * q_t_unsqueezed = ggml_reshape_4d(ctx0, q, 1, S_v, H_v, n_seqs); // unsqueeze q_t - ggml_tensor * state_q = ggml_mul(ctx0, state, q_t_unsqueezed); - // again, since it's over dim = -2, transpose, sum, transpose back - ggml_tensor * core_attn_out = - ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, state_q)))); + s_t = ggml_add(ctx0, s_t, kd); - // core_attn_out should be [S_v, 1, H_v, n_seqs] after this - cb(core_attn_out, "output_tokens", il); - cb(state, "new_state", il); + cb(s_t, "dnet_add_ar_state", il); - return {core_attn_out, state}; + ggml_tensor * s_q = ggml_mul (ctx0, s_t, q); + ggml_tensor * o = ggml_sum_rows(ctx0, s_q); + + o = ggml_permute (ctx0, o, 2, 0, 1, 3); // [S_v, H_v, n_tokens, n_seqs] + s = ggml_transpose(ctx0, s_t); // [S_v, S_v, H_v, n_seqs] + + return {o, s}; } ggml_tensor * llm_build_qwen3next::build_norm_gated( @@ -472,39 +433,29 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn( // Split Q projection into query and gate // The split should be along dimension 0 (the feature dimension) ggml_tensor * Qcur = ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1, - Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], 0); + Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], 0); + cb(Qcur, "Qcur_view", il); + ggml_tensor * gate = ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1, Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], n_embd_head * ggml_element_size(Qcur_full)); - cb(Qcur, "Qcur", il); cb(gate, "gate", il); - // Now reshape Qcur to [n_embd_head, n_head, n_tokens] for multi-head attention - Qcur = ggml_cont_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - cb(Qcur, "Qcur_reshaped", il); - - // Apply Q normalization - Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il); - cb(Qcur, "Qcur_normed", il); - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); cb(Kcur, "Kcur", il); ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); cb(Vcur, "Vcur", il); - // Apply K normalization Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il); cb(Kcur, "Kcur_normed", il); - // Reshape gate to [n_embd, n_tokens] for the sigmoid gating (flatten the heads) - gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens); - cb(gate, "gate_reshaped", il); - - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - // Apply RoPE Qcur = ggml_rope_ext( ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, @@ -519,7 +470,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn( cb(Kcur, "Kcur", il); cb(Vcur, "Vcur", il); - // Attention computation const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale; cur = build_attn(inp, @@ -527,10 +477,15 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn( Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); cb(cur, "attn_pregate", il); - ggml_tensor * gate_sigmoid = ggml_sigmoid(ctx0, gate); - cb(gate_sigmoid, "gate_sigmoid", il); + // TODO: CUDA is missing non-contiguous unary ops. when implemented: remove this cont + gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens); - cur = ggml_mul(ctx0, cur, gate_sigmoid); + gate = ggml_sigmoid(ctx0, gate); + cb(gate, "gate_sigmoid", il); + + gate = ggml_reshape_2d(ctx0, gate, n_embd_head * n_head, n_tokens); + + cur = ggml_mul(ctx0, cur, gate); cb(cur, "attn_gated", il); cur = build_lora_mm(model.layers[il].wo, cur); @@ -560,7 +515,6 @@ std::pair llm_build_qwen3next::build_qkvz( cb(z, "z", il); return { qkv_mixed, z }; - } else { // legacy (slower) path ggml_tensor * mixed_qkvz = build_lora_mm(model.layers[il].ssm_in, input); @@ -624,9 +578,6 @@ std::pair llm_build_qwen3next::build_qkvz( ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( llm_graph_input_rs * inp, ggml_tensor * cur, - ggml_tensor * causal_mask, - ggml_tensor * identity, - ggml_tensor * diag_mask, int il) { const auto * mctx_cur = inp->mctx; @@ -671,7 +622,12 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( split_sizes_ba[0] * ggml_element_size(mixed_ba_reshaped)); cb(a, "a", il); - ggml_tensor * beta = ggml_cont_4d(ctx0, b, num_v_heads, 1, n_seq_tokens, n_seqs); + // TODO: CUDA is missing non-contiguous unary ops. when implemented: remove this cont + b = ggml_cont(ctx0, b); + + ggml_tensor * beta = ggml_sigmoid(ctx0, b); + + beta = ggml_reshape_4d(ctx0, beta, num_v_heads, 1, n_seq_tokens, n_seqs); // Reshape a to merge head dimensions: [batch, seq_len, num_k_heads, num_v_heads/num_k_heads] -> [batch, seq_len, num_v_heads] ggml_tensor * alpha = ggml_cont_3d(ctx0, a, num_v_heads, n_seq_tokens, n_seqs); @@ -679,6 +635,7 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( ggml_tensor * alpha_biased = ggml_add(ctx0, alpha, model.layers[il].ssm_dt); ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased); cb(alpha_softplus, "a_softplus", il); + ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a); // -A_log.exp() * softplus cb(gate, "gate", il); @@ -686,8 +643,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( ggml_tensor * conv_states_all = mctx_cur->get_r_l(il); ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il); - // bool use_precomputed_states = n_seq_tokens == 1 && mctx_cur->has_previous_state(); - // Build the convolution states tensor ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs); cb(conv_states, "conv_states", il); @@ -696,11 +651,12 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( ggml_tensor * conv_kernel = model.layers[il].ssm_conv1d; const int64_t conv_kernel_size = conv_kernel->ne[0]; const int64_t conv_channels = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state; - conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs); + + conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs); cb(conv_states, "conv_states_reshaped", il); - qkv_mixed = ggml_permute(ctx0, qkv_mixed, 1, 0, 2, 3); - cb(qkv_mixed, "qkv_mixed_permuted", il); + qkv_mixed = ggml_transpose(ctx0, qkv_mixed); + cb(qkv_mixed, "qkv_mixed_transposed", il); ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0); cb(conv_input, "conv_input", il); @@ -720,7 +676,10 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target)); cb(conv_states_all, "conv_states_updated", il); - // Apply SSM convolution + ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs); + state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs); + cb(state, "state_predelta", il); + ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel); cb(conv_output_proper, "conv_output_raw", il); @@ -734,26 +693,36 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim); // Extract the convolved Q, K, V from conv_output - ggml_tensor * q_conv = - ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, 0); + ggml_tensor * q_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs, + ggml_row_size(conv_qkv_mix->type, head_k_dim), + nb1_qkv, + nb1_qkv * n_seq_tokens, + 0); + + ggml_tensor * k_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs, + ggml_row_size(conv_qkv_mix->type, head_k_dim), + nb1_qkv, + nb1_qkv * n_seq_tokens, + head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix)); + + ggml_tensor * v_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_v_dim, num_v_heads, n_seq_tokens, n_seqs, + ggml_row_size(conv_qkv_mix->type, head_v_dim), + nb1_qkv, + nb1_qkv * n_seq_tokens, + ggml_row_size(conv_qkv_mix->type, 2 * head_k_dim * num_k_heads)); + cb(q_conv, "q_conv", il); - ggml_tensor * k_conv = - ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, - head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix)); cb(k_conv, "k_conv", il); - ggml_tensor * v_conv = - ggml_view_2d(ctx0, conv_qkv_mix, head_v_dim * num_v_heads, n_seq_tokens * n_seqs, nb1_qkv, - 2 * head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix)); cb(v_conv, "v_conv", il); - // Unsqueeze them - q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); - k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); - v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs); + const float eps_norm = hparams.f_norm_rms_eps; - ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs); - state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim * num_v_heads, 1, n_seqs); - cb(state, "state_predelta", il); + q_conv = ggml_l2_norm(ctx0, q_conv, eps_norm); + k_conv = ggml_l2_norm(ctx0, k_conv, eps_norm); + + //q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); + //k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); + //v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs); // if head keys and value keys are different, repeat to force tensors into matching shapes if (num_k_heads != num_v_heads) { @@ -786,7 +755,7 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( if (n_seq_tokens == 1) { attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il); } else { - attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, diag_mask, il); + attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, il); } ggml_tensor * output = attn_out.first; ggml_tensor * new_state = attn_out.second; @@ -795,19 +764,15 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( // Update the recurrent states ggml_build_forward_expand(gf, - ggml_cpy(ctx0, new_state, - ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs, - kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all)))); - - // Reshape both attn_out_final and z to 2D tensors for normalization - // attn_out_final: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim] - ggml_tensor * attn_out_2d_final = ggml_reshape_2d(ctx0, output, head_v_dim, num_v_heads * n_seq_tokens * n_seqs); + ggml_cpy(ctx0, new_state, + ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs, + kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all)))); // z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim] - ggml_tensor * z_2d = ggml_reshape_2d(ctx0, z, head_v_dim, num_v_heads * n_seq_tokens * n_seqs); + ggml_tensor * z_2d = ggml_reshape_4d(ctx0, z, head_v_dim, num_v_heads, n_seq_tokens, n_seqs); // Apply gated normalization: self.norm(core_attn_out, z) - ggml_tensor * attn_out_norm = build_norm_gated(attn_out_2d_final, model.layers[il].ssm_norm, z_2d, il); + ggml_tensor * attn_out_norm = build_norm_gated(output, model.layers[il].ssm_norm, z_2d, il); // Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim] ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs); @@ -818,7 +783,8 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( cb(cur, "linear_attn_out", il); // Reshape back to original dimensions - cur = ggml_cont_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs); + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs); + return cur; } @@ -839,7 +805,7 @@ ggml_tensor * llm_build_qwen3next::build_layer_ffn(ggml_tensor * cur, const int if (model.layers[il].ffn_up_shexp != nullptr) { ggml_tensor * ffn_shexp = build_ffn(cur, - model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_up_shexp, NULL, NULL, model.layers[il].ffn_gate_shexp, NULL, NULL, model.layers[il].ffn_down_shexp, NULL, NULL, NULL, @@ -852,11 +818,9 @@ ggml_tensor * llm_build_qwen3next::build_layer_ffn(ggml_tensor * cur, const int ggml_tensor * shared_gate = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur); cb(shared_gate, "shared_expert_gate", il); - // Apply sigmoid to the gate shared_gate = ggml_sigmoid(ctx0, shared_gate); cb(shared_gate, "shared_expert_gate_sigmoid", il); - // Apply the gate to the shared expert output ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate); cb(ffn_shexp, "ffn_shexp_gated", il); diff --git a/tools/mtmd/clip-impl.h b/tools/mtmd/clip-impl.h index 2e369c52c..7bb93f44f 100644 --- a/tools/mtmd/clip-impl.h +++ b/tools/mtmd/clip-impl.h @@ -236,6 +236,7 @@ enum projector_type { PROJECTOR_TYPE_GLM4V, PROJECTOR_TYPE_YOUTUVL, PROJECTOR_TYPE_KIMIK25, + PROJECTOR_TYPE_NEMOTRON_V2_VL, PROJECTOR_TYPE_UNKNOWN, }; @@ -270,6 +271,7 @@ static std::map PROJECTOR_TYPE_NAMES = { { PROJECTOR_TYPE_GLM4V, "glm4v"}, { PROJECTOR_TYPE_YOUTUVL, "youtuvl"}, { PROJECTOR_TYPE_KIMIK25, "kimik25"}, + { PROJECTOR_TYPE_NEMOTRON_V2_VL, "nemotron_v2_vl"}, }; static projector_type clip_projector_type_from_string(const std::string & str) { diff --git a/tools/mtmd/clip-model.h b/tools/mtmd/clip-model.h index d4ff9151b..e0eb9b32c 100644 --- a/tools/mtmd/clip-model.h +++ b/tools/mtmd/clip-model.h @@ -15,6 +15,7 @@ enum ffn_op_type { FFN_GELU_ERF, FFN_SILU, FFN_GELU_QUICK, + FFN_RELU_SQR, }; enum norm_type { diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp index 099b4c2f3..58f7f8eba 100644 --- a/tools/mtmd/clip.cpp +++ b/tools/mtmd/clip.cpp @@ -56,6 +56,7 @@ #include "models/internvl.cpp" #include "models/kimivl.cpp" #include "models/kimik25.cpp" +#include "models/nemotron-v2-vl.cpp" #include "models/llama4.cpp" #include "models/llava.cpp" #include "models/minicpmv.cpp" @@ -607,6 +608,12 @@ ggml_tensor * clip_graph::build_ffn( cur = ggml_gelu_quick(ctx0, cur); cb(cur, "ffn_gelu_quick", il); } break; + case FFN_RELU_SQR: + { + cur = ggml_relu(ctx0, cur); + cur = ggml_sqr(ctx0, cur); + cb(cur, "ffn_relu_sqr", il); + } break; } if (down) { @@ -858,6 +865,10 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 { builder = std::make_unique(ctx, img); } break; + case PROJECTOR_TYPE_NEMOTRON_V2_VL: + { + builder = std::make_unique(ctx, img); + } break; case PROJECTOR_TYPE_LLAMA4: { builder = std::make_unique(ctx, img); @@ -1182,6 +1193,7 @@ struct clip_model_loader { } } break; case PROJECTOR_TYPE_INTERNVL: + case PROJECTOR_TYPE_NEMOTRON_V2_VL: { get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false); } break; @@ -1844,6 +1856,12 @@ struct clip_model_loader { model.mm_3_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "weight")); model.mm_3_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "bias")); } break; + case PROJECTOR_TYPE_NEMOTRON_V2_VL: + { + model.mm_0_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 0, "weight")); + model.mm_1_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 1, "weight")); + model.mm_3_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "weight")); + } break; case PROJECTOR_TYPE_GLMA: { model.conv1d_1_w = get_tensor(string_format(TN_CONV1D, 1, "weight")); @@ -3290,6 +3308,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str case PROJECTOR_TYPE_GLM_EDGE: case PROJECTOR_TYPE_GEMMA3: case PROJECTOR_TYPE_INTERNVL: // TODO @ngxson : support dynamic resolution + case PROJECTOR_TYPE_NEMOTRON_V2_VL: { clip_image_u8 resized_image; int sz = params.image_size; @@ -3599,6 +3618,7 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im case PROJECTOR_TYPE_GEMMA3: case PROJECTOR_TYPE_IDEFICS3: case PROJECTOR_TYPE_INTERNVL: + case PROJECTOR_TYPE_NEMOTRON_V2_VL: case PROJECTOR_TYPE_LLAMA4: { // both X and Y are downscaled by the scale factor @@ -4007,6 +4027,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima case PROJECTOR_TYPE_GEMMA3NV: case PROJECTOR_TYPE_IDEFICS3: case PROJECTOR_TYPE_INTERNVL: + case PROJECTOR_TYPE_NEMOTRON_V2_VL: case PROJECTOR_TYPE_QWEN2A: case PROJECTOR_TYPE_GLMA: case PROJECTOR_TYPE_ULTRAVOX: @@ -4367,6 +4388,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) { case PROJECTOR_TYPE_MUSIC_FLAMINGO: return ctx->model.mm_2_w->ne[1]; case PROJECTOR_TYPE_INTERNVL: + case PROJECTOR_TYPE_NEMOTRON_V2_VL: return ctx->model.mm_3_w->ne[1]; case PROJECTOR_TYPE_LLAMA4: return ctx->model.mm_model_proj->ne[1]; diff --git a/tools/mtmd/models/models.h b/tools/mtmd/models/models.h index c4c67ace6..0beff16c5 100644 --- a/tools/mtmd/models/models.h +++ b/tools/mtmd/models/models.h @@ -42,6 +42,11 @@ struct clip_graph_internvl : clip_graph { ggml_cgraph * build() override; }; +struct clip_graph_nemotron_v2_vl : clip_graph { + clip_graph_nemotron_v2_vl(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {} + ggml_cgraph * build() override; +}; + struct clip_graph_llama4 : clip_graph { clip_graph_llama4(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {} ggml_cgraph * build() override; diff --git a/tools/mtmd/models/nemotron-v2-vl.cpp b/tools/mtmd/models/nemotron-v2-vl.cpp new file mode 100644 index 000000000..03094be1b --- /dev/null +++ b/tools/mtmd/models/nemotron-v2-vl.cpp @@ -0,0 +1,35 @@ +#include "models.h" + +ggml_cgraph * clip_graph_nemotron_v2_vl::build() { + GGML_ASSERT(model.class_embedding != nullptr); + GGML_ASSERT(model.position_embeddings != nullptr); + + const int n_registers = model.class_embedding->ne[1]; + const int n_pos = n_patches + n_registers; + + ggml_tensor * inp = build_inp(); + + // add position embeddings (pre-downsampled during GGUF conversion for fixed 512x512 input) + inp = ggml_add(ctx0, inp, model.position_embeddings); + cb(inp, "inp_pos", -1); + + inp = ggml_concat(ctx0, model.class_embedding, inp, 1); + + ggml_tensor * cur = build_vit(inp, n_pos, NORM_TYPE_NORMAL, hparams.ffn_op, nullptr, nullptr); + + cur = ggml_view_2d(ctx0, cur, + n_embd, n_patches, + ggml_row_size(cur->type, n_embd), + n_registers * ggml_row_size(cur->type, n_embd)); + + cur = build_patch_merge_permute(cur, model.hparams.n_merge); + + { + cur = build_norm(cur, model.mm_0_w, nullptr, NORM_TYPE_RMS, 1e-6, -1); + cur = build_ffn(cur, model.mm_1_w, nullptr, nullptr, nullptr, model.mm_3_w, nullptr, FFN_RELU_SQR, -1); + } + + ggml_build_forward_expand(gf, cur); + + return gf; +} diff --git a/tools/server/public/index.html.gz b/tools/server/public/index.html.gz index f4ff57b4c..75fc856f5 100644 Binary files a/tools/server/public/index.html.gz and b/tools/server/public/index.html.gz differ diff --git a/tools/server/webui/.storybook/main.ts b/tools/server/webui/.storybook/main.ts index bfd16fa22..4f6945f21 100644 --- a/tools/server/webui/.storybook/main.ts +++ b/tools/server/webui/.storybook/main.ts @@ -1,17 +1,24 @@ import type { StorybookConfig } from '@storybook/sveltekit'; +import { dirname, resolve } from 'path'; +import { fileURLToPath } from 'url'; + +const __dirname = dirname(fileURLToPath(import.meta.url)); const config: StorybookConfig = { stories: ['../tests/stories/**/*.mdx', '../tests/stories/**/*.stories.@(js|ts|svelte)'], addons: [ '@storybook/addon-svelte-csf', '@chromatic-com/storybook', - '@storybook/addon-docs', + '@storybook/addon-vitest', '@storybook/addon-a11y', - '@storybook/addon-vitest' + '@storybook/addon-docs' ], - framework: { - name: '@storybook/sveltekit', - options: {} + framework: '@storybook/sveltekit', + viteFinal: async (config) => { + config.server = config.server || {}; + config.server.fs = config.server.fs || {}; + config.server.fs.allow = [...(config.server.fs.allow || []), resolve(__dirname, '../tests')]; + return config; } }; export default config; diff --git a/tools/server/webui/.storybook/preview.ts b/tools/server/webui/.storybook/preview.ts index 8d530e43e..566dbfd28 100644 --- a/tools/server/webui/.storybook/preview.ts +++ b/tools/server/webui/.storybook/preview.ts @@ -13,7 +13,7 @@ const preview: Preview = { }, backgrounds: { - disable: true + disabled: true }, a11y: { diff --git a/tools/server/webui/docs/flows/settings-flow.md b/tools/server/webui/docs/flows/settings-flow.md index 474aef01b..40ad3bd94 100644 --- a/tools/server/webui/docs/flows/settings-flow.md +++ b/tools/server/webui/docs/flows/settings-flow.md @@ -49,14 +49,20 @@ sequenceDiagram settingsStore->>serverStore: defaultParams serverStore-->>settingsStore: {temperature, top_p, top_k, ...} - settingsStore->>ParamSvc: extractServerDefaults(defaultParams) - ParamSvc-->>settingsStore: Record + loop each SYNCABLE_PARAMETER + alt key NOT in userOverrides + settingsStore->>settingsStore: config[key] = serverDefault[key] + Note right of settingsStore: Non-overridden params adopt server default + else key in userOverrides + Note right of settingsStore: Keep user value, skip server default + end + end - settingsStore->>ParamSvc: mergeWithServerDefaults(config, serverDefaults) - Note right of ParamSvc: For each syncable parameter:
- If NOT in userOverrides → use server default
- If in userOverrides → keep user value - ParamSvc-->>settingsStore: mergedConfig + alt serverStore.props has webuiSettings + settingsStore->>settingsStore: Apply webuiSettings from server + Note right of settingsStore: Server-provided UI settings
(e.g. showRawOutputSwitch) + end - settingsStore->>settingsStore: config = mergedConfig settingsStore->>settingsStore: saveConfig() deactivate settingsStore @@ -67,11 +73,18 @@ sequenceDiagram UI->>settingsStore: updateConfig(key, value) activate settingsStore settingsStore->>settingsStore: config[key] = value - settingsStore->>settingsStore: userOverrides.add(key) - Note right of settingsStore: Mark as user-modified (won't be overwritten by server) + + alt value matches server default for key + settingsStore->>settingsStore: userOverrides.delete(key) + Note right of settingsStore: Matches server default, remove override + else value differs from server default + settingsStore->>settingsStore: userOverrides.add(key) + Note right of settingsStore: Mark as user-modified (won't be overwritten) + end + settingsStore->>settingsStore: saveConfig() - settingsStore->>LS: set("llama-config", config) - settingsStore->>LS: set("llama-userOverrides", [...userOverrides]) + settingsStore->>LS: set(CONFIG_LOCALSTORAGE_KEY, config) + settingsStore->>LS: set(USER_OVERRIDES_LOCALSTORAGE_KEY, [...userOverrides]) deactivate settingsStore UI->>settingsStore: updateMultipleConfig({key1: val1, key2: val2}) @@ -88,10 +101,9 @@ sequenceDiagram UI->>settingsStore: resetConfig() activate settingsStore - settingsStore->>settingsStore: config = SETTING_CONFIG_DEFAULT + settingsStore->>settingsStore: config = {...SETTING_CONFIG_DEFAULT} settingsStore->>settingsStore: userOverrides.clear() - settingsStore->>settingsStore: syncWithServerDefaults() - Note right of settingsStore: Apply server defaults for syncable params + Note right of settingsStore: All params reset to defaults
Next syncWithServerDefaults will adopt server values settingsStore->>settingsStore: saveConfig() deactivate settingsStore diff --git a/tools/server/webui/src/lib/components/app/actions/ActionIconsCodeBlock.svelte b/tools/server/webui/src/lib/components/app/actions/ActionIconsCodeBlock.svelte index 54ff0af1a..b20e79b5e 100644 --- a/tools/server/webui/src/lib/components/app/actions/ActionIconsCodeBlock.svelte +++ b/tools/server/webui/src/lib/components/app/actions/ActionIconsCodeBlock.svelte @@ -1,6 +1,6 @@ + +
+ + + + + + + + +

{triggerTooltipText}

+ + + + + + {#each actions as item (item.id)} + {@const hasDisabledTooltip = !!item.disabled && !!item.disabledReason} + {@const hasEnabledTooltip = !item.disabled && !!item.tooltip} + + {#if hasDisabledTooltip} + + + + {#if item.id === 'images'} + + {:else if item.id === 'audio'} + + {:else if item.id === 'text'} + + {:else if item.id === 'pdf'} + + {:else} + + {/if} + + {item.label} + + + + +

{item.disabledReason}

+ + + {:else if hasEnabledTooltip} + + + handleActionClick(item.id)}> + {#if item.id === 'images'} + + {:else if item.id === 'audio'} + + {:else if item.id === 'text'} + + {:else if item.id === 'pdf'} + + {:else} + + {/if} + + {item.label} + + + + +

{item.tooltip}

+ + + {:else} + handleActionClick(item.id)}> + {#if item.id === 'images'} + + {:else if item.id === 'audio'} + + {:else if item.id === 'text'} + + {:else if item.id === 'pdf'} + + {:else} + + {/if} + + {item.label} + + {/if} + {/each} + + +
diff --git a/tools/server/webui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActions.svelte b/tools/server/webui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActions.svelte index c621a69e0..cf5aca42a 100644 --- a/tools/server/webui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActions.svelte +++ b/tools/server/webui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActions.svelte @@ -2,7 +2,7 @@ import { Square } from '@lucide/svelte'; import { Button } from '$lib/components/ui/button'; import { - ChatFormActionFileAttachments, + ChatFormActionAttachmentsDropdown, ChatFormActionRecord, ChatFormActionSubmit, ModelsSelector @@ -157,7 +157,7 @@ const { handleModelChange } = useModelChangeValidation({ getRequiredModalities: () => usedModalities(), - onValidationFailure: async (previousModelId) => { + onValidationFailure: async (previousModelId: string | null) => { if (previousModelId) { await modelsStore.selectModelById(previousModelId); } @@ -166,32 +166,39 @@
- +
+ +
- +
+ +
{#if isLoading} {:else if shouldShowRecordButton} diff --git a/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessage.svelte b/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessage.svelte index 3470e2f71..25895c83b 100644 --- a/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessage.svelte +++ b/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessage.svelte @@ -62,8 +62,8 @@ assistantMessages: number; messageTypes: string[]; } | null>(null); - let editedContent = $state(message.content); - let editedExtras = $state(message.extra ? [...message.extra] : []); + let editedContent = $derived(message.content); + let editedExtras = $derived(message.extra ? [...message.extra] : []); let editedUploadedFiles = $state([]); let isEditing = $state(false); let showDeleteDialog = $state(false); diff --git a/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageAssistant.svelte b/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageAssistant.svelte index 1cb6b274b..867def5fc 100644 --- a/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageAssistant.svelte +++ b/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageAssistant.svelte @@ -105,7 +105,7 @@ const { handleModelChange } = useModelChangeValidation({ getRequiredModalities: () => conversationsStore.getModalitiesUpToMessage(message.id), - onSuccess: (modelName) => onRegenerate(modelName) + onSuccess: (modelName: string) => onRegenerate(modelName) }); function handleCopyModel() { diff --git a/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageEditForm.svelte b/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageEditForm.svelte index f812ea2fd..c216ea690 100644 --- a/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageEditForm.svelte +++ b/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageEditForm.svelte @@ -133,7 +133,7 @@ const { handleModelChange } = useModelChangeValidation({ getRequiredModalities, - onValidationFailure: async (previousModelId) => { + onValidationFailure: async (previousModelId: string | null) => { if (previousModelId) { await modelsStore.selectModelById(previousModelId); } diff --git a/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageStatistics.svelte b/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageStatistics.svelte index d457e042f..b53e82aaf 100644 --- a/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageStatistics.svelte +++ b/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageStatistics.svelte @@ -28,7 +28,7 @@ initialView = ChatMessageStatsView.GENERATION }: Props = $props(); - let activeView: ChatMessageStatsView = $state(initialView); + let activeView: ChatMessageStatsView = $derived(initialView); let hasAutoSwitchedToGeneration = $state(false); // In live mode: auto-switch to GENERATION tab when prompt processing completes diff --git a/tools/server/webui/src/lib/components/app/chat/ChatScreen/ChatScreen.svelte b/tools/server/webui/src/lib/components/app/chat/ChatScreen/ChatScreen.svelte index a6e8d6123..7ecba030b 100644 --- a/tools/server/webui/src/lib/components/app/chat/ChatScreen/ChatScreen.svelte +++ b/tools/server/webui/src/lib/components/app/chat/ChatScreen/ChatScreen.svelte @@ -35,6 +35,7 @@ import { modelsStore, modelOptions, selectedModelId } from '$lib/stores/models.svelte'; import { isFileTypeSupported, filterFilesByModalities } from '$lib/utils'; import { parseFilesToMessageExtras, processFilesToChatUploaded } from '$lib/utils/browser-only'; + import { ErrorDialogType } from '$lib/enums'; import { onMount } from 'svelte'; import { fade, fly, slide } from 'svelte/transition'; import { Trash2, AlertTriangle, RefreshCw } from '@lucide/svelte'; @@ -616,7 +617,7 @@ contextInfo={activeErrorDialog?.contextInfo} onOpenChange={handleErrorDialogOpenChange} open={Boolean(activeErrorDialog)} - type={activeErrorDialog?.type ?? 'server'} + type={(activeErrorDialog?.type as ErrorDialogType) ?? ErrorDialogType.SERVER} /> diff --git a/tools/server/webui/src/lib/components/app/misc/RemoveButton.svelte b/tools/server/webui/src/lib/components/app/misc/RemoveButton.svelte deleted file mode 100644 index 173685510..000000000 --- a/tools/server/webui/src/lib/components/app/misc/RemoveButton.svelte +++ /dev/null @@ -1,26 +0,0 @@ - - - diff --git a/tools/server/webui/src/lib/components/app/misc/SyntaxHighlightedCode.svelte b/tools/server/webui/src/lib/components/app/misc/SyntaxHighlightedCode.svelte deleted file mode 100644 index bc42f9dd1..000000000 --- a/tools/server/webui/src/lib/components/app/misc/SyntaxHighlightedCode.svelte +++ /dev/null @@ -1,97 +0,0 @@ - - -
- - 
{@html highlightedHtml}
-
- - diff --git a/tools/server/webui/src/lib/components/app/models/ModelBadge.svelte b/tools/server/webui/src/lib/components/app/models/ModelBadge.svelte index bea1bf6e3..f98ba7d78 100644 --- a/tools/server/webui/src/lib/components/app/models/ModelBadge.svelte +++ b/tools/server/webui/src/lib/components/app/models/ModelBadge.svelte @@ -1,6 +1,6 @@