diff --git a/common/chat.cpp b/common/chat.cpp index 194299900a..0a426f4478 100644 --- a/common/chat.cpp +++ b/common/chat.cpp @@ -601,11 +601,6 @@ common_chat_templates_ptr common_chat_templates_init( "{%- if false %}"); } - // TODO @ngxson : hot fix for PaddleOCR - if (default_template_src.find("<|IMAGE_PLACEHOLDER|>") != std::string::npos) { - string_replace_all(default_template_src, "<|IMAGE_START|><|IMAGE_PLACEHOLDER|><|IMAGE_END|>", ""); - } - std::string token_bos = bos_token_override; std::string token_eos = eos_token_override; bool add_bos = false; diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 7391e51fce..83eb8a7e7c 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -3602,13 +3602,20 @@ class LLaDAModel(TextModel): yield from super().modify_tensors(data_torch, name, bid) -@ModelBase.register("Ernie4_5_ForCausalLM", "Ernie4_5ForCausalLM", "PaddleOCRVLForConditionalGeneration") +@ModelBase.register("Ernie4_5_ForCausalLM", "Ernie4_5ForCausalLM") class Ernie4_5Model(TextModel): model_arch = gguf.MODEL_ARCH.ERNIE4_5 def set_vocab(self): self._set_vocab_sentencepiece() + tokenizer_config_file = self.dir_model / 'tokenizer_config.json' + if tokenizer_config_file.is_file(): + with open(tokenizer_config_file, "r", encoding="utf-8") as f: + tokenizer_config_json = json.load(f) + if "add_prefix_space" in tokenizer_config_json: + self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"]) + def set_gguf_parameters(self): super().set_gguf_parameters() @@ -3739,6 +3746,12 @@ class Ernie4_5MoeModel(Ernie4_5Model): if len(experts) > 0: raise ValueError(f"Unprocessed experts: {experts}") + +@ModelBase.register("PaddleOCRVLForConditionalGeneration") +class PaddleOCRModel(Ernie4_5Model): + model_arch = gguf.MODEL_ARCH.PADDLEOCR + + @ModelBase.register("PaddleOCRVisionModel") class PaddleOCRVisionModel(MmprojModel): # PaddleOCR-VL uses a modified version of Siglip @@ -3776,6 +3789,7 @@ class PaddleOCRVisionModel(MmprojModel): return [(self.map_tensor_name(name), data_torch)] return [] # skip other tensors + @ModelBase.register( "Qwen2VLModel", "Qwen2VLForConditionalGeneration", diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 1ff70cbf49..197292206e 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -449,6 +449,7 @@ class MODEL_ARCH(IntEnum): RND1 = auto() PANGU_EMBED = auto() MISTRAL3 = auto() + PADDLEOCR = auto() class VISION_PROJECTOR_TYPE(IntEnum): @@ -826,6 +827,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.RND1: "rnd1", MODEL_ARCH.PANGU_EMBED: "pangu-embedded", MODEL_ARCH.MISTRAL3: "mistral3", + MODEL_ARCH.PADDLEOCR: "paddleocr", } VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = { @@ -2827,6 +2829,20 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, ], + MODEL_ARCH.PADDLEOCR: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], MODEL_ARCH.FALCON_H1: [ # Token embedding MODEL_TENSOR.TOKEN_EMBD, diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt index 4192af7c0c..353a4ed608 100644 --- a/src/CMakeLists.txt +++ b/src/CMakeLists.txt @@ -60,6 +60,7 @@ add_library(llama models/dream.cpp models/ernie4-5-moe.cpp models/ernie4-5.cpp + models/paddleocr.cpp models/exaone.cpp models/exaone4.cpp models/falcon-h1.cpp diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index 8caf80afcf..48eaefcb49 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -114,6 +114,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_RND1, "rnd1" }, { LLM_ARCH_PANGU_EMBED, "pangu-embedded" }, { LLM_ARCH_MISTRAL3, "mistral3" }, + { LLM_ARCH_PADDLEOCR, "paddleocr" }, { LLM_ARCH_UNKNOWN, "(unknown)" }, }; @@ -700,6 +701,7 @@ static std::set llm_get_tensor_names(llm_arch arch) { case LLM_ARCH_INTERNLM2: case LLM_ARCH_GRANITE: case LLM_ARCH_ERNIE4_5: + case LLM_ARCH_PADDLEOCR: case LLM_ARCH_SMOLLM3: case LLM_ARCH_DREAM: case LLM_ARCH_LLADA: diff --git a/src/llama-arch.h b/src/llama-arch.h index 6cbf9b1f89..0ce3993886 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -118,6 +118,7 @@ enum llm_arch { LLM_ARCH_RND1, LLM_ARCH_PANGU_EMBED, LLM_ARCH_MISTRAL3, + LLM_ARCH_PADDLEOCR, LLM_ARCH_UNKNOWN, }; diff --git a/src/llama-model.cpp b/src/llama-model.cpp index c9a3c5dfa2..f83991ef62 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -2056,7 +2056,11 @@ void llama_model::load_hparams(llama_model_loader & ml) { } break; case LLM_ARCH_ERNIE4_5: case LLM_ARCH_ERNIE4_5_MOE: + case LLM_ARCH_PADDLEOCR: { + // paddleocr need mrope_section + ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); if (arch == LLM_ARCH_ERNIE4_5_MOE) { ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); @@ -5964,6 +5968,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) { } break; case LLM_ARCH_ERNIE4_5: case LLM_ARCH_ERNIE4_5_MOE: + case LLM_ARCH_PADDLEOCR: { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -7574,6 +7579,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const { { llm = std::make_unique(*this, params); } break; + case LLM_ARCH_PADDLEOCR: + { + llm = std::make_unique(*this, params); + } break; case LLM_ARCH_HUNYUAN_MOE: { llm = std::make_unique(*this, params); @@ -7866,6 +7875,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { return LLAMA_ROPE_TYPE_NEOX; case LLM_ARCH_QWEN2VL: + case LLM_ARCH_PADDLEOCR: return LLAMA_ROPE_TYPE_MROPE; case LLM_ARCH_QWEN3VL: case LLM_ARCH_QWEN3VLMOE: diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp index 7b01a2edfe..b00885670e 100644 --- a/src/llama-vocab.cpp +++ b/src/llama-vocab.cpp @@ -2359,6 +2359,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { || t.first == "<|call|>" // o200k_harmony || t.first == "" || t.first == "<|endoftext|>" + || t.first == "" // paddleocr || t.first == "<|eom_id|>" || t.first == "" || t.first == "_" diff --git a/src/models/models.h b/src/models/models.h index ffb36acc61..dcef35473c 100644 --- a/src/models/models.h +++ b/src/models/models.h @@ -158,6 +158,10 @@ struct llm_build_ernie4_5_moe : public llm_graph_context { llm_build_ernie4_5_moe(const llama_model & model, const llm_graph_params & params); }; +struct llm_build_paddleocr : public llm_graph_context { + llm_build_paddleocr(const llama_model & model, const llm_graph_params & params); +}; + template struct llm_build_exaone4 : public llm_graph_context { llm_build_exaone4(const llama_model & model, const llm_graph_params & params); diff --git a/src/models/paddleocr.cpp b/src/models/paddleocr.cpp new file mode 100644 index 0000000000..1f6336eb97 --- /dev/null +++ b/src/models/paddleocr.cpp @@ -0,0 +1,119 @@ +#include "models.h" + +llm_build_paddleocr::llm_build_paddleocr(const llama_model & model, const llm_graph_params & params) : + llm_graph_context(params) { + 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); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + int sections[4]; + std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + { + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + } + // self-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + 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 = ggml_rope_multi( + ctx0, Qcur, inp_pos, nullptr, + n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_multi( + ctx0, Kcur, inp_pos, nullptr, + n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1) { + // skip computing output for unused tokens + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(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, il); + cb(cur, "ffn_out", il); + } + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/tools/mtmd/clip-model.h b/tools/mtmd/clip-model.h index 8ef04b6b2b..f5c41ff138 100644 --- a/tools/mtmd/clip-model.h +++ b/tools/mtmd/clip-model.h @@ -39,7 +39,6 @@ struct clip_hparams { int32_t image_min_pixels = -1; int32_t image_max_pixels = -1; int32_t n_merge = 0; // number of patch merges **per-side** - int32_t proj_scale_factor = 0; float image_mean[3]; float image_std[3]; diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp index cc022ad2e0..9142196639 100644 --- a/tools/mtmd/clip.cpp +++ b/tools/mtmd/clip.cpp @@ -1193,9 +1193,10 @@ struct clip_model_loader { } break; case PROJECTOR_TYPE_PADDLEOCR: { - hparams.proj_scale_factor = 2; - hparams.set_limit_image_tokens(8, 1024); - hparams.set_warmup_n_tokens(256); // avoid OOM on warmup + hparams.n_merge = 2; + // TODO(megemini): paddleocr vl not specified? + hparams.set_limit_image_tokens(8, 4096); + hparams.set_warmup_n_tokens(28*28); // avoid OOM on warmup } break; default: break; @@ -1460,7 +1461,7 @@ struct clip_model_loader { model.mm_model_ln_kv_w = get_tensor(string_format(TN_RESAMPL_LN, "kv", "weight")); model.mm_model_ln_kv_b = get_tensor(string_format(TN_RESAMPL_LN, "kv", "bias")); model.mm_model_ln_post_w = get_tensor(string_format(TN_RESAMPL_LN, "post", "weight")); - model.mm_model_ln_post_b = get_tensor(string_format(TN_RESAMPL_LN, "post", "bias")); + model.mm_model_ln_post_b = get_tensor(string_format(TN_RESAMPL_LN, "post", "bias")); } break; case PROJECTOR_TYPE_GLM_EDGE: { @@ -2869,6 +2870,7 @@ int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 * case PROJECTOR_TYPE_QWEN25VL: case PROJECTOR_TYPE_QWEN3VL: case PROJECTOR_TYPE_GLM4V: + case PROJECTOR_TYPE_PADDLEOCR: return (img->nx / params.patch_size) / 2; default: break; @@ -2884,6 +2886,7 @@ int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 * case PROJECTOR_TYPE_QWEN25VL: case PROJECTOR_TYPE_QWEN3VL: case PROJECTOR_TYPE_GLM4V: + case PROJECTOR_TYPE_PADDLEOCR: return (img->ny / params.patch_size) / 2; default: break; @@ -2971,8 +2974,8 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im case PROJECTOR_TYPE_PADDLEOCR: { // dynamic size - int scale_factor = ctx->model.hparams.proj_scale_factor; - int stride = scale_factor * scale_factor; + int n_merge = ctx->model.hparams.n_merge; + int stride = n_merge * n_merge; n_patches = CLIP_ALIGN(n_patches, stride) / stride; } break; case PROJECTOR_TYPE_PIXTRAL: @@ -3219,6 +3222,29 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima } } + set_input_i32("positions", positions); + } break; + case PROJECTOR_TYPE_PADDLEOCR: + { + const int merge_ratio = hparams.n_merge; + const int pw = image_size_width / patch_size; + const int ph = image_size_height / patch_size; + std::vector positions(n_pos * 4); + int ptr = 0; + for (int dy = 0; dy < 2; dy++) { + for (int y = 0; y < ph; y += merge_ratio) { + for (int x = 0; x < pw; x += merge_ratio) { + for (int dx = 0; dx < 2; dx++) { + positions[ ptr] = y + dy; + positions[ num_patches + ptr] = x + dx; + positions[2 * num_patches + ptr] = y + dy; + positions[3 * num_patches + ptr] = x + dx; + ptr++; + } + } + } + } + set_input_i32("positions", positions); } break; case PROJECTOR_TYPE_QWEN25VL: @@ -3304,7 +3330,6 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima } break; case PROJECTOR_TYPE_PIXTRAL: case PROJECTOR_TYPE_KIMIVL: - case PROJECTOR_TYPE_PADDLEOCR: case PROJECTOR_TYPE_LIGHTONOCR: { // set the 2D positions @@ -3499,6 +3524,7 @@ bool clip_is_mrope(const struct clip_ctx * ctx) { return ctx->proj_type() == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type() == PROJECTOR_TYPE_QWEN25VL || ctx->proj_type() == PROJECTOR_TYPE_QWEN3VL + || ctx->proj_type() == PROJECTOR_TYPE_PADDLEOCR || ctx->proj_type() == PROJECTOR_TYPE_GLM4V; } diff --git a/tools/mtmd/models/paddleocr.cpp b/tools/mtmd/models/paddleocr.cpp index ed8988ffef..92356f154b 100644 --- a/tools/mtmd/models/paddleocr.cpp +++ b/tools/mtmd/models/paddleocr.cpp @@ -1,23 +1,23 @@ #include "models.h" ggml_cgraph * clip_graph_paddleocr::build() { - // 2D input positions - ggml_tensor * pos_h = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches); - ggml_set_name(pos_h, "pos_h"); - ggml_set_input(pos_h); + const int n_pos = n_patches; + const int num_position_ids = n_pos * 4; // m-rope requires 4 dim per position - ggml_tensor * pos_w = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches); - ggml_set_name(pos_w, "pos_w"); - ggml_set_input(pos_w); + int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4}; - ggml_tensor * learned_pos_embd = resize_position_embeddings(); + ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_position_ids); + ggml_set_name(positions, "positions"); + ggml_set_input(positions); - // build ViT with 2D position embeddings auto add_pos = [&](ggml_tensor * cur, const clip_layer &) { - // first half is X axis and second half is Y axis - return build_rope_2d(ctx0, cur, pos_w, pos_h, hparams.rope_theta, false); + return ggml_rope_multi( + ctx0, cur, positions, nullptr, + d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, + 32768, 10000, 1, 0, 1, 32, 1); }; + ggml_tensor * learned_pos_embd = resize_position_embeddings(); ggml_tensor * inp = build_inp(); ggml_tensor * cur = build_vit( inp, n_patches, @@ -29,28 +29,16 @@ ggml_cgraph * clip_graph_paddleocr::build() { cb(cur, "vit_out", -1); { - // mlp_AR - float proj_norm_eps = 1e-5; // PaddleOCR uses hard-coded value eps=1e-5 for Projector + // mlp_AR paddleocr projector + float proj_norm_eps = 1e-5; cur = build_norm(cur, model.mm_input_norm_w, model.mm_input_norm_b, NORM_TYPE_NORMAL, proj_norm_eps, -1); - //cur = build_patch_merge_permute(cur, hparams.proj_scale_factor); - // stack and padding - int64_t stride = hparams.proj_scale_factor * hparams.proj_scale_factor; - int64_t n_embd = cur->ne[0]; - int64_t n_tokens = cur->ne[1]; - int64_t n_tokens_padded = CLIP_ALIGN(n_tokens, stride); - int64_t n_pad = n_tokens_padded - n_tokens; - if (n_pad > 0) { - cur = ggml_view_1d(ctx0, cur, ggml_nelements(cur), 0); - cur = ggml_pad(ctx0, cur, n_pad * n_embd, 0, 0, 0); - } - cur = ggml_view_2d(ctx0, cur, - n_embd * stride, - n_tokens_padded / stride, - ggml_row_size(cur->type, n_embd * stride), 0); - cb(cur, "after_stacked", -1); + const int scale_factor = model.hparams.n_merge; + int width = img.nx / patch_size; + int height = img.ny / patch_size; + cur = ggml_reshape_3d(ctx0, cur, n_embd * scale_factor * scale_factor, width / scale_factor * height / scale_factor, 1); cur = build_ffn(cur, model.mm_1_w, model.mm_1_b,