From 112337630968d832a1ce295814fb92b858164f9e Mon Sep 17 00:00:00 2001 From: caitianchi Date: Mon, 12 Aug 2024 21:11:25 +0800 Subject: [PATCH] fix convert script and readme --- examples/llava/README-minicpmv2.5.md | 4 +- examples/llava/README-minicpmv2.6.md | 4 +- ...minicpmv-convert-image-encoder-to-gguf.py} | 15 +- ...nicpmv2_5-convert-image-encoder-to-gguf.py | 385 ------------------ .../minicpmv-surgery.py | 0 5 files changed, 16 insertions(+), 392 deletions(-) rename examples/llava/{minicpmv-convert/minicpmv2_6-convert-image-encoder-to-gguf.py => minicpmv-convert-image-encoder-to-gguf.py} (98%) delete mode 100644 examples/llava/minicpmv-convert/minicpmv2_5-convert-image-encoder-to-gguf.py rename examples/llava/{minicpmv-convert => }/minicpmv-surgery.py (100%) diff --git a/examples/llava/README-minicpmv2.5.md b/examples/llava/README-minicpmv2.5.md index 166b9e2e98cf2..62009b0af3a9b 100644 --- a/examples/llava/README-minicpmv2.5.md +++ b/examples/llava/README-minicpmv2.5.md @@ -15,8 +15,8 @@ cd llama.cpp Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5-gguf) by us) ```bash -python ./examples/minicpmv/minicpmv-convert/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5 -python ./examples/minicpmv/minicpmv-convert/minicpmv2_5-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 +python ./examples/minicpmv/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5 +python ./examples/minicpmv/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 2 python ./convert_hf_to_gguf.py ../MiniCPM-Llama3-V-2_5/model # quantize int4 version diff --git a/examples/llava/README-minicpmv2.6.md b/examples/llava/README-minicpmv2.6.md index 6af84abb6794d..c4be5e5dd6484 100644 --- a/examples/llava/README-minicpmv2.6.md +++ b/examples/llava/README-minicpmv2.6.md @@ -16,8 +16,8 @@ git checkout minicpmv-main Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-V-2_6-gguf) by us) ```bash -python ./examples/llava/minicpmv-convert/minicpmv-surgery.py -m ../MiniCPM-V-2_6 -python ./examples/llava/minicpmv-convert/minicpmv2_6-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-2_6 --minicpmv-projector ../MiniCPM-V-2_6/minicpmv.projector --output-dir ../MiniCPM-V-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 +python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-V-2_6 +python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-2_6 --minicpmv-projector ../MiniCPM-V-2_6/minicpmv.projector --output-dir ../MiniCPM-V-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 3 python ./convert_hf_to_gguf.py ../MiniCPM-V-2_6/model # quantize int4 version diff --git a/examples/llava/minicpmv-convert/minicpmv2_6-convert-image-encoder-to-gguf.py b/examples/llava/minicpmv-convert-image-encoder-to-gguf.py similarity index 98% rename from examples/llava/minicpmv-convert/minicpmv2_6-convert-image-encoder-to-gguf.py rename to examples/llava/minicpmv-convert-image-encoder-to-gguf.py index 169b40508dbe1..7d68359656e67 100644 --- a/examples/llava/minicpmv-convert/minicpmv2_6-convert-image-encoder-to-gguf.py +++ b/examples/llava/minicpmv-convert-image-encoder-to-gguf.py @@ -500,6 +500,7 @@ def bytes_to_unicode(): default_image_std = [0.26862954, 0.26130258, 0.27577711] ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None) ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None) +ap.add_argument('--minicpmv_version', type=int, help='minicpmv_version: MiniCPM-V-2 use 1; MiniCPM-V-2.5 use 2; MiniCPM-V-2.6 use 3', default=2) # with proper args = ap.parse_args() @@ -565,7 +566,15 @@ def bytes_to_unicode(): has_text_encoder = True has_vision_encoder = True has_minicpmv_projector = False -minicpmv_version = 3 +minicpmv_version = args.minicpmv_version +emb_dim = 4096 +if minicpmv_version == 1: + emb_dim = 2304 +elif minicpmv_version == 2: + emb_dim = 4096 +elif minicpmv_version == 3: + emb_dim = 3584 + if args.text_only: fname_middle = "text-" has_vision_encoder = False @@ -684,11 +693,11 @@ def _replace_name_resampler(s, v): if re.match("resampler.pos_embed", s): return { s: v, - re.sub("pos_embed", "pos_embed_k", s): torch.from_numpy(get_2d_sincos_pos_embed(3584, (70, 70))), + re.sub("pos_embed", "pos_embed_k", s): torch.from_numpy(get_2d_sincos_pos_embed(emb_dim, (70, 70))), } if re.match("resampler.proj", s): return { - re.sub("proj", "pos_embed_k", s): torch.from_numpy(get_2d_sincos_pos_embed(3584, (70, 70))), + re.sub("proj", "pos_embed_k", s): torch.from_numpy(get_2d_sincos_pos_embed(emb_dim, (70, 70))), re.sub("proj", "proj.weight", s): v.transpose(-1, -2).contiguous(), } if re.match("resampler.attn.in_proj_.*", s): diff --git a/examples/llava/minicpmv-convert/minicpmv2_5-convert-image-encoder-to-gguf.py b/examples/llava/minicpmv-convert/minicpmv2_5-convert-image-encoder-to-gguf.py deleted file mode 100644 index 781a7f7a6f170..0000000000000 --- a/examples/llava/minicpmv-convert/minicpmv2_5-convert-image-encoder-to-gguf.py +++ /dev/null @@ -1,385 +0,0 @@ -import argparse -import os -import json -import re - -import torch -import numpy as np -from gguf import * -from transformers.models.idefics2.modeling_idefics2 import Idefics2VisionTransformer, Idefics2VisionConfig - -TEXT = "clip.text" -VISION = "clip.vision" - - -def add_key_str(raw_key: str, arch: str) -> str: - return raw_key.format(arch=arch) - - -def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_minicpmv: bool) -> bool: - if name in ( - "logit_scale", - "text_model.embeddings.position_ids", - "vision_model.embeddings.position_ids", - ): - return True - - if has_minicpmv and name in ["visual_projection.weight"]: - return True - - if name.startswith("v") and not has_vision: - return True - - if name.startswith("t") and not has_text: - return True - - return False - - -def get_tensor_name(name: str) -> str: - if "projection" in name: - return name - if "mm_projector" in name: - name = name.replace("model.mm_projector", "mm") - name = re.sub(r'mm\.mlp\.mlp', 'mm.model.mlp', name, count=1) - name = re.sub(r'mm\.peg\.peg', 'mm.model.peg', name, count=1) - return name - - return name.replace("text_model", "t").replace("vision_model", "v").replace("encoder.layers", "blk").replace("embeddings.", "").replace("_proj", "").replace("self_attn.", "attn_").replace("layer_norm", "ln").replace("layernorm", "ln").replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("embedding", "embd").replace("final", "post").replace("layrnorm", "ln") - - -def bytes_to_unicode(): - """ - Returns list of utf-8 byte and a corresponding list of unicode strings. - The reversible bpe codes work on unicode strings. - This means you need a large # of unicode characters in your vocab if you want to avoid UNKs. - When you're at something like a 10B token dataset you end up needing around 5K for decent coverage. - This is a significant percentage of your normal, say, 32K bpe vocab. - To avoid that, we want lookup tables between utf-8 bytes and unicode strings. - And avoids mapping to whitespace/control characters the bpe code barfs on. - """ - bs = ( - list(range(ord("!"), ord("~") + 1)) - + list(range(ord("¡"), ord("¬") + 1)) - + list(range(ord("®"), ord("ÿ") + 1)) - ) - cs = bs[:] - n = 0 - for b in range(2**8): - if b not in bs: - bs.append(b) - cs.append(2**8 + n) - n += 1 - cs = [chr(n) for n in cs] - return dict(zip(bs, cs)) - - -ap = argparse.ArgumentParser() -ap.add_argument("-m", "--model-dir", help="Path to model directory cloned from HF Hub", required=True) -ap.add_argument("--use-f32", action="store_true", default=False, help="Use f32 instead of f16") -ap.add_argument("--text-only", action="store_true", required=False, - help="Save a text-only model. It can't be used to encode images") -ap.add_argument("--vision-only", action="store_true", required=False, - help="Save a vision-only model. It can't be used to encode texts") -ap.add_argument("--clip-model-is-vision", action="store_true", required=False, - help="The clip model is a pure vision model (ShareGPT4V vision extract for example)") -ap.add_argument("--clip-model-is-openclip", action="store_true", required=False, - help="The clip model is from openclip (for ViT-SO400M type))") -ap.add_argument("--minicpmv-projector", help="Path to minicpmv.projector file. If specified, save an image encoder for MiniCPM-V models.") -ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp, ldpv2", choices=["mlp", "ldp", "ldpv2"], default="mlp") -ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None) -# Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711 -# Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5 -default_image_mean = [0.48145466, 0.4578275, 0.40821073] -default_image_std = [0.26862954, 0.26130258, 0.27577711] -ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None) -ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None) - -# with proper -args = ap.parse_args() - - -if args.text_only and args.vision_only: - print("--text-only and --image-only arguments cannot be specified at the same time.") - exit(1) - -if args.use_f32: - print("WARNING: Weights for the convolution op is always saved in f16, as the convolution op in GGML does not support 32-bit kernel weights yet.") - -# output in the same directory as the model if output_dir is None -dir_model = args.model_dir - -if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip: - vocab = None - tokens = None -else: - with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f: - vocab = json.load(f) - tokens = [key for key in vocab] - -# possible data types -# ftype == 0 -> float32 -# ftype == 1 -> float16 -# -# map from ftype to string -ftype_str = ["f32", "f16"] - -ftype = 1 -if args.use_f32: - ftype = 0 - -# if args.clip_model_is_vision or args.clip_model_is_openclip: -# model = CLIPVisionModel.from_pretrained(dir_model) -# processor = None -# else: -# model = CLIPModel.from_pretrained(dir_model) -# processor = CLIPProcessor.from_pretrained(dir_model) - -default_vision_config = { - "hidden_size": 1152, - "image_size": 980, - "intermediate_size": 4304, - "model_type": "idefics2", - "num_attention_heads": 16, - "num_hidden_layers": 27, - "patch_size": 14, - } -vision_config = Idefics2VisionConfig(**default_vision_config) -model = Idefics2VisionTransformer(vision_config) - -processor = None -# if model.attn_pool is not None: -# model.attn_pool = torch.nn.Identity() - -# model.blocks = model.blocks[:-1] -model.load_state_dict(torch.load(os.path.join(dir_model, "minicpmv.clip"))) - -fname_middle = None -has_text_encoder = True -has_vision_encoder = True -has_minicpmv_projector = False -minicpmv_version = 2 -if args.text_only: - fname_middle = "text-" - has_vision_encoder = False -elif args.minicpmv_projector is not None: - fname_middle = "mmproj-" - has_text_encoder = False - has_minicpmv_projector = True - minicpmv_version = 2 -elif args.vision_only: - fname_middle = "vision-" - has_text_encoder = False -else: - fname_middle = "" - -output_dir = args.output_dir if args.output_dir is not None else dir_model -os.makedirs(output_dir, exist_ok=True) -output_prefix = os.path.basename(output_dir).replace("ggml_", "") -fname_out = os.path.join(output_dir, f"{fname_middle}model-{ftype_str[ftype]}.gguf") -fout = GGUFWriter(path=fname_out, arch="clip") - -fout.add_bool("clip.has_text_encoder", has_text_encoder) -fout.add_bool("clip.has_vision_encoder", has_vision_encoder) -fout.add_bool("clip.has_minicpmv_projector", has_minicpmv_projector) -fout.add_file_type(ftype) -if args.text_only: - fout.add_description("text-only CLIP model") -elif args.vision_only and not has_minicpmv_projector: - fout.add_description("vision-only CLIP model") -elif has_minicpmv_projector: - fout.add_description("image encoder for MiniCPM-V") - # add projector type - fout.add_string("clip.projector_type", "resampler") - fout.add_int32("clip.minicpmv_version", minicpmv_version) -else: - fout.add_description("two-tower CLIP model") - -if has_vision_encoder: - # vision_model hparams - fout.add_uint32("clip.vision.image_size", 448) - fout.add_uint32("clip.vision.patch_size", 14) - fout.add_uint32(add_key_str(KEY_EMBEDDING_LENGTH, VISION), 1152) - fout.add_uint32(add_key_str(KEY_FEED_FORWARD_LENGTH, VISION), 4304) - fout.add_uint32("clip.vision.projection_dim", 0) - fout.add_uint32(add_key_str(KEY_ATTENTION_HEAD_COUNT, VISION), 16) - fout.add_float32(add_key_str(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6) - block_count = 26 - fout.add_uint32(add_key_str(KEY_BLOCK_COUNT, VISION), block_count) - - if processor is not None: - image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean - image_std = processor.image_processor.image_std if args.image_std is None or args.image_std == default_image_std else args.image_std - else: - image_mean = args.image_mean if args.image_mean is not None else default_image_mean - image_std = args.image_std if args.image_std is not None else default_image_std - fout.add_array("clip.vision.image_mean", image_mean) - fout.add_array("clip.vision.image_std", image_std) - -use_gelu = True -fout.add_bool("clip.use_gelu", use_gelu) - -def get_1d_sincos_pos_embed_from_grid(embed_dim, pos): - """ - embed_dim: output dimension for each position - pos: a list of positions to be encoded: size (M,) - out: (M, D) - """ - assert embed_dim % 2 == 0 - omega = np.arange(embed_dim // 2, dtype=np.float32) - omega /= embed_dim / 2. - omega = 1. / 10000 ** omega # (D/2,) - - pos = pos.reshape(-1) # (M,) - out = np.einsum('m,d->md', pos, omega) # (M, D/2), outer product - - emb_sin = np.sin(out) # (M, D/2) - emb_cos = np.cos(out) # (M, D/2) - - emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D) - return emb - -def get_2d_sincos_pos_embed_from_grid(embed_dim, grid): - assert embed_dim % 2 == 0 - - # use half of dimensions to encode grid_h - emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2) - emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2) - - emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D) - return emb - - -# https://github.com/facebookresearch/mae/blob/efb2a8062c206524e35e47d04501ed4f544c0ae8/util/pos_embed.py#L20 -def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False): - """ - grid_size: int of the grid height and width - return: - pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token) - """ - if isinstance(grid_size, int): - grid_h_size, grid_w_size = grid_size, grid_size - else: - grid_h_size, grid_w_size = grid_size[0], grid_size[1] - - grid_h = np.arange(grid_h_size, dtype=np.float32) - grid_w = np.arange(grid_w_size, dtype=np.float32) - grid = np.meshgrid(grid_w, grid_h) # here w goes first - grid = np.stack(grid, axis=0) - - grid = grid.reshape([2, 1, grid_h_size, grid_w_size]) - pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid) - if cls_token: - pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0) - return pos_embed - -def _replace_name_resampler(s, v): - if re.match("resampler.pos_embed", s): - return { - s: v, - re.sub("pos_embed", "pos_embed_k", s): torch.from_numpy(get_2d_sincos_pos_embed(4096, (70, 70))), - } - if re.match("resampler.proj", s): - return { - re.sub("proj", "pos_embed_k", s): torch.from_numpy(get_2d_sincos_pos_embed(4096, (70, 70))), - re.sub("proj", "proj.weight", s): v.transpose(-1, -2).contiguous(), - } - if re.match("resampler.attn.in_proj_.*", s): - return { - re.sub("attn.in_proj_", "attn.q.", s): v.chunk(3, dim=0)[0], - re.sub("attn.in_proj_", "attn.k.", s): v.chunk(3, dim=0)[1], - re.sub("attn.in_proj_", "attn.v.", s): v.chunk(3, dim=0)[2], - } - return {s: v} - -if has_minicpmv_projector: - projector = torch.load(args.minicpmv_projector) - new_state_dict = {} - for k, v in projector.items(): - kvs = _replace_name_resampler(k, v) - for nk, nv in kvs.items(): - new_state_dict[nk] = nv - projector = new_state_dict - ftype_cur = 0 - for name, data in projector.items(): - name = get_tensor_name(name) - data = data.squeeze().numpy() - - n_dims = len(data.shape) - if ftype == 1: - if name[-7:] == ".weight" and n_dims == 2: - print(" Converting to float16") - data = data.astype(np.float16) - ftype_cur = 1 - else: - print(" Converting to float32") - data = data.astype(np.float32) - ftype_cur = 0 - else: - if data.dtype != np.float32: - print(" Converting to float32") - data = data.astype(np.float32) - ftype_cur = 0 - - fout.add_tensor(name, data) - print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}") - - print("Projector tensors added\n") - -def _replace_name(s, v): - s = "vision_model." + s - if re.match("vision_model.embeddings.position_embedding", s): - v = v.unsqueeze(0) - return {s: v} - - return {s: v} - -state_dict = model.state_dict() -new_state_dict = {} -for k, v in state_dict.items(): - kvs = _replace_name(k, v) - for nk, nv in kvs.items(): - new_state_dict[nk] = nv -state_dict = new_state_dict -for name, data in state_dict.items(): - if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_minicpmv_projector): - # we don't need this - print(f"skipping parameter: {name}") - continue - - name = get_tensor_name(name) - data = data.squeeze().numpy() - - n_dims = len(data.shape) - - # ftype == 0 -> float32, ftype == 1 -> float16 - ftype_cur = 0 - if n_dims == 4: - print(f"tensor {name} is always saved in f16") - data = data.astype(np.float16) - ftype_cur = 1 - elif ftype == 1: - if name[-7:] == ".weight" and n_dims == 2: - print(" Converting to float16") - data = data.astype(np.float16) - ftype_cur = 1 - else: - print(" Converting to float32") - data = data.astype(np.float32) - ftype_cur = 0 - else: - if data.dtype != np.float32: - print(" Converting to float32") - data = data.astype(np.float32) - ftype_cur = 0 - - print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}") - fout.add_tensor(name, data) - - -fout.write_header_to_file() -fout.write_kv_data_to_file() -fout.write_tensors_to_file() -fout.close() - -print("Done. Output file: " + fname_out) diff --git a/examples/llava/minicpmv-convert/minicpmv-surgery.py b/examples/llava/minicpmv-surgery.py similarity index 100% rename from examples/llava/minicpmv-convert/minicpmv-surgery.py rename to examples/llava/minicpmv-surgery.py