modify convert script of minicpmv

examples/llava/minicpmv-convert/minicpmv2_5-convert-image-encoder-to-gguf.py

@@ -166,6 +166,7 @@ elif args.minicpmv_projector is not None:
     fname_middle = "mmproj-"
     has_text_encoder = False
     has_minicpmv_projector = True
+    minicpmv_version = 3
 elif args.vision_only:
     fname_middle = "vision-"
     has_text_encoder = False
@@ -190,6 +191,7 @@ 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")
 

examples/llava/minicpmv-convert/minicpmv2_5-surgery.py

@@ -41,7 +41,6 @@ config.auto_map = {
 model.llm.save_pretrained(f"{args.model}/model")
 tok = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
 tok.save_pretrained(f"{args.model}/model")
-# os.system(f"cp {args.model}/modeling_minicpm.py {args.model}/MiniCPM_l3/modeling_minicpm.py")
 
 print("Done!")
 print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")

examples/llava/minicpmv-convert/minicpmv2_6-surgery.py

@@ -0,0 +1,47 @@
+import argparse
+import glob
+import os
+import torch
+from transformers import AutoModel, AutoTokenizer
+
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model", help="Path to MiniCPM-V-2.6 model")
+args = ap.parse_args()
+
+# find the model part that includes the the multimodal projector weights
+model = AutoModel.from_pretrained(args.model, trust_remote_code=True, local_files_only=True)
+checkpoint = model.state_dict()
+
+# get a list of mm tensor names
+mm_tensors = [k for k, v in checkpoint.items() if k.startswith("resampler")]
+
+# store these tensors in a new dictionary and torch.save them
+projector = {name: checkpoint[name].float() for name in mm_tensors}
+torch.save(projector, f"{args.model}/minicpmv.projector")
+
+clip_tensors = [k for k, v in checkpoint.items() if k.startswith("vpm")]
+if len(clip_tensors) > 0:
+    clip = {name.replace("vpm.", ""): checkpoint[name].float() for name in clip_tensors}
+    torch.save(clip, f"{args.model}/minicpmv.clip")
+
+    # added tokens should be removed to be able to convert Mistral models
+    if os.path.exists(f"{args.model}/added_tokens.json"):
+        with open(f"{args.model}/added_tokens.json", "w") as f:
+            f.write("{}\n")
+
+config = model.llm.config
+config._name_or_path = "openbmb/MiniCPM-V-2.6"
+config.auto_map = {
+    "AutoConfig": "configuration_minicpm.MiniCPMConfig",
+    "AutoModel": "modeling_minicpm.MiniCPMModel",
+    "AutoModelForCausalLM": "modeling_minicpm.MiniCPMForCausalLM",
+    "AutoModelForSeq2SeqLM": "modeling_minicpm.MiniCPMForCausalLM",
+    "AutoModelForSequenceClassification": "modeling_minicpm.MiniCPMForSequenceClassification"
+}
+model.llm.save_pretrained(f"{args.model}/model")
+tok = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
+tok.save_pretrained(f"{args.model}/model")
+
+print("Done!")
+print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
+print(f"Also, use {args.model}/minicpmv.projector to prepare a minicpmv-encoder.gguf file.")