Initial model upload

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,88 @@
+---
+language: en
+tags:
+- vision-language-model
+- multimodal
+- vision
+- qwen
+- siglip
+license: apache-2.0
+datasets:
+- cc3m
+---
+
+# VLM Model: Qwen2.5 + SigLIP
+
+This model combines:
+- Vision encoder: google/siglip-base-patch16-224
+- Language model: Qwen/Qwen2.5-0.5B-Instruct
+
+## Model Details
+
+- **Developed by:** [Your name or organization]
+- **Model type:** Vision-Language Model (VLM)
+- **Language(s):** English
+- **License:** Apache 2.0
+
+## Usage
+
+```python
+from transformers import AutoProcessor, AutoTokenizer
+from vlm_model import VLMConfig, VLM
+
+# Load configuration, model, and tokenizer
+config = VLMConfig.from_pretrained("your-username/vlm-qwen-siglip")
+model = VLM.from_pretrained("your-username/vlm-qwen-siglip")
+tokenizer = AutoTokenizer.from_pretrained("your-username/vlm-qwen-siglip")
+processor = AutoProcessor.from_pretrained("google/siglip-base-patch16-224")
+
+# Usage example with an image
+from PIL import Image
+import torch
+
+# Load image
+image = Image.open("path/to/image.jpg").convert("RGB")
+
+# Process image
+processor_output = processor(text=None, images=image, return_tensors="pt")
+pixel_values = processor_output['pixel_values']
+
+# Create chat input
+chat = [
+    {"role": "system", "content": "You are a helpful assistant."},
+    {"role": "user", "content": "What's in this image?" + "<|image_pad|>" * config.image_pad_num}
+]
+
+# Apply chat template
+input_text = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
+input_ids = tokenizer(input_text, return_tensors="pt").input_ids
+
+# Generate response
+with torch.no_grad():
+    generated_ids = model.generate(
+        input_ids=input_ids,
+        pixel_values=pixel_values,
+        max_new_tokens=200,
+        do_sample=True,
+        temperature=0.7,
+        top_p=0.9,
+    )
+
+# Decode response
+response = tokenizer.decode(generated_ids[0][input_ids.shape[1]:], skip_special_tokens=True)
+print(response)
+```
+
+## Training Procedure
+
+This model was trained on [dataset details] using a custom training pipeline that:
+1. Processes images with SigLIP vision encoder
+2. Projects image features to the LLM embedding space 
+3. Inserts image features at image token positions in text prompts
+4. Fine-tunes the projection layers while keeping vision and language models frozen
+
+## Limitations
+
+- [List any known limitations of your model]
+- [Performance characteristics]
+- [Known issues]

added_tokens.json

@@ -0,0 +1,24 @@
+{
+  "</tool_call>": 151658,
+  "<tool_call>": 151657,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

config.json

@@ -0,0 +1,12 @@
+{
+  "architectures": [
+    "VLM"
+  ],
+  "freeze_vision_model": true,
+  "image_pad_num": 49,
+  "llm_model_id": "Qwen/Qwen2.5-0.5B-Instruct",
+  "model_type": "vlm_model",
+  "torch_dtype": "float32",
+  "transformers_version": "4.51.2",
+  "vision_model_id": "google/siglip-base-patch16-224"
+}

example.py

@@ -0,0 +1,38 @@
+from vlm_model import VLMConfig, VLM
+from transformers import AutoProcessor, AutoTokenizer
+from PIL import Image
+import torch
+
+# Load model and tokenizers
+config = VLMConfig.from_pretrained("YOUR_USERNAME/vlm-qwen-siglip")
+model = VLM.from_pretrained("YOUR_USERNAME/vlm-qwen-siglip")
+tokenizer = AutoTokenizer.from_pretrained("YOUR_USERNAME/vlm-qwen-siglip")
+processor = AutoProcessor.from_pretrained("google/siglip-base-patch16-224")
+
+# Load image
+image = Image.open("your_image.jpg").convert("RGB")
+processor_output = processor(text=None, images=image, return_tensors="pt")
+pixel_values = processor_output['pixel_values']
+
+# Create input with image placeholder
+chat = [
+    {"role": "system", "content": "You are a helpful assistant."},
+    {"role": "user", "content": f"What's in this image?{('<|image_pad|>' * config.image_pad_num)}"}
+]
+input_text = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
+input_ids = tokenizer(input_text, return_tensors="pt").input_ids
+
+# Generate response
+with torch.no_grad():
+    generated_ids = model.generate(
+        input_ids=input_ids,
+        pixel_values=pixel_values,
+        max_new_tokens=200,
+        do_sample=True,
+        temperature=0.7,
+        top_p=0.9,
+    )
+
+# Decode response
+response = tokenizer.decode(generated_ids[0][input_ids.shape[1]:], skip_special_tokens=True)
+print(response)

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a35ba549cd40c52fff2a7376cbee1b545f16ce0756df42e0591c6c8673e7f11
+size 2803230786

requirements.txt

@@ -0,0 +1,3 @@
+transformers>=4.35.0
+torch>=2.0.0
+pillow>=9.0.0

special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c5ae00e602b8860cbd784ba82a8aa14e8feecec692e7076590d014d7b7fdafa
+size 11421896

tokenizer_config.json

@@ -0,0 +1,208 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "chat_template": "{%- if tools %}\n    {{- '<|im_start|>system\\n' }}\n    {%- if messages[0]['role'] == 'system' %}\n        {{- messages[0]['content'] }}\n    {%- else %}\n        {{- 'You are Qwen, created by Alibaba Cloud. You are a helpful assistant.' }}\n    {%- endif %}\n    {{- \"\\n\\n# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within <tools></tools> XML tags:\\n<tools>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\n</tools>\\n\\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\\n<tool_call>\\n{\\\"name\\\": <function-name>, \\\"arguments\\\": <args-json-object>}\\n</tool_call><|im_end|>\\n\" }}\n{%- else %}\n    {%- if messages[0]['role'] == 'system' %}\n        {{- '<|im_start|>system\\n' + messages[0]['content'] + '<|im_end|>\\n' }}\n    {%- else %}\n        {{- '<|im_start|>system\\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\\n' }}\n    {%- endif %}\n{%- endif %}\n{%- for message in messages %}\n    {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) or (message.role == \"assistant\" and not message.tool_calls) %}\n        {{- '<|im_start|>' + message.role + '\\n' + message.content + '<|im_end|>' + '\\n' }}\n    {%- elif message.role == \"assistant\" %}\n        {{- '<|im_start|>' + message.role }}\n        {%- if message.content %}\n            {{- '\\n' + message.content }}\n        {%- endif %}\n        {%- for tool_call in message.tool_calls %}\n            {%- if tool_call.function is defined %}\n                {%- set tool_call = tool_call.function %}\n            {%- endif %}\n            {{- '\\n<tool_call>\\n{\"name\": \"' }}\n            {{- tool_call.name }}\n            {{- '\", \"arguments\": ' }}\n            {{- tool_call.arguments | tojson }}\n            {{- '}\\n</tool_call>' }}\n        {%- endfor %}\n        {{- '<|im_end|>\\n' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != \"tool\") %}\n            {{- '<|im_start|>user' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- message.content }}\n        {{- '\\n</tool_response>' }}\n        {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n            {{- '<|im_end|>\\n' }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n{%- endif %}\n",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94bb7a6bce525b1ac9a3402c658f0ad0bb7971ce8f7c676d0dcb804c19587b5b
+size 5304

vlm_model/__init__.py

@@ -0,0 +1 @@
+from .model import VLMConfig, VLM

vlm_model/model.py

@@ -0,0 +1,139 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import PreTrainedModel, PretrainedConfig, AutoTokenizer, AutoModelForCausalLM
+from transformers import AutoProcessor, AutoModel
+from transformers.modeling_outputs import CausalLMOutputWithPast
+
+# VLM Config and Model classes
+class VLMConfig(PretrainedConfig):
+    model_type = "vlm_model"
+    def __init__(self,
+                 llm_model_id = "Qwen/Qwen2.5-0.5B-Instruct",
+                 vision_model_id = "google/siglip-base-patch16-224",
+                 freeze_vision_model = True,
+                 image_pad_num = 49,
+                **kwargs):
+        self.vision_model_id = vision_model_id
+        self.llm_model_id = llm_model_id
+        self.freeze_vision_model = freeze_vision_model
+        self.image_pad_num = image_pad_num
+        super().__init__(**kwargs)
+        
+class VLM(PreTrainedModel):
+    config_class = VLMConfig
+    def __init__(self, config):
+        super().__init__(config)
+        self.config = config
+        
+        # Load models from Hugging Face
+        self.vision_model = AutoModel.from_pretrained(self.config.vision_model_id)
+        self.processor = AutoProcessor.from_pretrained(self.config.vision_model_id)
+        self.llm_model = AutoModelForCausalLM.from_pretrained(self.config.llm_model_id)
+        self.tokenizer = AutoTokenizer.from_pretrained(self.config.llm_model_id)
+        
+        # Projection layers
+        self.linear1 = nn.Linear(self.vision_model.config.vision_config.hidden_size*4, self.llm_model.config.hidden_size)
+        self.linear2 = nn.Linear(self.llm_model.config.hidden_size, self.llm_model.config.hidden_size)
+        
+        # Freeze models
+        if self.config.freeze_vision_model:
+            for param in self.vision_model.parameters():
+                param.requires_grad = False
+        
+        for param in self.llm_model.parameters():
+            param.requires_grad = False
+        
+    def forward(self, input_ids, labels=None, pixel_values=None, attention_mask=None):
+        # Get text embeddings
+        text_embeds = self.llm_model.get_input_embeddings()(input_ids)
+        
+        if pixel_values is not None:
+            # Ensure pixel_values has the right shape [batch_size, channels, height, width]
+            if len(pixel_values.shape) == 3:
+                pixel_values = pixel_values.unsqueeze(0)
+            # Handle case where pixel_values might have extra dimensions
+            elif len(pixel_values.shape) > 4:
+                # Reshape to expected 4D format (assuming first dim is batch)
+                b = pixel_values.shape[0]
+                pixel_values = pixel_values.view(b, 3, 224, 224)  # Assuming standard 224x224 image size
+            
+            # Get image embeddings
+            image_embeds = self.vision_model.vision_model(pixel_values).last_hidden_state 
+            b, s, d = image_embeds.shape
+            # Compress image tokens
+            image_embeds = image_embeds.view(b, -1, d*4)  
+            image_features = self.linear2(F.silu(self.linear1(image_embeds)))
+            
+            # Match dtype
+            text_embeds = text_embeds.to(image_features.dtype)
+            
+            # Merge embeddings
+            inputs_embeds = self.merge_input_ids_with_image_features(image_features, text_embeds, input_ids)
+        else:
+            inputs_embeds = text_embeds
+        
+        # Forward pass
+        outputs = self.llm_model(inputs_embeds=inputs_embeds, attention_mask=attention_mask)
+        logits = outputs[0]
+        
+        # Calculate loss
+        loss = None
+        if labels is not None:
+            loss_fct = nn.CrossEntropyLoss(ignore_index=self.tokenizer.pad_token_id)
+            loss = loss_fct(
+                logits.view(-1, logits.size(-1)), labels.view(-1).to(logits.device)
+            )
+        return CausalLMOutputWithPast(loss=loss, logits=logits)
+        
+    def merge_input_ids_with_image_features(self, image_features, inputs_embeds, input_ids):
+        # Replace image placeholder tokens with image features
+        batch_indices, image_indices = torch.where(input_ids == self.tokenizer('<|image_pad|>')['input_ids'][0])
+        
+        # Handle the case of multiple batches with multiple images
+        if len(batch_indices) > 0:
+            num_images, num_image_patches, embed_dim = image_features.shape
+            # Group indices by batch
+            for b_idx in range(input_ids.shape[0]):
+                batch_mask = (batch_indices == b_idx)
+                if batch_mask.sum() > 0:
+                    # Get indices for this batch
+                    img_indices = image_indices[batch_mask]
+                    # Replace tokens with image features
+                    img_idx = min(b_idx, num_images-1)  # Prevent out of bounds
+                    inputs_embeds[b_idx, img_indices] = image_features[img_idx].repeat(len(img_indices), 1)[:len(img_indices)]
+                    
+        return inputs_embeds
+    
+    def generate(self, input_ids=None, pixel_values=None, attention_mask=None, **kwargs):
+        # Process the input just like in forward pass
+        text_embeds = self.llm_model.get_input_embeddings()(input_ids)
+        
+        if pixel_values is not None:
+            # Ensure pixel_values has the right shape
+            if len(pixel_values.shape) == 3:
+                pixel_values = pixel_values.unsqueeze(0)
+            elif len(pixel_values.shape) > 4:
+                b = pixel_values.shape[0]
+                pixel_values = pixel_values.view(b, 3, 224, 224)
+                
+            # Get image embeddings and project
+            image_embeds = self.vision_model.vision_model(pixel_values).last_hidden_state
+            b, s, d = image_embeds.shape
+            image_embeds = image_embeds.view(b, -1, d*4)
+            image_features = self.linear2(F.silu(self.linear1(image_embeds)))
+            
+            # Match dtype
+            text_embeds = text_embeds.to(image_features.dtype)
+            
+            # Merge embeddings
+            inputs_embeds = self.merge_input_ids_with_image_features(image_features, text_embeds, input_ids)
+        else:
+            inputs_embeds = text_embeds
+        
+        # Use the LLM's generate method with the processed inputs
+        return self.llm_model.generate(
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            **kwargs
+        )