Add conversion for interleave llava (#31858)

* add conversion for interleave llava

* remove debug lines

* remove unused imports

* Update src/transformers/models/llava/convert_llava_weights_to_hf.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* small changes + docs

---------

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

docs/source/en/model_doc/llava.md

@@ -40,8 +40,20 @@ The original code can be found [here](https://github.com/haotian-liu/LLaVA/tree/
 
 - Note the model has not been explicitly trained to process multiple images in the same prompt, although this is technically possible, you may experience inaccurate results.
 
-- For better results, we recommend users to prompt the model with the correct prompt format: 
+- For better results, we recommend users to prompt the model with the correct prompt format. Below is a list of prompt formats accepted by each llava checkpoint:
 
+[llava-interleave models](https://huggingface.co/collections/llava-hf/llava-interleave-668e19a97da0036aad4a2f19) requires the following format:
+```bash
+"<|im_start|>user <image>\nWhat is shown in this image?<|im_end|><|im_start|>assistant"
+```
+
+For multiple turns conversation:
+
+```bash
+"<|im_start|>user <image>\n<prompt1><|im_end|><|im_start|>assistant <answer1><|im_end|><|im_start|>user <image>\n<prompt1><|im_end|><|im_start|>assistant "
+```
+
+[llava-1.5 models](https://huggingface.co/collections/llava-hf/llava-15-65f762d5b6941db5c2ba07e0) requires the following format:
 ```bash
 "USER: <image>\n<prompt> ASSISTANT:"
 ```

src/transformers/models/llava/convert_llava_weights_to_hf.py

@@ -12,18 +12,21 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import argparse
+import glob
 
 import torch
-from huggingface_hub import hf_hub_download
+from huggingface_hub import hf_hub_download, snapshot_download
+from safetensors import safe_open
 
 from transformers import (
     AddedToken,
     AutoConfig,
+    AutoImageProcessor,
     AutoTokenizer,
-    CLIPImageProcessor,
     LlavaConfig,
     LlavaForConditionalGeneration,
     LlavaProcessor,
+    SiglipVisionConfig,
 )
 
 
@@ -48,6 +51,7 @@ Example for creating the old state dict file with Python:
 
 KEYS_TO_MODIFY_MAPPING = {
     "model.vision_tower.": "",
+    ".vision_resampler": "",  # all lmms-lab models do avg pooling, so no vision_resampler
     "model.mm_projector": "multi_modal_projector",
     "model": "model.model",
     "vision_model.model": "vision_model",
@@ -58,6 +62,26 @@ KEYS_TO_MODIFY_MAPPING = {
 }
 
 
+def load_original_state_dict(model_id):
+    directory_path = snapshot_download(repo_id=model_id, allow_patterns=["*.safetensors"])
+
+    original_state_dict = {}
+    for path in glob.glob(f"{directory_path}/*"):
+        if path.endswith(".safetensors"):
+            with safe_open(path, framework="pt", device="cpu") as f:
+                for key in f.keys():
+                    original_state_dict[key] = f.get_tensor(key)
+
+    # tied wieghts so lm.head is not saved. Let's clone to load state dict
+    if "lm_head.weight" not in original_state_dict:
+        original_state_dict["lm_head.weight"] = original_state_dict["model.embed_tokens.weight"].clone()
+
+    del original_state_dict["model.image_newline"]  # not used in the original implementation because "merge_type=flat"
+    return original_state_dict
+
+
+# used only for llava-interlave
+# for ex: Qwen/Qwen1.5-0.5B-Chat google/siglip-so400m-patch14-384 lmms-lab/llava-next-interleave-qwen-0.5b
 def convert_state_dict_to_hf(state_dict):
     new_state_dict = {}
     for key, value in state_dict.items():
@@ -77,24 +101,48 @@ def convert_llava_llama_to_hf(text_model_id, vision_model_id, output_hub_path, o
 
     tokenizer = AutoTokenizer.from_pretrained(text_model_id)
     tokenizer.add_tokens(AddedToken("<image>", special=True, normalized=False), special_tokens=True)
-    tokenizer.add_special_tokens({"pad_token": "<pad>"})
-
-    image_processor = CLIPImageProcessor.from_pretrained(vision_model_id)
+    if "Qwen" not in text_model_id:  # qwen already has a pad token
+        tokenizer.add_special_tokens({"pad_token": "<pad>"})
 
+    image_processor = AutoImageProcessor.from_pretrained(vision_model_id)
     processor = LlavaProcessor(tokenizer=tokenizer, image_processor=image_processor)
 
-    config = LlavaConfig(text_config=text_config)
-    config.pad_token_id = 32001
+    if "Qwen" in text_model_id:
+        vision_config = SiglipVisionConfig(
+            hidden_size=1152,
+            image_size=384,
+            intermediate_size=4304,
+            num_attention_heads=16,
+            num_hidden_layers=26,
+            patch_size=14,
+            vision_use_head=False,
+        ).to_dict()
+    else:
+        vision_config = None
+
+    config = LlavaConfig(
+        text_config=text_config,
+        vision_config=vision_config,
+    )
+
+    # llms-lab interleeave models do not use any selection startegy except for last hidden state
+    if "Qwen" in text_model_id:
+        config.image_token_index = 151646
+        config.vision_feature_select_strategy = "full"
+        config.vision_feature_layer = -1
+    else:
+        config.pad_token_id = 32001
+        config.image_token_index = 32000
 
     with torch.device("meta"):
         model = LlavaForConditionalGeneration(config)
 
-    # Pad to 64 for performance reasons
-    pad_shape = 64
+    if "Qwen" in text_model_id:
+        state_dict = load_original_state_dict(old_state_dict_id)
+    else:
+        state_dict_path = hf_hub_download(old_state_dict_id, "model_state_dict.bin")
+        state_dict = torch.load(state_dict_path, map_location="cpu")
 
-    state_dict_path = hf_hub_download(old_state_dict_id, "model_state_dict.bin")
-
-    state_dict = torch.load(state_dict_path, map_location="cpu")
     state_dict = convert_state_dict_to_hf(state_dict)
     model.load_state_dict(state_dict, strict=True, assign=True)
 
@@ -104,14 +152,18 @@ def convert_llava_llama_to_hf(text_model_id, vision_model_id, output_hub_path, o
     sigma = ((pre_expansion_embeddings - mu).T @ (pre_expansion_embeddings - mu)) / n
     dist = torch.distributions.multivariate_normal.MultivariateNormal(mu, covariance_matrix=1e-5 * sigma)
 
-    # We add an image token so we resize the model
+    # We add an image token so we resize the model and pad to 64 for performance reasons
+    pad_shape = 64
+    vocab_size = config.text_config.vocab_size
     model.resize_token_embeddings(config.text_config.vocab_size + 2, pad_shape)
-    model.language_model.model.embed_tokens.weight.data[32000:] = torch.stack(
-        tuple((dist.sample() for _ in range(model.language_model.model.embed_tokens.weight.data[32000:].shape[0]))),
+    model.language_model.model.embed_tokens.weight.data[vocab_size:] = torch.stack(
+        tuple(
+            (dist.sample() for _ in range(model.language_model.model.embed_tokens.weight.data[vocab_size:].shape[0]))
+        ),
         dim=0,
     )
-    model.language_model.lm_head.weight.data[32000:] = torch.stack(
-        tuple((dist.sample() for _ in range(model.language_model.lm_head.weight.data[32000:].shape[0]))),
+    model.language_model.lm_head.weight.data[vocab_size:] = torch.stack(
+        tuple((dist.sample() for _ in range(model.language_model.lm_head.weight.data[vocab_size:].shape[0]))),
         dim=0,
     )