Add files using upload-large-folder tool

.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

@@ -1,3 +1,201 @@
----
-license: mit
----
+---
+license: mit
+library_name: transformers
+pipeline_tag: image-text-to-text
+language:
+- en
+tags:
+- multimodal
+- vision
+- video
+- long-video
+- token-selection
+- compression
+- qwen2.5-vl
+- qtsplus
+---
+
+[![arXiv](https://img.shields.io/badge/arXiv-2511.11910-grey?labelColor=B31B1B&logo=arxiv&logoColor=white)](https://arxiv.org/abs/2511.11910)
+[![Website](https://img.shields.io/badge/Website-QTSplus-grey?labelColor=3776AB&logo=GoogleChrome&logoColor=white)](https://qtsplus.github.io/)
+[![Github](https://img.shields.io/badge/Github-QTSplus-grey?labelColor=000&logo=github)](https://github.com/Siyou-Li/QTSplus)
+
+## Model Description
+![](./assets/qtsplus.svg)
+
+QTSplus-7B is a Qwen2.5-VL–based multimodal LLM finetuned with Query‑Aware Token Selector (QTSplus), a lightweight visual token selection module that acts as an information gate between the vision encoder and the LLM.
+
+- Query‑aware selection: scores vision tokens via cross‑attention against the input text query.
+- Adaptive retention: predicts an instance‑specific budget and keeps only the most relevant tokens.
+- Temporal reasoning: a small re‑encoder preserves temporal order with absolute time cues.
+- Efficient long‑video understanding: up to 89% vision token compression and 28% end‑to‑end latency reduction on long videos (see paper for details).
+
+## Intended Uses & Limitations
+
+Intended uses
+- Long‑video question answering and captioning
+- Multi‑image reasoning and story understanding
+- Efficient multimodal chat with reduced latency on long inputs
+
+Limitations
+- May miss fine details if the predicted retention budget is too small.
+- Inherits biases and failure modes from the base Qwen2.5‑VL model and training data.
+- Not a safety‑aligned system; outputs may be inaccurate or unsafe without human oversight.
+
+## Quick Start
+
+The repository is designed around a conda‑based Python 3.11 environment with a CUDA‑enabled GPU.
+
+1. **Create and activate the conda environment**
+
+```bash
+conda create -n qtsplus python=3.11 -y
+conda activate qtsplus
+```
+
+2. **Install toolchain and CUDA toolkit**
+
+```bash
+conda install conda-forge::gcc=11 conda-forge::gxx=11 -y
+conda install nvidia/label/cuda-12.8.1::cuda-toolkit -y
+conda install av -c conda-forge -y
+```
+
+3. **Install PyTorch with CUDA 12.8 support**
+
+```bash
+pip3 install torch==2.9.0 torchvision --index-url https://download.pytorch.org/whl/cu128
+```
+
+4. **Install core Python libraries**
+
+```bash
+pip install transformers==4.57.1
+DS_BUILD_CUTLASS_OPS=0 DS_BUILD_RAGGED_DEVICE_OPS=0 DS_BUILD_EVOFORMER_ATTN=0 pip install deepspeed
+pip install accelerate pandas wandb matplotlib scikit-learn datasets evaluate ftfy sentencepiece bitsandbytes
+```
+
+5. **Install FlashAttention (prebuilt wheel)**
+
+```bash
+pip install https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.4.22/flash_attn-2.8.1+cu128torch2.9-cp311-cp311-linux_x86_64.whl
+```
+
+This wheel is specific to Linux x86_64, CUDA 12.8, PyTorch 2.9.0 and Python 3.11; if you deviate from this configuration, you will need to install a compatible FlashAttention build instead.
+
+6. **Verify installation**
+
+After installation, you should be able to run:
+
+```bash
+python -c "import torch, transformers, deepspeed, accelerate; print(torch.cuda.is_available())"
+```
+
+which should print `True` on a correctly configured GPU machine.
+
+Video example
+```python
+import torch, glob, os
+from transformers import AutoModelForCausalLM, AutoProcessor
+from qwen_vl_utils import process_vision_info
+
+model_id = "AlpachinoNLP/QTSplus-7B"
+device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+dtype = torch.bfloat16 if torch.cuda.is_available() else torch.float16
+
+model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True).to(dtype=dtype, device=device).eval()
+processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
+
+question = "Summarize the key events in this video."
+video_path = "/path/to/video.mp4"
+
+messages = [{
+    "role": "user",
+    "content": [
+        {"type": "video", "video": video_path, "max_pixels": 360*420, "fps": 1.0},
+        {"type": "text", "text": question},
+    ],
+}]
+
+chat = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+_, video_inputs, video_kwargs = process_vision_info(messages, return_video_kwargs=True)
+
+inputs = processor(text=[chat], images=None, videos=video_inputs, padding=True, return_tensors="pt", **video_kwargs)
+inputs = inputs.to(dtype=torch.float16, device=device)
+
+# Pack vision inputs for QTSplus
+pixel_values_videos = inputs.pop("pixel_values_videos", None)
+video_grid_thw = inputs.pop("video_grid_thw", None)
+inputs.pop("second_per_grid_ts", None)
+vision_input = None
+if pixel_values_videos is not None and video_grid_thw is not None:
+    vision_input = {"pixel_values_videos": pixel_values_videos, "video_grid_thw": video_grid_thw}
+
+# Text ids from the question only (exclude special/system/vision tokens)
+question_ids = processor.tokenizer(question, return_tensors="pt", add_special_tokens=False).input_ids.to(dtype=torch.long, device=device)
+
+out_ids = model.generate(vision_input=vision_input, input_ids=inputs.input_ids, question_input_ids=question_ids, max_new_tokens=256)
+trimmed = [o[len(i):] for i, o in zip(inputs.input_ids, out_ids)]
+text = processor.batch_decode(trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=True)
+print(text[0])
+```
+
+Multiple images (treated as a video sequence)
+```python
+images_dir = "/path/to/images"
+image_list = sorted(glob.glob(os.path.join(images_dir, "*.jpg"))) or sorted(glob.glob(os.path.join(images_dir, "*.jpeg")))
+messages = [{
+    "role": "user",
+    "content": [
+        {"type": "video", "video": image_list},
+        {"type": "text", "text": "What story do these images tell?"},
+    ],
+}]
+
+chat = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+_, video_inputs, video_kwargs = process_vision_info(messages, return_video_kwargs=True)
+inputs = processor(text=[chat], images=None, videos=video_inputs, padding=True, return_tensors="pt", **video_kwargs).to(dtype=torch.float16, device=device)
+
+pixel_values_videos = inputs.pop("pixel_values_videos", None)
+video_grid_thw = inputs.pop("video_grid_thw", None)
+inputs.pop("second_per_grid_ts", None)
+vision_input = {"pixel_values_videos": pixel_values_videos, "video_grid_thw": video_grid_thw}
+
+out = model.generate(vision_input=vision_input, input_ids=inputs.input_ids, max_new_tokens=256)
+print(processor.decode(out[0], skip_special_tokens=True))
+```
+
+Notes
+- The chat template is applied via `processor.apply_chat_template` and expects the messages schema shown above.
+- QTSplus expects the vision payload under the `vision_input` keyword argument during generation.
+- For fully offline use, pass `local_files_only=True` to `from_pretrained` calls once the files are cached locally.
+
+## Efficiency & Controls
+
+The following QTSplus hyperparameters in `config.json` control compression and selection behavior:
+- `qts_plus_rho_min` / `qts_plus_rho_max`: min/max retention ratio bounds.(default: 0.05 / 0.5)
+- `qts_plus_tau_s`: scoring temperature for cross‑attention.(default: 0.5)
+- `qts_plus_nmax`: hard cap on selected tokens per sample. （default: 25600)
+These trade off detail vs. speed/memory. See the paper for guidance, ablations, and latency/throughput measurements.
+
+
+## Safety, Bias, and Limitations
+
+- Outputs may be factually incorrect, biased, or unsafe. Do not use without human oversight.
+- QTSplus compresses the vision stream; extremely small budgets may drop rare but important details.
+- Inherits safety/bias characteristics from the underlying Qwen2.5‑VL model and training data.
+
+## Citation
+
+If you find this work helpful, please cite:
+
+```bibtex
+@misc{li2025seeingforesttreesqueryaware,
+  title        = {Seeing the Forest and the Trees: Query-Aware Tokenizer for Long-Video Multimodal Language Models},
+  author       = {Siyou Li and Huanan Wu and Juexi Shao and Yinghao Ma and Yujian Gan and Yihao Luo and Yuwei Wang and Dong Nie and Lu Wang and Wengqing Wu and Le Zhang and Massimo Poesio and Juntao Yu},
+  year         = {2025},
+  eprint       = {2511.11910},
+  archivePrefix= {arXiv},
+  primaryClass = {cs.CV},
+  url          = {https://arxiv.org/abs/2511.11910}
+}
+```

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
+}

chat_template.jinja

@@ -0,0 +1,7 @@
+{% set image_count = namespace(value=0) %}{% set video_count = namespace(value=0) %}{% for message in messages %}{% if loop.first and message['role'] != 'system' %}<|im_start|>system
+You are a helpful assistant.<|im_end|>
+{% endif %}<|im_start|>{{ message['role'] }}
+{% if message['content'] is string %}{{ message['content'] }}<|im_end|>
+{% else %}{% for content in message['content'] %}{% if content['type'] == 'image' or 'image' in content or 'image_url' in content %}{% set image_count.value = image_count.value + 1 %}{% if add_vision_id %}Picture {{ image_count.value }}: {% endif %}<|vision_start|><|image_pad|><|vision_end|>{% elif content['type'] == 'video' or 'video' in content %}{% set video_count.value = video_count.value + 1 %}{% if add_vision_id %}Video {{ video_count.value }}: {% endif %}<|vision_start|><|video_pad|><|vision_end|>{% elif 'text' in content %}{{ content['text'] }}{% endif %}{% endfor %}<|im_end|>
+{% endif %}{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant
+{% endif %}

config.json

@@ -0,0 +1,82 @@
+{
+  "architectures": [
+    "QTSplusQwen2_5_VLTextForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_qts_plus_qwen2_5_vl.QTSplusQwen2_5_VL_CausalLM_Config",
+    "AutoModelForCausalLM": "modeling_qts_plus_qwen2_5_vl.QTSplusQwen2_5_VLTextForCausalLM",
+    "AutoProcessor": "processing_qts_plus_qwen2_5_vl.QTSplusQwen2_5_VLProcessor"
+  },
+  "vision_tower": "qwen2_5_vl_vision",
+  "enable_qts_plus": true,
+  "qts_plus_n_heads": 28,
+  "qts_plus_tau_s": 0.5,
+  "qts_plus_nmax": 25600,
+  "qts_plus_rho_min": 0.05,
+  "qts_plus_rho_max": 0.5,
+  "qts_plus_block_dropout": 0.0,
+  "qts_plus_reencode": true,
+  "qts_plus_scoring_layers": 1,
+  "qts_plus_reencode_layers": 2,
+  "project_text_if_needed": false,
+  "freeze_qts_scoring_layers": false,
+  "lambda_t": 0,
+  "lambda_m": 0,
+  "lambda_s": 0,
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "eos_token_id": 151645,
+  "vision_start_token_id": 151652,
+  "vision_end_token_id": 151653,
+  "vision_token_id": 151654,
+  "image_token_id": 151655,
+  "video_token_id": 151656,
+  "hidden_act": "silu",
+  "hidden_size": 3584,
+  "initializer_range": 0.02,
+  "intermediate_size": 18944,
+  "max_position_embeddings": 128000,
+  "max_window_layers": 28,
+  "model_type": "qts_plus_qwen2_5_vl_causal_lm",
+  "num_attention_heads": 28,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 4,
+  "rms_norm_eps": 1e-06,
+  "rope_theta": 1000000.0,
+  "sliding_window": null,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.57.1",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vision_config": {
+    "depth": 32,
+    "hidden_act": "silu",
+    "hidden_size": 1280,
+    "intermediate_size": 3420,
+    "num_heads": 16,
+    "in_chans": 3,
+    "out_hidden_size": 3584,
+    "patch_size": 14,
+    "spatial_merge_size": 2,
+    "spatial_patch_size": 14,
+    "window_size": 112,
+    "fullatt_block_indexes": [
+      7,
+      15,
+      23,
+      31
+    ],
+    "tokens_per_second": 2,
+    "temporal_patch_size": 2
+  },
+  "rope_scaling": {
+    "type": "mrope",
+    "mrope_section": [
+      16,
+      24,
+      24
+    ]
+  },
+  "vocab_size": 152064
+}

configuration_qts_plus_qwen2_5_vl.py

@@ -0,0 +1,21 @@
+"""
+Self-contained config shim for trust_remote_code.
+
+This file defines the minimal configuration class expected by
+`config.json` without importing from a local `src` package.
+"""
+from transformers import AutoConfig
+from transformers.models.qwen2_5_vl.configuration_qwen2_5_vl import Qwen2_5_VLTextConfig
+
+class QTSplusQwen2_5_VL_CausalLM_Config(Qwen2_5_VLTextConfig):
+    """Config alias for QTS+ Qwen2.5-VL Causal LM.
+
+    It inherits from the upstream Qwen2.5-VL text config and only sets a
+    distinct `model_type` so that Transformers can resolve the proper
+    architecture via `auto_map`.
+    """
+
+    model_type = "qts_plus_qwen2_5_vl_causal_lm"
+
+AutoConfig.register("qts_plus_qwen2_5_vl_causal_lm", QTSplusQwen2_5_VL_CausalLM_Config)
+__all__ = ["QTSplusQwen2_5_VL_CausalLM_Config"]

generation_config.json

@@ -0,0 +1,9 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 151643,
+  "eos_token_id": [
+    151645
+  ],
+  "pad_token_id": 151643,
+  "transformers_version": "4.57.1"
+}

latest

@@ -0,0 +1 @@
+global_step30000

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1366365f50c6d1464cec1b067f4a1d600b8fa05c4a22911c1e46aa8b81fd5b14
+size 17379187170

modeling_qts_plus_qwen2_5_vl.py

@@ -0,0 +1,1090 @@
+"""
+Self-contained modeling shim for trust_remote_code.
+
+Implements the QTS+ Qwen2.5‑VL Causal LM architecture locally by
+composing upstream Transformers' Qwen2.5‑VL text and vision modules
+with a lightweight QTS+ selector. This avoids importing any local `src`
+package while preserving checkpoint compatibility (including
+`model.vision_tower.*` and `model.qts_plus.selector.*` parameters).
+"""
+import json
+import os
+from typing import Optional
+import torch
+import torch.nn as nn
+from transformers import AutoConfig, AutoModelForCausalLM, logging
+from transformers.modeling_flash_attention_utils import is_flash_attn_available
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.generation import GenerationMixin
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VisionTransformerPretrainedModel as Qwen2_5_VisionTransformerPretrainedModelBase
+from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
+    Qwen2_5_VLTextModel,
+    Qwen2_5_VLPreTrainedModel,
+
+)
+from transformers.models.qwen2_5_vl.configuration_qwen2_5_vl import Qwen2_5_VLTextConfig, Qwen2_5_VLVisionConfig
+from .configuration_qts_plus_qwen2_5_vl import (
+    QTSplusQwen2_5_VL_CausalLM_Config
+)
+logger = logging.get_logger(__name__)
+# ------------------------------
+# Utilities: embedding integration
+# ------------------------------
+def qts_integrate_embeddings(
+    vision_features: torch.Tensor,
+    input_ids: torch.Tensor,
+    attention_mask: torch.Tensor,
+    labels: Optional[torch.Tensor] = None,
+    image_token_id: Optional[int] = None,
+    video_token_id: Optional[int] = None,
+    image_grid_thw: Optional[torch.Tensor] = None,
+    video_grid_thw: Optional[torch.Tensor] = None,
+    text_model_embed_layer: Optional[nn.Embedding] = None,
+    kept_indices: Optional[torch.Tensor] = None,
+) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
+    """Integrate visual features into text embeddings (single-sample batch).
+
+    This mirrors the behavior of the full Qwen2.5‑VL generation path, but
+    works with pre-computed visual features and placeholder tokens in the
+    text sequence. It supports both the single <|video_pad|> token case and
+    multi-placeholder templates.
+    """
+    if text_model_embed_layer is None:
+        raise ValueError("text_model_embed_tokens is required for text embedding integration")
+    if input_ids.dtype is not torch.long:
+        input_ids = input_ids.long()
+
+    inputs_embeds = text_model_embed_layer(input_ids)
+    if vision_features.shape[0] <= 0:
+        raise ValueError("vision_features must contain at least one feature vector")
+    if video_token_id is None:
+        raise ValueError("video_token_id must be provided for video feature integration")
+
+    B, S = input_ids.shape
+    assert B == 1, "Sequence-trimming currently assumes batch_size == 1."
+
+    vid_pos = (input_ids[0] == video_token_id).nonzero(as_tuple=False).flatten()
+    n_feats = int(vision_features.shape[0])
+
+    if vid_pos.numel() == 1 and n_feats >= 1:
+        insert_idx = int(vid_pos.item())
+        vision_features = vision_features.to(inputs_embeds.device, inputs_embeds.dtype)
+
+        pre_embeds = inputs_embeds[:, :insert_idx, :]
+        post_embeds = inputs_embeds[:, insert_idx + 1 :, :]
+
+        feats_embeds = vision_features.unsqueeze(0)
+        inputs_embeds = torch.cat([pre_embeds, feats_embeds, post_embeds], dim=1)
+
+        feats_mask = torch.ones((1, n_feats), dtype=attention_mask.dtype, device=attention_mask.device)
+        pre_mask = attention_mask[:, :insert_idx]
+        post_mask = attention_mask[:, insert_idx + 1 :]
+        attention_mask = torch.cat([pre_mask, feats_mask, post_mask], dim=1)
+
+        if labels is not None:
+            labels = labels.clone()
+            if labels.size(1) > insert_idx:
+                pre_labels = labels[:, :insert_idx]
+                post_labels = labels[:, insert_idx + 1 :]
+                pad = torch.full((1, n_feats), -100, dtype=labels.dtype, device=labels.device)
+                labels = torch.cat([pre_labels, pad, post_labels], dim=1)
+        return inputs_embeds, attention_mask, labels
+
+    # Fallback: multi-placeholder handling
+    M = int(vid_pos.numel())
+    if M == 0:
+        raise ValueError("No video placeholder tokens found in input_ids for provided vision_features")
+
+    vision_features = vision_features.to(inputs_embeds.device, inputs_embeds.dtype)
+    N = n_feats
+    if N > M:
+        raise NotImplementedError(
+            "Number of vision features exceeds video placeholders; use a single <|video_pad|> token template."
+        )
+    if N < M:
+        drop_pos = vid_pos[N:]
+        if drop_pos.numel() > 0:
+            keep_seq = torch.ones(S, dtype=torch.bool, device=input_ids.device)
+            keep_seq[drop_pos] = False
+            input_ids = input_ids[:, keep_seq]
+            attention_mask = attention_mask[:, keep_seq]
+            inputs_embeds = inputs_embeds[:, keep_seq, :]
+            if labels is not None:
+                labels = labels[:, keep_seq]
+            vid_pos = (input_ids[0] == video_token_id).nonzero(as_tuple=False).flatten()
+            M = int(vid_pos.numel())
+
+    for i in range(N):
+        pos = int(vid_pos[i].item())
+        inputs_embeds[0, pos, :] = vision_features[i, :]
+    if labels is not None and N > 0:
+        labels = labels.clone()
+        labels[0, vid_pos[:N]] = -100
+
+    return inputs_embeds, attention_mask.to(inputs_embeds.device), labels
+
+
+# ------------------------------
+# QTS+ modules (selector + tokenizer)
+# ------------------------------
+class RMSNorm(nn.Module):
+    def __init__(self, d: int, eps: float = 1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(d))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        norm = x.pow(2).mean(dim=-1, keepdim=True)
+        x = x * torch.rsqrt(norm + self.eps)
+        return self.weight * x
+
+
+class FeedForward(nn.Module):
+    def __init__(self, d_model: int, d_ff: int, dropout: float = 0.0):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Linear(d_model, d_ff),
+            nn.GELU(),
+            nn.Linear(d_ff, d_model),
+            nn.Dropout(dropout),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.net(x)
+
+
+class Qwen2_5_ScoringCrossAttentionLayer(nn.Module):
+    """Qwen2.5-style cross-attention used in QTS+ scoring.
+
+    Separate q/k/v projections (with optional multi-query kv heads) followed by
+    an output projection and a small FFN on the query path.
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        num_heads: int,
+        num_key_value_heads: Optional[int] = None,
+        dropout: float = 0.0,
+        d_ff: Optional[int] = None,
+        rms_norm_eps: float = 1e-6,
+        use_qwen_rms: bool = True,
+    ) -> None:
+        super().__init__()
+        assert d_model % num_heads == 0
+        self.hidden_size = d_model
+        self.num_heads = int(num_heads)
+        self.head_dim = d_model // self.num_heads
+        self.num_key_value_heads = int(num_key_value_heads) if num_key_value_heads else self.num_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.attention_dropout = dropout
+
+        # Minimal Qwen-like RMS norms
+        class _Qwen2RMSNorm(nn.Module):
+            def __init__(self, hidden_size: int, eps: float = 1e-6):
+                super().__init__()
+                self.weight = nn.Parameter(torch.ones(hidden_size))
+                self.eps = float(eps)
+
+            def forward(self, x: torch.Tensor) -> torch.Tensor:
+                dtype = x.dtype
+                x = x.float()
+                variance = x.pow(2).mean(-1, keepdim=True)
+                x = x * torch.rsqrt(variance + self.eps)
+                x = x.to(dtype)
+                return self.weight * x
+
+        self.q_norm = _Qwen2RMSNorm(d_model, eps=rms_norm_eps) if use_qwen_rms else RMSNorm(d_model, eps=rms_norm_eps)
+        self.kv_norm = _Qwen2RMSNorm(d_model, eps=rms_norm_eps) if use_qwen_rms else RMSNorm(d_model, eps=rms_norm_eps)
+        self.ffn_norm = _Qwen2RMSNorm(d_model, eps=rms_norm_eps) if use_qwen_rms else RMSNorm(d_model, eps=rms_norm_eps)
+
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+
+        self.ffn = FeedForward(d_model, d_ff or (4 * d_model), dropout=dropout)
+
+    @staticmethod
+    def _repeat_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor:
+        b, h_kv, t, dh = x.shape
+        if n_rep == 1:
+            return x
+        x = x[:, :, None, :, :].expand(b, h_kv, n_rep, t, dh)
+        return x.reshape(b, h_kv * n_rep, t, dh)
+
+    def forward(
+        self,
+        q: torch.Tensor,  # [B, L, D]
+        kv: torch.Tensor,  # [B, M, D]
+        kv_key_padding_mask: Optional[torch.Tensor] = None,  # [B, M]
+        need_weights: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        B, L, _ = q.shape
+        _, M, _ = kv.shape
+
+        qn = self.q_norm(q)
+        kvn = self.kv_norm(kv)
+
+        q_states = self.q_proj(qn)
+        k_states = self.k_proj(kvn)
+        v_states = self.v_proj(kvn)
+
+        q_states = q_states.view(B, L, self.num_heads, self.head_dim).transpose(1, 2)
+        k_states = k_states.view(B, M, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        v_states = v_states.view(B, M, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        if self.num_key_value_groups > 1:
+            k_states = self._repeat_kv(k_states, self.num_key_value_groups)
+            v_states = self._repeat_kv(v_states, self.num_key_value_groups)
+
+        attn_weights = torch.matmul(q_states, k_states.transpose(2, 3)) / (self.head_dim ** 0.5)
+        if kv_key_padding_mask is not None:
+            mask = kv_key_padding_mask[:, None, None, :].to(dtype=attn_weights.dtype)
+            attn_weights = attn_weights.masked_fill(mask > 0.5, float("-inf"))
+        attn_dtype = attn_weights.dtype
+        attn_weights = torch.softmax(attn_weights, dim=-1, dtype=torch.float32).to(attn_dtype)
+        attn_output = torch.matmul(attn_weights, v_states)
+        attn_output = attn_output.transpose(1, 2).contiguous().view(B, L, self.num_heads * self.head_dim)
+
+        out = self.o_proj(attn_output)
+        q = q + out
+        q = q + self.ffn(self.ffn_norm(q))
+        return q, (attn_weights if need_weights else None)
+
+
+class Qwen2_5_SelfReencodeLayer(nn.Module):
+    def __init__(
+        self,
+        d_model: int,
+        num_heads: int,
+        num_key_value_heads: Optional[int] = None,
+        dropout: float = 0.0,
+        d_ff: Optional[int] = None,
+        rms_norm_eps: float = 1e-6,
+        use_qwen_rms: bool = True,
+    ) -> None:
+        super().__init__()
+        self.core = Qwen2_5_ScoringCrossAttentionLayer(
+            d_model=d_model,
+            num_heads=num_heads,
+            num_key_value_heads=num_key_value_heads or num_heads,
+            dropout=dropout,
+            d_ff=d_ff,
+            rms_norm_eps=rms_norm_eps,
+            use_qwen_rms=use_qwen_rms,
+        )
+
+    def forward(self, x: torch.Tensor, key_padding_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
+        y, _ = self.core(x, x, kv_key_padding_mask=key_padding_mask, need_weights=False)
+        return y
+
+    def init_from_qwen_attn(self, qwen_attn: nn.Module, qwen_input_ln: Optional[nn.Module] = None) -> None:
+        self.core.init_from_qwen_attn(qwen_attn, qwen_input_ln)
+
+
+class BudgetHead(nn.Module):
+    def __init__(self, d_model: int, hidden: int = 256, rho_min: float = 0.05, rho_max: float = 0.5) -> None:
+        super().__init__()
+        self.rho_min = rho_min
+        self.rho_max = rho_max
+        self.mlp = nn.Sequential(
+            nn.Linear(d_model + 3, hidden),
+            nn.GELU(),
+            nn.Linear(hidden, 1),
+        )
+
+    def forward(self, sq: torch.Tensor, logM: torch.Tensor, r_max: torch.Tensor, H: torch.Tensor) -> torch.Tensor:
+        B, D = sq.shape
+        x = torch.cat([sq, logM.view(B, 1), r_max.view(B, 1), H.view(B, 1)], dim=1)
+        # Ensure input dtype matches layer weights to avoid Float/Half mismatch
+        x = x.to(dtype=self.mlp[0].weight.dtype)
+        logits = self.mlp(x).squeeze(1)
+        rho = self.rho_min + (self.rho_max - self.rho_min) * torch.sigmoid(logits)
+        return rho
+
+
+class QTSplus(nn.Module):
+    """Query‑Aware Token Selector with Adaptive Budget."""
+
+    def __init__(
+        self,
+        d_model: int,
+        n_heads: int = 8,
+        n_kv_heads: Optional[int] = None,
+        tau_s: float = 0.1,
+        nmax: int = 2560,
+        rho_min: float = 0.05,
+        rho_max: float = 0.5,
+        block_dropout: float = 0.0,
+        use_reencode: bool = True,
+        n_scoring_layers: int = 1,
+        n_reencode_layers: int = 1,
+    ) -> None:
+        super().__init__()
+        assert d_model % n_heads == 0
+        self.d_model = d_model
+        self.n_heads = int(n_heads)
+        self.d_head = d_model // self.n_heads
+        self.tau_s = float(tau_s)
+        self.nmax = int(nmax)
+        self.use_reencode = bool(use_reencode)
+        self.n_scoring_layers = max(int(n_scoring_layers), 1)
+        self.n_reencode_layers = max(int(n_reencode_layers), 1)
+
+        n_kv_heads_eff = int(n_kv_heads) if (n_kv_heads is not None and int(n_kv_heads) > 0) else self.n_heads
+        self.scoring_layers = nn.ModuleList(
+            [
+                Qwen2_5_ScoringCrossAttentionLayer(
+                    d_model,
+                    num_heads=self.n_heads,
+                    num_key_value_heads=n_kv_heads_eff,
+                    dropout=0.0,
+                    rms_norm_eps=1e-6,
+                    use_qwen_rms=True,
+                )
+                for _ in range(self.n_scoring_layers)
+            ]
+        )
+
+        self.budget = BudgetHead(d_model, rho_min=rho_min, rho_max=rho_max)
+
+        if self.use_reencode:
+            self.reencode_layers = nn.ModuleList(
+                [
+                    Qwen2_5_SelfReencodeLayer(
+                        d_model,
+                        num_heads=self.n_heads,
+                        num_key_value_heads=n_kv_heads_eff,
+                        dropout=0.0,
+                        rms_norm_eps=1e-6,
+                        use_qwen_rms=True,
+                    )
+                    for _ in range(self.n_reencode_layers)
+                ]
+            )
+        else:
+            self.reencode_layers = None
+
+    def _score(self, Xv: torch.Tensor, Qt: torch.Tensor) -> torch.Tensor:
+        # Simple cross-attention based scoring aggregated across heads and query positions
+        B, M, D = Xv.shape
+        q = Qt
+        kv = Xv
+        for layer in self.scoring_layers:
+            q, attn = layer(q, kv, need_weights=True)
+        # attn: [B, H, L, M]; aggregate -> [B, M]
+        r = attn.amax(dim=2).mean(dim=1)
+        return r
+
+    def _predict_budget(self, q: torch.Tensor, r: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        B, L, D = q.shape
+        M = r.shape[-1]
+        sq = q.mean(dim=1)
+        # Create logM with same dtype/device as q to keep types consistent
+        logM = torch.log(torch.tensor(float(M), device=q.device, dtype=q.dtype)).expand(B)
+        r_max = r.max(dim=-1).values
+        # entropy H over token scores after softmax
+        p = torch.softmax(r, dim=-1)
+        H = -(p * (p.clamp(min=1e-12).log())).sum(dim=-1)
+        rho = self.budget(sq, logM, r_max, H)
+        # n = clamp(round(rho * M), 1, nmax)
+        n = torch.clamp((rho * float(M)).round(), min=1.0, max=float(self.nmax)).to(torch.long)
+        return rho, n
+
+    def forward(self, Xv: torch.Tensor, Qt: torch.Tensor, mode: str = "train") -> Dict[str, Any]:
+        assert mode in ("train", "infer")
+        B, M, D = Xv.shape
+        r = self._score(Xv, Qt)
+        rho, n = self._predict_budget(Qt, r)
+
+        # Hard top-n with original order preserved
+        kept_idx_list: List[torch.Tensor] = []
+        Z_out: List[torch.Tensor] = []
+        for b in range(B):
+            kb = torch.topk(r[b], k=int(n[b].item()), dim=0).indices
+            kb, _ = torch.sort(kb)
+            kept_idx_list.append(kb)
+            Z_out.append(Xv[b, kb])
+
+        if self.use_reencode:
+            max_keep = int(max(z.size(0) for z in Z_out))
+            Zb = []
+            for z in Z_out:
+                if z.size(0) < max_keep:
+                    pad = z[-1:].repeat(max_keep - z.size(0), 1)
+                    z = torch.cat([z, pad], dim=0)
+                Zb.append(z.unsqueeze(0))
+            Zb = torch.cat(Zb, dim=0)
+            for layer in self.reencode_layers or []:
+                Zb = layer(Zb)
+            Z_final = [Zb[b, : kept_idx_list[b].numel()] for b in range(B)]
+        else:
+            Z_final = Z_out
+
+        # Simple training proxies
+        p = torch.softmax(r, dim=-1)
+        flops_proxy = ((rho * float(M)) ** 2) / float(self.nmax ** 2)
+        kv_proxy = (rho * float(M)) / float(self.nmax)
+
+        return {
+            "indices": kept_idx_list,
+            "Z": Z_final,
+            "rho": rho,
+            "r": r,
+            "n": n,
+            "add_loss": {
+                "flops": flops_proxy.mean(),
+                "kv": kv_proxy.mean(),
+                "smooth": torch.tensor(0.0, device=Xv.device, dtype=Xv.dtype),
+            },
+        }
+
+
+class QTSplusTokenizerConfig:
+    def __init__(
+        self,
+        embedding_dim: int,
+        n_heads: int = 8,
+        num_kv_heads: Optional[int] = None,
+        tau_s: float = 0.1,
+        nmax: int = 2560,
+        rho_min: float = 0.05,
+        rho_max: float = 0.5,
+        block_dropout: float = 0.0,
+        reencode: bool = True,
+        scoring_layers: int = 1,
+        reencode_layers: int = 1,
+        lambda_t: float = 1.0,
+        lambda_m: float = 1.7,
+        lambda_s: float = 0.05,
+        project_text_if_needed: bool = False,
+    ) -> None:
+        self.embedding_dim = embedding_dim
+        self.n_heads = n_heads
+        self.num_kv_heads = num_kv_heads
+        self.tau_s = tau_s
+        self.nmax = nmax
+        self.rho_min = rho_min
+        self.rho_max = rho_max
+        self.block_dropout = block_dropout
+        self.reencode = reencode
+        self.scoring_layers = scoring_layers
+        self.reencode_layers = reencode_layers
+        self.lambda_t = lambda_t
+        self.lambda_m = lambda_m
+        self.lambda_s = lambda_s
+        self.project_text_if_needed = project_text_if_needed
+
+
+class QTSplusTokenizer(nn.Module):
+    def __init__(self, cfg: QTSplusTokenizerConfig) -> None:
+        super().__init__()
+        self.cfg = cfg
+        self.selector = QTSplus(
+            d_model=cfg.embedding_dim,
+            n_heads=cfg.n_heads,
+            n_kv_heads=cfg.num_kv_heads or cfg.n_heads,
+            tau_s=cfg.tau_s,
+            nmax=cfg.nmax,
+            rho_min=cfg.rho_min,
+            rho_max=cfg.rho_max,
+            block_dropout=cfg.block_dropout,
+            use_reencode=cfg.reencode,
+            n_scoring_layers=cfg.scoring_layers,
+            n_reencode_layers=cfg.reencode_layers,
+        )
+        self.text_proj: Optional[nn.Linear] = None
+
+    def forward(self, X_v: torch.Tensor, Q_t: torch.Tensor, mode: str = "train") -> Dict[str, Any]:
+        B, M, D = X_v.shape
+        D_txt = Q_t.shape[-1]
+        if D_txt != D:
+            if self.cfg.project_text_if_needed:
+                if self.text_proj is None:
+                    self.text_proj = nn.Linear(D_txt, D, bias=False).to(device=Q_t.device, dtype=Q_t.dtype)
+                Q_proj = self.text_proj(Q_t)
+            else:
+                raise ValueError(f"QTS+ expects text dim {D}, got {D_txt}. Set project_text_if_needed=True.")
+        else:
+            Q_proj = Q_t
+        sel = self.selector(X_v, Q_proj, mode=mode)
+        # Add simple proxies for train-time regularization
+        M_tensor = torch.tensor(float(M), device=X_v.device, dtype=X_v.dtype)
+        rho = sel["rho"]
+        flops_proxy = ((rho * M_tensor) ** 2) / float(self.cfg.nmax ** 2)
+        kv_proxy = (rho * M_tensor) / float(self.cfg.nmax)
+        sel["add_loss"] = {
+            "flops": flops_proxy.mean() * self.cfg.lambda_t,
+            "kv": kv_proxy.mean() * self.cfg.lambda_m,
+            "smooth": torch.tensor(0.0, device=X_v.device, dtype=X_v.dtype),
+        }
+        return sel
+
+class Qwen2_5_VisionTransformerPretrainedModel(Qwen2_5_VisionTransformerPretrainedModelBase):
+    def __init__(self, config, *inputs, **kwargs) -> None:
+        super().__init__(config, *inputs, **kwargs)
+    
+    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs) -> torch.Tensor:
+        # Return the output from the base implementation.
+        # Without this return, callers receive None and downstream code fails.
+        return super().forward(hidden_states, grid_thw, **kwargs)
+
+    def get_video_features(
+        self, pixel_values_videos: torch.FloatTensor, video_grid_thw: Optional[torch.LongTensor] = None
+    ):
+        """
+        Encodes videos into continuous embeddings that can be forwarded to the language model.
+
+        Args:
+            pixel_values_videos (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
+                The tensors corresponding to the input videos.
+            video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
+                The temporal, height and width of feature shape of each video in LLM.
+        """
+        pixel_values_videos = pixel_values_videos.type(self.dtype)
+        video_embeds = self.forward(pixel_values_videos, grid_thw=video_grid_thw)
+        # split_sizes = (video_grid_thw.prod(-1) // self.spatial_merge_size**2).tolist()
+        # video_embeds = torch.split(video_embeds, split_sizes)
+        return video_embeds
+    
+def _try_load_vision_config_from_path(path: str) -> Optional[Dict[str, Any]]:
+    """Best-effort load of Qwen2.5-VL vision `config.json`.
+
+    Accepts either a directory containing `config.json` or a file path to a
+    weights file. In the latter case, attempts to locate a sibling
+    `config.json` in the same directory.
+    """
+    if not path:
+        return None
+
+    cfg_path = None
+    if os.path.isdir(path):
+        candidate = os.path.join(path, "config.json")
+        if os.path.isfile(candidate):
+            cfg_path = candidate
+    else:
+        # If a file is given (e.g., .../model.safetensors), look next to it
+        base_dir = os.path.dirname(path)
+        candidate = os.path.join(base_dir, "config.json")
+        if os.path.isfile(candidate):
+            cfg_path = candidate
+
+    if cfg_path is None:
+        return None
+
+    try:
+        with open(cfg_path, "r", encoding="utf-8") as f:
+            return json.load(f)
+    except Exception:
+        return None
+
+
+def build_vision_tower(vision_tower_cfg, **kwargs):
+    vision_tower = getattr(vision_tower_cfg, 'mm_vision_tower', getattr(vision_tower_cfg, 'vision_tower', None))
+    if vision_tower != "qwen2_5_vl_vision":
+        raise ValueError(f"Unknown vision tower type: {vision_tower}")
+
+    # Attempt to infer correct dimensions from the provided pretrained path
+    pretrained_path = getattr(vision_tower_cfg, 'pretrain_vision_model', None)
+    cfg_json = _try_load_vision_config_from_path(pretrained_path) if pretrained_path else None
+
+    if cfg_json is not None:
+        # Map json fields to Qwen2_5_VLVisionConfig kwargs (use json defaults when available)
+        config = Qwen2_5_VLVisionConfig(
+            hidden_size=cfg_json.get("hidden_size", 1280),
+            out_hidden_size=cfg_json.get("out_hidden_size", cfg_json.get("hidden_size", 1280)),
+            depth=cfg_json.get("depth", 32),
+            intermediate_size=cfg_json.get("intermediate_size", 3420),
+            num_heads=cfg_json.get("num_heads", 16),
+            fullatt_block_indexes=cfg_json.get("fullatt_block_indexes", [7, 15, 23, 31]),
+            in_channels=cfg_json.get("in_channels", cfg_json.get("in_chans", 3)),
+            patch_size=cfg_json.get("patch_size", cfg_json.get("spatial_patch_size", 14)),
+            spatial_merge_size=cfg_json.get("spatial_merge_size", 2),
+            temporal_patch_size=cfg_json.get("temporal_patch_size", 2),
+            tokens_per_second=cfg_json.get("tokens_per_second", 2),
+            window_size=cfg_json.get("window_size", 112),
+            initializer_range=cfg_json.get("initializer_range", 0.02),
+        )
+    else:
+        # Fallback to a safe default (3B) when no config file is available
+        # This keeps backwards-compatibility but different-scale checkpoints
+        # should always provide a config.json alongside the weights.
+        config = Qwen2_5_VLVisionConfig(
+            hidden_size=1280,
+            out_hidden_size=2048,
+            depth=32,
+            intermediate_size=3420,
+            num_heads=16,
+            fullatt_block_indexes=[7, 15, 23, 31],
+        )
+
+    return Qwen2_5_VisionTransformerPretrainedModel(config)
+
+# ------------------------------
+# Builders used by the meta model
+# ------------------------------
+def build_vision_tower(config: Qwen2_5_VLTextConfig) -> Qwen2_5_VisionTransformerPretrainedModel:
+    vcfg_dict = getattr(config, "vision_config", None) or {}
+    vcfg = Qwen2_5_VLVisionConfig(**vcfg_dict) if vcfg_dict else Qwen2_5_VLVisionConfig()
+    return Qwen2_5_VisionTransformerPretrainedModel(vcfg)
+
+
+def build_qts_plus_tower(config: Qwen2_5_VLTextConfig) -> QTSplusTokenizer:
+    lm_heads = getattr(config, "num_attention_heads", None)
+    vision_dim = getattr(config, "vision_embed_size", None)
+    if not isinstance(lm_heads, int) or lm_heads <= 0:
+        raise ValueError("num_attention_heads must be provided by the Qwen2.5‑VL config")
+    if not isinstance(vision_dim, int) or vision_dim <= 0:
+        raise ValueError("vision_embed_size must be a positive int before building QTS+")
+    if vision_dim % lm_heads != 0:
+        raise ValueError(
+            f"vision_embed_size ({vision_dim}) must be divisible by LM num_attention_heads ({lm_heads})"
+        )
+    kv_heads = getattr(config, "num_key_value_heads", None)
+    cfg = QTSplusTokenizerConfig(
+        embedding_dim=vision_dim,
+        n_heads=lm_heads,
+        num_kv_heads=kv_heads if isinstance(kv_heads, int) and kv_heads > 0 else None,
+        tau_s=getattr(config, "qts_plus_tau_s", 0.1),
+        nmax=getattr(config, "qts_plus_nmax", 2560),
+        rho_min=getattr(config, "qts_plus_rho_min", 0.05),
+        rho_max=getattr(config, "qts_plus_rho_max", 0.5),
+        block_dropout=getattr(config, "qts_plus_block_dropout", 0.0),
+        reencode=getattr(config, "qts_plus_reencode", True),
+        scoring_layers=getattr(config, "qts_plus_scoring_layers", 1),
+        reencode_layers=getattr(config, "qts_plus_reencode_layers", 1),
+        lambda_t=getattr(config, "lambda_t", 1.0),
+        lambda_m=getattr(config, "lambda_m", 1.7),
+        lambda_s=getattr(config, "lambda_s", 0.05),
+        project_text_if_needed=getattr(config, "project_text_if_needed", False),
+    )
+    return QTSplusTokenizer(cfg)
+
+
+# ------------------------------
+# Meta classes to build vision/QTS+ towers and preprocessing hook
+# ------------------------------
+class QTSplusMetaModel:
+    def __init__(self, config):
+        super(QTSplusMetaModel, self).__init__(config)
+        self.config = config
+
+        # Vision tower: build early so weights under `model.vision_tower.*` load
+        if hasattr(config, "vision_tower"):
+            self.vision_tower = build_vision_tower(config)
+            try:
+                vt = getattr(self, "vision_tower", None)
+                out_hidden = getattr(getattr(vt, "config", None), "out_hidden_size", None)
+                if isinstance(out_hidden, int) and out_hidden > 0:
+                    self.config.vision_embed_size = out_hidden
+            except Exception:
+                pass
+
+        # QTS+ tower: build early if enabled so parameters exist during load
+        if getattr(self.config, "enable_qts_plus", False) and getattr(self, "qts_plus", None) is None:
+            try:
+                self.qts_plus = build_qts_plus_tower(self.config)
+            except Exception:
+                pass
+
+    def get_qts_plus_tower(self):
+        return getattr(self, "qts_plus", None)
+
+    def get_vision_tower(self):
+        return getattr(self, "vision_tower", None)
+
+
+class QTSplusMetaForCausalLM:
+    def get_model(self):  # pragma: no cover - abstract in practice
+        raise NotImplementedError
+
+    def get_vision_tower(self):
+        return self.get_model().get_vision_tower()
+
+    def get_qts_plus_tower(self):
+        return self.get_model().get_qts_plus_tower()
+
+    def encode_visions(self, vision):
+        return self.get_model().get_vision_tower()(vision)
+
+    def prepare_inputs_for_multimodal(
+        self,
+        vision_input,
+        input_ids,
+        position_ids,
+        attention_mask,
+        past_key_values,
+        labels,
+        question_input_ids: Optional[torch.Tensor] = None,
+        video_token_id: Optional[int] = None,
+        mode: str = "train",
+    ):
+        vision_tower = self.get_vision_tower()
+        qts_plus_tower = self.get_qts_plus_tower()
+        text_embed_layer = self.get_model().get_input_embeddings()
+
+        if vision_tower is None or vision_input is None or input_ids.shape[1] == 1:
+            # Match text embedding dtype for scalar placeholders
+            z = torch.tensor(0.0, device=input_ids.device, dtype=text_embed_layer.weight.dtype)
+            return input_ids, position_ids, attention_mask, past_key_values, None, labels, z, z, z
+
+        if self.config.enable_qts_plus:
+            if self.config.vision_tower == "qwen2_5_vl_vision":
+                if isinstance(vision_input, list):
+                    if len(vision_input) == 0:
+                        z = torch.tensor(0.0, device=input_ids.device, dtype=text_embed_layer.weight.dtype)
+                        return input_ids, position_ids, attention_mask, past_key_values, None, labels, z, z, z
+                    vision_input = vision_input[0]
+
+                vision_features = vision_tower.get_video_features(
+                    vision_input["pixel_values_videos"].to(vision_tower.device),
+                    vision_input["video_grid_thw"].to(vision_tower.device),
+                )
+                video_grid_thw = vision_input["video_grid_thw"]
+                if isinstance(vision_features, list) and len(vision_features) > 0:
+                    vision_features = vision_features[0]
+                if vision_features.ndim == 2:
+                    vision_features = vision_features.unsqueeze(0)
+
+                if question_input_ids is None:
+                    raise AssertionError("question_input_ids must be provided in training to avoid data leakage")
+                if question_input_ids.dtype is not torch.long:
+                    question_input_ids = question_input_ids.long()
+
+                text_embeddings = text_embed_layer(question_input_ids)
+                vision_features = vision_features.to(dtype=text_embeddings.dtype)
+
+                qts_plus_out = qts_plus_tower(vision_features, text_embeddings, mode=mode)
+                vision_features = qts_plus_out["Z"]
+                flops_loss = qts_plus_out["add_loss"]["flops"]
+                kv_loss = qts_plus_out["add_loss"]["kv"]
+                smooth_loss = qts_plus_out["add_loss"]["smooth"]
+
+                if video_token_id is None:
+                    video_token_id = getattr(self.config, "video_token_id", None) or 151656
+
+                inputs_embeds, attention_mask, labels = qts_integrate_embeddings(
+                    vision_features=vision_features[0],
+                    input_ids=input_ids,
+                    attention_mask=attention_mask,
+                    labels=labels,
+                    video_token_id=video_token_id,
+                    text_model_embed_layer=text_embed_layer,
+                    video_grid_thw=video_grid_thw,
+                )
+                return (
+                    vision_input,
+                    position_ids,
+                    attention_mask,
+                    past_key_values,
+                    inputs_embeds,
+                    labels,
+                    flops_loss,
+                    kv_loss,
+                    smooth_loss,
+                )
+            else:
+                raise ValueError("Not support this model")
+
+        # QTS+ disabled: just embed tokens
+        z = torch.tensor(0.0, device=input_ids.device, dtype=text_embed_layer.weight.dtype)
+        return input_ids, position_ids, attention_mask, past_key_values, None, labels, z, z, z
+
+    def vision_features_count_qtsplus(
+        self,
+        pixel_values_videos: Optional[torch.Tensor],
+        video_grid_thw: Optional[torch.Tensor],
+        question_input_ids: Optional[torch.Tensor],
+    ) -> int:
+        try:
+            if pixel_values_videos is None or video_grid_thw is None or question_input_ids is None:
+                return 0
+            vision_tower = self.get_vision_tower()
+            qts_tower = self.get_qts_plus_tower()
+            text_embed = self.get_model().get_input_embeddings()
+            if vision_tower is None or qts_tower is None or text_embed is None:
+                return 0
+            if question_input_ids.dtype is not torch.long:
+                question_input_ids = question_input_ids.long()
+            try:
+                vt_device = next(vision_tower.parameters()).device
+            except StopIteration:
+                vt_device = text_embed.weight.device
+            vf = vision_tower.get_video_features(
+                pixel_values_videos.to(vt_device),
+                video_grid_thw.to(vt_device),
+            )
+            if isinstance(vf, list) and len(vf) > 0:
+                vf = vf[0]
+            if isinstance(vf, torch.Tensor) and vf.ndim == 2:
+                vf = vf.unsqueeze(0)
+            te = text_embed(question_input_ids.to(text_embed.weight.device))
+            if isinstance(vf, torch.Tensor):
+                vf = vf.to(device=te.device, dtype=te.dtype)
+            with torch.inference_mode():
+                qpo = qts_tower(vf, te, mode="infer")
+            Z = qpo.get("Z")
+            if isinstance(Z, list) and len(Z) > 0:
+                return int(Z[0].shape[0])
+            if isinstance(Z, torch.Tensor):
+                return int(Z.shape[1] if Z.ndim == 3 else Z.shape[0])
+            return 0
+        except Exception:
+            return 0
+
+
+# ------------------------------
+# Base text-only CausalLM for Qwen2.5‑VL (local copy)
+# ------------------------------
+class Qwen2_5_VL_CausalLM_Config(Qwen2_5_VLTextConfig):
+    model_type = "qwen2_5_vl_causal_lm"
+
+
+class Qwen2_5_VLTextForCausalLM(Qwen2_5_VLPreTrainedModel, GenerationMixin):
+    config_class = Qwen2_5_VL_CausalLM_Config
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config: Qwen2_5_VL_CausalLM_Config):
+        super().__init__(config)
+        self.model = Qwen2_5_VLTextModel(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def get_decoder(self):
+        return self.model
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        num_logits_to_keep: int = 0,
+        **loss_kwargs,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+
+        loss = None
+        if labels is not None:
+            # Defer to simple cross-entropy with ignore_index set by caller
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+# ------------------------------
+# QTS+ Qwen2.5‑VL Causal LM (text model + QTS+ + vision)
+# ------------------------------
+class QTSplusQwen2_5_VLModel(QTSplusMetaModel, Qwen2_5_VLTextModel):
+    config_class = QTSplusQwen2_5_VL_CausalLM_Config
+
+    def __init__(self, config: Qwen2_5_VLTextConfig):
+        super(QTSplusQwen2_5_VLModel, self).__init__(config)
+
+
+class QTSplusQwen2_5_VLTextForCausalLM(QTSplusMetaForCausalLM, Qwen2_5_VLTextForCausalLM):
+    config_class = QTSplusQwen2_5_VL_CausalLM_Config
+
+    def __init__(self, config):
+        try:
+            cfg_attn = getattr(config, "attn_implementation", None)
+            if (cfg_attn is None or str(cfg_attn) == "auto") and is_flash_attn_available():
+                setattr(config, "attn_implementation", "flash_attention_2")
+                setattr(config, "_attn_implementation", "flash_attention_2")
+        except Exception:
+            pass
+
+        super(Qwen2_5_VLTextForCausalLM, self).__init__(config)
+        self.model = QTSplusQwen2_5_VLModel(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.post_init()
+
+    def get_model(self):
+        return self.model
+
+    def forward(
+        self,
+        vision_input: Optional[torch.FloatTensor] = None,
+        input_ids: torch.LongTensor = None,
+        labels: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        question_input_ids: Optional[torch.LongTensor] = None,
+        video_token_id: Optional[int] = None,
+    ):
+        if inputs_embeds is not None:
+            input_ids = None
+
+        if inputs_embeds is None:
+            (
+                vision_input,
+                position_ids,
+                attention_mask,
+                past_key_values,
+                inputs_embeds,
+                labels,
+                flops_loss,
+                kv_loss,
+                smooth_loss,
+            ) = self.prepare_inputs_for_multimodal(
+                vision_input,
+                input_ids,
+                position_ids,
+                attention_mask,
+                past_key_values,
+                labels,
+                question_input_ids,
+                video_token_id,
+                mode="train" if self.training else "infer",
+            )
+            if inputs_embeds is None:
+                inputs_embeds = self.get_model().embed_tokens(input_ids)
+
+        input_ids = None
+        try:
+            outputs = super().forward(
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                inputs_embeds=inputs_embeds,
+                labels=labels,
+                use_cache=use_cache,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+                cache_position=cache_position,
+            )
+        except ValueError as error:
+            raise ValueError(
+                f"{error} (input_ids is None: {input_ids is None}, inputs_embeds is None: {inputs_embeds is None})"
+            ) from error
+
+        add_loss = {
+            "flops_loss": flops_loss if vision_input is not None else 0.0,
+            "kv_loss": kv_loss if vision_input is not None else 0.0,
+            "smooth_loss": smooth_loss if vision_input is not None else 0.0,
+        }
+        if labels is None and not self.training:
+            return outputs
+        return (outputs, add_loss)
+
+    @torch.no_grad()
+    def generate(
+        self,
+        vision_input: Optional[torch.Tensor] = None,
+        input_ids: Optional[torch.Tensor] = None,
+        question_input_ids: Optional[torch.Tensor] = None,
+        video_token_id: Optional[int] = None,
+        **kwargs,
+    ):
+        position_ids = kwargs.pop("position_ids", None)
+        attention_mask = kwargs.pop("attention_mask", None)
+        if attention_mask is None and input_ids is not None:
+            attention_mask = torch.ones_like(input_ids, dtype=torch.long, device=input_ids.device)
+        if "inputs_embeds" in kwargs:
+            raise NotImplementedError("`inputs_embeds` is not supported")
+
+        if vision_input is not None:
+            (
+                vision_input,
+                position_ids,
+                attention_mask,
+                _,
+                inputs_embeds,
+                _,
+                *_unused_losses,
+            ) = self.prepare_inputs_for_multimodal(
+                vision_input,
+                input_ids,
+                position_ids,
+                attention_mask,
+                None,
+                None,
+                question_input_ids,
+                video_token_id,
+                mode="infer",
+            )
+        else:
+            inputs_embeds = self.get_model().embed_tokens(input_ids)
+
+        kwargs["attention_mask"] = attention_mask
+        if position_ids is not None:
+            kwargs["position_ids"] = position_ids
+        kwargs.pop("input_ids", None)
+        if "use_cache" not in kwargs:
+            kwargs["use_cache"] = True
+        output_ids = super().generate(inputs_embeds=inputs_embeds, **kwargs)
+        if input_ids is not None:
+            input_ids = input_ids.to(output_ids.device)
+            output_ids = torch.cat([input_ids, output_ids], dim=1)
+        return output_ids
+    
+# Register for Auto* resolution
+AutoModelForCausalLM.register(QTSplusQwen2_5_VL_CausalLM_Config, QTSplusQwen2_5_VLTextForCausalLM)
+__all__ = ["QTSplusQwen2_5_VLTextForCausalLM"]

preprocessor_config.json

@@ -0,0 +1,39 @@
+{
+  "crop_size": null,
+  "data_format": "channels_first",
+  "default_to_square": true,
+  "device": null,
+  "disable_grouping": null,
+  "do_center_crop": null,
+  "do_convert_rgb": true,
+  "do_normalize": true,
+  "do_pad": null,
+  "do_rescale": true,
+  "do_resize": true,
+  "image_mean": [
+    0.48145466,
+    0.4578275,
+    0.40821073
+  ],
+  "image_processor_type": "Qwen2VLImageProcessorFast",
+  "image_std": [
+    0.26862954,
+    0.26130258,
+    0.27577711
+  ],
+  "input_data_format": null,
+  "max_pixels": 12845056,
+  "merge_size": 2,
+  "min_pixels": 3136,
+  "pad_size": null,
+  "patch_size": 14,
+  "processor_class": "Qwen2_5_VLVisionProcessor",
+  "resample": 3,
+  "rescale_factor": 0.00392156862745098,
+  "return_tensors": null,
+  "size": {
+    "longest_edge": 12845056,
+    "shortest_edge": 3136
+  },
+  "temporal_patch_size": 2
+}

processing_qts_plus_qwen2_5_vl.py

@@ -0,0 +1,260 @@
+"""
+Self-contained processor shim for trust_remote_code.
+
+Exports `QTSplusQwen2_5_VLProcessor` by aliasing the upstream
+Qwen2.5-VL processor from Transformers. This avoids importing a local
+`src` package while keeping the same class name referenced in
+`processor_config.json`.
+"""
+from typing import Optional, Union
+
+import numpy as np
+
+from transformers.feature_extraction_utils import BatchFeature
+from transformers.image_utils import ImageInput
+from transformers.processing_utils import ImagesKwargs, MultiModalData, ProcessingKwargs, ProcessorMixin, Unpack, VideosKwargs
+from transformers.tokenization_utils_base import PreTokenizedInput, TextInput
+from transformers.video_utils import VideoInput
+from transformers import AutoProcessor
+
+class Qwen2_5_VLVideosProcessorKwargs(VideosKwargs, total=False):
+    fps: Union[list[float], float]
+
+
+class Qwen2_5_VLImagesKwargs(ImagesKwargs):
+    min_pixels: Optional[int]
+    max_pixels: Optional[int]
+    patch_size: Optional[int]
+    temporal_patch_size: Optional[int]
+    merge_size: Optional[int]
+
+
+class Qwen2_5_VLProcessorKwargs(ProcessingKwargs, total=False):
+    images_kwargs: Qwen2_5_VLImagesKwargs
+    videos_kwargs: Qwen2_5_VLVideosProcessorKwargs
+    _defaults = {
+        "text_kwargs": {
+            "padding": False,
+            "return_mm_token_type_ids": False,
+        },
+    }
+
+
+class QTSplusQwen2_5_VLProcessor(ProcessorMixin):
+    r"""
+    Constructs a Qwen2.5-VL processor which wraps a Qwen2.5-VL image processor and a Qwen2 tokenizer into a single processor.
+    [`Qwen2_5_VLProcessor`] offers all the functionalities of [`Qwen2VLImageProcessor`] and [`Qwen2TokenizerFast`]. See the
+    [`~Qwen2_5_VLProcessor.__call__`] and [`~Qwen2_5_VLProcessor.decode`] for more information.
+    Args:
+        image_processor ([`Qwen2VLImageProcessor`], *optional*):
+            The image processor is a required input.
+        tokenizer ([`Qwen2TokenizerFast`], *optional*):
+            The tokenizer is a required input.
+        video_processor ([`Qwen2_5_VLVideoProcessor`], *optional*):
+            The video processor is a required input.
+        chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages
+            in a chat into a tokenizable string.
+    """
+
+    attributes = ["image_processor", "tokenizer", "video_processor"]
+
+    image_processor_class = "AutoImageProcessor"
+    video_processor_class = "AutoVideoProcessor"
+    tokenizer_class = ("Qwen2Tokenizer", "Qwen2TokenizerFast")
+
+    def __init__(self, image_processor=None, tokenizer=None, video_processor=None, chat_template=None, **kwargs):
+        self.image_token = "<|image_pad|>" if not hasattr(tokenizer, "image_token") else tokenizer.image_token
+        self.video_token = "<|video_pad|>" if not hasattr(tokenizer, "video_token") else tokenizer.video_token
+        self.image_token_id = (
+            tokenizer.image_token_id
+            if getattr(tokenizer, "image_token_id", None)
+            else tokenizer.convert_tokens_to_ids(self.image_token)
+        )
+        self.video_token_id = (
+            tokenizer.video_token_id
+            if getattr(tokenizer, "video_token_id", None)
+            else tokenizer.convert_tokens_to_ids(self.video_token)
+        )
+        super().__init__(image_processor, tokenizer, video_processor, chat_template=chat_template)
+
+    def __call__(
+        self,
+        images: Optional[ImageInput] = None,
+        text: Union[TextInput, PreTokenizedInput, list[TextInput], list[PreTokenizedInput]] = None,
+        videos: Optional[VideoInput] = None,
+        **kwargs: Unpack[Qwen2_5_VLProcessorKwargs],
+    ) -> BatchFeature:
+        """
+        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        and `kwargs` arguments to Qwen2TokenizerFast's [`~Qwen2TokenizerFast.__call__`] if `text` is not `None` to encode
+        the text. To prepare the vision inputs, this method forwards the `vision_infos` and `kwargs` arguments to
+        Qwen2VLImageProcessor's [`~Qwen2VLImageProcessor.__call__`] if `vision_infos` is not `None`.
+
+        Args:
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `list[PIL.Image.Image]`, `list[np.ndarray]`, `list[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. Both channels-first and channels-last formats are supported.
+            text (`str`, `list[str]`, `list[list[str]]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            videos (`np.ndarray`, `torch.Tensor`, `list[np.ndarray]`, `list[torch.Tensor]`):
+                The image or batch of videos to be prepared. Each video can be a 4D NumPy array or PyTorch
+                tensor, or a nested list of 3D frames. Both channels-first and channels-last formats are supported.
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+                - `'jax'`: Return JAX `jnp.ndarray` objects.
+
+        Returns:
+            [`BatchFeature`]: A [`BatchFeature`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+            - **pixel_values_videos** -- Pixel values of videos to be fed to a model. Returned when `videos` is not `None`.
+            - **image_grid_thw** -- List of image 3D grid in LLM. Returned when `images` is not `None`.
+            - **video_grid_thw** -- List of video 3D grid in LLM. Returned when `videos` is not `None`.
+            - **second_per_grid_ts** -- List of video seconds per time grid. Returned when `videos` is not `None`.
+        """
+        output_kwargs = self._merge_kwargs(
+            Qwen2_5_VLProcessorKwargs,
+            tokenizer_init_kwargs=self.tokenizer.init_kwargs,
+            **kwargs,
+        )
+
+        image_inputs = videos_inputs = {}
+        if images is not None:
+            image_inputs = self.image_processor(images=images, **output_kwargs["images_kwargs"])
+            image_grid_thw = image_inputs["image_grid_thw"]
+
+        if videos is not None:
+            fps = output_kwargs["videos_kwargs"].get("fps", 2.0)
+            videos_inputs = self.video_processor(videos=videos, **output_kwargs["videos_kwargs"])
+            video_grid_thw = videos_inputs["video_grid_thw"]
+
+            if isinstance(fps, (int, float)):
+                second_per_grid_ts = [self.video_processor.temporal_patch_size / fps] * len(video_grid_thw)
+            elif hasattr(fps, "__len__") and len(fps) == len(video_grid_thw):
+                second_per_grid_ts = [self.video_processor.temporal_patch_size / tmp for tmp in fps]
+            else:
+                raise ValueError(
+                    f"The length of fps ({len(fps) if hasattr(fps, '__len__') else fps}) must be equal to the length of video_grid_thw ({len(video_grid_thw)}) or fps should be a single number."
+                )
+            videos_inputs.update({"second_per_grid_ts": second_per_grid_ts})
+
+        if not isinstance(text, list):
+            text = [text]
+
+        text = text.copy()  # below lines change text in-place
+        if images is not None:
+            merge_length = self.image_processor.merge_size**2
+            index = 0
+            for i in range(len(text)):
+                while self.image_token in text[i]:
+                    num_image_tokens = image_grid_thw[index].prod() // merge_length
+                    text[i] = text[i].replace(self.image_token, "<|placeholder|>" * num_image_tokens, 1)
+                    index += 1
+                text[i] = text[i].replace("<|placeholder|>", self.image_token)
+
+        if videos is not None:
+            merge_length = self.video_processor.merge_size**2
+            index = 0
+            for i in range(len(text)):
+                while self.video_token in text[i]:
+                    num_video_tokens = video_grid_thw[index].prod() // merge_length
+                    text[i] = text[i].replace(self.video_token, "<|placeholder|>" * num_video_tokens, 1)
+                    index += 1
+                text[i] = text[i].replace("<|placeholder|>", self.video_token)
+
+        return_tensors = output_kwargs["text_kwargs"].pop("return_tensors", None)
+        return_mm_token_type_ids = output_kwargs["text_kwargs"].pop("return_mm_token_type_ids", None)
+        text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"])
+        self._check_special_mm_tokens(text, text_inputs, modalities=["image", "video"])
+
+        if return_mm_token_type_ids:
+            array_ids = np.array(text_inputs["input_ids"])
+            mm_token_type_ids = np.zeros_like(text_inputs["input_ids"])
+            mm_token_type_ids[array_ids == self.image_token_id] = 1
+            text_inputs["mm_token_type_ids"] = mm_token_type_ids.tolist()
+
+        return BatchFeature(data={**text_inputs, **image_inputs, **videos_inputs}, tensor_type=return_tensors)
+
+    def _get_num_multimodal_tokens(self, image_sizes=None, video_sizes=None, **kwargs):
+        """
+        Computes the number of placeholder tokens needed for multimodal inputs with the given sizes.
+        Args:
+            image_sizes (`list[list[int]]`, *optional*):
+                The input sizes formatted as (height, width) per each image.
+            video_sizes (`list[list[int]]`, *optional*):
+                The input sizes formatted as (num_frames, height, width) per each video.
+        Returns:
+            `MultiModalData`: A `MultiModalData` object holding number of tokens per each of the provided
+            input modalities, along with other useful data.
+        """
+
+        vision_data = {}
+        if image_sizes is not None:
+            images_kwargs = Qwen2_5_VLProcessorKwargs._defaults.get("images_kwargs", {})
+            images_kwargs.update(kwargs)
+            merge_size = images_kwargs.get("merge_size", None) or self.image_processor.merge_size
+
+            num_image_patches = [
+                self.image_processor.get_number_of_image_patches(*image_size, images_kwargs)
+                for image_size in image_sizes
+            ]
+            num_image_tokens = [(num_patches // merge_size**2) for num_patches in num_image_patches]
+            vision_data.update({"num_image_tokens": num_image_tokens, "num_image_patches": num_image_patches})
+
+        if video_sizes is not None:
+            videos_kwargs = Qwen2_5_VLProcessorKwargs._defaults.get("videos_kwargs", {})
+            videos_kwargs.update(kwargs)
+            num_video_patches = [
+                self.video_processor.get_number_of_video_patches(*video_size, videos_kwargs)
+                for video_size in video_sizes
+            ]
+            num_video_tokens = [(num_patches // merge_size**2) for num_patches in num_video_patches]
+            vision_data["num_video_tokens"] = num_video_tokens
+
+        return MultiModalData(**vision_data)
+
+    def post_process_image_text_to_text(
+        self, generated_outputs, skip_special_tokens=True, clean_up_tokenization_spaces=False, **kwargs
+    ):
+        """
+        Post-process the output of the model to decode the text.
+
+        Args:
+            generated_outputs (`torch.Tensor` or `np.ndarray`):
+                The output of the model `generate` function. The output is expected to be a tensor of shape `(batch_size, sequence_length)`
+                or `(sequence_length,)`.
+            skip_special_tokens (`bool`, *optional*, defaults to `True`):
+                Whether or not to remove special tokens in the output. Argument passed to the tokenizer's `batch_decode` method.
+            clean_up_tokenization_spaces (`bool`, *optional*, defaults to `False`):
+                Whether or not to clean up the tokenization spaces. Argument passed to the tokenizer's `batch_decode` method.
+            **kwargs:
+                Additional arguments to be passed to the tokenizer's `batch_decode method`.
+
+        Returns:
+            `list[str]`: The decoded text.
+        """
+        return self.tokenizer.batch_decode(
+            generated_outputs,
+            skip_special_tokens=skip_special_tokens,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            **kwargs,
+        )
+
+    @property
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+        names_from_processor = list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
+        return names_from_processor + ["second_per_grid_ts"]
+
+AutoProcessor.register("QTSplusQwen2_5_VLProcessor", QTSplusQwen2_5_VLProcessor)
+__all__ = ["QTSplusQwen2_5_VLProcessor"]

processor_config.json

@@ -0,0 +1,9 @@
+{
+  "auto_map": {
+    "AutoProcessor": "processing_qts_plus_qwen2_5_vl.QTSplusQwen2_5_VLProcessor"
+  },
+  "image_processor_type": "Qwen2VLImageProcessorFast",
+  "processor_class": "QTSplusQwen2_5_VLProcessor",
+  "tokenizer_class": "Qwen2Tokenizer",
+  "video_processor_type": "Qwen2VLVideoProcessor"
+}

special_tokens_map.json

@@ -0,0 +1,20 @@
+{
+  "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": "<|endoftext|>",
+  "eos_token": "<|im_end|>",
+  "pad_token": "<|endoftext|>"
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:04c5a433a454dcf945e826dea381181827c01c6a9f99c5d1eb969b77c364d6da
+size 11422174

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": "<|endoftext|>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "processor_class": "Qwen2_5_VLVisionProcessor",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

video_preprocessor_config.json

@@ -0,0 +1,43 @@
+{
+  "crop_size": null,
+  "data_format": "channels_first",
+  "default_to_square": true,
+  "device": null,
+  "do_center_crop": null,
+  "do_convert_rgb": true,
+  "do_normalize": true,
+  "do_rescale": true,
+  "do_resize": true,
+  "do_sample_frames": false,
+  "fps": null,
+  "image_mean": [
+    0.48145466,
+    0.4578275,
+    0.40821073
+  ],
+  "image_std": [
+    0.26862954,
+    0.26130258,
+    0.27577711
+  ],
+  "input_data_format": null,
+  "max_frames": 768,
+  "max_pixels": 12845056,
+  "merge_size": 2,
+  "min_frames": 4,
+  "min_pixels": 3136,
+  "num_frames": null,
+  "pad_size": null,
+  "patch_size": 14,
+  "processor_class": "Qwen2_5_VLVisionProcessor",
+  "resample": 3,
+  "rescale_factor": 0.00392156862745098,
+  "return_metadata": false,
+  "size": {
+    "longest_edge": 12845056,
+    "shortest_edge": 3136
+  },
+  "temporal_patch_size": 2,
+  "video_metadata": null,
+  "video_processor_type": "Qwen2VLVideoProcessor"
+}

zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if ZERO_STAGE not in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info("Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info("Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)