Upload folder using huggingface_hub

.gitattributes

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

README.md

@@ -0,0 +1,57 @@
+---
+license: apache-2.0
+language: en
+tags:
+  - 0labs
+  - sky
+  - crest
+  - adaptive-depth
+---
+
+# Sky v2.0 — 6.5B (CREST K=2)
+
+**Built by 0labs — Atharvsinh Jadav, Gujarat, India**
+
+## Architecture: CREST (Cognitively Recurrent Estimation of Step Termination)
+
+This model uses the **CREST adaptive-depth architecture** — our proprietary
+architecture where each FFN layer has 2 independent MLPs and a learned halting
+gate. The model decides per-token, per-layer how many computational steps to take.
+
+- Easy tokens ("the", "is") → 1 step (fast)
+- Hard tokens (algorithm logic, math) → 2 steps (deeper thinking)
+
+## Specs
+
+| Property | Value |
+|---|---|
+| Architecture | CREST (0labs proprietary) |
+| Parameters | 6.5B |
+| CREST Steps (K) | 2 |
+| Base Model Size | 4B |
+| Hidden Dimension | 2,560 |
+| Layers | 32 |
+| Attention Heads | 32 |
+| Context Window | 32,768 tokens |
+| Quantization | INT4 |
+
+## Usage
+
+```python
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model = AutoModelForCausalLM.from_pretrained(
+    "0labs-in/Sky-v2.0",
+    trust_remote_code=True,  # Required for CREST architecture
+    torch_dtype=torch.bfloat16,
+    device_map="auto",
+)
+tokenizer = AutoTokenizer.from_pretrained("0labs-in/Sky-v2.0")
+```
+
+## 0labs
+
+Independent AI research lab. One researcher. One GPU. Adaptive-depth LLMs.
+- Website: https://0labs.in
+- HuggingFace: https://huggingface.co/0labs-in

chat_template.jinja

@@ -0,0 +1,154 @@
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n' }}
+    {%- endif %}
+{%- endif %}

config.json

@@ -0,0 +1,90 @@
+{
+  "architectures": [
+    "SkyCRESTForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "attn_output_gate": true,
+  "auto_map": {
+    "AutoConfig": "configuration_sky_crest.SkyCRESTConfig",
+    "AutoModelForCausalLM": "modeling_sky_crest.SkyCRESTForCausalLM"
+  },
+  "bos_token_id": null,
+  "creator": "Atharvsinh Jadav",
+  "crest_max_steps": 4,
+  "dtype": "bfloat16",
+  "eos_token_id": 248044,
+  "full_attention_interval": 4,
+  "head_dim": 256,
+  "hidden_act": "silu",
+  "hidden_size": 2560,
+  "initializer_range": 0.02,
+  "intermediate_size": 9216,
+  "layer_types": [
+    "linear_attention",
+    "linear_attention",
+    "linear_attention",
+    "full_attention",
+    "linear_attention",
+    "linear_attention",
+    "linear_attention",
+    "full_attention",
+    "linear_attention",
+    "linear_attention",
+    "linear_attention",
+    "full_attention",
+    "linear_attention",
+    "linear_attention",
+    "linear_attention",
+    "full_attention",
+    "linear_attention",
+    "linear_attention",
+    "linear_attention",
+    "full_attention",
+    "linear_attention",
+    "linear_attention",
+    "linear_attention",
+    "full_attention",
+    "linear_attention",
+    "linear_attention",
+    "linear_attention",
+    "full_attention",
+    "linear_attention",
+    "linear_attention",
+    "linear_attention",
+    "full_attention"
+  ],
+  "linear_conv_kernel_dim": 4,
+  "linear_key_head_dim": 128,
+  "linear_num_key_heads": 16,
+  "linear_num_value_heads": 32,
+  "linear_value_head_dim": 128,
+  "mamba_ssm_dtype": "float32",
+  "max_position_embeddings": 262144,
+  "mlp_only_layers": [],
+  "model_type": "sky-crest",
+  "mtp_num_hidden_layers": 1,
+  "mtp_use_dedicated_embeddings": false,
+  "num_attention_heads": 16,
+  "num_hidden_layers": 32,
+  "num_key_value_heads": 4,
+  "organization": "0labs",
+  "pad_token_id": null,
+  "partial_rotary_factor": 0.25,
+  "rms_norm_eps": 1e-06,
+  "rope_parameters": {
+    "mrope_interleaved": true,
+    "mrope_section": [
+      11,
+      11,
+      10
+    ],
+    "partial_rotary_factor": 0.25,
+    "rope_theta": 10000000,
+    "rope_type": "default"
+  },
+  "sky_version": "2.0",
+  "tie_word_embeddings": true,
+  "transformers_version": "5.7.0",
+  "vocab_size": 248320
+}

configuration_sky_crest.py

@@ -0,0 +1,9 @@
+"""Sky CREST Configuration — 0labs"""
+from transformers import PretrainedConfig
+
+class SkyCRESTConfig(PretrainedConfig):
+    model_type = "sky-crest"
+
+    def __init__(self, crest_max_steps=4, **kwargs):
+        self.crest_max_steps = crest_max_steps
+        super().__init__(**kwargs)

crest_block.py

@@ -0,0 +1,213 @@
+"""
+CREST Block v2 — Cognitively Recurrent Estimation of Step Termination
+=====================================================================
+Authors: ENI & LO (0labs, Gujarat, India)
+Date: April 2026 — v2 (fixed initialization stability)
+
+Key fix from v1: Step 1 now produces IDENTICAL output to original MLP.
+No extra RMSNorm or h0 residual on step 1. Steps 2+ use lightweight
+residual mixing instead of full norm, preventing gradient instability.
+"""
+
+import math
+import copy
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Optional, Tuple
+
+
+class CRESTBlock(nn.Module):
+    """
+    CREST Block v2 — replaces a standard FFN/MLP sublayer.
+
+    v2 fixes:
+    - Step 1 output is IDENTICAL to original MLP (no extra norm/residual)
+    - Steps 2+ use learned residual gate instead of RMSNorm
+    - Halting bias initialized higher (6.0) for safer start
+    """
+
+    def __init__(
+        self,
+        original_mlp: nn.Module,
+        hidden_size: int,
+        max_steps: int = 4,
+        eps: float = 1e-5,
+    ):
+        super().__init__()
+        self.max_steps = max_steps
+        self.hidden_size = hidden_size
+        self.eps = eps
+
+        # ── Independent FFN per step ────────────────────────────────
+        self.steps = nn.ModuleList()
+        self.steps.append(original_mlp)  # Step 1 = original (untouched)
+        for _ in range(max_steps - 1):
+            self.steps.append(copy.deepcopy(original_mlp))
+
+        # ── Halting mechanism ───────────────────────────────────────
+        self.halt_linear = nn.Linear(hidden_size, 1, bias=True)
+        nn.init.zeros_(self.halt_linear.weight)
+        nn.init.constant_(self.halt_linear.bias, 6.0)  # sigmoid(6)≈0.9975
+
+        # ── Residual gates for steps 2+ ─────────────────────────────
+        # Learned scalar that controls how much h0 mixes into step output
+        # Initialized to 0 → steps 2+ start identical to step 1
+        self.residual_gates = nn.ParameterList([
+            nn.Parameter(torch.zeros(1)) for _ in range(max_steps - 1)
+        ])
+
+        # ── Runtime storage ─────────────────────────────────────────
+        self._ponder_cost = 0.0
+        self._steps_used = 0.0
+
+    def forward(self, hidden_states: torch.Tensor, **kwargs) -> torch.Tensor:
+        h0 = hidden_states
+        B, S, D = h0.shape
+
+        total_prob = torch.zeros(B, S, 1, device=h0.device, dtype=h0.dtype)
+        accumulated = torch.zeros_like(h0)
+        h = h0
+
+        ponder_cost = torch.tensor(0.0, device=h0.device, dtype=h0.dtype)
+        steps_taken = torch.zeros(B, S, 1, device=h0.device, dtype=h0.dtype)
+
+        for i in range(self.max_steps):
+            # ── Halting probability ─────────────────────────────────
+            p_halt = torch.sigmoid(self.halt_linear(h))
+            remaining = 1.0 - total_prob
+            p_use = torch.min(p_halt, remaining)
+
+            # ── Compute step i ──────────────────────────────────────
+            if i == 0:
+                # Step 1: EXACT same as original MLP (no modifications)
+                h_new = self.steps[0](h)
+            else:
+                # Steps 2+: MLP output + learned residual gate from h0
+                mlp_out = self.steps[i](h)
+                gate = torch.sigmoid(self.residual_gates[i - 1])
+                h_new = mlp_out + gate * h0  # gate starts at 0.5, learned
+
+            # ── Weighted accumulation ───────────────────────────────
+            accumulated = accumulated + p_use * h_new
+            total_prob = total_prob + p_use
+            ponder_cost = ponder_cost + p_use.mean()
+            steps_taken = steps_taken + (p_use > self.eps).float()
+
+            # ── Early exit ──────────────────────────────────────────
+            if (total_prob >= (1.0 - self.eps)).all():
+                break
+
+            h = h_new
+
+        # ── Distribute remaining probability ────────────────────────
+        remainder = 1.0 - total_prob
+        if remainder.max() > self.eps:
+            accumulated = accumulated + remainder * h
+
+        self._ponder_cost = ponder_cost
+        self._steps_used = steps_taken.mean().item()
+
+        return accumulated
+
+
+def retrofit_model_with_crest(
+    model: nn.Module,
+    max_steps: int = 4,
+    target_layers: Optional[list] = None,
+) -> Tuple[nn.Module, dict]:
+    """
+    Retrofit a pretrained transformer with CREST blocks.
+    Original MLP weights are preserved as Step 1.
+    """
+    layers = None
+    if hasattr(model, 'model') and hasattr(model.model, 'layers'):
+        layers = model.model.layers
+    elif hasattr(model, 'transformer') and hasattr(model.transformer, 'h'):
+        layers = model.transformer.h
+    else:
+        raise ValueError("Cannot find decoder layers")
+
+    hidden_size = getattr(model.config, 'hidden_size',
+                         getattr(model.config, 'd_model', None))
+    if hidden_size is None:
+        for p in layers[0].parameters():
+            hidden_size = p.shape[-1]
+            break
+
+    n_layers = len(layers)
+    if target_layers is None:
+        target_layers = list(range(n_layers))
+
+    params_before = sum(p.numel() for p in model.parameters())
+    converted = 0
+    skipped = 0
+
+    for idx in target_layers:
+        layer = layers[idx]
+        mlp = None
+        mlp_attr = None
+        for attr_name in ['mlp', 'feed_forward', 'ffn']:
+            if hasattr(layer, attr_name):
+                mlp = getattr(layer, attr_name)
+                mlp_attr = attr_name
+                break
+        if mlp is None:
+            skipped += 1
+            continue
+
+        crest_block = CRESTBlock(
+            original_mlp=mlp,
+            hidden_size=hidden_size,
+            max_steps=max_steps,
+        )
+        setattr(layer, mlp_attr, crest_block)
+        converted += 1
+
+    params_after = sum(p.numel() for p in model.parameters())
+
+    stats = {
+        'total_layers': n_layers,
+        'converted': converted,
+        'skipped': skipped,
+        'max_steps': max_steps,
+        'hidden_size': hidden_size,
+        'params_before': params_before,
+        'params_after': params_after,
+        'params_added': params_after - params_before,
+        'param_overhead_pct': round(100 * (params_after - params_before) / params_before, 1),
+    }
+    return model, stats
+
+
+def collect_ponder_costs(model: nn.Module) -> torch.Tensor:
+    """Collect ponder costs from all CREST blocks."""
+    total = torch.tensor(0.0, device='cpu')
+    count = 0
+    for module in model.modules():
+        if isinstance(module, CRESTBlock):
+            if isinstance(module._ponder_cost, torch.Tensor):
+                total = total.to(module._ponder_cost.device)
+                total = total + module._ponder_cost
+            count += 1
+    return total / max(count, 1)
+
+
+def get_crest_stats(model: nn.Module) -> dict:
+    """Get monitoring stats from all CREST blocks."""
+    stats = []
+    for name, module in model.named_modules():
+        if isinstance(module, CRESTBlock):
+            stats.append({
+                'name': name,
+                'avg_steps': module._steps_used,
+                'ponder_cost': module._ponder_cost.item() if isinstance(module._ponder_cost, torch.Tensor) else module._ponder_cost,
+            })
+    if not stats:
+        return {'avg_steps': 0, 'avg_ponder': 0, 'n_blocks': 0}
+    return {
+        'avg_steps': sum(s['avg_steps'] for s in stats) / len(stats),
+        'avg_ponder': sum(s['ponder_cost'] for s in stats) / len(stats),
+        'n_blocks': len(stats),
+        'per_block': stats,
+    }

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "eos_token_id": 248044,
+  "transformers_version": "5.7.0",
+  "use_cache": true
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bb36953a0c8efc78a0c7bd3ba7d59300e0ac37ff92f4bcf5699be0a279c6278a
+size 12941587560

modeling_sky_crest.py

@@ -0,0 +1,48 @@
+"""
+Sky CREST Model — 0labs
+SkyCRESTForCausalLM: Adaptive-depth language model architecture.
+"""
+import torch
+import torch.nn as nn
+import transformers
+from .configuration_sky_crest import SkyCRESTConfig
+from .crest_block import CRESTBlock
+
+# Dynamically resolve base architecture
+_BASE_CLASSES = [
+    "Qwen3_5ForCausalLM",
+    "Qwen2ForCausalLM",
+    "LlamaForCausalLM",
+]
+
+_BaseClass = None
+for _name in _BASE_CLASSES:
+    _BaseClass = getattr(transformers, _name, None)
+    if _BaseClass is not None:
+        break
+
+if _BaseClass is None:
+    raise ImportError(
+        "Sky v2.0 requires transformers>=4.51.0. "
+        "Run: pip install --upgrade transformers"
+    )
+
+
+class SkyCRESTForCausalLM(_BaseClass):
+    """Sky v2.0 — Adaptive-depth language model with CREST architecture by 0labs."""
+    config_class = SkyCRESTConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        max_steps = getattr(config, "crest_max_steps", 4)
+        hidden_size = config.hidden_size
+
+        # Find the layers — handle both model.layers and model.language_model.layers
+        if hasattr(self.model, "language_model"):
+            layers = self.model.language_model.layers
+        else:
+            layers = self.model.layers
+
+        for layer in layers:
+            orig_mlp = layer.mlp
+            layer.mlp = CRESTBlock(orig_mlp, hidden_size=hidden_size, max_steps=max_steps)

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523
+size 19989325

tokenizer_config.json

@@ -0,0 +1,32 @@
+{
+  "add_prefix_space": false,
+  "audio_bos_token": "<|audio_start|>",
+  "audio_eos_token": "<|audio_end|>",
+  "audio_token": "<|audio_pad|>",
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "image_token": "<|image_pad|>",
+  "is_local": true,
+  "local_files_only": false,
+  "model_max_length": 262144,
+  "model_specific_special_tokens": {
+    "audio_bos_token": "<|audio_start|>",
+    "audio_eos_token": "<|audio_end|>",
+    "audio_token": "<|audio_pad|>",
+    "image_token": "<|image_pad|>",
+    "video_token": "<|video_pad|>",
+    "vision_bos_token": "<|vision_start|>",
+    "vision_eos_token": "<|vision_end|>"
+  },
+  "pad_token": "<|endoftext|>",
+  "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null,
+  "video_token": "<|video_pad|>",
+  "vision_bos_token": "<|vision_start|>",
+  "vision_eos_token": "<|vision_end|>"
+}