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

added_tokens.json

@@ -0,0 +1,30 @@
+{
+  "</think>": 151668,
+  "</tool_call>": 151658,
+  "</tool_response>": 151666,
+  "<think>": 151667,
+  "<tool_call>": 151657,
+  "<tool_response>": 151665,
+  "<ts/>": 151670,
+  "<ts>": 151669,
+  "<|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 +1,89 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].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" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if message.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|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 %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {%- if loop.last or (not loop.last and reasoning_content) %}
+                {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+            {%- else %}
+                {{- '<|im_start|>' + message.role + '\n' + content }}
+            {%- endif %}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- 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' }}
+    {%- endif %}
+{%- endif %}

config.json

@@ -0,0 +1,90 @@
+{
+  "architectures": [
+    "Qwen3TSForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_qwen3_ts.Qwen3TSConfig",
+    "AutoModel": "modeling_qwen3_ts.Qwen3TSForCausalLM",
+    "AutoModelForCausalLM": "modeling_qwen3_ts.Qwen3TSForCausalLM",
+    "AutoProcessor": "processing_qwen3_ts.Qwen3TSProcessor"
+  },
+  "bos_token_id": 151643,
+  "eos_token_id": 151645,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "ignore_index": -100,
+  "initializer_range": 0.02,
+  "intermediate_size": 12288,
+  "layer_types": [
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 40960,
+  "max_window_layers": 36,
+  "model_type": "qwen3ts",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 36,
+  "num_key_value_heads": 8,
+  "pad_token_id": 151643,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "sliding_window": null,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float16",
+  "transformers_version": "4.52.4",
+  "ts": {
+    "embedding_dim": 16,
+    "hidden_size": 4096,
+    "max_length": 32768,
+    "max_sequence_length": 8192,
+    "num_features": 2,
+    "num_layers": 5,
+    "patch_size": 8,
+    "use_layer_norm": false,
+    "use_position_embedding": true,
+    "use_position_idx": false
+  },
+  "ts_token_end_index": 151670,
+  "ts_token_start_index": 151669,
+  "use_cache": false,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

configuration_qwen3_ts.py

@@ -0,0 +1,225 @@
+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Qwen3 model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_rope_utils import rope_config_validation
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class Qwen3TSConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Qwen3Model`]. It is used to instantiate a
+    Qwen3 model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of
+    Qwen3-8B [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B).
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 151936):
+            Vocabulary size of the Qwen3 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Qwen3Model`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 22016):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*, defaults to 32):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details, check out [this
+            paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to `32`.
+        head_dim (`int`, *optional*, defaults to 128):
+            The attention head dimension.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 32768):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
+            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
+            accordingly.
+            Expected contents:
+                `rope_type` (`str`):
+                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+                    'llama3'], with 'default' being the original RoPE implementation.
+                `factor` (`float`, *optional*):
+                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+                    original maximum pre-trained length.
+                `original_max_position_embeddings` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
+                    pretraining.
+                `attention_factor` (`float`, *optional*):
+                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+                    computation. If unspecified, it defaults to value recommended by the implementation, using the
+                    `factor` field to infer the suggested value.
+                `beta_fast` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 32.
+                `beta_slow` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 1.
+                `short_factor` (`list[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `long_factor` (`list[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `low_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
+        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        use_sliding_window (`bool`, *optional*, defaults to `False`):
+            Whether to use sliding window attention.
+        sliding_window (`int`, *optional*, defaults to 4096):
+            Sliding window attention (SWA) window size. If not specified, will default to `4096`.
+        max_window_layers (`int`, *optional*, defaults to 28):
+            The number of layers using full attention. The first `max_window_layers` layers will use full attention, while any
+            additional layer afterwards will use SWA (Sliding Window Attention).
+        layer_types (`list`, *optional*):
+            Attention pattern for each layer.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+
+    ```python
+    >>> from transformers import Qwen3Model, Qwen3Config
+
+    >>> # Initializing a Qwen3 style configuration
+    >>> configuration = Qwen3Config()
+
+    >>> # Initializing a model from the Qwen3-8B style configuration
+    >>> model = Qwen3Model(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "qwen3ts"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    # Default tensor parallel plan for base model `Qwen3`
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+
+    def __init__(
+        self,
+        vocab_size=151936,
+        hidden_size=4096,
+        intermediate_size=22016,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=32,
+        head_dim=128,
+        hidden_act="silu",
+        max_position_embeddings=32768,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        use_sliding_window=False,
+        sliding_window=4096,
+        max_window_layers=28,
+        layer_types=None,
+        attention_dropout=0.0,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.use_sliding_window = use_sliding_window
+        self.sliding_window = sliding_window if self.use_sliding_window else None
+        self.max_window_layers = max_window_layers
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.head_dim = head_dim
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        # Validate the correctness of rotary position embeddings parameters
+        # BC: if there is a 'type' field, move it to 'rope_type'.
+        if self.rope_scaling is not None and "type" in self.rope_scaling:
+            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
+        rope_config_validation(self)
+
+        self.layer_types = layer_types
+        if self.layer_types is None:
+            self.layer_types = [
+                "sliding_attention"
+                if self.sliding_window is not None and i >= self.max_window_layers
+                else "full_attention"
+                for i in range(self.num_hidden_layers)
+            ]
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+
+__all__ = ["Qwen3TSConfig"]

generation_config.json

@@ -0,0 +1,13 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "pad_token_id": 151643,
+  "temperature": 0.6,
+  "top_k": 20,
+  "top_p": 0.95,
+  "transformers_version": "4.52.4"
+}

modeling_qwen3_ts.py

@@ -0,0 +1,622 @@
+# Copyright 2025 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Callable, Optional, Union
+
+import torch
+from torch import nn
+from dataclasses import dataclass
+
+from transformers.cache_utils import Cache
+from transformers.generation import GenerationMixin
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast
+)
+from transformers.models.qwen3.modeling_qwen3 import Qwen3PreTrainedModel, Qwen3Model
+from transformers.modeling_utils import PreTrainedModel
+from transformers.processing_utils import Unpack
+from transformers.utils import LossKwargs, auto_docstring, can_return_tuple, logging, ModelOutput
+from .configuration_qwen3_ts import Qwen3TSConfig
+from typing import Any, Dict
+
+
+logger = logging.get_logger(__name__)
+
+
+########################MLP TS Embedding#####################
+class TimeSeriesEmbedding(nn.Module):
+    def __init__(self, config):
+        super(TimeSeriesEmbedding, self).__init__()
+        self.patch_size = config['patch_size']
+        self.num_layers = config['num_layers']
+        self.hidden_size = config['hidden_size']
+        self.num_features = config['num_features']
+        self.max_sequence_length = config['max_sequence_length']  # Maximum time series length
+        self.use_position_embedding = config.get('use_position_embedding', False)
+        self.use_position_idx = config.get('use_position_idx', False)
+        self.use_layer_norm = config.get('use_layer_norm', False)
+        self.embedding_dim = config.get('embedding_dim', 16)  # Embedding dimension
+        
+        if self.use_position_embedding:
+            # Extended vocabulary: [0, max_sequence_length) for real positions, max_sequence_length for padding
+            self.position_embedding = nn.Embedding(self.max_sequence_length + 1, self.embedding_dim)
+            self.padding_idx = self.max_sequence_length  # Special index for padding
+            input_size = 1 * self.patch_size + self.embedding_dim * self.patch_size
+        elif self.use_position_idx:
+            input_size = 2 * self.patch_size
+        else:
+            input_size = 1 * self.patch_size
+        
+        # Build MLP layers
+        layers = []
+        for _ in range(self.num_layers - 1):
+            layers.append(nn.Linear(input_size, self.hidden_size))
+            layers.append(nn.GELU())
+            input_size = self.hidden_size
+
+        layers.append(nn.Linear(input_size, self.hidden_size))
+        if self.use_layer_norm:
+            layers.append(nn.LayerNorm(self.hidden_size))
+
+        self.mlp = nn.Sequential(*layers)
+
+    def forward(self, x: torch.Tensor):
+        batch_size = x.size(0)
+        x = x.reshape(batch_size, -1, self.num_features)
+
+        # Extract mask and calculate valid lengths
+        mask = x[:, :, -1].long()
+        valid_lengths = mask.sum(dim=1).long()
+        patch_cnt = (valid_lengths + self.patch_size - 1) // self.patch_size
+
+        patches_list = []
+        # Collect position indices for batch embedding lookup
+        all_position_indices = []
+        patch_info_list = []  # Store metadata for each patch group
+        
+        for i in range(batch_size):
+            vl = valid_lengths[i].item()
+            pc = patch_cnt[i].item()
+            if pc == 0:
+                continue
+            
+            # Extract time series data (excluding mask)
+            xi = x[i, :vl, :1]  # Time-series data
+            total_padded_length = pc * self.patch_size
+            padding_length = total_padded_length - vl
+            
+            # Create position indices: real positions for actual data, special index for padding
+            position_indices = torch.arange(vl, device=x.device)
+            
+            if padding_length > 0:
+                # Pad with last value
+                last_value = xi[-1:, :]
+                padding = last_value.repeat(padding_length, 1)
+                xi = torch.cat([xi, padding], dim=0)
+                
+                # Use special padding index for padding positions
+                padding_positions = torch.full((padding_length,), self.padding_idx, device=x.device)
+                position_indices = torch.cat([position_indices, padding_positions], dim=0)
+
+            # Reshape to patches
+            xi = xi.reshape(pc, self.patch_size)  # (num_patches, patch_size)
+            position_indices = position_indices.reshape(pc, self.patch_size)  # (num_patches, patch_size)
+
+            if self.use_position_embedding:
+                # Collect position indices instead of calling embedding immediately
+                all_position_indices.append(position_indices)
+                patch_info_list.append({
+                    'xi': xi,
+                    'pc': pc,
+                    'sample_idx': i
+                })
+            elif self.use_position_idx:
+                # Normalize position indices
+                pos_indices = torch.arange(vl, device=x.device).unsqueeze(1)
+                pos_indices = pos_indices / max(1, valid_lengths.max().item() - 1)
+                if padding_length > 0:
+                    # Use -1 for padding positions
+                    padding_indices = torch.full((padding_length, 1), -1, device=x.device)
+                    pos_indices = torch.cat([pos_indices, padding_indices], dim=0)
+                # Combine time series data with position indices
+                xi_combined = torch.cat([xi.reshape(-1, 1), pos_indices], dim=1)
+                patch_input = xi_combined.reshape(pc, self.patch_size * 2)
+                patches_list.append(patch_input)
+            else:
+                # No position embedding, use raw patches
+                patch_input = xi
+                patches_list.append(patch_input)
+
+        # Batch process position embeddings if needed
+        if self.use_position_embedding and all_position_indices:
+            # Concatenate all position indices for batch embedding lookup
+            batch_position_indices = torch.cat(all_position_indices, dim=0)
+            # print(f"{x.shape=}, {x.device=}, {len(all_position_indices)=}, {batch_position_indices=}")
+            batch_pos_emb = self.position_embedding(batch_position_indices)  # Single embedding call
+            
+            # Split embeddings back and create patch inputs
+            emb_start_idx = 0
+            for patch_info in patch_info_list:
+                xi = patch_info['xi']
+                pc = patch_info['pc']
+                
+                # Extract corresponding embeddings
+                pos_emb = batch_pos_emb[emb_start_idx:emb_start_idx + pc]
+                emb_start_idx += pc
+                
+                # Flatten and concatenate
+                xi = xi.unsqueeze(-1)  # (num_patches, patch_size, 1)
+                patch_input = torch.cat([
+                    xi.flatten(1),  # (num_patches, patch_size)
+                    pos_emb.flatten(1)  # (num_patches, patch_size * embedding_dim)
+                ], dim=1)
+                patches_list.append(patch_input)
+
+        # Process all patches through MLP
+        if patches_list:
+            x_patches = torch.cat(patches_list, dim=0)
+            x = self.mlp(x_patches)
+        else:
+            # Handle empty case
+            x = torch.empty(0, self.hidden_size, device=x.device)
+
+        return x, patch_cnt
+
+
+@dataclass
+class Qwen3TSCausalLMOutputWithPast(CausalLMOutputWithPast):
+    """
+    Output type of Qwen3TSForCausalLM that includes additional fields for timeseries processing.
+    
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+            Language modeling loss (for next-token prediction).
+        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed):
+            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`.
+        attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            The attention mask used in the forward pass, potentially expanded to accommodate timeseries patches.
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss.
+        new_token_positions (`torch.LongTensor` of shape `(batch_size, num_new_tokens)`, *optional*):
+            Positions where new tokens (not from timeseries) are located in the expanded sequence.
+    """
+    attention_mask: Optional[torch.FloatTensor] = None
+    labels: Optional[torch.LongTensor] = None
+    new_token_positions: Optional[torch.LongTensor] = None
+
+
+class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ...
+
+
+class Qwen3TSGenerationMixin(GenerationMixin):
+    """
+    Generation mixin for Qwen3 models with timeseries support.
+    
+    This mixin handles the special case where timeseries embeddings expand the sequence length
+    during the first forward pass, requiring special attention mask management.
+    """
+    
+    def prepare_inputs_for_generation(
+        self,
+        input_ids: torch.LongTensor,
+        past_key_values: Optional[Cache] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        timeseries: Optional[torch.FloatTensor] = None,
+        **kwargs,
+    ) -> Dict[str, Any]:
+        """
+        Prepare inputs for generation with timeseries support.
+        
+        Timeseries are only processed during the first forward pass. In subsequent
+        generation steps, they are already embedded in the past_key_values.
+        """
+        # Check if we have timeseries data
+        has_ts = timeseries is not None and len(timeseries) > 0
+        
+        # Handle timeseries for generation with past_key_values
+        if has_ts and past_key_values is not None:
+            # Get the number of tokens already processed
+            if isinstance(past_key_values, Cache):
+                past_length = past_key_values.seen_tokens
+            else:
+                past_length = past_key_values[0][0].shape[2] if past_key_values[0] is not None else 0
+            
+            # If we have processed tokens, timeseries have already been embedded
+            if past_length > 0:
+                # Only keep the last token for next token prediction
+                input_ids = input_ids[:, -1:]
+                # Clear timeseries as they've been processed
+                timeseries = None
+                has_ts = False
+        
+        # Call parent's prepare_inputs_for_generation to handle all standard logic
+        model_inputs = super().prepare_inputs_for_generation(
+            input_ids=input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            cache_position=cache_position,
+            **kwargs
+        )
+        
+        # Add timeseries to model inputs
+        model_inputs["timeseries"] = timeseries
+        
+        return model_inputs
+    
+    def _update_model_kwargs_for_generation(
+        self,
+        outputs: ModelOutput,
+        model_kwargs: Dict[str, Any],
+        is_encoder_decoder: bool = False,
+        num_new_tokens: int = 1,
+    ) -> Dict[str, Any]:
+        """
+        Update model keyword arguments for generation, handling attention mask from outputs.
+        
+        This is necessary because timeseries processing can expand the sequence length,
+        and we need to use the expanded attention_mask from the model outputs.
+        """
+        # Handle special case: use attention_mask from outputs if available
+        # This is crucial for timeseries models where the sequence length is expanded
+        if hasattr(outputs, "attention_mask") and outputs.attention_mask is not None:
+            model_kwargs["attention_mask"] = outputs.attention_mask
+        
+        # Call parent's implementation for standard updates
+        model_kwargs = super()._update_model_kwargs_for_generation(
+            outputs=outputs,
+            model_kwargs=model_kwargs,
+            is_encoder_decoder=is_encoder_decoder,
+            num_new_tokens=num_new_tokens,
+        )
+        
+        return model_kwargs
+    
+    def generate(
+        self,
+        inputs: Optional[torch.Tensor] = None,
+        timeseries: Optional[torch.FloatTensor] = None,
+        generation_config=None,
+        **kwargs,
+    ):
+        """
+        Generate sequences with timeseries support.
+        
+        Args:
+            inputs: Input token ids
+            timeseries: Optional timeseries data to be processed in the first forward pass
+            generation_config: Generation configuration
+            **kwargs: Additional keyword arguments for generation
+        """
+        # Add timeseries to kwargs if provided
+        if timeseries is not None:
+            kwargs["timeseries"] = timeseries
+        
+        # Call parent's generate method
+        return super().generate(
+            inputs=inputs,
+            generation_config=generation_config,
+            **kwargs,
+        )
+    
+    def _validate_model_kwargs(self, model_kwargs: Dict[str, Any]) -> None:
+        """
+        Validate model kwargs, allowing timeseries as a valid argument.
+        """
+        # Remove timeseries from model_kwargs temporarily for validation
+        timeseries = model_kwargs.pop("timeseries", None)
+        
+        # Call parent's validation
+        super()._validate_model_kwargs(model_kwargs)
+        
+        # Restore timeseries
+        if timeseries is not None:
+            model_kwargs["timeseries"] = timeseries
+
+class Qwen3TSPreTrainedModel(Qwen3PreTrainedModel):
+    config_class = Qwen3TSConfig
+
+@auto_docstring
+class Qwen3TSForCausalLM(Qwen3TSPreTrainedModel, Qwen3TSGenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = Qwen3Model(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # TS embedding
+        self.ts_encoder = TimeSeriesEmbedding(config.ts)
+
+        # Initialize weights and apply final processing
+        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 set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+    
+    def _merge_input_ids_with_time_series_features(self, time_series_features, inputs_embeds, input_ids, attention_mask, labels, patch_cnt):
+        batch_size, sequence_length = input_ids.shape
+        _left_padding = torch.any(attention_mask[:, 0] == 0)
+        _right_padding = torch.any(attention_mask[:, -1] == 0)
+        left_padding = False
+        if batch_size > 1:
+            if _left_padding and not _right_padding:
+                left_padding = True
+            elif not _left_padding and _right_padding:
+                left_padding = False
+            elif not _left_padding and not _right_padding:
+                left_padding = False
+            else:
+                raise ValueError(f"both side of attention_mask has zero, invalid. {attention_mask}")
+        else:
+            if _left_padding and not _right_padding:
+                left_padding = True
+            else:
+                left_padding = False
+
+        # 1. Create a mask to know where special time series tokens are
+        special_ts_token_mask_start = input_ids == self.config.ts_token_start_index
+        special_ts_token_mask_end = input_ids == self.config.ts_token_end_index
+        special_ts_token_mask = special_ts_token_mask_start | special_ts_token_mask_end
+
+        # 2. Calculate patch count
+        num_special_ts_tokens = torch.sum(special_ts_token_mask_start, dim=-1)
+        total_time_steps, embed_dim = time_series_features.shape
+
+        # Correctly calculate the total number of patches per batch
+        patch_index = 0
+        num_total_patches = torch.zeros(batch_size, dtype=patch_cnt.dtype, device=patch_cnt.device)
+        special_ts_token_mask_start_nonzero = special_ts_token_mask_start.nonzero()
+        special_ts_token_mask_start_with_size = special_ts_token_mask_start.clone().long()
+
+        attn_mask_cnt = attention_mask.sum(dim=-1)
+        for i in range(batch_size):
+            num_ts_in_batch = num_special_ts_tokens[i]
+            num_total_patches[i] = patch_cnt[patch_index : patch_index + num_ts_in_batch].sum() - 2 * num_ts_in_batch
+            for idx in range(patch_index, patch_index + num_ts_in_batch):
+                b_idx, pos = special_ts_token_mask_start_nonzero[idx]
+                special_ts_token_mask_start_with_size[b_idx, pos] *= (patch_cnt[idx].item() - 2)
+            patch_index += num_ts_in_batch
+            attn_mask_cnt[i] += num_total_patches[i].item()
+
+        # 3. Embeding length
+        max_embed_dim = sequence_length + num_total_patches.max()
+
+        # 4. Non ts tokens
+        batch_indices, non_ts_indices = torch.where(~special_ts_token_mask)
+        attn_batch_indices, attn_indices = torch.where(attention_mask == 1)
+
+        # 5. Text token in final text positions
+        new_token_positions = torch.cumsum((special_ts_token_mask_start_with_size + 1), dim=-1) - 1
+
+        # nb_ts_pad
+        nb_ts_pad = max_embed_dim - 1 - new_token_positions[:, -1]
+        if left_padding:
+            new_token_positions += nb_ts_pad[:, None]
+
+        text_to_overwrite = new_token_positions[batch_indices, non_ts_indices]
+
+        # 6. Final embedding and attention masks
+        final_embedding = torch.zeros(
+            batch_size, max_embed_dim, embed_dim, dtype=inputs_embeds.dtype, device=inputs_embeds.device
+        )
+
+        final_attention_mask = torch.zeros(batch_size, max_embed_dim, dtype=attention_mask.dtype, device=inputs_embeds.device)
+        for i in range(attention_mask.size(0)):
+            if left_padding:
+                final_attention_mask[i, max_embed_dim - attn_mask_cnt[i] :] = 1
+            else:
+                final_attention_mask[i, : attn_mask_cnt[i]] = 1
+
+        final_labels = None
+        if labels is not None:
+            final_labels = torch.full(
+                (batch_size, max_embed_dim), self.config.ignore_index, dtype=input_ids.dtype, device=input_ids.device
+            )
+
+        target_device = inputs_embeds.device
+        batch_indices, non_ts_indices, text_to_overwrite = (
+            batch_indices.to(target_device),
+            non_ts_indices.to(target_device),
+            text_to_overwrite.to(target_device),
+        )
+
+        # 7. Move embedding and labels to final positions
+        final_embedding[batch_indices, text_to_overwrite] = inputs_embeds[batch_indices, non_ts_indices]
+        if labels is not None:
+            final_labels[batch_indices, text_to_overwrite] = labels[batch_indices, non_ts_indices]
+
+        # 8. Move time series to final positions
+        ts_to_overwrite = torch.full(
+            (batch_size, max_embed_dim), True, dtype=torch.bool, device=inputs_embeds.device
+        )
+        ts_to_overwrite[batch_indices, text_to_overwrite] = False
+
+        reversed_cumsum = ts_to_overwrite.flip(dims=[-1]).cumsum(-1).flip(dims=[-1]) - 1
+        ts_to_overwrite &= reversed_cumsum >= nb_ts_pad[:, None].to(target_device)
+
+        # Check that the number of time series tokens is correct
+        if ts_to_overwrite.sum() != time_series_features.shape[:-1].numel():
+            raise ValueError(
+                f"The input provided to the model are wrong. The number of time series tokens is {torch.sum(special_ts_token_mask_start)} while"
+                f" the number of time series given to the model is {len(patch_cnt)}. This prevents correct indexing and breaks batch generation."
+            )
+        final_embedding[ts_to_overwrite] = time_series_features.contiguous().reshape(-1, embed_dim).to(target_device)
+        # if str(input_ids.device) == 'cuda:0':
+        #     print(f"[EMBED] {final_embedding[ts_to_overwrite][0]=}")
+
+        # 9. Calculate position ids
+        position_ids = (final_attention_mask.cumsum(-1) - 1).masked_fill_((final_attention_mask == 0), 1)
+        if position_ids.size(-1) < input_ids.size(-1):
+            position_ids = position_ids[:, -input_ids.size(-1) :]
+
+        # 10. Move attention mask to final positions
+        # print(f"{type(input_ids)=}, {input_ids.shape=}, {self.config.pad_token_id=}")
+        pad_batch_indices, pad_indices = torch.where(input_ids == self.config.pad_token_id)
+        if len(pad_batch_indices) > 0:
+            indices_to_mask = new_token_positions[pad_batch_indices, pad_indices]
+            final_embedding[pad_batch_indices, indices_to_mask] = 0
+
+        # 11. Post-process new_token_positions (set -1 for padding positions)
+        new_token_positions = new_token_positions.masked_fill(attention_mask == 0, -1)
+
+        return final_embedding, final_attention_mask, position_ids, final_labels, new_token_positions
+
+    @can_return_tuple
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        timeseries: torch.FloatTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = 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,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **kwargs: Unpack[KwargsForCausalLM],
+    ) -> Qwen3TSCausalLMOutputWithPast:
+        r"""
+        Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary.
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings.
+        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
+            Pre-computed hidden-states (key and values in the attention blocks).
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        timeseries (`torch.FloatTensor` of shape `(batch_size, num_patches, patch_size)`, *optional*):
+            Timeseries data to be encoded and merged with text embeddings.
+    
+    Returns:
+        [`Qwen3TSCausalLMOutputWithPast`] or `tuple(torch.FloatTensor)`:
+        The model outputs with potential timeseries-expanded attention mask.
+        """
+
+        # if input_ids is not None and timeseries is not None:
+        #     # Print the input ts
+        #     print("=================================================================")
+        #     print("Timeseries shape:", timeseries.shape)
+        #     print("=================================================================\n\n")
+        # else:
+        #     print("Time series is None!!!!")
+
+        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
+        )
+
+        if inputs_embeds is None:
+            inputs_embeds = self.get_input_embeddings()(input_ids)
+
+            if timeseries is not None and timeseries.shape[0] > 0:
+                # use_cache = False
+                # print(f"timeseries shape: {timeseries.shape=}, {input_ids.shape=}")
+                ts_features, patch_cnt = self.ts_encoder(timeseries)
+                inputs_embeds = inputs_embeds.to(ts_features.dtype)
+                # if str(input_ids.device) == 'cuda:0':
+                #     print(f"-------------------------------------------------------\n[before] {input_ids.shape=}, timeseries={timeseries.shape if timeseries is not None else None}, {attention_mask.sum(-1)=}")
+                inputs_embeds, attention_mask, position_ids, labels, new_token_positions = self._merge_input_ids_with_time_series_features(
+                    ts_features, inputs_embeds, input_ids, attention_mask, labels, patch_cnt
+                )
+                # print(f"{inputs_embeds.shape=}, {attention_mask.shape=}, {position_ids.shape=}, {labels.shape=}")
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        # if str(input_ids.device) == 'cuda:0':
+        #     print(f"[after] {inputs_embeds.shape=}, {attention_mask.sum(-1)=}, {position_ids.max(-1)=}, {cache_position=}")
+        outputs: BaseModelOutputWithPast = self.model(
+            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,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+        hidden_states = outputs.last_hidden_state
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+        # print(f"{logits.shape=}, {hidden_states.shape=}, {slice_indices=}, {logits_to_keep=}, labels={(labels != self.config.ignore_index).sum() if labels is not None else None}, {loss=}, {torch.sum(torch.isnan(logits))=}, {torch.sum(torch.isinf(logits))=}, {torch.sum(torch.isnan(hidden_states))=}, {torch.sum(torch.isinf(hidden_states))=}")
+
+        return Qwen3TSCausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            attention_mask=attention_mask,
+            labels=labels,
+            new_token_positions=new_token_positions if timeseries is not None and timeseries.shape[0] > 0 else None,
+        )
+
+
+__all__ = [
+    "Qwen3TSForCausalLM"
+]

processing_qwen3_ts.py

@@ -0,0 +1,224 @@
+# coding=utf-8
+# Copyright 2024 Tsinghua University and ByteDance.
+#
+# Licensed under the MIT License (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://opensource.org/license/mit
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+from typing import List, Union, Tuple, Optional
+import torch
+
+from transformers.feature_extraction_utils import BatchFeature
+from transformers.processing_utils import ProcessorMixin
+from transformers.tokenization_utils_base import PaddingStrategy
+
+def sp_encoding(timeseries: np.ndarray, eots_token: bool = True) -> Tuple[np.ndarray, str, dict]:
+    """
+    Encodes a time series with scalar normalization.
+
+    Args:
+        timeseries (np.ndarray): The raw time series data (1D or 2D).
+
+    Returns:
+        result_timeseries (np.ndarray): The encoded time series, shape [seq_len, 1].
+        prompt (str): The placeholder string with offset and scaling info.
+        metadata (dict): Metadata containing the offset and scaling factor.
+    """
+    timeseries = np.array(timeseries)
+    mean = np.mean(timeseries)
+    scaled_timeseries = timeseries - mean
+    scale_factor = 1.0
+    if np.any(np.abs(scaled_timeseries) >= 3.0):
+        scale_factor = np.max(np.abs(scaled_timeseries)) / 3.0
+        scaled_timeseries /= scale_factor
+
+    prompt = f"[offset={-mean:.4f}|scaling={scale_factor:.4f}|length={len(timeseries)}|max={max(timeseries):.4f}|min={min(timeseries):.4f}|left={timeseries[0]:.4f}|right={timeseries[-1]:.4f}]<ts>"
+    if eots_token:
+        prompt += '<ts/>'
+
+    result_timeseries = np.stack([scaled_timeseries, np.ones_like(scaled_timeseries)], axis=-1).reshape(-1, 1)
+
+    return result_timeseries, prompt, {"offset": float(-mean), "scale_factor": float(scale_factor)}
+
+class Qwen3TSProcessor(ProcessorMixin):
+    """
+    A processor for ChatTS that integrates text prompt processing and time series encoding.
+    """
+
+    attributes = ["tokenizer"]
+    feature_extractor_class = None  # You can add a feature extractor if needed
+    tokenizer_class = "AutoTokenizer"
+
+    def __init__(self, tokenizer=None, chat_template=None, **kwargs):
+        """
+        Args:
+            tokenizer: An optional tokenizer to process text prompts.
+        """
+        if chat_template is None and tokenizer is not None and tokenizer.chat_template is not None:
+            chat_template = tokenizer.chat_template
+        self.chat_template = chat_template
+
+        super().__init__(tokenizer=tokenizer, chat_template=chat_template, **kwargs)
+
+    def __call__(
+        self,
+        text: Union[str, List[str]],
+        timeseries: Optional[List[List[np.ndarray]]] = None,
+        padding: Union[bool, str, PaddingStrategy] = False,
+        padding_side: str = 'left',
+        vllm_flag: bool = False,
+        tokenize: bool = True,
+        **kwargs,
+    ) -> BatchFeature:
+        """
+        Encodes a prompt and its associated time series.
+
+        Args:
+            prompt (List[str]): The input prompt containing <ts><ts/> placeholders.
+            timeseries (List[np.ndarray]): A list of time series matched to placeholders in the prompt.
+            padding (bool or str or PaddingStrategy, optional): Passed to the tokenizer for text padding.
+            return_tensors (str, optional): "pt" to return PyTorch tensors; None to return NumPy arrays.
+            **kwargs: Additional tokenizer parameters.
+
+        Returns:
+            BatchFeature: Contains processed prompt, encoded time series, and tokenizer outputs.
+        """
+        if type(text) == str:
+            text = [text]
+        if timeseries is None:
+            timeseries = []
+
+        reconstructed_prompts = []
+        concatenated_ts = None
+        ts_tokens = []
+
+        if vllm_flag:
+            # All prompt modifications have to be done inside of the vLLM
+            # to work correctly with its caching mechanism.
+            reconstructed_prompts = text
+            
+            # Process timeseries data
+            encoded_ts_arrays = []
+            for ts in timeseries:
+                # Get the normalized data and prompt text
+                encoded_ts, ts_prompt, _ = sp_encoding(ts, eots_token=False)
+                # Tokenize the ts_prompt and add to the tokens list
+                if self.tokenizer is not None:
+                    tokens = self.tokenizer.encode(ts_prompt, add_special_tokens=False)
+                    ts_tokens.append(tokens)
+                encoded_ts_arrays.append(encoded_ts[None, ...])
+        else:
+            encoded_ts_arrays = []
+            total_ts_cnt = 0
+            for idx, prompt in enumerate(text):
+                # Split prompt by <ts><ts/> placeholders
+                last_ts_cnt = total_ts_cnt
+                prompt_segments = prompt.split("<ts><ts/>")
+                total_ts_cnt = total_ts_cnt + len(prompt_segments) - 1
+
+                # Encode each time series and rebuild the prompt
+                reconstructed_prompt = prompt_segments[0]
+
+                for i, ts in enumerate(timeseries[last_ts_cnt:total_ts_cnt]):
+                    encoded_ts, ts_prompt, _ = sp_encoding(ts, eots_token=not vllm_flag)
+                    reconstructed_prompt += ts_prompt + prompt_segments[i + 1]
+                    # Ensure time series shape [1, seq_len, feature_dim] for batch concatenation
+                    encoded_ts_arrays.append(encoded_ts[None, ...])
+
+                reconstructed_prompts.append(reconstructed_prompt)
+
+            if len(timeseries) != len(encoded_ts_arrays):
+                raise ValueError(
+                    f"Mismatch between <ts><ts/> placeholders ({total_ts_cnt}) "
+                    f"and time series ({len(encoded_ts_arrays)})."
+                )
+
+            if len(encoded_ts_arrays) > 0:
+                # Pad time series to the same length
+                max_length = max(ts.shape[1] for ts in encoded_ts_arrays)
+                padded_ts_arrays = [
+                    np.pad(ts, ((0, 0), (0, max_length - ts.shape[1]), (0, 0)), mode="constant", constant_values=0.0)
+                    for ts in encoded_ts_arrays
+                ]
+                concatenated_ts = np.concatenate(padded_ts_arrays, axis=0)  # Shape: [batch_size, max_length, feature_dim]
+                
+                # Convert to torch
+                concatenated_ts = torch.from_numpy(concatenated_ts).half()
+
+        # Tokenize the processed prompt
+        tokenizer_outputs = {}
+        if tokenize and self.tokenizer is not None:
+            tokenizer_outputs = self.tokenizer(reconstructed_prompts, padding=padding, padding_side=padding_side, **kwargs)
+        else:
+            tokenizer_outputs = {"text": reconstructed_prompts}
+
+        # Create the final output
+        outputs = tokenizer_outputs
+        if vllm_flag:
+            outputs["timeseries"] = zip(ts_tokens, encoded_ts_arrays)
+        elif concatenated_ts is not None:
+            outputs["timeseries"] = concatenated_ts
+
+        return BatchFeature(data=outputs)
+
+    def encode_timeseries(
+        self,
+        timeseries: Optional[List[List[np.ndarray]]] = None,
+    ) -> np.ndarray:
+        if timeseries is None:
+            timeseries = []
+
+        concatenated_ts = None
+        encoded_ts_arrays = []
+
+        for i, ts in enumerate(timeseries):
+            encoded_ts, _, _ = sp_encoding(ts)
+            # Ensure time series shape [1, seq_len, feature_dim] for batch concatenation
+            encoded_ts_arrays.append(encoded_ts[None, ...])
+
+        if len(encoded_ts_arrays) > 0:
+            # Pad time series to the same length
+            max_length = max(ts.shape[1] for ts in encoded_ts_arrays)
+            padded_ts_arrays = [
+                np.pad(ts, ((0, 0), (0, max_length - ts.shape[1]), (0, 0)), mode="constant", constant_values=0.0)
+                for ts in encoded_ts_arrays
+            ]
+            concatenated_ts = np.concatenate(padded_ts_arrays, axis=0)  # Shape: [batch_size, max_length, feature_dim]
+            
+            # Convert to torch
+            concatenated_ts = torch.from_numpy(concatenated_ts).half()
+
+        return concatenated_ts
+
+    @property
+    def model_input_names(self):
+        """
+        Define the input names expected by the model.
+        """
+        tokenizer_input_names = []
+        if self.tokenizer and hasattr(self.tokenizer, "model_input_names"):
+            tokenizer_input_names = self.tokenizer.model_input_names
+        return list(dict.fromkeys(["processed_prompt", "time_series"] + tokenizer_input_names))
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to Qwen3TokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to Qwen3TokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)

processor_config.json

@@ -0,0 +1,6 @@
+{
+  "auto_map": {
+    "AutoProcessor": "processing_qwen3_ts.Qwen3TSProcessor"
+  },
+  "processor_class": "Qwen3TSProcessor"
+}

pytorch_model-00001-of-00004.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f99f086fd1bf92bd3aab1bb91fee2c840df0af525f8038cae1d7b1b8900480ee
+size 4902281803

pytorch_model-00002-of-00004.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b8700532091097ce1c635bfcd6b3c3fe0da628a7d7def61fa16015a72f52a535
+size 4915992380

pytorch_model-00003-of-00004.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0fb0a51248089c6b0a05086e5f32497cc14661530afca9aa42c70c1319577cc0
+size 4983100114

pytorch_model-00004-of-00004.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b961b2c4237ebfb312192e27bfd3999d3488120fe152a98ad7f1294f179bcfa
+size 1715871460

pytorch_model.bin.index.json

@@ -0,0 +1,417 @@
+{
+  "metadata": {
+    "total_size": 16517105696
+  },
+  "weight_map": {
+    "lm_head.weight": "pytorch_model-00004-of-00004.bin",
+    "model.embed_tokens.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.0.input_layernorm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.0.mlp.down_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.0.mlp.gate_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.0.mlp.up_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.0.post_attention_layernorm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.0.self_attn.k_norm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.0.self_attn.k_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.0.self_attn.o_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.0.self_attn.q_norm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.0.self_attn.q_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.0.self_attn.v_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.1.input_layernorm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.1.mlp.down_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.1.mlp.gate_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.1.mlp.up_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.1.post_attention_layernorm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.1.self_attn.k_norm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.1.self_attn.k_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.1.self_attn.o_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.1.self_attn.q_norm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.1.self_attn.q_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.1.self_attn.v_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.10.input_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.10.mlp.down_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.10.mlp.gate_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.10.mlp.up_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.10.post_attention_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.10.self_attn.k_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.10.self_attn.k_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.10.self_attn.o_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.10.self_attn.q_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.10.self_attn.q_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.10.self_attn.v_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.11.input_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.11.mlp.down_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.11.mlp.gate_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.11.mlp.up_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.11.post_attention_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.11.self_attn.k_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.11.self_attn.k_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.11.self_attn.o_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.11.self_attn.q_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.11.self_attn.q_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.11.self_attn.v_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.12.input_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.12.mlp.down_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.12.mlp.gate_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.12.mlp.up_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.12.post_attention_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.12.self_attn.k_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.12.self_attn.k_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.12.self_attn.o_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.12.self_attn.q_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.12.self_attn.q_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.12.self_attn.v_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.13.input_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.13.mlp.down_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.13.mlp.gate_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.13.mlp.up_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.13.post_attention_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.13.self_attn.k_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.13.self_attn.k_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.13.self_attn.o_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.13.self_attn.q_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.13.self_attn.q_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.13.self_attn.v_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.14.input_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.14.mlp.down_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.14.mlp.gate_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.14.mlp.up_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.14.post_attention_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.14.self_attn.k_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.14.self_attn.k_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.14.self_attn.o_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.14.self_attn.q_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.14.self_attn.q_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.14.self_attn.v_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.15.input_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.15.mlp.down_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.15.mlp.gate_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.15.mlp.up_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.15.post_attention_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.15.self_attn.k_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.15.self_attn.k_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.15.self_attn.o_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.15.self_attn.q_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.15.self_attn.q_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.15.self_attn.v_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.16.input_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.16.mlp.down_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.16.mlp.gate_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.16.mlp.up_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.16.post_attention_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.16.self_attn.k_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.16.self_attn.k_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.16.self_attn.o_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.16.self_attn.q_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.16.self_attn.q_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.16.self_attn.v_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.17.input_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.17.mlp.down_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.17.mlp.gate_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.17.mlp.up_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.17.post_attention_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.17.self_attn.k_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.17.self_attn.k_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.17.self_attn.o_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.17.self_attn.q_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.17.self_attn.q_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.17.self_attn.v_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.18.input_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.18.mlp.down_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.18.mlp.gate_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.18.mlp.up_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.18.post_attention_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.18.self_attn.k_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.18.self_attn.k_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.18.self_attn.o_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.18.self_attn.q_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.18.self_attn.q_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.18.self_attn.v_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.19.input_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.19.mlp.down_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.19.mlp.gate_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.19.mlp.up_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.19.post_attention_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.19.self_attn.k_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.19.self_attn.k_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.19.self_attn.o_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.19.self_attn.q_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.19.self_attn.q_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.19.self_attn.v_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.2.input_layernorm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.2.mlp.down_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.2.mlp.gate_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.2.mlp.up_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.2.post_attention_layernorm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.2.self_attn.k_norm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.2.self_attn.k_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.2.self_attn.o_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.2.self_attn.q_norm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.2.self_attn.q_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.2.self_attn.v_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.20.input_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.20.mlp.down_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.20.mlp.gate_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.20.mlp.up_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.20.post_attention_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.20.self_attn.k_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.20.self_attn.k_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.20.self_attn.o_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.20.self_attn.q_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.20.self_attn.q_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.20.self_attn.v_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.21.input_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.21.mlp.down_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.21.mlp.gate_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.21.mlp.up_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.21.post_attention_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.21.self_attn.k_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.21.self_attn.k_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.21.self_attn.o_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.21.self_attn.q_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.21.self_attn.q_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.21.self_attn.v_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.22.input_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.22.mlp.down_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.22.mlp.gate_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.22.mlp.up_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.22.post_attention_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.22.self_attn.k_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.22.self_attn.k_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.22.self_attn.o_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.22.self_attn.q_norm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.22.self_attn.q_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.22.self_attn.v_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.23.input_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.23.mlp.down_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.23.mlp.gate_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.23.mlp.up_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.23.post_attention_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.23.self_attn.k_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.23.self_attn.k_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.23.self_attn.o_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.23.self_attn.q_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.23.self_attn.q_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.23.self_attn.v_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.24.input_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.24.mlp.down_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.24.mlp.gate_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.24.mlp.up_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.24.post_attention_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.24.self_attn.k_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.24.self_attn.k_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.24.self_attn.o_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.24.self_attn.q_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.24.self_attn.q_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.24.self_attn.v_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.25.input_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.25.mlp.down_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.25.mlp.gate_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.25.mlp.up_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.25.post_attention_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.25.self_attn.k_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.25.self_attn.k_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.25.self_attn.o_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.25.self_attn.q_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.25.self_attn.q_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.25.self_attn.v_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.26.input_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.26.mlp.down_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.26.mlp.gate_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.26.mlp.up_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.26.post_attention_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.26.self_attn.k_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.26.self_attn.k_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.26.self_attn.o_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.26.self_attn.q_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.26.self_attn.q_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.26.self_attn.v_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.27.input_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.27.mlp.down_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.27.mlp.gate_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.27.mlp.up_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.27.post_attention_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.27.self_attn.k_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.27.self_attn.k_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.27.self_attn.o_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.27.self_attn.q_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.27.self_attn.q_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.27.self_attn.v_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.28.input_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.28.mlp.down_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.28.mlp.gate_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.28.mlp.up_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.28.post_attention_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.28.self_attn.k_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.28.self_attn.k_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.28.self_attn.o_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.28.self_attn.q_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.28.self_attn.q_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.28.self_attn.v_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.29.input_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.29.mlp.down_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.29.mlp.gate_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.29.mlp.up_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.29.post_attention_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.29.self_attn.k_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.29.self_attn.k_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.29.self_attn.o_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.29.self_attn.q_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.29.self_attn.q_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.29.self_attn.v_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.3.input_layernorm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.3.mlp.down_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.3.mlp.gate_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.3.mlp.up_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.3.post_attention_layernorm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.3.self_attn.k_norm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.3.self_attn.k_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.3.self_attn.o_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.3.self_attn.q_norm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.3.self_attn.q_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.3.self_attn.v_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.30.input_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.30.mlp.down_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.30.mlp.gate_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.30.mlp.up_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.30.post_attention_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.30.self_attn.k_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.30.self_attn.k_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.30.self_attn.o_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.30.self_attn.q_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.30.self_attn.q_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.30.self_attn.v_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.31.input_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.31.mlp.down_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.31.mlp.gate_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.31.mlp.up_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.31.post_attention_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.31.self_attn.k_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.31.self_attn.k_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.31.self_attn.o_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.31.self_attn.q_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.31.self_attn.q_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.31.self_attn.v_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.32.input_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.32.mlp.down_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.32.mlp.gate_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.32.mlp.up_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.32.post_attention_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.32.self_attn.k_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.32.self_attn.k_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.32.self_attn.o_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.32.self_attn.q_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.32.self_attn.q_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.32.self_attn.v_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.33.input_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.33.mlp.down_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.33.mlp.gate_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.33.mlp.up_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.33.post_attention_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.33.self_attn.k_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.33.self_attn.k_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.33.self_attn.o_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.33.self_attn.q_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.33.self_attn.q_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.33.self_attn.v_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.34.input_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.34.mlp.down_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.34.mlp.gate_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.34.mlp.up_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.34.post_attention_layernorm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.34.self_attn.k_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.34.self_attn.k_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.34.self_attn.o_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.34.self_attn.q_norm.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.34.self_attn.q_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.34.self_attn.v_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.35.input_layernorm.weight": "pytorch_model-00004-of-00004.bin",
+    "model.layers.35.mlp.down_proj.weight": "pytorch_model-00004-of-00004.bin",
+    "model.layers.35.mlp.gate_proj.weight": "pytorch_model-00004-of-00004.bin",
+    "model.layers.35.mlp.up_proj.weight": "pytorch_model-00004-of-00004.bin",
+    "model.layers.35.post_attention_layernorm.weight": "pytorch_model-00004-of-00004.bin",
+    "model.layers.35.self_attn.k_norm.weight": "pytorch_model-00004-of-00004.bin",
+    "model.layers.35.self_attn.k_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.35.self_attn.o_proj.weight": "pytorch_model-00004-of-00004.bin",
+    "model.layers.35.self_attn.q_norm.weight": "pytorch_model-00004-of-00004.bin",
+    "model.layers.35.self_attn.q_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.35.self_attn.v_proj.weight": "pytorch_model-00003-of-00004.bin",
+    "model.layers.4.input_layernorm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.4.mlp.down_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.4.mlp.gate_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.4.mlp.up_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.4.post_attention_layernorm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.4.self_attn.k_norm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.4.self_attn.k_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.4.self_attn.o_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.4.self_attn.q_norm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.4.self_attn.q_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.4.self_attn.v_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.5.input_layernorm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.5.mlp.down_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.5.mlp.gate_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.5.mlp.up_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.5.post_attention_layernorm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.5.self_attn.k_norm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.5.self_attn.k_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.5.self_attn.o_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.5.self_attn.q_norm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.5.self_attn.q_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.5.self_attn.v_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.6.input_layernorm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.6.mlp.down_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.6.mlp.gate_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.6.mlp.up_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.6.post_attention_layernorm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.6.self_attn.k_norm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.6.self_attn.k_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.6.self_attn.o_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.6.self_attn.q_norm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.6.self_attn.q_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.6.self_attn.v_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.7.input_layernorm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.7.mlp.down_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.7.mlp.gate_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.7.mlp.up_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.7.post_attention_layernorm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.7.self_attn.k_norm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.7.self_attn.k_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.7.self_attn.o_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.7.self_attn.q_norm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.7.self_attn.q_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.7.self_attn.v_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.8.input_layernorm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.8.mlp.down_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.8.mlp.gate_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.8.mlp.up_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.8.post_attention_layernorm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.8.self_attn.k_norm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.8.self_attn.k_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.8.self_attn.o_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.8.self_attn.q_norm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.8.self_attn.q_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.8.self_attn.v_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.9.input_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.9.mlp.down_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.9.mlp.gate_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.9.mlp.up_proj.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.9.post_attention_layernorm.weight": "pytorch_model-00002-of-00004.bin",
+    "model.layers.9.self_attn.k_norm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.9.self_attn.k_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.9.self_attn.o_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.9.self_attn.q_norm.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.9.self_attn.q_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.layers.9.self_attn.v_proj.weight": "pytorch_model-00001-of-00004.bin",
+    "model.norm.weight": "pytorch_model-00004-of-00004.bin",
+    "ts_encoder.mlp.0.bias": "pytorch_model-00004-of-00004.bin",
+    "ts_encoder.mlp.0.weight": "pytorch_model-00004-of-00004.bin",
+    "ts_encoder.mlp.2.bias": "pytorch_model-00004-of-00004.bin",
+    "ts_encoder.mlp.2.weight": "pytorch_model-00004-of-00004.bin",
+    "ts_encoder.mlp.4.bias": "pytorch_model-00004-of-00004.bin",
+    "ts_encoder.mlp.4.weight": "pytorch_model-00004-of-00004.bin",
+    "ts_encoder.mlp.6.bias": "pytorch_model-00004-of-00004.bin",
+    "ts_encoder.mlp.6.weight": "pytorch_model-00004-of-00004.bin",
+    "ts_encoder.mlp.8.bias": "pytorch_model-00004-of-00004.bin",
+    "ts_encoder.mlp.8.weight": "pytorch_model-00004-of-00004.bin",
+    "ts_encoder.position_embedding.weight": "pytorch_model-00004-of-00004.bin"
+  }
+}

special_tokens_map.json

@@ -0,0 +1,20 @@
+{
+  "additional_special_tokens": [
+    "<ts>",
+    "<ts/>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f58249c804f02dc95986d90822c6ca889b91cfdd40befa5066ff7d50fce80342
+size 11423017

tokenizer_config.json

@@ -0,0 +1,249 @@
+{
+  "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
+    },
+    "151665": {
+      "content": "<tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151666": {
+      "content": "</tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151667": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151668": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151669": {
+      "content": "<ts>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151670": {
+      "content": "<ts/>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "additional_special_tokens": [
+    "<ts>",
+    "<ts/>"
+  ],
+  "auto_map": {
+    "AutoProcessor": "processing_qwen3_ts.Qwen3TSProcessor"
+  },
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "processor_class": "Qwen3TSProcessor",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2875c51ada563e5aedfdeacfcafff7e1c038724a25627f0d30a0b6b74d6c18bc
+size 7480