Upload sdar_4b_multi_block_causal-final

README.md

@@ -0,0 +1,61 @@
+---
+library_name: transformers
+license: other
+base_model: JetLM/SDAR-4B-Chat
+tags:
+- llama-factory
+- full
+- generated_from_trainer
+model-index:
+- name: sft
+  results: []
+---
+
+<!-- This model card has been generated automatically according to the information the Trainer had access to. You
+should probably proofread and complete it, then remove this comment. -->
+
+# sft
+
+This model is a fine-tuned version of [./training/model/SDAR-4B-Chat](https://huggingface.co/./training/model/SDAR-4B-Chat) on an unknown dataset.
+
+## Model description
+
+More information needed
+
+## Intended uses & limitations
+
+More information needed
+
+## Training and evaluation data
+
+More information needed
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 1e-05
+- train_batch_size: 4
+- eval_batch_size: 8
+- seed: 42
+- distributed_type: multi-GPU
+- num_devices: 8
+- gradient_accumulation_steps: 2
+- total_train_batch_size: 64
+- total_eval_batch_size: 64
+- optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
+- lr_scheduler_type: cosine
+- lr_scheduler_warmup_ratio: 0.03
+- num_epochs: 3.0
+
+### Training results
+
+
+
+### Framework versions
+
+- Transformers 4.52.4
+- Pytorch 2.8.0+cu128
+- Datasets 3.6.0
+- Tokenizers 0.21.1

added_tokens.json

@@ -0,0 +1,29 @@
+{
+  "</think>": 151668,
+  "</tool_call>": 151658,
+  "</tool_response>": 151666,
+  "<think>": 151667,
+  "<tool_call>": 151657,
+  "<tool_response>": 151665,
+  "<|MASK|>": 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
+}

all_results.json

@@ -0,0 +1,9 @@
+{
+    "effective_tokens_per_sec": 2340.6354596146743,
+    "epoch": 3.0,
+    "total_flos": 2.73366675523371e+17,
+    "train_loss": 0.053853037605258475,
+    "train_runtime": 459.8341,
+    "train_samples_per_second": 48.755,
+    "train_steps_per_second": 0.763
+}

chat_template.jinja

@@ -0,0 +1,85 @@
+{%- 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 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.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set content = message.content %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is defined and message.reasoning_content is not none %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in message.content %}
+                {%- set content = message.content.split('</think>')[-1].lstrip('\n') %}
+                {%- set reasoning_content = message.content.split('</think>')[0].rstrip('\n').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' }}
+        {{- message.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 +1,44 @@
+{
+  "architectures": [
+    "SDARForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_sdar.SDARConfig",
+    "AutoModel": "modeling_sdar.SDARForCausalLM",
+    "AutoModelForCausalLM": "modeling_sdar.SDARForCausalLM"
+  },
+  "block_causal_prompt": true,
+  "block_size": 4,
+  "bos_token_id": 151643,
+  "debug": false,
+  "eos_token_id": 151643,
+  "ep_size": 1,
+  "fuse_cross_entropy": true,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 2560,
+  "initializer_range": 0.02,
+  "intermediate_size": 9728,
+  "mask_token_id": 151669,
+  "max_position_embeddings": 32768,
+  "max_window_layers": 36,
+  "micro_forward": false,
+  "model_type": "sdar",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 36,
+  "num_key_value_heads": 8,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "skip_checkpoint": false,
+  "sliding_window": null,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.4",
+  "use_cache": false,
+  "use_deepep": false,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

configuration_sdar.py

@@ -0,0 +1,212 @@
+# 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.
+"""SDAR 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 SDARConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`SDARModel`]. It is used to instantiate a
+    SDAR model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of
+    SDAR-1.7B [DiffuOpen/SDAR-1.7B-Chat](https://huggingface.co/DiffuOpen/SDAR-1.7B-Chat/).
+
+    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 SDAR model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`SDARModel`]
+        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 checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). 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 that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+
+    ```python
+    >>> from transformers import SDARModel, SDARConfig
+
+    >>> # Initializing a SDAR style configuration
+    >>> configuration = SDARConfig()
+
+    >>> # Initializing a model from the SDAR-8B style configuration
+    >>> model = SDARModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "sdar"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    # Default tensor parallel plan for base model `SDAR`
+    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,
+        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  # we check `use_sliding_window` in the modeling code
+        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)
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+
+__all__ = ["SDARConfig"]

fused_linear_diffusion_cross_entropy.py

@@ -0,0 +1,682 @@
+# -*- coding: utf-8 -*-
+
+# Code adapted from
+# https://github.com/fla-org/flash-linear-attention/blob/main/fla/modules/fused_linear_cross_entropy.py
+# Implementation of element-wise division of cross entropy loss
+
+
+# Code adapted from
+# https://github.com/linkedin/Liger-Kernel/blob/main/src/liger_kernel/ops/fused_linear_cross_entropy.py
+
+from functools import partial
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import triton
+import triton.language as tl
+from torch.distributed import DeviceMesh
+from torch.distributed.tensor import DTensor, Replicate, Shard, distribute_module
+from torch.distributed.tensor.parallel import ParallelStyle
+
+# The hard limit of TRITON_MAX_TENSOR_NUMEL is 1048576
+# https://github.com/triton-lang/triton/blob/ba42a5c68fd0505f8c42f4202d53be0f8d9a5fe0/python/triton/language/core.py#L19
+# However, setting limit as 65536 as in LayerNorm tutorial is faster because of less register spilling
+# The optimal maximum block size depends on your hardware, your kernel, and your dtype
+MAX_FUSED_SIZE = 65536 // 2
+
+
+@triton.heuristics({
+    'HAS_SCALE': lambda args: args['scale'] is not None
+})
+@triton.autotune(
+    configs=[
+        triton.Config({}, num_warps=num_warps)
+        for num_warps in [1, 2, 4, 8, 16, 32]
+    ],
+    key=['D']
+)
+@triton.jit
+def logsumexp_fwd_kernel(
+    x,
+    z,
+    scale,
+    D: tl.constexpr,
+    B: tl.constexpr,
+    HAS_SCALE: tl.constexpr
+):
+    i_n, i_d = tl.program_id(0).to(tl.int64), tl.program_id(1).to(tl.int64)
+    o_d = i_d * B + tl.arange(0, B)
+    m_d = o_d < D
+
+    b_x = tl.load(x + i_n * D + o_d, mask=m_d, other=-float('inf'))
+    if HAS_SCALE:
+        b_x = b_x * scale
+    b_m = tl.max(b_x, 0)
+    b_z = tl.log(tl.sum(tl.exp(b_x - b_m), 0)) + b_m
+    tl.store(z + i_n * tl.cdiv(D, B) + i_d, b_z)
+
+
+def logsumexp_fwd(
+    x,
+    scale: Optional[float] = None,
+    dtype: Optional[torch.dtype] = None
+):
+    r"""
+    Compute the logsumexp of the input tensor over the last dimension.
+
+    Args:
+        x (Tensor):
+            The input tensor of any shape.
+        scale (Optional[float]):
+            The scale applied to the input tensor. Default: `None`.
+        dtype (Optional[torch.dtype]):
+            The data type of the output tensor. Default: `None`.
+    Returns:
+        Tensor: The logsumexp of the input tensor.
+    """
+
+    shape = x.shape
+    x = x.view(-1, shape[-1])
+    N, D = x.shape
+    B = min(triton.next_power_of_2(D), 64 * 1024)
+    ND = triton.cdiv(D, B)
+
+    z = x.new_empty(N, ND, dtype=torch.float)
+    logsumexp_fwd_kernel[(N, ND)](
+        x=x,
+        z=z,
+        scale=scale,
+        D=D,
+        B=B
+    )
+    z = z.logsumexp(-1).view(*shape[:-1])
+    if dtype is not None and dtype != torch.float:
+        z = z.to(dtype)
+    return z
+
+@triton.jit
+def cross_entropy_kernel(
+    logits,
+    lse,
+    target,
+    p_mask,
+    loss,
+    total,
+    ignore_index,
+    label_smoothing: tl.constexpr,
+    logit_scale: tl.constexpr,
+    reduction: tl.constexpr,
+    V: tl.constexpr,
+    BV: tl.constexpr
+):
+    """
+    This kernel computes both cross entropy loss and the gradient of the input.
+    We only consider hard label + mean reduction for now.
+    Please refer to https://pytorch.org/docs/stable/generated/torch.nn.CrossEntropyLoss.html for the math.
+
+    Args:
+        logits:
+            Pointer to logits tensor.
+        lse:
+            Pointer to logsumexp tensor.
+        target: Pointer to target tensor.
+        loss:
+            Pointer to tensor to store the loss.
+        V (int):
+            The number of columns in the input tensor.
+        total (int):
+            The number of non-ignored classes.
+        ignore_index (int):
+            The index to ignore in the target.
+        label_smoothing (float):
+            The amount of smoothing when computing the loss, where 0.0 means no smoothing.
+        reduction (str):
+            The string for the reduction to apply
+        BV (int):
+            The block size for vocab.
+    """
+
+    # https://github.com/triton-lang/triton/issues/1058
+    # If B*T*V is too large, i_n * stride will overflow out of int32, so we convert to int64
+    i_n = tl.program_id(0).to(tl.int64)
+    NV = tl.cdiv(V, BV)
+
+    # 1. Load target first because if the target is ignore_index, we can return right away
+    b_y = tl.load(target + i_n)
+    # load p_mask
+    b_p_mask = tl.load(p_mask + i_n)
+
+    # 2. locate the start index
+    logits += i_n * V
+
+    if b_y == ignore_index:
+        # set all x as 0
+        for i in range(0, V, BV):
+            o_v = i + tl.arange(0, BV)
+            tl.store(logits + o_v, 0.0, mask=o_v < V)
+        return
+
+    # Online softmax: 2 loads + 1 store (compared with 3 loads + 1 store for the safe softmax)
+    # Refer to Algorithm 3 in the paper: https://arxiv.org/pdf/1805.02867
+
+    # 3. [Online softmax] first pass: compute logsumexp
+    # we did this in anouter kernel
+    b_l = tl.load(logits + b_y) * logit_scale
+    b_lse = tl.load(lse + i_n)
+
+    # 4. Calculate the loss
+    # loss = lse - logits_l
+    # celoss = -log(q_y) = -log(softmax(x_y))
+    b_loss = (b_lse - b_l) / b_p_mask  # Diffusion Scaled '1/t'
+
+    # Label smoothing is a general case of normal cross entropy
+    # See the full derivation at https://github.com/linkedin/Liger-Kernel/pull/198#issue-2503665310
+    b_z = 0.0
+    eps = label_smoothing / V
+
+    # We need tl.debug_barrier() as mentioned in
+    # https://github.com/triton-lang/triton/blob/ba42a5c68fd0505f8c42f4202d53be0f8d9a5fe0/python/triton/ops/cross_entropy.py#L34
+    tl.debug_barrier()
+
+    # 5. [Online Softmax] Second pass: compute gradients
+    # For 'mean' reduction, gradients are normalized by number of non-ignored elements
+    # dx_y = (softmax(x_y) - 1) / N
+    # dx_i = softmax(x_i) / N, i != y
+    # For label smoothing:
+    # dx_i = (softmax(x_y) - label_smoothing / V) / N, i != y
+    # dx_y = (softmax(x_y) - label_smoothing / V - (1 - label_smoothing)) / N
+    #      = dx_i - (1 - label_smoothing) / N
+    for iv in range(0, NV):
+        o_v = iv * BV + tl.arange(0, BV)
+        b_logits = tl.load(logits + o_v, mask=o_v < V, other=float('-inf')) * logit_scale
+        if label_smoothing > 0:
+            # scale X beforehand to avoid overflow
+            b_z += tl.sum(tl.where(o_v < V, -eps * b_logits, 0.0))
+        b_p = (tl.exp(b_logits - b_lse) - eps) * logit_scale
+        b_p /= b_p_mask # 修改
+        if reduction == "mean":
+            b_p = b_p / total
+        tl.store(logits + o_v, b_p, mask=o_v < V)
+
+        tl.debug_barrier()
+
+    # Orginal loss = H(q, p),  with label smoothing regularization = H(q', p) and (label_smoothing / V) = eps
+    # H(q', p) = (1 - label_smoothing) * H(q, p) + label_smoothing * H(u, p)
+    #          = (1 - label_smoothing) * H(q, p) + eps * sum(logsoftmax(x_i))
+    # By using m (global max of xi) and d (sum of e^(xi-m)), we can simplify as:
+    #          = (1 - label_smoothing) * H(q, p) + (-sum(x_i * eps) + label_smoothing * (m + logd))
+    # Refer to H(q', p) in section 7 of the paper:
+    # https://arxiv.org/pdf/1512.00567
+    # pytorch:
+    # https://github.com/pytorch/pytorch/blob/2981534f54d49fa3a9755c9b0855e7929c2527f0/aten/src/ATen/native/LossNLL.cpp#L516
+    # See full derivation at https://github.com/linkedin/Liger-Kernel/pull/198#issuecomment-2333753087
+    if label_smoothing > 0:
+        b_loss = b_loss * (1 - label_smoothing) + (b_z + label_smoothing * b_lse)
+
+    # 6. Specially handle the i==y case where `dx_y = (softmax(x_y) - (1 - label_smoothing) / N`
+    b_l = tl.load(logits + b_y)
+
+    # Normalize the loss by the number of non-ignored elements if reduction is "mean"
+    if reduction == 'mean':
+        b_loss = b_loss / total
+        # b_l += (label_smoothing - 1) / total * logit_scale
+        # b_l has already been divided by b_p_mask and total
+        b_l += (label_smoothing - 1) / b_p_mask / total * logit_scale
+    else:
+        # b_l += (label_smoothing - 1) * logit_scale
+        b_l += (label_smoothing - 1) / b_p_mask * logit_scale
+
+    tl.store(loss + i_n, b_loss)
+    tl.store(logits + b_y, b_l)
+
+
+@triton.jit
+def elementwise_mul_kernel(
+    x,
+    g,
+    N: tl.constexpr,
+    B: tl.constexpr
+):
+    """
+    This function multiplies each element of the tensor pointed by x with the value pointed by g.
+    The multiplication is performed in-place on the tensor pointed by x.
+
+    Parameters:
+    x:
+        Pointer to the input tensor.
+    g:
+        Pointer to the gradient output value.
+    N (int):
+        The number of columns in the input tensor.
+    B (int):
+        The block size for Triton operations.
+    """
+
+    # Get the program ID and convert it to int64 to avoid overflow
+    i_x = tl.program_id(0).to(tl.int64)
+    o_x = i_x * B + tl.arange(0, B)
+
+    # Load the gradient output value
+    b_g = tl.load(g)
+    b_x = tl.load(x + o_x, mask=o_x < N)
+    tl.store(x + o_x, b_x * b_g, mask=o_x < N)
+
+
+def fused_linear_cross_entropy_forward(
+    x: torch.Tensor,
+    target: torch.LongTensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor = None,
+    p_mask: torch.Tensor = None,
+    ignore_index: int = -100,
+    label_smoothing: float = 0.0,
+    logit_scale: float = 1.0,
+    num_chunks: int = 8,
+    reduction: str = "mean"
+):
+    device = x.device
+    # inputs have shape: [N, H]
+    # materialized activations will have shape: [N, V]
+    # the increase in memory = [N, V]
+    # reduction can be achieved by partitioning the number of tokens N into smaller chunks.
+
+    # ideally, we would like to achieve the same memory consumption as [N, H],
+    # so the expected chunk size should be:
+    # NC = ceil(V / H)
+    # C = ceil(N / NC)
+    # for ex: N = 4096*4, V = 32000, H = 4096 ==> NC = 8, C = ceil(N / NC) = 2048
+    N, H, V = *x.shape, weight.shape[0]
+    BV = min(MAX_FUSED_SIZE, triton.next_power_of_2(V))
+    # TODO: in real cases, we may need to limit the number of chunks NC to
+    # ensure the precisions of accumulated gradients
+    NC = min(num_chunks, triton.cdiv(V, H))
+    C = triton.next_power_of_2(triton.cdiv(N, NC))
+    NC = triton.cdiv(N, C)
+
+    # [N, H]
+    dx = torch.zeros_like(x, device=device)
+    # [V, H]
+    dw = torch.zeros_like(weight, device=device, dtype=torch.float) if weight is not None else None
+    # [V]
+    db = torch.zeros_like(bias, device=device, dtype=torch.float) if bias is not None else None
+    # [N]
+    loss = torch.zeros(N, device=device, dtype=torch.float)
+
+    total = target.ne(ignore_index).sum().item()
+
+    for ic in range(NC):
+        start, end = ic * C, min((ic + 1) * C, N)
+        # [C, N]
+        c_x = x[start:end]
+        # when doing matmul, use the original precision
+        # [C, V]
+        c_logits = F.linear(c_x, weight, bias)
+        c_target = target[start:end]
+        c_p_mask = p_mask[start:end]
+        # [C]
+        # keep lse in fp32 to maintain precision
+        c_lse = logsumexp_fwd(c_logits, scale=logit_scale, dtype=torch.float)
+
+        # unreduced loss
+        c_loss = loss[start:end]
+
+        # Here we calculate the gradient of c_logits in place so we can save memory.
+        cross_entropy_kernel[(c_logits.shape[0],)](
+            logits=c_logits,
+            lse=c_lse,
+            target=c_target,
+            p_mask=c_p_mask,
+            loss=c_loss,
+            total=total,
+            ignore_index=ignore_index,
+            label_smoothing=label_smoothing,
+            logit_scale=logit_scale,
+            reduction=reduction,
+            V=V,
+            BV=BV,
+            num_warps=32
+        )
+
+        # gradient of logits is computed in-place by the above triton kernel and is of shape: C x V
+        # thus dx should be of shape: C x H
+        dx[start:end] = torch.mm(c_logits, weight)
+
+        # keep dw in fp32 to maintain precision
+        if weight is not None:
+            dw += c_logits.t() @ c_x
+
+        if bias is not None:
+            torch.add(input=db, other=c_logits.sum(0), out=db)
+
+    loss = loss.sum()
+    if dw is not None:
+        dw = dw.to(weight)
+    if db is not None:
+        db = db.to(bias)
+    return loss, dx, dw, db
+
+
+def fused_linear_cross_entropy_backward(
+    do: torch.Tensor,
+    dx: torch.Tensor,
+    dw: torch.Tensor,
+    db: torch.Tensor
+):
+    # If cross entropy is the last layer, do is 1.0. Skip the mul to save time
+    if torch.ne(do, torch.tensor(1.0, device=do.device)):
+        # We use a Triton kernel instead of a PyTorch operation because modifying inputs in-place
+        # for gradient storage and backward multiple times causes anomalies with PyTorch but not with Triton.
+        N, H = dx.shape
+        B = min(MAX_FUSED_SIZE, triton.next_power_of_2(H))
+
+        elementwise_mul_kernel[(triton.cdiv(N * H, B),)](
+            x=dx,
+            g=do,
+            N=N*H,
+            B=B,
+            num_warps=32,
+        )
+
+        # handle dw
+        if dw is not None:
+            V, H = dw.shape
+            elementwise_mul_kernel[(triton.cdiv(V * H, B),)](
+                x=dw,
+                g=do,
+                N=V*H,
+                B=B,
+                num_warps=32,
+            )
+
+        if db is not None:
+            V = db.shape[0]
+            elementwise_mul_kernel[(triton.cdiv(V, B),)](
+                x=db,
+                g=do,
+                N=V,
+                B=B,
+                num_warps=32,
+            )
+    return dx, dw, db
+
+
+class FusedLinearCrossEntropyFunction(torch.autograd.Function):
+
+    @staticmethod
+    def forward(
+        ctx,
+        x: torch.Tensor,
+        target: torch.LongTensor,
+        weight: torch.Tensor,
+        bias: torch.Tensor = None,
+        p_mask: torch.Tensor = None,
+        ignore_index: int = -100,
+        label_smoothing: float = 0.0,
+        logit_scale: float = 1.0,
+        num_chunks: int = 8,
+        reduction: str = "mean"
+    ):
+        """
+        Fusing the last linear layer with cross-entropy loss
+            Reference: https://github.com/mgmalek/efficient_cross_entropy
+
+        Handle the forward and backward pass of the final linear layer via cross-entropy loss by avoiding
+        the materialization of the large logits tensor. Since Cross Entropy Loss is the last layer, we can
+        compute the gradient at the forward pass. By doing so, we don't have to store the x and target
+        for the backward pass.
+
+        x (torch.Tensor): [batch_size * seq_len, hidden_size]
+        target (torch.LongTensor): [batch_size * seq_len]
+            where each value is in [0, vocab_size).
+        weight (torch.Tensor): [vocab_size, hidden_size]
+            where `vocab_size` is the number of classes.
+        bias (Optional[torch.Tensor]): [vocab_size]
+            where `vocab_size` is the number of classes.
+        p_mask(torch.Tensor): [batch_size * seq_len]
+            Its shape should be same as target. 
+        ignore_index:
+            the index to ignore in the target.
+        label_smoothing:
+            the amount of smoothing when computing the loss, where 0.0 means no smoothing.
+        logit_scale: float = 1.0,
+            A scaling factor applied to the logits. Default: 1.0
+        num_chunks: int
+            The number of chunks to split the input tensor into for processing.
+            This can help optimize memory usage and computation speed.
+            Default: 8
+        reduction:
+            Specifies the reduction to apply to the output: 'mean' | 'sum'.
+            'mean': the weighted mean of the output is taken,
+            'sum': the output will be summed.
+            Default: 'mean'.
+        """
+        loss, dx, dw, db = fused_linear_cross_entropy_forward(
+            x,
+            target,
+            weight,
+            bias,
+            p_mask,
+            ignore_index,
+            label_smoothing,
+            logit_scale,
+            num_chunks,
+            reduction
+        )
+        # downcast to dtype and store for backward
+        ctx.save_for_backward(
+            dx.detach(),
+            dw.detach() if weight is not None else None,
+            db.detach() if bias is not None else None,
+        )
+        return loss
+
+    @staticmethod
+    def backward(ctx, do):
+        dx, dw, db = ctx.saved_tensors
+        dx, dw, db = fused_linear_cross_entropy_backward(do, dx, dw, db)
+        # 10 gradients should be returned, with `p_mask` having no grads
+        # Check the number of arguments in the `forward` method
+        return dx, None, dw, db, None, None, None, None, None, None
+
+
+def fused_linear_cross_entropy_loss(
+    x: torch.Tensor,
+    target: torch.LongTensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor = None,
+    p_mask: torch.Tensor = None,
+    ignore_index: int = -100,
+    label_smoothing: float = 0.0,
+    logit_scale: float = 1.0,
+    num_chunks: int = 8,
+    reduction: str = "mean"
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Args:
+        x (torch.Tensor): [batch_size * seq_len, hidden_size]
+        target (torch.LongTensor): [batch_size * seq_len]
+            where each value is in [0, vocab_size).
+        weight (torch.Tensor): [vocab_size, hidden_size]
+            where `vocab_size` is the number of classes.
+        bias (Optional[torch.Tensor]): [vocab_size]
+            where `vocab_size` is the number of classes.
+        p_mask(torch.Tensor): [batch_size * seq_len]
+            Its shape should be same as target. 
+        ignore_index: int.
+            If target == ignore_index, the loss is set to 0.0.
+        label_smoothing: float
+        logit_scale: float
+            A scaling factor applied to the logits. Default: 1.0
+        num_chunks: int
+            The number of chunks to split the input tensor into for processing.
+            This can help optimize memory usage and computation speed.
+            Default: 8
+        reduction:
+            Specifies the reduction to apply to the output: 'mean' | 'sum'.
+            'mean': the weighted mean of the output is taken,
+            'sum': the output will be summed.
+            Default: 'mean'.
+    Returns:
+        losses: [batch,], float
+    """
+    return FusedLinearCrossEntropyFunction.apply(
+        x,
+        target,
+        weight,
+        bias,
+        p_mask,
+        ignore_index,
+        label_smoothing,
+        logit_scale,
+        num_chunks,
+        reduction
+    )
+
+
+class FusedLinearDiffusionCrossEntropyLoss(nn.Module):
+
+    def __init__(
+        self,
+        ignore_index: int = -100,
+        label_smoothing: float = 0.0,
+        logit_scale: float = 1.0,
+        num_chunks: int = 8,
+        reduction: str = "mean"
+    ):
+        """
+        Args:
+            ignore_index: int.
+                If target == ignore_index, the loss is set to 0.0.
+            label_smoothing: float
+            logit_scale: float
+                A scaling factor applied to the logits. Default: 1.0
+            num_chunks: int
+                The number of chunks to split the input tensor into for processing.
+                This can help optimize memory usage and computation speed.
+                Default: 8
+            reduction:
+                Specifies the reduction to apply to the output: 'mean' | 'sum'.
+                'mean': the weighted mean of the output is taken,
+                'sum': the output will be summed.
+                Default: 'mean'.
+        """
+        super().__init__()
+
+        assert reduction in ["mean", "sum"], f"reduction: {reduction} is not supported"
+
+        self.ignore_index = ignore_index
+        self.label_smoothing = label_smoothing
+        self.logit_scale = logit_scale
+        self.num_chunks = num_chunks
+        self.reduction = reduction
+
+    @torch.compiler.disable
+    def forward(
+        self,
+        x: torch.Tensor,
+        target: torch.LongTensor,
+        weight: torch.Tensor,
+        bias: Optional[torch.Tensor] = None,
+        p_mask: torch.Tensor = None
+    ):
+        """
+        Args:
+            x (torch.Tensor): [batch_size, seq_len, hidden_size]
+            target (torch.LongTensor): [batch_size, seq_len]
+                where each value is in [0, V).
+            weight (torch.Tensor): [vocab_size, hidden_size]
+                where `vocab_size` is the number of classes.
+            bias (Optional[torch.Tensor]): [vocab_size]
+                where `vocab_size` is the number of classes.
+            p_mask(torch.Tensor): [batch_size, seq_len]
+                Its shape is same as target. 
+                Shape: (1, packed_length) when varlen attn is used.
+        Returns:
+            loss
+            
+        TODO:
+            follow https://github.com/ML-GSAI/LLaDA/blob/main/GUIDELINES.md#pre-training
+            ```py
+            unreduced_loss /= p_mask
+            ```
+            Scale the values of `unreduced_loss at different positions
+        """
+        if p_mask is None:
+            p_mask = torch.ones_like(target, dtype=torch.float, device=x.device)
+        
+        x = x.contiguous().view(-1, x.shape[-1])
+        target = target.contiguous().view(-1)
+        weight = weight.contiguous()
+        bias = bias.contiguous() if bias else None
+        p_mask = p_mask.contiguous().view(-1)
+        l, d = x.shape
+        assert l == target.shape[0] == p_mask.shape[0], f"{x.shape=}, {target.shape=}, {p_mask.shape=}"
+            
+        loss = fused_linear_cross_entropy_loss(
+            x,
+            target,
+            weight=weight,
+            bias=bias,
+            p_mask=p_mask,
+            ignore_index=self.ignore_index,
+            label_smoothing=self.label_smoothing,
+            logit_scale=self.logit_scale,
+            num_chunks=self.num_chunks,
+            reduction=self.reduction
+        )
+        return loss
+
+
+class LinearLossParallel(ParallelStyle):
+    def __init__(
+        self,
+        *,
+        sequence_dim: int = 1,
+        use_local_output: bool = False,
+    ):
+        super().__init__()
+
+        self.sequence_sharding = (Shard(sequence_dim),)
+        self.use_local_output = use_local_output
+
+    @staticmethod
+    def _prepare_input_fn(sequence_sharding, mod, inputs, device_mesh):
+        x, target, weight, bias = inputs
+
+        if not isinstance(x, DTensor):
+            # assume the input passed in already sharded on the sequence dim and create the DTensor
+            x = DTensor.from_local(x, device_mesh, sequence_sharding)
+        if x.placements != sequence_sharding:
+            x = x.redistribute(placements=sequence_sharding, async_op=True)
+        if not isinstance(target, DTensor):
+            target = DTensor.from_local(target, device_mesh, [Replicate()])
+        if target.placements != sequence_sharding:
+            target = target.redistribute(placements=sequence_sharding, async_op=True)
+
+        if not isinstance(weight, DTensor):
+            weight = DTensor.from_local(weight, device_mesh, [Replicate()])
+        if weight.placements != [Replicate()]:
+            # we replicate the weight/bias in FLCE
+            weight = weight.redistribute(placements=[Replicate()], async_op=True)
+
+        if bias is not None and not isinstance(bias, DTensor):
+            bias = DTensor.from_local(bias, device_mesh, [Replicate()])
+        if bias is not None and bias.placements != [Replicate()]:
+            bias = bias.redistribute(placements=[Replicate()], async_op=True)
+
+        return x.to_local(), target.to_local(), weight.to_local(), bias.to_local() if bias is not None else bias
+
+    @staticmethod
+    def _prepare_output_fn(use_local_output, mod, outputs, device_mesh):
+        return outputs.to_local() if use_local_output else outputs
+
+    def _apply(self, module: nn.Module, device_mesh: DeviceMesh) -> nn.Module:
+        return distribute_module(
+            module,
+            device_mesh,
+            partition_fn=None,
+            input_fn=partial(self._prepare_input_fn, self.sequence_sharding),
+            output_fn=partial(self._prepare_output_fn, self.use_local_output)
+        )

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"
+}

model-00001-of-00002.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:644d52de1a2695e24391a147ada735b028d3b5325991b50eeb5a601af1bd3c2b
+size 4967215360

model-00002-of-00002.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:596b253b261e4d5005b64a7fd23038a019c9f0a123aaaf13563c69b550bb68fb
+size 3855679144

model.safetensors.index.json

@@ -0,0 +1,406 @@
+{
+  "metadata": {
+    "total_size": 8822848512
+  },
+  "weight_map": {
+    "lm_head.weight": "model-00002-of-00002.safetensors",
+    "model.embed_tokens.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.14.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.15.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.16.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.17.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.18.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.19.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.20.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.20.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.20.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.20.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.21.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.21.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.21.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.21.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.21.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.21.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.21.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.21.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.21.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.21.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.21.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.22.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.22.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.22.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.22.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.22.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.22.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.22.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.22.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.22.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.22.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.22.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.23.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.23.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.23.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.23.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.23.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.23.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.23.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.23.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.23.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.23.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.23.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.24.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.24.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.24.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.24.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.24.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.24.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.24.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.24.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.24.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.24.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.24.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.25.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.25.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.25.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.25.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.25.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.25.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.25.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.25.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.25.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.25.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.25.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.26.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.26.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.26.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.26.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.26.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.26.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.26.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.26.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.26.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.26.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.26.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.27.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.27.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.27.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.27.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.27.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.27.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.27.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.27.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.27.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.27.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.27.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.28.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.28.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.28.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.28.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.28.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.28.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.28.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.28.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.28.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.28.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.28.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.29.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.29.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.29.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.29.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.29.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.29.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.29.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.29.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.29.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.29.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.29.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.3.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.30.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.30.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.30.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.30.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.30.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.30.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.30.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.30.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.30.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.30.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.30.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.31.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.31.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.31.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.31.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.31.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.31.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.31.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.31.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.31.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.31.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.31.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.32.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.32.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.32.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.32.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.32.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.32.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.32.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.32.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.32.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.32.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.32.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.33.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.33.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.33.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.33.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.33.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.33.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.33.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.33.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.33.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.33.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.33.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.34.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.34.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.34.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.34.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.34.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.34.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.34.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.34.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.34.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.34.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.34.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.35.input_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.35.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.35.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.35.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.35.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.35.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.35.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.35.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.35.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.35.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.35.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.4.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.input_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.norm.weight": "model-00002-of-00002.safetensors"
+  }
+}

modeling_sdar.py

@@ -0,0 +1,1519 @@
+# This file is modified based on https://github.com/huggingface/transformers/blob/v4.52.4/src/transformers/models/qwen3/modeling_qwen3.py.
+#
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/qwen3/modular_qwen3.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_qwen3.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# 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, Tuple, Union, List
+
+import torch
+from torch import nn
+from einops import rearrange
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache
+from transformers.generation import GenerationMixin
+from transformers.integrations import use_kernel_forward_from_hub
+from transformers.modeling_attn_mask_utils import AttentionMaskConverter
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.modeling_layers import GradientCheckpointingLayer
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.processing_utils import Unpack
+from transformers.utils import LossKwargs, auto_docstring, can_return_tuple, is_torch_flex_attn_available, logging
+from .configuration_sdar import SDARConfig
+from .fused_linear_diffusion_cross_entropy import FusedLinearDiffusionCrossEntropyLoss
+
+from flash_attn.ops.triton.layer_norm import rms_norm_fn as flash_rms_norm
+
+import torch.nn.functional as F
+try:
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input
+except:
+    pass
+
+try:
+    from liger_kernel.ops.swiglu import LigerSiLUMulFunction  # noqa: F401
+    liger_kernel_is_available = True
+except ImportError:
+    liger_kernel_is_available = False
+
+
+if is_torch_flex_attn_available():
+    from torch.nn.attention.flex_attention import BlockMask, create_block_mask, flex_attention
+    from transformers.integrations.flex_attention import make_flex_block_causal_mask
+
+
+logger = logging.get_logger(__name__)
+
+
+def modify_padded_position_ids_2d(position_ids: torch.LongTensor) -> torch.LongTensor:
+    """
+    使用完全向量化的 PyTorch 操作修改一个 batch 的 packed position_ids。
+    这个函数假设输入是一个 2D Tensor，形状为 (batch_size, sequence_length)。
+    它会独立地处理 batch 中的每一行。
+
+    Args:
+        position_ids: 二维 PyTorch Tensor, shape (batch_size, sequence_length).
+
+    Returns:
+        修改后的 position_ids Tensor, shape (batch_size, sequence_length).
+    """
+    if position_ids.dim() != 2:
+        raise ValueError(f"Input tensor must be 2D, but got {position_ids.dim()} dimensions.")
+        
+    batch_size, seq_len = position_ids.shape
+    device = position_ids.device
+
+    col_indices = torch.arange(seq_len, device=device, dtype=position_ids.dtype).expand(batch_size, -1)
+    mask = (position_ids != 0)
+
+    masked_indices = col_indices * mask
+    last_nonzero_idx = torch.max(masked_indices, dim=1).values
+    has_nonzero = torch.any(mask, dim=1)
+    pad_start_idx = torch.where(has_nonzero, last_nonzero_idx + 1, torch.tensor(0, device=device, dtype=position_ids.dtype))
+
+    padding_mask = col_indices >= pad_start_idx.unsqueeze(1)
+    new_pad_values = col_indices - pad_start_idx.unsqueeze(1)
+    position_ids = torch.where(padding_mask, new_pad_values, position_ids)
+
+    return position_ids
+
+
+def calculate_token_nums(position_ids: torch.Tensor):
+    """
+    使用 PyTorch 高效计算一个批次中每个打包序列的长度。
+
+    Args:
+        position_ids (torch.Tensor): 一个 2D Tensor，形状为 (batch_size, sequence_length)。
+                                     例如：tensor([[0,1,2,3,4,0,1,2,3,4,5,0,1,2,3,0,0,0]])
+    Returns:
+        list[list[int]]: 一个嵌套列表，包含每个批次项中各个序列的长度。
+                         例如：[[5, 6, 4, 1, 1, 1]]
+    """
+    # 检查输入是否为 2D Tensor
+    if position_ids.dim() != 2:
+        raise ValueError(f"输入必须是 2D Tensor，但得到了 {position_ids.dim()}D")
+
+    all_lengths = []
+    
+    # 我们按批次逐行处理。因为每行的序列长度数量不同（ragged），
+    # 所以 Python 循环在批次维度上是最高效且最清晰的写法。
+    # 循环内部的操作是完全向量化的。
+    for pids_row in position_ids:
+        # 获取当前行的总长度
+        seq_len = pids_row.shape[0]
+        
+        # 1. 找到所有值为 0 的元素的索引
+        # pids_row == 0 会返回一个布尔 Tensor: [True, False, ..., True, ...]
+        # torch.nonzero 会返回这些 True 值的索引
+        # .flatten() 将其从 (N, 1) 形状的 Tensor 变为 (N,) 形状
+        zero_indices = torch.nonzero(pids_row == 0).flatten()
+        
+        # 2. 将序列的总长度作为一个额外的切分点添加到末尾
+        # 这对于计算最后一个序列的长度至关重要
+        # 注意：要确保新创建的 tensor 和原始 tensor 在同一个设备上 (cpu/cuda)
+        split_points = torch.cat([
+            zero_indices,
+            torch.tensor([seq_len], device=pids_row.device, dtype=zero_indices.dtype)
+        ])
+        
+        # 3. 计算相邻切分点之间的差值，这就是我们想要的长度
+        # torch.diff([a, b, c, d]) 会返回 [b-a, c-b, d-c]
+        lengths = torch.diff(split_points)
+
+        all_lengths.append(lengths)
+
+    return all_lengths
+
+
+def forward_add_noise_packed(
+    inputs_ids: torch.Tensor,
+    num_tokens_list: List[torch.Tensor],
+    prompt_mask: torch.Tensor,
+    mask_id: int,
+    eps: float = 1e-3,
+    max_tries: int = 10,
+) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    """
+    为一批打包（packed）序列的 token ID 添加噪声。
+
+    此函数保留了为每个逻辑样本（在每个批次项内拼接）生成独立随机噪声率的逻辑。
+    它会随机将一部分 token 的 ID 替换为 mask_id。
+    这个过程会避开被 prompt_mask 标记的位置。
+
+    Args:
+        inputs_ids (torch.Tensor): 
+            输入的 token ID 张量，形状为 (bsz, total_tokens)。
+        num_tokens_list (List[torch.Tensor]): 
+            一个张量列表，长度为 bsz。列表中的每个张量记录了对应批次项中
+            每个逻辑样本的长度。例如: [tensor([len1, len2]), tensor([len3, len4, len5])].
+        prompt_mask (torch.Tensor): 
+            布尔型张量，形状为 (bsz, total_tokens)，值为 True 的位置表示是 prompt，
+            不应添加噪声。
+        mask_id (int): 
+            用于替换的 mask token 的 ID。
+        eps (float): 
+            微小值，用于防止噪声率 t 恰好为 0，确保 p_mask > 0。
+        max_tries (int): 
+            为确保至少一个非 prompt token 被 mask，对每个批次项尝试的最大次数。
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        - noisy_input_ids (torch.Tensor): 
+            添加噪声后的 token ID 张量，形状为 (bsz, total_tokens)。
+        - final_masked_indices (torch.Tensor): 
+            布尔型张量，标记了哪些位置被实际 mask 了，形状为 (bsz, total_tokens)。
+        - p_masks (torch.Tensor): 
+            一个一维张量，包含了被 mask 的 token 对应的实际噪声率。
+    """
+    # 1. 验证和获取形状
+    bsz, total_tokens = inputs_ids.shape
+    device = inputs_ids.device
+
+    # 检查输入的一致性
+    assert len(num_tokens_list) == bsz, f"num_tokens_list 的长度 ({len(num_tokens_list)}) 必须等于 bsz ({bsz})"
+    assert prompt_mask.shape == (bsz, total_tokens), f"prompt_mask 形状不匹配, 期望 {(bsz, total_tokens)}, 得到 {prompt_mask.shape}"
+
+    # 准备结果容器
+    noisy_ids_list = []
+    final_masked_indices_list = []
+    p_masks_per_token_list = []
+
+    # 2. 在批次维度上迭代
+    # 这是处理不同打包结构最直接有效的方法
+    for i in range(bsz):
+        # 提取当前批次项的数据
+        current_ids = inputs_ids[i:i+1] # shape: (1, total_tokens)
+        current_num_tokens = num_tokens_list[i]
+        current_prompt_mask = prompt_mask[i:i+1] # shape: (1, total_tokens)
+        
+        num_samples_in_item = len(current_num_tokens)
+        # 验证当前批次项的 token 总数是否匹配
+        assert total_tokens == torch.sum(current_num_tokens), \
+            f"批次项 {i} 的 num_tokens 之和 ({torch.sum(current_num_tokens)}) 与 total_tokens ({total_tokens}) 不匹配"
+
+        eligible_for_masking = ~current_prompt_mask
+
+        # 如果没有任何 token 可以被 mask，直接使用原始输入，并设置 p_mask 为 eps
+        if not eligible_for_masking.any():
+            noisy_ids_list.append(current_ids)
+            final_masked_indices_list.append(torch.zeros_like(current_prompt_mask, dtype=torch.bool))
+            # p_mask_per_token 的形状应为 (1, total_tokens) 以便后续拼接
+            p_masks_per_token_list.append(torch.full((1, total_tokens), eps, device=device, dtype=torch.float))
+            continue
+
+        # --- 尝试生成 mask，确保至少 mask 一个 token ---
+        final_masked_indices_item = torch.zeros_like(current_prompt_mask, dtype=torch.bool)
+        p_mask_per_token = None
+        
+        for _ in range(max_tries):
+            # 为每个逻辑样本生成一个独立的噪声率 t
+            t = torch.rand(num_samples_in_item, device=device)
+            p_mask_per_sample = (1 - eps) * t + eps
+
+            # 将每个样本的噪声率扩展到其所有 token 上
+            p_mask_per_token_1d = torch.repeat_interleave(p_mask_per_sample, current_num_tokens)
+            p_mask_per_token = p_mask_per_token_1d.unsqueeze(0) # shape: (1, total_tokens)
+
+            # 根据噪声率生成随机 mask
+            masked_indices = torch.rand_like(p_mask_per_token) < p_mask_per_token
+            # 应用 prompt mask，确保 prompt 不被 mask
+            final_masked_indices_item = masked_indices & eligible_for_masking
+
+            # 如果成功 mask 了至少一个 token，则跳出尝试循环
+            if final_masked_indices_item.any():
+                break
+        
+        # 如果 max_tries 之后仍然没有 mask 任何 token (极小概率)，就强制 mask 一个可 mask 的 token
+        if not final_masked_indices_item.any():
+            eligible_indices = torch.nonzero(eligible_for_masking.squeeze(0), as_tuple=True)[0]
+            if len(eligible_indices) > 0:
+                # 随机选择一个可 mask 的位置
+                random_choice = torch.randint(0, len(eligible_indices), (1,)).item()
+                force_mask_idx = eligible_indices[random_choice]
+                final_masked_indices_item[0, force_mask_idx] = True
+
+
+        # --- 根据最终的 mask 生成带噪声的 IDs ---
+        noisy_ids_item = torch.where(
+            final_masked_indices_item,
+            mask_id,
+            current_ids
+        )
+        
+        # 保存这个批次项的结果
+        noisy_ids_list.append(noisy_ids_item)
+        final_masked_indices_list.append(final_masked_indices_item)
+        p_masks_per_token_list.append(p_mask_per_token)
+
+    # 3. 将列表中的结果堆叠成最终的批处理张量
+    noisy_input_ids = torch.cat(noisy_ids_list, dim=0)
+    final_masked_indices = torch.cat(final_masked_indices_list, dim=0)
+    p_mask_full = torch.cat(p_masks_per_token_list, dim=0)
+    
+    # 4. 提取被 mask 位置对应的噪声率
+    p_masks = p_mask_full[final_masked_indices]
+    
+    return noisy_input_ids, final_masked_indices, p_masks
+
+
+def block_diff_mask(b, h, q_idx, kv_idx, block_size=None, n=None):
+    """
+    Constructs the specialized block diffusion attention mask for training
+    composed of three masks:
+    - **Block Diagonal Mask (M_BD)**: Self-attention within noised blocks
+    - **Offset Block Causal Mask (M_OBC)**: Cross-attention for conditional context
+    - **Block Causal Mask (M_BC)**: Attention to update x0
+
+    Args:
+        b, h: Batch and head indices (ignored for mask logic).
+        q_idx, kv_idx: Query and Key indices.
+        seq_len: Total sequence length.
+        block_size: Defines the block structure.
+
+    Returns:
+        A boolean attention mask.
+    """
+
+    # Indicate whether token belongs to xt or x0
+    x0_flag_q = q_idx >= n
+    x0_flag_kv = kv_idx >= n
+
+    # Compute block indices
+    block_q = torch.where(
+        x0_flag_q == 1, (q_idx - n) // block_size, q_idx // block_size
+    )
+    block_kv = torch.where(
+        x0_flag_kv == 1, (kv_idx - n) // block_size, kv_idx // block_size
+    )
+
+    # **1. Block Diagonal Mask (M_BD) **
+    block_diagonal = (block_q == block_kv) & (x0_flag_q == x0_flag_kv)
+
+    # **2. Offset Block-Causal Mask (M_OBC) **
+    offset_block_causal = (block_q > block_kv) & (
+        x0_flag_kv == 1) & (x0_flag_q == 0)
+
+    # **3. Block-Causal Mask (M_BC) **
+    block_causal = (block_q >= block_kv) & (x0_flag_kv == 1) & (x0_flag_q == 1)
+
+    # **4. Combine Masks **
+    return block_diagonal | offset_block_causal | block_causal
+
+
+def block_attn_mask(num_tokens, block_size, device):
+    masks = []
+    for i in range(len(num_tokens)):
+        cur_masks = []
+        for num in num_tokens[i]:
+            # 全部返回 n*n 而非 2n*2n
+            single_mask = block_diff_mask(
+                b=None,
+                h=None,
+                q_idx=torch.arange(num * 2, device=device)[:, None],
+                kv_idx=torch.arange(num * 2, device=device)[None, :],
+                block_size=block_size,
+                n=num,
+            )
+            cur_masks.append(single_mask)
+        masks.append(torch.block_diag(*cur_masks))
+    masks = torch.stack(masks, dim=0)
+    return masks
+
+
+def create_causal_mask_from_labels(token_labels: torch.LongTensor, block_size: int) -> torch.Tensor:
+    """
+    Build a causal mask from token_labels for token-label SFT.
+
+    token_labels shape: (batch_size, seq_len)
+      - 0: prompt
+      - 1..block_size: clean block labels (generation steps)
+      - block_size + 1: mask block labels
+      - -1: padding
+    """
+    if token_labels.dim() != 2:
+        raise ValueError(f"`token_labels` must be 2D, got shape {tuple(token_labels.shape)}.")
+
+    bsz, _ = token_labels.shape
+    device = token_labels.device
+
+    is_prompt = token_labels == 0
+    is_data = (token_labels > 0) & (token_labels <= block_size)
+    is_mask = token_labels == (block_size + 1)
+    is_pad = token_labels == -1
+
+    time_steps = token_labels.clone().float()
+    for b in range(bsz):
+        data_vals = time_steps[b, is_data[b]]
+        mask_indices = torch.nonzero(is_mask[b], as_tuple=True)[0]
+        if mask_indices.numel() == 0:
+            continue
+        if mask_indices.numel() == data_vals.numel():
+            time_steps[b, mask_indices] = data_vals
+        else:
+            min_len = min(mask_indices.numel(), data_vals.numel())
+            time_steps[b, mask_indices[:min_len]] = data_vals[:min_len]
+
+    time_steps[is_pad] = float("inf")
+
+    type_i = torch.zeros_like(token_labels)  # 1=data, 2=mask
+    type_i[is_data] = 1
+    type_i[is_mask] = 2
+    type_i = type_i.unsqueeze(1).unsqueeze(2)  # (B, 1, L, 1)
+    type_j = type_i.view(bsz, 1, 1, -1)  # (B, 1, 1, L)
+
+    time_i = time_steps.unsqueeze(1).unsqueeze(2)
+    time_j = time_steps.unsqueeze(1).unsqueeze(1)
+
+    is_prompt_j = is_prompt.view(bsz, 1, 1, -1)
+    is_pad_i = is_pad.view(bsz, 1, -1, 1)
+    is_pad_j = is_pad.view(bsz, 1, 1, -1)
+
+    mask_prompt = is_prompt_j
+    mask_data_data = (type_i == 1) & (type_j == 1) & (time_j <= time_i)
+    mask_data_mask = (type_i == 1) & (type_j == 2) & (time_j > time_i)
+    mask_mask_data = (type_i == 2) & (type_j == 1) & (time_j < time_i)
+    mask_mask_mask = (type_i == 2) & (type_j == 2) & (time_j >= time_i)
+    mask_prompt_internal = (token_labels.unsqueeze(1).unsqueeze(2) == 0) & is_prompt_j
+
+    final_mask = (
+        mask_prompt
+        | mask_data_data
+        | mask_data_mask
+        | mask_mask_data
+        | mask_mask_mask
+        | mask_prompt_internal
+    )
+    final_mask = final_mask & (~is_pad_i) & (~is_pad_j)
+    return final_mask.squeeze(1).to(dtype=torch.bool, device=device)
+
+
+def create_multi_block_causal_mask(
+    token_labels: torch.LongTensor,
+    block_ids: torch.LongTensor,
+    block_size: int,
+    block_causal_prompt: bool = True,
+) -> torch.Tensor:
+    """
+    Generate attention mask for multi-block causal mask training.
+
+    Args:
+        token_labels: (B, L) — 0=prompt, 1..block_size=data step, block_size+1=mask, -1=pad
+        block_ids: (B, L) — -1=prompt/pad, 0,1,2,...=block index
+        block_size: denoising steps per block
+        block_causal_prompt:
+            True:  prompt block-level causal (SDAR-style), prompt 不看 data/mask
+            False: 原版, prompt 看所有 prompt + mask, 所有人看所有 prompt
+
+    Returns:
+        attn_mask: (B, L, L) bool tensor (squeezed from (B,1,L,L)), True = visible
+    """
+    B, L = token_labels.shape
+    device = token_labels.device
+
+    is_prompt = (token_labels == 0)
+    is_data = (token_labels > 0) & (token_labels <= block_size)
+    is_mask = (token_labels == (block_size + 1))
+    is_pad = (token_labels == -1)
+
+    time_steps = token_labels.clone().float()
+    time_steps[is_pad] = float("inf")
+    time_steps[is_prompt] = 0
+
+    for b in range(B):
+        blk_vals = block_ids[b][block_ids[b] >= 0].unique()
+        for blk in blk_vals:
+            blk_mask = (block_ids[b] == blk)
+            data_in_blk = blk_mask & is_data[b]
+            mask_in_blk = blk_mask & is_mask[b]
+            data_steps = time_steps[b, data_in_blk]
+            mask_indices = torch.nonzero(mask_in_blk, as_tuple=True)[0]
+            n_data = data_steps.shape[0]
+            n_mask = mask_indices.shape[0]
+            if n_mask > 0 and n_data > 0:
+                min_len = min(n_data, n_mask)
+                time_steps[b, mask_indices[:min_len]] = data_steps[:min_len]
+
+    type_vals = torch.zeros_like(token_labels)
+    type_vals[is_data] = 1
+    type_vals[is_mask] = 2
+
+    type_i = type_vals[:, None, :, None]
+    type_j = type_vals[:, None, None, :]
+    time_i = time_steps[:, None, :, None]
+    time_j = time_steps[:, None, None, :]
+    blkid_i = block_ids[:, None, :, None].float()
+    blkid_j = block_ids[:, None, None, :].float()
+
+    is_prompt_i = is_prompt.view(B, 1, L, 1)
+    is_prompt_j = is_prompt.view(B, 1, 1, L)
+    is_pad_i = is_pad.view(B, 1, L, 1)
+    is_pad_j = is_pad.view(B, 1, 1, L)
+
+    if block_causal_prompt:
+        # 新版: prompt block-level causal (SDAR-style)
+        # prompt 按 block_size 分块, 每个 block 只看前面的 block, 不看 data/mask
+        prompt_cumpos = is_prompt.long().cumsum(dim=1) - 1
+        prompt_blk = prompt_cumpos // block_size
+        prompt_blk_i = prompt_blk[:, None, :, None]
+        prompt_blk_j = prompt_blk[:, None, None, :]
+        rule_prompt = is_prompt_i & is_prompt_j & (prompt_blk_j <= prompt_blk_i)
+        rule_see_prompt = (~is_prompt_i) & (~is_pad_i) & is_prompt_j
+    else:
+        # 原版: prompt 看所有 prompt + mask, 所有人看所有 prompt
+        rule_prompt = is_prompt_i & is_prompt_j
+        is_mask_j = is_mask.view(B, 1, 1, L)
+        rule_prompt = rule_prompt | (is_prompt_i & is_mask_j)
+        rule_see_prompt = is_prompt_j.expand(B, 1, L, L)
+
+    same_block = (blkid_i == blkid_j) & (blkid_i >= 0)
+    intra_dd = same_block & (type_i == 1) & (type_j == 1) & (time_j <= time_i)
+    intra_dm = same_block & (type_i == 1) & (type_j == 2) & (time_j > time_i)
+    intra_md = same_block & (type_i == 2) & (type_j == 1) & (time_j < time_i)
+    intra_mm = same_block & (type_i == 2) & (type_j == 2) & (time_j >= time_i)
+
+    cross_block_data = (blkid_i > blkid_j) & (blkid_j >= 0) & (type_j == 1)
+
+    final_mask = (
+        rule_prompt
+        | rule_see_prompt
+        | intra_dd | intra_dm | intra_md | intra_mm
+        | cross_block_data
+    )
+    final_mask = final_mask & (~is_pad_i) & (~is_pad_j)
+    return final_mask.squeeze(1).to(dtype=torch.bool, device=device)
+
+
+@torch.compile(fullgraph=True, mode="max-autotune-no-cudagraphs")
+def fused_flex_attention(query, key, value, attention_mask, **kwargs):
+    return flex_attention(query, key, value, block_mask=attention_mask, **kwargs)
+
+
+@use_kernel_forward_from_hub("RMSNorm")
+class SDARRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        SDARRMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        return flash_rms_norm(
+            hidden_states, weight=self.weight, bias=None, eps=self.variance_epsilon)
+        '''
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * \
+            torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+        '''
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+class SDARMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_proj = nn.Linear(
+            self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(
+            self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(
+            self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        if liger_kernel_is_available:
+            return self.down_proj(LigerSiLUMulFunction.apply(self.gate_proj(x), self.up_proj(x)))
+        else:
+            down_proj = self.down_proj(self.act_fn(
+                self.gate_proj(x)) * self.up_proj(x))
+            return down_proj
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2:]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(
+        batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs,
+):
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = nn.functional.softmax(
+        attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = nn.functional.dropout(
+        attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+class SDARAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: SDARConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.head_dim = getattr(
+            config, "head_dim", config.hidden_size // config.num_attention_heads)
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.is_causal = True
+
+        self.hidden_size = config.hidden_size
+        self.num_attention_heads = config.num_attention_heads
+        self.num_key_value_heads = config.num_key_value_heads
+
+        self.q_proj = nn.Linear(
+            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.k_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.v_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.o_proj = nn.Linear(
+            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
+        )
+        # unlike olmo, only on the head dim!
+        self.q_norm = SDARRMSNorm(self.head_dim, eps=config.rms_norm_eps)
+        # thus post q_norm does not need reshape
+        self.k_norm = SDARRMSNorm(self.head_dim, eps=config.rms_norm_eps)
+        self.sliding_window = config.sliding_window
+        if not (
+            self.config.use_sliding_window
+            and getattr(self.config, "sliding_window", None) is not None
+            and self.layer_idx >= self.config.max_window_layers
+        ):
+            self.sliding_window = None
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_value: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        input_shape = hidden_states.shape[:-1]
+        bsz, q_len = input_shape
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_norm(self.q_proj(
+            hidden_states).view(hidden_shape)).transpose(1, 2)
+        key_states = self.k_norm(self.k_proj(
+            hidden_states).view(hidden_shape)).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(
+            hidden_shape).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(
+            query_states, key_states, cos, sin)
+
+        if past_key_value is not None and kwargs.get("store_kv", False):
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            key_states, value_states = past_key_value.update(
+                key_states, value_states, self.layer_idx)
+        elif past_key_value is not None and not kwargs.get("store_kv", False) and len(past_key_value) > self.layer_idx:
+            # only retrive, do not store kv
+            past_key_states, past_value_states = past_key_value[self.layer_idx]
+            key_states = torch.cat(
+                [past_key_states, key_states], dim=-2)
+            value_states = torch.cat(
+                [past_value_states, value_states], dim=-2)
+
+        if self.training:
+            attn_output, attn_weights = fused_flex_attention(
+                query=query_states,
+                key=key_states,
+                value=value_states,
+                attention_mask=attention_mask,
+                enable_gqa=True,
+                scale=self.scaling,
+                return_lse=True
+            )
+            attn_weights = attn_weights.to(
+                value_states.dtype) if attn_weights is not None else None
+            attn_output = rearrange(attn_output, 'b h l d -> b l (h d)')
+        else:
+            attention_mask = attention_mask.bool() if attention_mask is not None else None
+            attn_weights = None
+            if torch.all(attention_mask):  # decoding
+                query_states = query_states.transpose(1, 2)
+                key_states = key_states.transpose(1, 2)
+                value_states = value_states.transpose(1, 2)
+                attn_output = flash_attn_func(
+                    query_states,
+                    key_states,
+                    value_states,
+                    causal=False,
+                    softmax_scale=self.scaling
+                )
+                attn_output = rearrange(attn_output, 'b l h d -> b l (h d)')
+            else:  # prefilling
+                attn_output = F.scaled_dot_product_attention(
+                    query=query_states,
+                    key=key_states,
+                    value=value_states,
+                    attn_mask=attention_mask,
+                    is_causal=False,
+                    scale=self.scaling,
+                    enable_gqa=True
+                )
+                attn_output = rearrange(attn_output, 'b h l d -> b l (h d)')
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights  # , attn_weights
+
+
+class SDARDecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: SDARConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = SDARAttention(config=config, layer_idx=layer_idx)
+        self.mlp = SDARMLP(config)
+        self.input_layernorm = SDARRMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = SDARRMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps)
+        if (
+            config.sliding_window and config._attn_implementation != "flash_attention_2"
+        ):  # diff with Llama is this warning
+            logger.warning_once(
+                f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
+                "unexpected results may be encountered."
+            )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        store_kv: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        # necessary, but kept here for BC
+        position_embeddings: Optional[Tuple[torch.Tensor,
+                                            torch.Tensor]] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            store_kv=store_kv,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        return outputs
+
+
+@auto_docstring
+class SDARPreTrainedModel(PreTrainedModel):
+    config_class = SDARConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["SDARDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _supports_cache_class = True
+    _supports_quantized_cache = True
+    _supports_static_cache = True
+    _supports_attention_backend = True
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, SDARRMSNorm):
+            module.weight.data.fill_(1.0)
+
+
+class SDARRotaryEmbedding(nn.Module):
+    def __init__(self, config: SDARConfig, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+            self.rope_type = config.rope_scaling.get(
+                "rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(
+            self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    @torch.no_grad()
+    # power user: used with advanced RoPE types (e.g. dynamic rope)
+    @dynamic_rope_update
+    def forward(self, x, position_ids, token_labels: Optional[torch.LongTensor] = None):
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(
+            position_ids.shape[0], -1, 1).to(x.device)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        device_type = x.device.type if isinstance(
+            x.device.type, str) and x.device.type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+            freqs = (inv_freq_expanded.float() @
+                     position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+
+        cos = cos.to(dtype=x.dtype)
+        sin = sin.to(dtype=x.dtype)
+
+        if token_labels is not None:
+            if token_labels.shape != position_ids.shape:
+                raise ValueError(
+                    f"`token_labels` shape {tuple(token_labels.shape)} must match `position_ids` shape {tuple(position_ids.shape)}."
+                )
+
+            clean_min_label = 1
+            clean_max_label = self.config.block_size
+            mask_label = self.config.block_size + 1
+            token_labels = token_labels.to(position_ids.device)
+            for batch_idx in range(token_labels.size(0)):
+                clean_indices = torch.nonzero(
+                    (token_labels[batch_idx] >= clean_min_label) & (token_labels[batch_idx] <= clean_max_label),
+                    as_tuple=True,
+                )[0]
+                mask_indices = torch.nonzero(token_labels[batch_idx] == mask_label, as_tuple=True)[0]
+                if mask_indices.numel() == 0:
+                    continue
+                if clean_indices.numel() != mask_indices.numel():
+                    raise ValueError(
+                        "The clean block and mask block must have equal lengths for RoPE frequency copy."
+                    )
+                cos[batch_idx, mask_indices] = cos[batch_idx, clean_indices]
+                sin[batch_idx, mask_indices] = sin[batch_idx, clean_indices]
+
+        return cos, sin
+
+
+@auto_docstring
+class SDARModel(SDARPreTrainedModel):
+    def __init__(self, config: SDARConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(
+            config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [SDARDecoderLayer(config, layer_idx)
+             for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = SDARRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = SDARRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        token_labels: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        store_kv: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
+    ) -> BaseModelOutputWithPast:
+        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
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError(
+                "You must specify exactly one of input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        # TODO (joao): remove this exception in v4.56 -- it exists for users that try to pass a legacy cache
+        if not isinstance(past_key_values, (type(None), Cache)):
+            raise ValueError(
+                "The `past_key_values` should be either a `Cache` object or `None`.")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length(
+            ) if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0).expand(inputs_embeds.shape[0], -1)
+
+        # causal_mask = self._update_causal_mask(
+        #     attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+        # )
+
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        # RoPE frequency copy: for single-block training (without block_ids),
+        # data and mask have different position_ids, so copy data's RoPE to mask.
+        # For multi-block training (with block_ids), position_ids are already shared
+        # between data and mask, so no copy is needed.
+        if token_labels is not None and not hasattr(self, '_skip_rope_copy'):
+            cos, sin = position_embeddings
+            block_size = self.config.block_size
+            clean_min_label, clean_max_label = 1, block_size
+            mask_label = block_size + 1
+            tl = token_labels.to(position_ids.device)
+            for b_idx in range(tl.size(0)):
+                clean_idx = torch.nonzero(
+                    (tl[b_idx] >= clean_min_label) & (tl[b_idx] <= clean_max_label), as_tuple=True
+                )[0]
+                mask_idx = torch.nonzero(tl[b_idx] == mask_label, as_tuple=True)[0]
+                if mask_idx.numel() > 0 and clean_idx.numel() == mask_idx.numel():
+                    cos[b_idx, mask_idx] = cos[b_idx, clean_idx]
+                    sin[b_idx, mask_idx] = sin[b_idx, clean_idx]
+            position_embeddings = (cos, sin)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            layer_outputs = decoder_layer(
+                hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_values,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+                store_kv=store_kv,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+                **flash_attn_kwargs,
+            )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values if use_cache else None,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+    def _update_causal_mask(
+        self,
+        attention_mask: Union[torch.Tensor, "BlockMask"],
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool = False,
+    ):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and past_key_values is not None:
+                is_padding_right = attention_mask[:, -
+                                                  1].sum().item() != input_tensor.size()[0]
+                if is_padding_right:
+                    raise ValueError(
+                        "You are attempting to perform batched generation with padding_side='right'"
+                        " this may lead to unexpected behaviour for Flash Attention version of Qwen3. Make sure to "
+                        " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
+                    )
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+        if self.config._attn_implementation == "flex_attention":
+            if isinstance(attention_mask, torch.Tensor):
+                seq_len_q, seq_len_kv = attention_mask.shape
+                assert seq_len_q == seq_len_kv, f"got {attention_mask.shape=}"
+                attention_mask = create_block_mask(
+                    # 2d bool tensor, shape: [2*seqlen, 2*seqlen]
+                    lambda b, h, q_idx, kv_idx: attention_mask[q_idx, kv_idx],
+                    B=None, H=None, Q_LEN=seq_len_q, KV_LEN=seq_len_kv,
+                )
+            else:
+                # Here we pass in flex mask computed externally
+                assert isinstance(attention_mask, BlockMask)
+            return attention_mask
+
+        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+        # to infer the attention mask.
+        past_seen_tokens = past_key_values.get_seq_length(
+        ) if past_key_values is not None else 0
+        using_static_cache = isinstance(past_key_values, StaticCache)
+        using_sliding_window_cache = isinstance(
+            past_key_values, SlidingWindowCache)
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if (
+            self.config._attn_implementation == "sdpa"
+            and not (using_static_cache or using_sliding_window_cache)
+            and not output_attentions
+        ):
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                sliding_window=self.config.sliding_window,
+                is_training=self.training,
+            ):
+                return None
+
+        dtype = input_tensor.dtype
+        min_dtype = torch.finfo(dtype).min
+        sequence_length = input_tensor.shape[1]
+        # SlidingWindowCache or StaticCache
+        if using_sliding_window_cache or using_static_cache:
+            target_length = past_key_values.get_max_cache_shape()
+        # DynamicCache or no cache
+        else:
+            target_length = (
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
+            )
+
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+            config=self.config,
+            past_key_values=past_key_values,
+        )
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type in ["cuda", "xpu", "npu"]
+            and not output_attentions
+        ):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            causal_mask = AttentionMaskConverter._unmask_unattended(
+                causal_mask, min_dtype)
+
+        return causal_mask
+
+    @staticmethod
+    def _prepare_4d_causal_attention_mask_with_cache_position(
+        attention_mask: torch.Tensor,
+        sequence_length: int,
+        target_length: int,
+        dtype: torch.dtype,
+        cache_position: torch.Tensor,
+        batch_size: int,
+        config: SDARConfig,
+        past_key_values: Cache,
+    ):
+        """
+        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+        Args:
+            attention_mask (`torch.Tensor`):
+                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+            sequence_length (`int`):
+                The sequence length being processed.
+            target_length (`int`):
+                The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+            dtype (`torch.dtype`):
+                The dtype to use for the 4D attention mask.
+            cache_position (`torch.Tensor`):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            batch_size (`torch.Tensor`):
+                Batch size.
+            config (`SDARConfig`):
+                The model's configuration class
+            past_key_values (`Cache`):
+                The cache class that is being used currently to generate
+        """
+        if attention_mask is not None and attention_mask.dim() == 4:
+            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+            causal_mask = attention_mask
+        else:
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = torch.full(
+                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=cache_position.device
+            )
+            diagonal_attend_mask = torch.arange(target_length, device=cache_position.device) > cache_position.reshape(
+                -1, 1
+            )
+            text_config = config.get_text_config()
+            if getattr(text_config, "use_sliding_window", True) and text_config.sliding_window is not None:
+                # if we have sliding window, we should not attend to tokens beyond sliding window length, so we mask them out also
+                # the check is needed to verify is current checkpoint was trained with sliding window or not
+                if not isinstance(past_key_values, SlidingWindowCache) or sequence_length > target_length:
+                    sliding_attend_mask = torch.arange(target_length, device=cache_position.device) <= (
+                        cache_position.reshape(-1, 1) -
+                        text_config.sliding_window
+                    )
+                    diagonal_attend_mask.bitwise_or_(sliding_attend_mask)
+            causal_mask *= diagonal_attend_mask
+            causal_mask = causal_mask[None, None,
+                                      :, :].expand(batch_size, 1, -1, -1)
+            if attention_mask is not None:
+                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+                if attention_mask.shape[-1] > target_length:
+                    attention_mask = attention_mask[:, :target_length]
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
+                    causal_mask.device
+                )
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                    padding_mask, min_dtype
+                )
+        return causal_mask
+
+
+class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs):
+    ...
+
+
+@auto_docstring
+class SDARForCausalLM(SDARPreTrainedModel, GenerationMixin):
+    _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 = SDARModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(
+            config.hidden_size, config.vocab_size, bias=False)
+
+        # 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 prepare_for_bd_training(self, inputs_ids, position_ids, prompt_mask):
+        bsz, seq_len = inputs_ids.shape
+        num_tokens = calculate_token_nums(position_ids) # List[torch.Tensor]
+        noisy_inputs_ids, logits_to_keep_half, p_mask = forward_add_noise_packed(
+            inputs_ids=inputs_ids,
+            num_tokens_list=num_tokens,
+            prompt_mask=prompt_mask,
+            mask_id=self.config.mask_token_id,
+        )
+        router_noisy_part_list = []
+        for i in range(bsz):
+            cur_router_noisy_part = (torch.arange(num_tokens[i].shape[0] *2) % 2 == 0).to(inputs_ids.device)
+            cur_router_noisy_part = cur_router_noisy_part.repeat_interleave(num_tokens[i].repeat_interleave(2))
+            router_noisy_part_list.append(cur_router_noisy_part)
+        router_noisy_part = torch.stack(router_noisy_part_list, dim=0)
+
+        # concated inputs_ids: (bzs, seq_len x 2)
+        concat_inputs_ids = inputs_ids.repeat(1, 2)
+        # concated logits_to_keep: (bsz, seq_len x 2)
+        logits_to_keep = torch.zeros(
+                    bsz, 2 * seq_len, dtype=torch.bool, device=inputs_ids.device)
+        # concated position_ids: (bsz, seq_len x 2)
+        concat_position_ids = torch.zeros(
+                    bsz, 2 * seq_len, dtype=position_ids.dtype, device=position_ids.device)
+        for i in range(bsz):
+            concat_inputs_ids[i][router_noisy_part[i]] = noisy_inputs_ids[i]
+            concat_inputs_ids[i][~router_noisy_part[i]] = inputs_ids[i]
+
+            logits_to_keep[i][router_noisy_part[i]] = logits_to_keep_half[i]
+
+            concat_position_ids[i][router_noisy_part[i]] = position_ids[i]
+            concat_position_ids[i][~router_noisy_part[i]] = position_ids[i]
+
+        # create flex_attention mask
+        attention_mask = block_attn_mask(num_tokens, self.config.block_size, inputs_ids.device)
+        flex_attention_mask_3d = create_block_mask(
+                            lambda b, h, q_idx, kv_idx: attention_mask[b, q_idx, kv_idx],
+                            B=attention_mask.size(0), H=None,
+                            Q_LEN=attention_mask.size(1), KV_LEN=attention_mask.size(2),
+        )
+
+        return concat_inputs_ids, concat_position_ids, flex_attention_mask_3d, logits_to_keep_half, logits_to_keep, p_mask
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        token_labels: Optional[torch.LongTensor] = None,
+        block_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],
+    ) -> CausalLMOutputWithPast:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, SDARForCausalLM
+
+        >>> model = SDARForCausalLM.from_pretrained("DiffuOpen/SDAR-1.7B-Chat")
+        >>> tokenizer = AutoTokenizer.from_pretrained("DiffuOpen/SDAR-1.7B-Chat")
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+        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 self.training:
+            assert inputs_embeds is None, "only support input_ids during training"
+            assert labels is not None, "Labels must be provided for training."
+            # Trace SFT path: pre-computed block attention mask provided via kwargs
+            block_attention_mask = kwargs.pop("block_attention_mask", None)
+            if block_attention_mask is not None:
+                # block_attention_mask: (B, L, L) boolean tensor
+                flex_attention_mask_3d = create_block_mask(
+                    lambda b, h, q_idx, kv_idx: block_attention_mask[b, q_idx, kv_idx],
+                    B=block_attention_mask.size(0),
+                    H=None,
+                    Q_LEN=block_attention_mask.size(1),
+                    KV_LEN=block_attention_mask.size(2),
+                )
+                outputs = self.model(
+                    input_ids=input_ids,
+                    attention_mask=flex_attention_mask_3d,
+                    position_ids=position_ids,
+                    output_attentions=output_attentions,
+                    output_hidden_states=output_hidden_states,
+                    return_dict=True,
+                    cache_position=cache_position,
+                )
+                hidden_states = outputs.last_hidden_state
+                logits = self.lm_head(hidden_states)
+                # Unshifted cross-entropy loss (diffusion-style)
+                loss = nn.CrossEntropyLoss(ignore_index=-100)(
+                    logits.view(-1, self.config.vocab_size), labels.view(-1)
+                )
+                logits = None
+            elif token_labels is not None:
+                if input_ids is None:
+                    raise ValueError("`input_ids` is required in token-label SFT training.")
+                if token_labels.shape != input_ids.shape:
+                    raise ValueError(
+                        f"`token_labels` shape {tuple(token_labels.shape)} must match `input_ids` shape {tuple(input_ids.shape)}."
+                    )
+                # Multi-block mask when block_ids provided, else single-block
+                if block_ids is not None:
+                    bcp = getattr(self.config, "block_causal_prompt", True)
+                    token_label_mask = create_multi_block_causal_mask(
+                        token_labels, block_ids, self.config.block_size,
+                        block_causal_prompt=bcp,
+                    )
+                else:
+                    token_label_mask = create_causal_mask_from_labels(token_labels, self.config.block_size)
+                flex_attention_mask_3d = create_block_mask(
+                    lambda b, h, q_idx, kv_idx: token_label_mask[b, q_idx, kv_idx],
+                    B=token_label_mask.size(0),
+                    H=None,
+                    Q_LEN=token_label_mask.size(1),
+                    KV_LEN=token_label_mask.size(2),
+                )
+                outputs = self.model(
+                    input_ids=input_ids,
+                    attention_mask=flex_attention_mask_3d,
+                    position_ids=position_ids,
+                    token_labels=token_labels,
+                    output_attentions=output_attentions,
+                    output_hidden_states=output_hidden_states,
+                    return_dict=True,
+                    cache_position=cache_position,
+                    **kwargs,
+                )
+                hidden_states = outputs.last_hidden_state
+                logits = self.lm_head(hidden_states)
+                masked_labels = labels.masked_fill(token_labels != (self.config.block_size + 1), -100)
+                if not torch.any(masked_labels != -100):
+                    raise ValueError("No valid supervision token found for token-label SFT loss.")
+                loss = nn.CrossEntropyLoss(ignore_index=-100)(
+                    logits.view(-1, self.config.vocab_size), masked_labels.view(-1)
+                )
+                logits = None
+            else:
+                prompt_mask = labels == -100
+                position_ids = modify_padded_position_ids_2d(position_ids)
+                concat_inputs_ids, concat_position_ids, flex_attention_mask_3d, logits_to_keep_half, logits_to_keep, p_mask = self.prepare_for_bd_training(input_ids, position_ids, prompt_mask)
+                outputs = self.model(
+                    input_ids=concat_inputs_ids,
+                    attention_mask=flex_attention_mask_3d,
+                    position_ids=concat_position_ids,
+                    output_attentions=output_attentions,
+                    output_hidden_states=output_hidden_states,
+                    return_dict=True,
+                    cache_position=cache_position,
+                    **kwargs,
+                )
+                hidden_states = outputs.last_hidden_state
+                hidden_states = hidden_states[logits_to_keep].contiguous()
+                answer_len = (labels != -100).sum()
+                loss_fct = FusedLinearDiffusionCrossEntropyLoss(reduction='sum')
+                loss = loss_fct(  # it will return (sum_loss, unreduced_loss)
+                        # conduct `view(-1, V)` inside the function
+                        x=hidden_states,
+                        target=labels[logits_to_keep_half].contiguous(),
+                        weight=self.lm_head.weight,
+                        bias=self.lm_head.bias,
+                        p_mask=p_mask,
+                    )
+                loss = loss / answer_len
+                logits = None
+        else:
+            # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+            outputs: BaseModelOutputWithPast = self.model(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                token_labels=token_labels,
+                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
+            hidden_states = hidden_states[:, slice_indices, :].contiguous()
+            fuse_linear_and_cross_entropy = self.config.fuse_cross_entropy and self.training
+            if fuse_linear_and_cross_entropy:
+                # When using fused_linear_ce_loss, we do not compute the whole logits on HBM
+                logits = None
+            else:
+                logits = self.lm_head(hidden_states)
+
+            loss = None
+            if labels is not None:
+                # FusedLinearCrossEntropyLoss will be implemented by monkey patch when training
+                # We don't use it when inferencing
+                loss_fct = nn.CrossEntropyLoss()  # nn.CE
+                loss = loss_fct(
+                    logits.view(-1, self.config.vocab_size), labels.view(-1))
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+__all__ = [
+    "SDARForCausalLM",
+    "SDARModel",
+    "SDARPreTrainedModel",
+]

special_tokens_map.json

@@ -0,0 +1,39 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>",
+    "<|MASK|>"
+  ],
+  "eos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "mask_token": {
+    "content": "<|MASK|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenization_qwen2.py

@@ -0,0 +1,342 @@
+# 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.
+"""Tokenization classes for Qwen2."""
+
+import json
+import os
+import unicodedata
+from functools import lru_cache
+from typing import Optional, Tuple
+
+import regex as re
+
+from transformers.tokenization_utils import AddedToken, PreTrainedTokenizer
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {
+    "vocab_file": "vocab.json",
+    "merges_file": "merges.txt",
+}
+
+
+MAX_MODEL_INPUT_SIZES = {"qwen/qwen-tokenizer": 32768}
+
+PRETOKENIZE_REGEX = r"""(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+"""
+
+
+@lru_cache()
+# Copied from transformers.models.gpt2.tokenization_gpt2.bytes_to_unicode
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a mapping to unicode strings. We specifically avoids mapping to whitespace/control
+    characters the bpe code barfs on.
+
+    The reversible bpe codes work on unicode strings. This means you need a large # of unicode characters in your vocab
+    if you want to avoid UNKs. When you're at something like a 10B token dataset you end up needing around 5K for
+    decent coverage. This is a significant percentage of your normal, say, 32K bpe vocab. To avoid that, we want lookup
+    tables between utf-8 bytes and unicode strings.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1)) + list(range(ord("¡"), ord("¬") + 1)) + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+# Copied from transformers.models.gpt2.tokenization_gpt2.get_pairs
+def get_pairs(word):
+    """
+    Return set of symbol pairs in a word.
+
+    Word is represented as tuple of symbols (symbols being variable-length strings).
+    """
+    pairs = set()
+    prev_char = word[0]
+    for char in word[1:]:
+        pairs.add((prev_char, char))
+        prev_char = char
+    return pairs
+
+
+class Qwen2Tokenizer(PreTrainedTokenizer):
+    """
+    Construct a Qwen2 tokenizer. Based on byte-level Byte-Pair-Encoding.
+
+    Same with GPT2Tokenizer, this tokenizer has been trained to treat spaces like parts of the tokens so a word will
+    be encoded differently whether it is at the beginning of the sentence (without space) or not:
+
+    ```python
+    >>> from transformers import Qwen2Tokenizer
+
+    >>> tokenizer = Qwen2Tokenizer.from_pretrained("Qwen/Qwen-tokenizer")
+    >>> tokenizer("Hello world")["input_ids"]
+    [9707, 1879]
+
+    >>> tokenizer(" Hello world")["input_ids"]
+    [21927, 1879]
+    ```
+    This is expected.
+
+    You should not use GPT2Tokenizer instead, because of the different pretokenization rules.
+
+    This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
+    this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (`str`):
+            Path to the vocabulary file.
+        merges_file (`str`):
+            Path to the merges file.
+        errors (`str`, *optional*, defaults to `"replace"`):
+            Paradigm to follow when decoding bytes to UTF-8. See
+            [bytes.decode](https://docs.python.org/3/library/stdtypes.html#bytes.decode) for more information.
+        unk_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        bos_token (`str`, *optional*):
+            The beginning of sequence token. Not applicable for this tokenizer.
+        eos_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The end of sequence token.
+        pad_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `False`):
+            Whether or not the model should cleanup the spaces that were added when splitting the input text during the
+            tokenization process. Not applicable to this tokenizer, since tokenization does not add spaces.
+        split_special_tokens (`bool`, *optional*, defaults to `False`):
+            Whether or not the special tokens should be split during the tokenization process. The default behavior is
+            to not split special tokens. This means that if `<|endoftext|>` is the `eos_token`, then `tokenizer.tokenize("<|endoftext|>") =
+            ['<|endoftext|>`]. Otherwise, if `split_special_tokens=True`, then `tokenizer.tokenize("<|endoftext|>")` will be give `['<',
+            '|', 'endo', 'ft', 'ext', '|', '>']`. This argument is only supported for `slow` tokenizers for the moment.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file,
+        merges_file,
+        errors="replace",
+        unk_token="<|endoftext|>",
+        bos_token=None,
+        eos_token="<|endoftext|>",
+        pad_token="<|endoftext|>",
+        clean_up_tokenization_spaces=False,
+        split_special_tokens=False,
+        **kwargs,
+    ):
+        # Qwen vocab does not contain control tokens; added tokens need to be special
+        bos_token = (
+            AddedToken(bos_token, lstrip=False, rstrip=False, special=True, normalized=False)
+            if isinstance(bos_token, str)
+            else bos_token
+        )
+        eos_token = (
+            AddedToken(eos_token, lstrip=False, rstrip=False, special=True, normalized=False)
+            if isinstance(eos_token, str)
+            else eos_token
+        )
+        unk_token = (
+            AddedToken(unk_token, lstrip=False, rstrip=False, special=True, normalized=False)
+            if isinstance(unk_token, str)
+            else unk_token
+        )
+        pad_token = (
+            AddedToken(pad_token, lstrip=False, rstrip=False, special=True, normalized=False)
+            if isinstance(pad_token, str)
+            else pad_token
+        )
+
+        with open(vocab_file, encoding="utf-8") as vocab_handle:
+            self.encoder = json.load(vocab_handle)
+        self.decoder = {v: k for k, v in self.encoder.items()}
+        self.errors = errors  # how to handle errors in decoding
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+        bpe_merges = []
+        with open(merges_file, encoding="utf-8") as merges_handle:
+            for i, line in enumerate(merges_handle):
+                line = line.strip()
+                if (i == 0 and line.startswith("#version:")) or not line:
+                    continue
+                bpe_merges.append(tuple(line.split()))
+        self.bpe_ranks = dict(zip(bpe_merges, range(len(bpe_merges))))
+        # NOTE: the cache can grow without bound and will get really large for long running processes
+        # (esp. for texts of language that do not use space between word, e.g. Chinese); technically
+        # not a memory leak but appears as one.
+        # GPT2Tokenizer has the same problem, so let's be consistent.
+        self.cache = {}
+
+        self.pat = re.compile(PRETOKENIZE_REGEX)
+
+        if kwargs.get("add_prefix_space", False):
+            logger.warning_once(
+                f"{self.__class__.__name} does not support `add_prefix_space`, setting it to True has no effect."
+            )
+
+        super().__init__(
+            errors=errors,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            pad_token=pad_token,
+            unk_token=unk_token,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            split_special_tokens=split_special_tokens,
+            **kwargs,
+        )
+
+    @property
+    def vocab_size(self) -> int:
+        return len(self.encoder)
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer.get_vocab
+    def get_vocab(self):
+        return dict(self.encoder, **self.added_tokens_encoder)
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer.bpe
+    def bpe(self, token):
+        if token in self.cache:
+            return self.cache[token]
+        word = tuple(token)
+        pairs = get_pairs(word)
+
+        if not pairs:
+            return token
+
+        while True:
+            bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
+            if bigram not in self.bpe_ranks:
+                break
+            first, second = bigram
+            new_word = []
+            i = 0
+            while i < len(word):
+                try:
+                    j = word.index(first, i)
+                except ValueError:
+                    new_word.extend(word[i:])
+                    break
+                else:
+                    new_word.extend(word[i:j])
+                    i = j
+
+                if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
+                    new_word.append(first + second)
+                    i += 2
+                else:
+                    new_word.append(word[i])
+                    i += 1
+            new_word = tuple(new_word)
+            word = new_word
+            if len(word) == 1:
+                break
+            else:
+                pairs = get_pairs(word)
+        word = " ".join(word)
+        self.cache[token] = word
+        return word
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer._tokenize
+    def _tokenize(self, text):
+        """Tokenize a string."""
+        bpe_tokens = []
+        for token in re.findall(self.pat, text):
+            token = "".join(
+                self.byte_encoder[b] for b in token.encode("utf-8")
+            )  # Maps all our bytes to unicode strings, avoiding control tokens of the BPE (spaces in our case)
+            bpe_tokens.extend(bpe_token for bpe_token in self.bpe(token).split(" "))
+        return bpe_tokens
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer._convert_token_to_id
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        return self.encoder.get(token, self.encoder.get(self.unk_token))
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer._convert_id_to_token
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        return self.decoder.get(index)
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer.convert_tokens_to_string
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        text = "".join(tokens)
+        text = bytearray([self.byte_decoder[c] for c in text]).decode("utf-8", errors=self.errors)
+        return text
+
+    def decode(
+        self,
+        token_ids,
+        skip_special_tokens: bool = False,
+        clean_up_tokenization_spaces: Optional[bool] = False,
+        spaces_between_special_tokens: bool = False,
+        **kwargs,
+    ) -> str:
+        # `spaces_between_special_tokens` defaults to True for _decode in slow tokenizers
+        # and cannot be configured elsewhere, but it should default to False for Qwen2Tokenizer
+        return super().decode(
+            token_ids,
+            skip_special_tokens=skip_special_tokens,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            spaces_between_special_tokens=spaces_between_special_tokens,
+            **kwargs,
+        )
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer.save_vocabulary
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+        merge_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["merges_file"]
+        )
+
+        with open(vocab_file, "w", encoding="utf-8") as f:
+            f.write(json.dumps(self.encoder, indent=2, sort_keys=True, ensure_ascii=False) + "\n")
+
+        index = 0
+        with open(merge_file, "w", encoding="utf-8") as writer:
+            writer.write("#version: 0.2\n")
+            for bpe_tokens, token_index in sorted(self.bpe_ranks.items(), key=lambda kv: kv[1]):
+                if index != token_index:
+                    logger.warning(
+                        f"Saving vocabulary to {merge_file}: BPE merge indices are not consecutive."
+                        " Please check that the tokenizer is not corrupted!"
+                    )
+                    index = token_index
+                writer.write(" ".join(bpe_tokens) + "\n")
+                index += 1
+
+        return vocab_file, merge_file
+
+    def prepare_for_tokenization(self, text, **kwargs):
+        text = unicodedata.normalize("NFC", text)
+        return (text, kwargs)
+
+
+__all__ = ["Qwen2Tokenizer"]

tokenizer_config.json

@@ -0,0 +1,256 @@
+{
+  "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": "<|MASK|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>",
+    "<|MASK|>"
+  ],
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenization_qwen2.Qwen2Tokenizer",
+      null
+    ]
+  },
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "mask_token": "<|MASK|>",
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

train_results.json

@@ -0,0 +1,9 @@
+{
+    "effective_tokens_per_sec": 2340.6354596146743,
+    "epoch": 3.0,
+    "total_flos": 2.73366675523371e+17,
+    "train_loss": 0.053853037605258475,
+    "train_runtime": 459.8341,
+    "train_samples_per_second": 48.755,
+    "train_steps_per_second": 0.763
+}

trainer_log.jsonl

@@ -0,0 +1,71 @@
+{"current_steps": 5, "total_steps": 351, "loss": 0.2434, "lr": 3.6363636363636366e-06, "epoch": 0.042735042735042736, "percentage": 1.42, "elapsed_time": "0:00:30", "remaining_time": "0:35:17"}
+{"current_steps": 10, "total_steps": 351, "loss": 0.1924, "lr": 8.181818181818183e-06, "epoch": 0.08547008547008547, "percentage": 2.85, "elapsed_time": "0:00:34", "remaining_time": "0:19:40"}
+{"current_steps": 15, "total_steps": 351, "loss": 0.174, "lr": 9.998079135987437e-06, "epoch": 0.1282051282051282, "percentage": 4.27, "elapsed_time": "0:00:38", "remaining_time": "0:14:15"}
+{"current_steps": 20, "total_steps": 351, "loss": 0.1488, "lr": 9.98634586692894e-06, "epoch": 0.17094017094017094, "percentage": 5.7, "elapsed_time": "0:00:41", "remaining_time": "0:11:28"}
+{"current_steps": 25, "total_steps": 351, "loss": 0.1332, "lr": 9.963971484502247e-06, "epoch": 0.21367521367521367, "percentage": 7.12, "elapsed_time": "0:00:45", "remaining_time": "0:09:47"}
+{"current_steps": 30, "total_steps": 351, "loss": 0.1146, "lr": 9.931003736767013e-06, "epoch": 0.2564102564102564, "percentage": 8.55, "elapsed_time": "0:00:48", "remaining_time": "0:08:40"}
+{"current_steps": 35, "total_steps": 351, "loss": 0.1175, "lr": 9.887512978558329e-06, "epoch": 0.29914529914529914, "percentage": 9.97, "elapsed_time": "0:00:52", "remaining_time": "0:07:51"}
+{"current_steps": 40, "total_steps": 351, "loss": 0.1114, "lr": 9.833592021345938e-06, "epoch": 0.3418803418803419, "percentage": 11.4, "elapsed_time": "0:00:55", "remaining_time": "0:07:11"}
+{"current_steps": 45, "total_steps": 351, "loss": 0.1092, "lr": 9.76935593516989e-06, "epoch": 0.38461538461538464, "percentage": 12.82, "elapsed_time": "0:00:58", "remaining_time": "0:06:40"}
+{"current_steps": 50, "total_steps": 351, "loss": 0.1005, "lr": 9.694941803075285e-06, "epoch": 0.42735042735042733, "percentage": 14.25, "elapsed_time": "0:01:02", "remaining_time": "0:06:15"}
+{"current_steps": 55, "total_steps": 351, "loss": 0.1079, "lr": 9.610508428570122e-06, "epoch": 0.4700854700854701, "percentage": 15.67, "elapsed_time": "0:01:06", "remaining_time": "0:05:55"}
+{"current_steps": 60, "total_steps": 351, "loss": 0.1003, "lr": 9.516235996730645e-06, "epoch": 0.5128205128205128, "percentage": 17.09, "elapsed_time": "0:01:09", "remaining_time": "0:05:36"}
+{"current_steps": 65, "total_steps": 351, "loss": 0.1005, "lr": 9.41232568967728e-06, "epoch": 0.5555555555555556, "percentage": 18.52, "elapsed_time": "0:01:12", "remaining_time": "0:05:20"}
+{"current_steps": 70, "total_steps": 351, "loss": 0.1029, "lr": 9.298999257241862e-06, "epoch": 0.5982905982905983, "percentage": 19.94, "elapsed_time": "0:01:16", "remaining_time": "0:05:06"}
+{"current_steps": 75, "total_steps": 351, "loss": 0.0988, "lr": 9.176498543742328e-06, "epoch": 0.6410256410256411, "percentage": 21.37, "elapsed_time": "0:01:19", "remaining_time": "0:04:54"}
+{"current_steps": 80, "total_steps": 351, "loss": 0.0936, "lr": 9.045084971874738e-06, "epoch": 0.6837606837606838, "percentage": 22.79, "elapsed_time": "0:01:23", "remaining_time": "0:04:42"}
+{"current_steps": 85, "total_steps": 351, "loss": 0.1036, "lr": 8.905038984824079e-06, "epoch": 0.7264957264957265, "percentage": 24.22, "elapsed_time": "0:01:26", "remaining_time": "0:04:30"}
+{"current_steps": 90, "total_steps": 351, "loss": 0.1, "lr": 8.756659447784367e-06, "epoch": 0.7692307692307693, "percentage": 25.64, "elapsed_time": "0:01:29", "remaining_time": "0:04:20"}
+{"current_steps": 95, "total_steps": 351, "loss": 0.0956, "lr": 8.600263010165275e-06, "epoch": 0.811965811965812, "percentage": 27.07, "elapsed_time": "0:01:33", "remaining_time": "0:04:11"}
+{"current_steps": 100, "total_steps": 351, "loss": 0.1028, "lr": 8.436183429846314e-06, "epoch": 0.8547008547008547, "percentage": 28.49, "elapsed_time": "0:01:36", "remaining_time": "0:04:03"}
+{"current_steps": 105, "total_steps": 351, "loss": 0.1077, "lr": 8.264770860920722e-06, "epoch": 0.8974358974358975, "percentage": 29.91, "elapsed_time": "0:01:40", "remaining_time": "0:03:54"}
+{"current_steps": 110, "total_steps": 351, "loss": 0.0976, "lr": 8.086391106448965e-06, "epoch": 0.9401709401709402, "percentage": 31.34, "elapsed_time": "0:01:43", "remaining_time": "0:03:46"}
+{"current_steps": 115, "total_steps": 351, "loss": 0.1008, "lr": 7.90142483781658e-06, "epoch": 0.9829059829059829, "percentage": 32.76, "elapsed_time": "0:01:46", "remaining_time": "0:03:39"}
+{"current_steps": 120, "total_steps": 351, "loss": 0.0635, "lr": 7.710266782362248e-06, "epoch": 1.0256410256410255, "percentage": 34.19, "elapsed_time": "0:02:53", "remaining_time": "0:05:34"}
+{"current_steps": 125, "total_steps": 351, "loss": 0.0403, "lr": 7.513324881009769e-06, "epoch": 1.0683760683760684, "percentage": 35.61, "elapsed_time": "0:02:56", "remaining_time": "0:05:19"}
+{"current_steps": 130, "total_steps": 351, "loss": 0.036, "lr": 7.311019417701567e-06, "epoch": 1.1111111111111112, "percentage": 37.04, "elapsed_time": "0:03:00", "remaining_time": "0:05:06"}
+{"current_steps": 135, "total_steps": 351, "loss": 0.0361, "lr": 7.103782122491577e-06, "epoch": 1.1538461538461537, "percentage": 38.46, "elapsed_time": "0:03:03", "remaining_time": "0:04:54"}
+{"current_steps": 140, "total_steps": 351, "loss": 0.0343, "lr": 6.892055250211552e-06, "epoch": 1.1965811965811965, "percentage": 39.89, "elapsed_time": "0:03:07", "remaining_time": "0:04:42"}
+{"current_steps": 145, "total_steps": 351, "loss": 0.0353, "lr": 6.67629063667697e-06, "epoch": 1.2393162393162394, "percentage": 41.31, "elapsed_time": "0:03:10", "remaining_time": "0:04:30"}
+{"current_steps": 150, "total_steps": 351, "loss": 0.0356, "lr": 6.456948734446624e-06, "epoch": 1.282051282051282, "percentage": 42.74, "elapsed_time": "0:03:13", "remaining_time": "0:04:19"}
+{"current_steps": 155, "total_steps": 351, "loss": 0.0373, "lr": 6.234497630193666e-06, "epoch": 1.3247863247863247, "percentage": 44.16, "elapsed_time": "0:03:17", "remaining_time": "0:04:09"}
+{"current_steps": 160, "total_steps": 351, "loss": 0.0356, "lr": 6.009412045785051e-06, "epoch": 1.3675213675213675, "percentage": 45.58, "elapsed_time": "0:03:20", "remaining_time": "0:03:59"}
+{"current_steps": 165, "total_steps": 351, "loss": 0.0348, "lr": 5.782172325201155e-06, "epoch": 1.4102564102564101, "percentage": 47.01, "elapsed_time": "0:03:24", "remaining_time": "0:03:50"}
+{"current_steps": 170, "total_steps": 351, "loss": 0.0322, "lr": 5.553263409457504e-06, "epoch": 1.452991452991453, "percentage": 48.43, "elapsed_time": "0:03:27", "remaining_time": "0:03:41"}
+{"current_steps": 175, "total_steps": 351, "loss": 0.036, "lr": 5.323173801716222e-06, "epoch": 1.4957264957264957, "percentage": 49.86, "elapsed_time": "0:03:30", "remaining_time": "0:03:32"}
+{"current_steps": 180, "total_steps": 351, "loss": 0.0316, "lr": 5.09239452479565e-06, "epoch": 1.5384615384615383, "percentage": 51.28, "elapsed_time": "0:03:34", "remaining_time": "0:03:23"}
+{"current_steps": 185, "total_steps": 351, "loss": 0.0326, "lr": 4.861418073302919e-06, "epoch": 1.5811965811965814, "percentage": 52.71, "elapsed_time": "0:03:37", "remaining_time": "0:03:15"}
+{"current_steps": 190, "total_steps": 351, "loss": 0.0333, "lr": 4.630737362625631e-06, "epoch": 1.623931623931624, "percentage": 54.13, "elapsed_time": "0:03:41", "remaining_time": "0:03:07"}
+{"current_steps": 195, "total_steps": 351, "loss": 0.0361, "lr": 4.400844677025585e-06, "epoch": 1.6666666666666665, "percentage": 55.56, "elapsed_time": "0:03:44", "remaining_time": "0:02:59"}
+{"current_steps": 200, "total_steps": 351, "loss": 0.0356, "lr": 4.17223061907935e-06, "epoch": 1.7094017094017095, "percentage": 56.98, "elapsed_time": "0:03:48", "remaining_time": "0:02:52"}
+{"current_steps": 205, "total_steps": 351, "loss": 0.033, "lr": 3.945383062707652e-06, "epoch": 1.7521367521367521, "percentage": 58.4, "elapsed_time": "0:03:51", "remaining_time": "0:02:44"}
+{"current_steps": 210, "total_steps": 351, "loss": 0.0306, "lr": 3.720786112027822e-06, "epoch": 1.7948717948717947, "percentage": 59.83, "elapsed_time": "0:03:54", "remaining_time": "0:02:37"}
+{"current_steps": 215, "total_steps": 351, "loss": 0.0332, "lr": 3.498919068251237e-06, "epoch": 1.8376068376068377, "percentage": 61.25, "elapsed_time": "0:03:58", "remaining_time": "0:02:30"}
+{"current_steps": 220, "total_steps": 351, "loss": 0.0321, "lr": 3.2802554068303595e-06, "epoch": 1.8803418803418803, "percentage": 62.68, "elapsed_time": "0:04:01", "remaining_time": "0:02:23"}
+{"current_steps": 225, "total_steps": 351, "loss": 0.0321, "lr": 3.0652617670382745e-06, "epoch": 1.9230769230769231, "percentage": 64.1, "elapsed_time": "0:04:05", "remaining_time": "0:02:17"}
+{"current_steps": 230, "total_steps": 351, "loss": 0.0299, "lr": 2.8543969561369556e-06, "epoch": 1.965811965811966, "percentage": 65.53, "elapsed_time": "0:04:08", "remaining_time": "0:02:10"}
+{"current_steps": 235, "total_steps": 351, "loss": 0.0271, "lr": 2.648110970259454e-06, "epoch": 2.0085470085470085, "percentage": 66.95, "elapsed_time": "0:05:14", "remaining_time": "0:02:35"}
+{"current_steps": 240, "total_steps": 351, "loss": 0.01, "lr": 2.4468440340954664e-06, "epoch": 2.051282051282051, "percentage": 68.38, "elapsed_time": "0:05:17", "remaining_time": "0:02:26"}
+{"current_steps": 245, "total_steps": 351, "loss": 0.0091, "lr": 2.2510256614296638e-06, "epoch": 2.094017094017094, "percentage": 69.8, "elapsed_time": "0:05:21", "remaining_time": "0:02:18"}
+{"current_steps": 250, "total_steps": 351, "loss": 0.0084, "lr": 2.061073738537635e-06, "epoch": 2.1367521367521367, "percentage": 71.23, "elapsed_time": "0:05:24", "remaining_time": "0:02:11"}
+{"current_steps": 255, "total_steps": 351, "loss": 0.0093, "lr": 1.8773936323955055e-06, "epoch": 2.1794871794871793, "percentage": 72.65, "elapsed_time": "0:05:27", "remaining_time": "0:02:03"}
+{"current_steps": 260, "total_steps": 351, "loss": 0.008, "lr": 1.7003773256063882e-06, "epoch": 2.2222222222222223, "percentage": 74.07, "elapsed_time": "0:05:31", "remaining_time": "0:01:55"}
+{"current_steps": 265, "total_steps": 351, "loss": 0.0075, "lr": 1.5304025798897521e-06, "epoch": 2.264957264957265, "percentage": 75.5, "elapsed_time": "0:05:34", "remaining_time": "0:01:48"}
+{"current_steps": 270, "total_steps": 351, "loss": 0.0069, "lr": 1.3678321299188802e-06, "epoch": 2.3076923076923075, "percentage": 76.92, "elapsed_time": "0:05:38", "remaining_time": "0:01:41"}
+{"current_steps": 275, "total_steps": 351, "loss": 0.0076, "lr": 1.213012909226786e-06, "epoch": 2.3504273504273505, "percentage": 78.35, "elapsed_time": "0:05:41", "remaining_time": "0:01:34"}
+{"current_steps": 280, "total_steps": 351, "loss": 0.0066, "lr": 1.066275309832584e-06, "epoch": 2.393162393162393, "percentage": 79.77, "elapsed_time": "0:05:44", "remaining_time": "0:01:27"}
+{"current_steps": 285, "total_steps": 351, "loss": 0.007, "lr": 9.279324771682586e-07, "epoch": 2.435897435897436, "percentage": 81.2, "elapsed_time": "0:05:48", "remaining_time": "0:01:20"}
+{"current_steps": 290, "total_steps": 351, "loss": 0.0069, "lr": 7.98279641810537e-07, "epoch": 2.4786324786324787, "percentage": 82.62, "elapsed_time": "0:05:51", "remaining_time": "0:01:14"}
+{"current_steps": 295, "total_steps": 351, "loss": 0.0083, "lr": 6.775934894439606e-07, "epoch": 2.5213675213675213, "percentage": 84.05, "elapsed_time": "0:05:55", "remaining_time": "0:01:07"}
+{"current_steps": 300, "total_steps": 351, "loss": 0.0087, "lr": 5.661315703996905e-07, "epoch": 2.564102564102564, "percentage": 85.47, "elapsed_time": "0:05:58", "remaining_time": "0:01:00"}
+{"current_steps": 305, "total_steps": 351, "loss": 0.0079, "lr": 4.641317500301173e-07, "epoch": 2.606837606837607, "percentage": 86.89, "elapsed_time": "0:06:03", "remaining_time": "0:00:54"}
+{"current_steps": 310, "total_steps": 351, "loss": 0.0074, "lr": 3.71811701092219e-07, "epoch": 2.6495726495726495, "percentage": 88.32, "elapsed_time": "0:06:07", "remaining_time": "0:00:48"}
+{"current_steps": 315, "total_steps": 351, "loss": 0.0065, "lr": 2.893684392229185e-07, "epoch": 2.6923076923076925, "percentage": 89.74, "elapsed_time": "0:06:10", "remaining_time": "0:00:42"}
+{"current_steps": 320, "total_steps": 351, "loss": 0.0063, "lr": 2.1697790249779638e-07, "epoch": 2.735042735042735, "percentage": 91.17, "elapsed_time": "0:06:13", "remaining_time": "0:00:36"}
+{"current_steps": 325, "total_steps": 351, "loss": 0.0068, "lr": 1.547945759703623e-07, "epoch": 2.7777777777777777, "percentage": 92.59, "elapsed_time": "0:06:17", "remaining_time": "0:00:30"}
+{"current_steps": 330, "total_steps": 351, "loss": 0.0055, "lr": 1.0295116199317057e-07, "epoch": 2.8205128205128203, "percentage": 94.02, "elapsed_time": "0:06:20", "remaining_time": "0:00:24"}
+{"current_steps": 335, "total_steps": 351, "loss": 0.0088, "lr": 6.15582970243117e-08, "epoch": 2.8632478632478633, "percentage": 95.44, "elapsed_time": "0:06:24", "remaining_time": "0:00:18"}
+{"current_steps": 340, "total_steps": 351, "loss": 0.0073, "lr": 3.0704315523631956e-08, "epoch": 2.905982905982906, "percentage": 96.87, "elapsed_time": "0:06:27", "remaining_time": "0:00:12"}
+{"current_steps": 345, "total_steps": 351, "loss": 0.0091, "lr": 1.0455061442548597e-08, "epoch": 2.948717948717949, "percentage": 98.29, "elapsed_time": "0:06:30", "remaining_time": "0:00:06"}
+{"current_steps": 350, "total_steps": 351, "loss": 0.0082, "lr": 8.537477097364522e-10, "epoch": 2.9914529914529915, "percentage": 99.72, "elapsed_time": "0:06:34", "remaining_time": "0:00:01"}
+{"current_steps": 351, "total_steps": 351, "epoch": 3.0, "percentage": 100.0, "elapsed_time": "0:07:37", "remaining_time": "0:00:00"}

trainer_state.json

@@ -0,0 +1,533 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 3.0,
+  "eval_steps": 500,
+  "global_step": 351,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.042735042735042736,
+      "grad_norm": 3.711402177810669,
+      "learning_rate": 3.6363636363636366e-06,
+      "loss": 0.2434,
+      "step": 5
+    },
+    {
+      "epoch": 0.08547008547008547,
+      "grad_norm": 3.794971227645874,
+      "learning_rate": 8.181818181818183e-06,
+      "loss": 0.1924,
+      "step": 10
+    },
+    {
+      "epoch": 0.1282051282051282,
+      "grad_norm": 2.4459402561187744,
+      "learning_rate": 9.998079135987437e-06,
+      "loss": 0.174,
+      "step": 15
+    },
+    {
+      "epoch": 0.17094017094017094,
+      "grad_norm": 2.0912649631500244,
+      "learning_rate": 9.98634586692894e-06,
+      "loss": 0.1488,
+      "step": 20
+    },
+    {
+      "epoch": 0.21367521367521367,
+      "grad_norm": 1.9830565452575684,
+      "learning_rate": 9.963971484502247e-06,
+      "loss": 0.1332,
+      "step": 25
+    },
+    {
+      "epoch": 0.2564102564102564,
+      "grad_norm": 1.4833670854568481,
+      "learning_rate": 9.931003736767013e-06,
+      "loss": 0.1146,
+      "step": 30
+    },
+    {
+      "epoch": 0.29914529914529914,
+      "grad_norm": 1.437637448310852,
+      "learning_rate": 9.887512978558329e-06,
+      "loss": 0.1175,
+      "step": 35
+    },
+    {
+      "epoch": 0.3418803418803419,
+      "grad_norm": 1.591968297958374,
+      "learning_rate": 9.833592021345938e-06,
+      "loss": 0.1114,
+      "step": 40
+    },
+    {
+      "epoch": 0.38461538461538464,
+      "grad_norm": 1.383669376373291,
+      "learning_rate": 9.76935593516989e-06,
+      "loss": 0.1092,
+      "step": 45
+    },
+    {
+      "epoch": 0.42735042735042733,
+      "grad_norm": 1.2138135433197021,
+      "learning_rate": 9.694941803075285e-06,
+      "loss": 0.1005,
+      "step": 50
+    },
+    {
+      "epoch": 0.4700854700854701,
+      "grad_norm": 1.317816138267517,
+      "learning_rate": 9.610508428570122e-06,
+      "loss": 0.1079,
+      "step": 55
+    },
+    {
+      "epoch": 0.5128205128205128,
+      "grad_norm": 1.1774816513061523,
+      "learning_rate": 9.516235996730645e-06,
+      "loss": 0.1003,
+      "step": 60
+    },
+    {
+      "epoch": 0.5555555555555556,
+      "grad_norm": 1.3384287357330322,
+      "learning_rate": 9.41232568967728e-06,
+      "loss": 0.1005,
+      "step": 65
+    },
+    {
+      "epoch": 0.5982905982905983,
+      "grad_norm": 1.0372443199157715,
+      "learning_rate": 9.298999257241862e-06,
+      "loss": 0.1029,
+      "step": 70
+    },
+    {
+      "epoch": 0.6410256410256411,
+      "grad_norm": 1.1703890562057495,
+      "learning_rate": 9.176498543742328e-06,
+      "loss": 0.0988,
+      "step": 75
+    },
+    {
+      "epoch": 0.6837606837606838,
+      "grad_norm": 1.1787683963775635,
+      "learning_rate": 9.045084971874738e-06,
+      "loss": 0.0936,
+      "step": 80
+    },
+    {
+      "epoch": 0.7264957264957265,
+      "grad_norm": 1.180239200592041,
+      "learning_rate": 8.905038984824079e-06,
+      "loss": 0.1036,
+      "step": 85
+    },
+    {
+      "epoch": 0.7692307692307693,
+      "grad_norm": 1.1398251056671143,
+      "learning_rate": 8.756659447784367e-06,
+      "loss": 0.1,
+      "step": 90
+    },
+    {
+      "epoch": 0.811965811965812,
+      "grad_norm": 1.2935456037521362,
+      "learning_rate": 8.600263010165275e-06,
+      "loss": 0.0956,
+      "step": 95
+    },
+    {
+      "epoch": 0.8547008547008547,
+      "grad_norm": 1.1012020111083984,
+      "learning_rate": 8.436183429846314e-06,
+      "loss": 0.1028,
+      "step": 100
+    },
+    {
+      "epoch": 0.8974358974358975,
+      "grad_norm": 1.0556975603103638,
+      "learning_rate": 8.264770860920722e-06,
+      "loss": 0.1077,
+      "step": 105
+    },
+    {
+      "epoch": 0.9401709401709402,
+      "grad_norm": 1.069830298423767,
+      "learning_rate": 8.086391106448965e-06,
+      "loss": 0.0976,
+      "step": 110
+    },
+    {
+      "epoch": 0.9829059829059829,
+      "grad_norm": 1.019631028175354,
+      "learning_rate": 7.90142483781658e-06,
+      "loss": 0.1008,
+      "step": 115
+    },
+    {
+      "epoch": 1.0256410256410255,
+      "grad_norm": 0.7582312226295471,
+      "learning_rate": 7.710266782362248e-06,
+      "loss": 0.0635,
+      "step": 120
+    },
+    {
+      "epoch": 1.0683760683760684,
+      "grad_norm": 0.7042792439460754,
+      "learning_rate": 7.513324881009769e-06,
+      "loss": 0.0403,
+      "step": 125
+    },
+    {
+      "epoch": 1.1111111111111112,
+      "grad_norm": 0.9356122016906738,
+      "learning_rate": 7.311019417701567e-06,
+      "loss": 0.036,
+      "step": 130
+    },
+    {
+      "epoch": 1.1538461538461537,
+      "grad_norm": 0.9737276434898376,
+      "learning_rate": 7.103782122491577e-06,
+      "loss": 0.0361,
+      "step": 135
+    },
+    {
+      "epoch": 1.1965811965811965,
+      "grad_norm": 0.9380191564559937,
+      "learning_rate": 6.892055250211552e-06,
+      "loss": 0.0343,
+      "step": 140
+    },
+    {
+      "epoch": 1.2393162393162394,
+      "grad_norm": 0.7957591414451599,
+      "learning_rate": 6.67629063667697e-06,
+      "loss": 0.0353,
+      "step": 145
+    },
+    {
+      "epoch": 1.282051282051282,
+      "grad_norm": 0.8701767921447754,
+      "learning_rate": 6.456948734446624e-06,
+      "loss": 0.0356,
+      "step": 150
+    },
+    {
+      "epoch": 1.3247863247863247,
+      "grad_norm": 0.7799380421638489,
+      "learning_rate": 6.234497630193666e-06,
+      "loss": 0.0373,
+      "step": 155
+    },
+    {
+      "epoch": 1.3675213675213675,
+      "grad_norm": 0.8129424452781677,
+      "learning_rate": 6.009412045785051e-06,
+      "loss": 0.0356,
+      "step": 160
+    },
+    {
+      "epoch": 1.4102564102564101,
+      "grad_norm": 0.8334594964981079,
+      "learning_rate": 5.782172325201155e-06,
+      "loss": 0.0348,
+      "step": 165
+    },
+    {
+      "epoch": 1.452991452991453,
+      "grad_norm": 0.8230361938476562,
+      "learning_rate": 5.553263409457504e-06,
+      "loss": 0.0322,
+      "step": 170
+    },
+    {
+      "epoch": 1.4957264957264957,
+      "grad_norm": 1.0665380954742432,
+      "learning_rate": 5.323173801716222e-06,
+      "loss": 0.036,
+      "step": 175
+    },
+    {
+      "epoch": 1.5384615384615383,
+      "grad_norm": 1.0307265520095825,
+      "learning_rate": 5.09239452479565e-06,
+      "loss": 0.0316,
+      "step": 180
+    },
+    {
+      "epoch": 1.5811965811965814,
+      "grad_norm": 0.8220146894454956,
+      "learning_rate": 4.861418073302919e-06,
+      "loss": 0.0326,
+      "step": 185
+    },
+    {
+      "epoch": 1.623931623931624,
+      "grad_norm": 1.0208584070205688,
+      "learning_rate": 4.630737362625631e-06,
+      "loss": 0.0333,
+      "step": 190
+    },
+    {
+      "epoch": 1.6666666666666665,
+      "grad_norm": 0.9469585418701172,
+      "learning_rate": 4.400844677025585e-06,
+      "loss": 0.0361,
+      "step": 195
+    },
+    {
+      "epoch": 1.7094017094017095,
+      "grad_norm": 0.8382295370101929,
+      "learning_rate": 4.17223061907935e-06,
+      "loss": 0.0356,
+      "step": 200
+    },
+    {
+      "epoch": 1.7521367521367521,
+      "grad_norm": 0.7920061945915222,
+      "learning_rate": 3.945383062707652e-06,
+      "loss": 0.033,
+      "step": 205
+    },
+    {
+      "epoch": 1.7948717948717947,
+      "grad_norm": 0.7468324899673462,
+      "learning_rate": 3.720786112027822e-06,
+      "loss": 0.0306,
+      "step": 210
+    },
+    {
+      "epoch": 1.8376068376068377,
+      "grad_norm": 0.969772219657898,
+      "learning_rate": 3.498919068251237e-06,
+      "loss": 0.0332,
+      "step": 215
+    },
+    {
+      "epoch": 1.8803418803418803,
+      "grad_norm": 0.9195359349250793,
+      "learning_rate": 3.2802554068303595e-06,
+      "loss": 0.0321,
+      "step": 220
+    },
+    {
+      "epoch": 1.9230769230769231,
+      "grad_norm": 1.082493782043457,
+      "learning_rate": 3.0652617670382745e-06,
+      "loss": 0.0321,
+      "step": 225
+    },
+    {
+      "epoch": 1.965811965811966,
+      "grad_norm": 0.8420332074165344,
+      "learning_rate": 2.8543969561369556e-06,
+      "loss": 0.0299,
+      "step": 230
+    },
+    {
+      "epoch": 2.0085470085470085,
+      "grad_norm": 0.44002777338027954,
+      "learning_rate": 2.648110970259454e-06,
+      "loss": 0.0271,
+      "step": 235
+    },
+    {
+      "epoch": 2.051282051282051,
+      "grad_norm": 0.449357271194458,
+      "learning_rate": 2.4468440340954664e-06,
+      "loss": 0.01,
+      "step": 240
+    },
+    {
+      "epoch": 2.094017094017094,
+      "grad_norm": 0.47545549273490906,
+      "learning_rate": 2.2510256614296638e-06,
+      "loss": 0.0091,
+      "step": 245
+    },
+    {
+      "epoch": 2.1367521367521367,
+      "grad_norm": 0.623047411441803,
+      "learning_rate": 2.061073738537635e-06,
+      "loss": 0.0084,
+      "step": 250
+    },
+    {
+      "epoch": 2.1794871794871793,
+      "grad_norm": 0.376277893781662,
+      "learning_rate": 1.8773936323955055e-06,
+      "loss": 0.0093,
+      "step": 255
+    },
+    {
+      "epoch": 2.2222222222222223,
+      "grad_norm": 0.676085889339447,
+      "learning_rate": 1.7003773256063882e-06,
+      "loss": 0.008,
+      "step": 260
+    },
+    {
+      "epoch": 2.264957264957265,
+      "grad_norm": 0.5928175449371338,
+      "learning_rate": 1.5304025798897521e-06,
+      "loss": 0.0075,
+      "step": 265
+    },
+    {
+      "epoch": 2.3076923076923075,
+      "grad_norm": 0.47083908319473267,
+      "learning_rate": 1.3678321299188802e-06,
+      "loss": 0.0069,
+      "step": 270
+    },
+    {
+      "epoch": 2.3504273504273505,
+      "grad_norm": 0.6990512609481812,
+      "learning_rate": 1.213012909226786e-06,
+      "loss": 0.0076,
+      "step": 275
+    },
+    {
+      "epoch": 2.393162393162393,
+      "grad_norm": 0.46621406078338623,
+      "learning_rate": 1.066275309832584e-06,
+      "loss": 0.0066,
+      "step": 280
+    },
+    {
+      "epoch": 2.435897435897436,
+      "grad_norm": 0.5243452787399292,
+      "learning_rate": 9.279324771682586e-07,
+      "loss": 0.007,
+      "step": 285
+    },
+    {
+      "epoch": 2.4786324786324787,
+      "grad_norm": 0.5895669460296631,
+      "learning_rate": 7.98279641810537e-07,
+      "loss": 0.0069,
+      "step": 290
+    },
+    {
+      "epoch": 2.5213675213675213,
+      "grad_norm": 0.6112285852432251,
+      "learning_rate": 6.775934894439606e-07,
+      "loss": 0.0083,
+      "step": 295
+    },
+    {
+      "epoch": 2.564102564102564,
+      "grad_norm": 0.5791566371917725,
+      "learning_rate": 5.661315703996905e-07,
+      "loss": 0.0087,
+      "step": 300
+    },
+    {
+      "epoch": 2.606837606837607,
+      "grad_norm": 0.574300229549408,
+      "learning_rate": 4.641317500301173e-07,
+      "loss": 0.0079,
+      "step": 305
+    },
+    {
+      "epoch": 2.6495726495726495,
+      "grad_norm": 0.4943869113922119,
+      "learning_rate": 3.71811701092219e-07,
+      "loss": 0.0074,
+      "step": 310
+    },
+    {
+      "epoch": 2.6923076923076925,
+      "grad_norm": 0.4588395059108734,
+      "learning_rate": 2.893684392229185e-07,
+      "loss": 0.0065,
+      "step": 315
+    },
+    {
+      "epoch": 2.735042735042735,
+      "grad_norm": 0.48180243372917175,
+      "learning_rate": 2.1697790249779638e-07,
+      "loss": 0.0063,
+      "step": 320
+    },
+    {
+      "epoch": 2.7777777777777777,
+      "grad_norm": 0.45809242129325867,
+      "learning_rate": 1.547945759703623e-07,
+      "loss": 0.0068,
+      "step": 325
+    },
+    {
+      "epoch": 2.8205128205128203,
+      "grad_norm": 0.5285813212394714,
+      "learning_rate": 1.0295116199317057e-07,
+      "loss": 0.0055,
+      "step": 330
+    },
+    {
+      "epoch": 2.8632478632478633,
+      "grad_norm": 0.44899120926856995,
+      "learning_rate": 6.15582970243117e-08,
+      "loss": 0.0088,
+      "step": 335
+    },
+    {
+      "epoch": 2.905982905982906,
+      "grad_norm": 0.7157308459281921,
+      "learning_rate": 3.0704315523631956e-08,
+      "loss": 0.0073,
+      "step": 340
+    },
+    {
+      "epoch": 2.948717948717949,
+      "grad_norm": 0.7962710857391357,
+      "learning_rate": 1.0455061442548597e-08,
+      "loss": 0.0091,
+      "step": 345
+    },
+    {
+      "epoch": 2.9914529914529915,
+      "grad_norm": 0.7418946623802185,
+      "learning_rate": 8.537477097364522e-10,
+      "loss": 0.0082,
+      "step": 350
+    },
+    {
+      "epoch": 3.0,
+      "step": 351,
+      "total_flos": 2.73366675523371e+17,
+      "train_loss": 0.053853037605258475,
+      "train_runtime": 459.8341,
+      "train_samples_per_second": 48.755,
+      "train_steps_per_second": 0.763
+    }
+  ],
+  "logging_steps": 5,
+  "max_steps": 351,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 3,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": true
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 2.73366675523371e+17,
+  "train_batch_size": 4,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b94e3aa7c8c3122b3bba4ab7efe4a5b72b6efbec734ab3fb35ed2938c134e4b8
+size 7889