Upload folder using huggingface_hub

SDAR-4B-Math-R1-CoT-SFT/README.md

@@ -0,0 +1,61 @@
+---
+library_name: transformers
+license: other
+base_model: models/SDAR-4B-Base
+tags:
+- llama-factory
+- full
+- generated_from_trainer
+model-index:
+- name: sft_block_4
+  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_block_4
+
+This model is a fine-tuned version of [models/SDAR-4B-Base](https://huggingface.co/models/SDAR-4B-Base) on the open_thoughts 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: 5e-05
+- train_batch_size: 2
+- eval_batch_size: 8
+- seed: 42
+- distributed_type: multi-GPU
+- num_devices: 2
+- gradient_accumulation_steps: 16
+- total_train_batch_size: 64
+- total_eval_batch_size: 16
+- optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED 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: 0.05
+- num_epochs: 4.0
+
+### Training results
+
+
+
+### Framework versions
+
+- Transformers 4.52.4
+- Pytorch 2.8.0+cu128
+- Datasets 3.6.0
+- Tokenizers 0.21.4

SDAR-4B-Math-R1-CoT-SFT/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
+}

SDAR-4B-Math-R1-CoT-SFT/all_results.json

@@ -0,0 +1,9 @@
+{
+    "effective_tokens_per_sec": 16342.776477265266,
+    "epoch": 3.9988184324537217,
+    "total_flos": 4.3565600076592054e+19,
+    "train_loss": 1.895302517482642,
+    "train_runtime": 61070.1581,
+    "train_samples_per_second": 2.661,
+    "train_steps_per_second": 0.042
+}

SDAR-4B-Math-R1-CoT-SFT/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 %}

SDAR-4B-Math-R1-CoT-SFT/config.json

@@ -0,0 +1,43 @@
+{
+  "architectures": [
+    "SDARForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_sdar.SDARConfig",
+    "AutoModel": "modeling_sdar.SDARForCausalLM",
+    "AutoModelForCausalLM": "modeling_sdar.SDARForCausalLM"
+  },
+  "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": true,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.4",
+  "use_cache": false,
+  "use_deepep": false,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

SDAR-4B-Math-R1-CoT-SFT/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"]

SDAR-4B-Math-R1-CoT-SFT/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)
+        )

SDAR-4B-Math-R1-CoT-SFT/generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 151643,
+  "eos_token_id": 151643,
+  "transformers_version": "4.52.4",
+  "use_cache": false
+}

SDAR-4B-Math-R1-CoT-SFT/model-00001-of-00002.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3cf721838f56081281667205fd4e2d29858a804aad837977f092eb179b428b35
+size 4967215360

SDAR-4B-Math-R1-CoT-SFT/model-00002-of-00002.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:309f36d853f5db682e8771eceee6fb90e8a963e9f64e0f8e8bb12178c7ce5baa
+size 3855679144

SDAR-4B-Math-R1-CoT-SFT/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"
+  }
+}

SDAR-4B-Math-R1-CoT-SFT/modeling_sdar.py

@@ -0,0 +1,1233 @@
+# 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
+
+
+@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):
+        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
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+@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,
+        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)
+
+        # 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)
+
+        # 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,
+        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"
+            prompt_mask = (labels == -100) if labels is not None else None
+            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()
+            assert labels is not None, "Labels must be provided for training."
+            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,
+                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",
+]

SDAR-4B-Math-R1-CoT-SFT/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": "<|im_end|>",
+    "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
+  }
+}

SDAR-4B-Math-R1-CoT-SFT/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"]

SDAR-4B-Math-R1-CoT-SFT/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": "<|im_end|>",
+  "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
+}

SDAR-4B-Math-R1-CoT-SFT/train_results.json

@@ -0,0 +1,9 @@
+{
+    "effective_tokens_per_sec": 16342.776477265266,
+    "epoch": 3.9988184324537217,
+    "total_flos": 4.3565600076592054e+19,
+    "train_loss": 1.895302517482642,
+    "train_runtime": 61070.1581,
+    "train_samples_per_second": 2.661,
+    "train_steps_per_second": 0.042
+}

SDAR-4B-Math-R1-CoT-SFT/trainer_state.json

@@ -0,0 +1,3592 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 3.9988184324537217,
+  "eval_steps": 500,
+  "global_step": 2539,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.007877116975187082,
+      "grad_norm": 3399.454833984375,
+      "learning_rate": 1.5748031496062992e-06,
+      "loss": 131.9922,
+      "step": 5
+    },
+    {
+      "epoch": 0.015754233950374164,
+      "grad_norm": 1452.986572265625,
+      "learning_rate": 3.5433070866141735e-06,
+      "loss": 88.8359,
+      "step": 10
+    },
+    {
+      "epoch": 0.023631350925561245,
+      "grad_norm": 534.6153564453125,
+      "learning_rate": 5.511811023622048e-06,
+      "loss": 48.5337,
+      "step": 15
+    },
+    {
+      "epoch": 0.03150846790074833,
+      "grad_norm": 281.75048828125,
+      "learning_rate": 7.4803149606299226e-06,
+      "loss": 35.3826,
+      "step": 20
+    },
+    {
+      "epoch": 0.03938558487593541,
+      "grad_norm": 87.65968322753906,
+      "learning_rate": 9.448818897637795e-06,
+      "loss": 27.0295,
+      "step": 25
+    },
+    {
+      "epoch": 0.04726270185112249,
+      "grad_norm": 65.45500183105469,
+      "learning_rate": 1.141732283464567e-05,
+      "loss": 23.913,
+      "step": 30
+    },
+    {
+      "epoch": 0.05513981882630957,
+      "grad_norm": 65.12550354003906,
+      "learning_rate": 1.3385826771653545e-05,
+      "loss": 20.433,
+      "step": 35
+    },
+    {
+      "epoch": 0.06301693580149666,
+      "grad_norm": 20.998428344726562,
+      "learning_rate": 1.535433070866142e-05,
+      "loss": 18.3731,
+      "step": 40
+    },
+    {
+      "epoch": 0.07089405277668373,
+      "grad_norm": 29.879304885864258,
+      "learning_rate": 1.732283464566929e-05,
+      "loss": 16.9914,
+      "step": 45
+    },
+    {
+      "epoch": 0.07877116975187082,
+      "grad_norm": 22.706512451171875,
+      "learning_rate": 1.9291338582677166e-05,
+      "loss": 15.2466,
+      "step": 50
+    },
+    {
+      "epoch": 0.08664828672705789,
+      "grad_norm": 18.713308334350586,
+      "learning_rate": 2.125984251968504e-05,
+      "loss": 13.814,
+      "step": 55
+    },
+    {
+      "epoch": 0.09452540370224498,
+      "grad_norm": 29.742462158203125,
+      "learning_rate": 2.3228346456692916e-05,
+      "loss": 14.2114,
+      "step": 60
+    },
+    {
+      "epoch": 0.10240252067743207,
+      "grad_norm": 29.92841339111328,
+      "learning_rate": 2.5196850393700788e-05,
+      "loss": 14.0749,
+      "step": 65
+    },
+    {
+      "epoch": 0.11027963765261914,
+      "grad_norm": 17.866037368774414,
+      "learning_rate": 2.7165354330708666e-05,
+      "loss": 13.3448,
+      "step": 70
+    },
+    {
+      "epoch": 0.11815675462780623,
+      "grad_norm": 26.47714614868164,
+      "learning_rate": 2.9133858267716534e-05,
+      "loss": 13.2347,
+      "step": 75
+    },
+    {
+      "epoch": 0.1260338716029933,
+      "grad_norm": 22.94607162475586,
+      "learning_rate": 3.110236220472441e-05,
+      "loss": 12.4648,
+      "step": 80
+    },
+    {
+      "epoch": 0.13391098857818037,
+      "grad_norm": 14.325417518615723,
+      "learning_rate": 3.3070866141732284e-05,
+      "loss": 12.8074,
+      "step": 85
+    },
+    {
+      "epoch": 0.14178810555336746,
+      "grad_norm": 14.305035591125488,
+      "learning_rate": 3.5039370078740156e-05,
+      "loss": 12.1508,
+      "step": 90
+    },
+    {
+      "epoch": 0.14966522252855455,
+      "grad_norm": 14.955709457397461,
+      "learning_rate": 3.7007874015748034e-05,
+      "loss": 12.5528,
+      "step": 95
+    },
+    {
+      "epoch": 0.15754233950374164,
+      "grad_norm": 14.336165428161621,
+      "learning_rate": 3.8976377952755905e-05,
+      "loss": 11.6424,
+      "step": 100
+    },
+    {
+      "epoch": 0.16541945647892872,
+      "grad_norm": 14.17673110961914,
+      "learning_rate": 4.0944881889763784e-05,
+      "loss": 12.1279,
+      "step": 105
+    },
+    {
+      "epoch": 0.17329657345411578,
+      "grad_norm": 12.128265380859375,
+      "learning_rate": 4.2913385826771655e-05,
+      "loss": 11.7178,
+      "step": 110
+    },
+    {
+      "epoch": 0.18117369042930287,
+      "grad_norm": 13.574759483337402,
+      "learning_rate": 4.488188976377953e-05,
+      "loss": 11.9381,
+      "step": 115
+    },
+    {
+      "epoch": 0.18905080740448996,
+      "grad_norm": 11.064770698547363,
+      "learning_rate": 4.6850393700787405e-05,
+      "loss": 11.1822,
+      "step": 120
+    },
+    {
+      "epoch": 0.19692792437967704,
+      "grad_norm": 13.90650463104248,
+      "learning_rate": 4.881889763779528e-05,
+      "loss": 12.0308,
+      "step": 125
+    },
+    {
+      "epoch": 0.20480504135486413,
+      "grad_norm": 12.934874534606934,
+      "learning_rate": 4.999991524704281e-05,
+      "loss": 11.8371,
+      "step": 130
+    },
+    {
+      "epoch": 0.2126821583300512,
+      "grad_norm": 11.595288276672363,
+      "learning_rate": 4.999896178287378e-05,
+      "loss": 11.3976,
+      "step": 135
+    },
+    {
+      "epoch": 0.22055927530523828,
+      "grad_norm": 10.271947860717773,
+      "learning_rate": 4.999694895387824e-05,
+      "loss": 11.6096,
+      "step": 140
+    },
+    {
+      "epoch": 0.22843639228042537,
+      "grad_norm": 12.979105949401855,
+      "learning_rate": 4.999387684535255e-05,
+      "loss": 11.5108,
+      "step": 145
+    },
+    {
+      "epoch": 0.23631350925561245,
+      "grad_norm": 12.064156532287598,
+      "learning_rate": 4.998974558748146e-05,
+      "loss": 11.7928,
+      "step": 150
+    },
+    {
+      "epoch": 0.24419062623079954,
+      "grad_norm": 12.464822769165039,
+      "learning_rate": 4.9984555355332604e-05,
+      "loss": 11.2361,
+      "step": 155
+    },
+    {
+      "epoch": 0.2520677432059866,
+      "grad_norm": 11.160205841064453,
+      "learning_rate": 4.997830636884911e-05,
+      "loss": 11.118,
+      "step": 160
+    },
+    {
+      "epoch": 0.2599448601811737,
+      "grad_norm": 10.261472702026367,
+      "learning_rate": 4.997099889284021e-05,
+      "loss": 11.6838,
+      "step": 165
+    },
+    {
+      "epoch": 0.26782197715636075,
+      "grad_norm": 9.86904525756836,
+      "learning_rate": 4.996263323697011e-05,
+      "loss": 11.4614,
+      "step": 170
+    },
+    {
+      "epoch": 0.27569909413154786,
+      "grad_norm": 10.730037689208984,
+      "learning_rate": 4.9953209755744784e-05,
+      "loss": 11.69,
+      "step": 175
+    },
+    {
+      "epoch": 0.2835762111067349,
+      "grad_norm": 10.000442504882812,
+      "learning_rate": 4.9942728848496995e-05,
+      "loss": 11.3298,
+      "step": 180
+    },
+    {
+      "epoch": 0.29145332808192204,
+      "grad_norm": 9.375894546508789,
+      "learning_rate": 4.993119095936937e-05,
+      "loss": 11.5506,
+      "step": 185
+    },
+    {
+      "epoch": 0.2993304450571091,
+      "grad_norm": 11.61759090423584,
+      "learning_rate": 4.991859657729557e-05,
+      "loss": 11.0027,
+      "step": 190
+    },
+    {
+      "epoch": 0.30720756203229616,
+      "grad_norm": 8.72737979888916,
+      "learning_rate": 4.9904946235979574e-05,
+      "loss": 11.1883,
+      "step": 195
+    },
+    {
+      "epoch": 0.31508467900748327,
+      "grad_norm": 8.50748348236084,
+      "learning_rate": 4.9890240513873064e-05,
+      "loss": 11.0641,
+      "step": 200
+    },
+    {
+      "epoch": 0.32296179598267033,
+      "grad_norm": 8.511228561401367,
+      "learning_rate": 4.9874480034150915e-05,
+      "loss": 11.1942,
+      "step": 205
+    },
+    {
+      "epoch": 0.33083891295785745,
+      "grad_norm": 9.39171028137207,
+      "learning_rate": 4.985766546468477e-05,
+      "loss": 10.6584,
+      "step": 210
+    },
+    {
+      "epoch": 0.3387160299330445,
+      "grad_norm": 7.405202865600586,
+      "learning_rate": 4.983979751801475e-05,
+      "loss": 11.1103,
+      "step": 215
+    },
+    {
+      "epoch": 0.34659314690823156,
+      "grad_norm": 7.648876667022705,
+      "learning_rate": 4.982087695131928e-05,
+      "loss": 11.0276,
+      "step": 220
+    },
+    {
+      "epoch": 0.3544702638834187,
+      "grad_norm": 7.407284259796143,
+      "learning_rate": 4.9800904566382984e-05,
+      "loss": 9.9736,
+      "step": 225
+    },
+    {
+      "epoch": 0.36234738085860574,
+      "grad_norm": 7.161754608154297,
+      "learning_rate": 4.9779881209562675e-05,
+      "loss": 10.8461,
+      "step": 230
+    },
+    {
+      "epoch": 0.37022449783379285,
+      "grad_norm": 9.168039321899414,
+      "learning_rate": 4.975780777175154e-05,
+      "loss": 10.7467,
+      "step": 235
+    },
+    {
+      "epoch": 0.3781016148089799,
+      "grad_norm": 8.102660179138184,
+      "learning_rate": 4.9734685188341376e-05,
+      "loss": 10.8515,
+      "step": 240
+    },
+    {
+      "epoch": 0.385978731784167,
+      "grad_norm": 8.496058464050293,
+      "learning_rate": 4.971051443918294e-05,
+      "loss": 10.1448,
+      "step": 245
+    },
+    {
+      "epoch": 0.3938558487593541,
+      "grad_norm": 8.0912446975708,
+      "learning_rate": 4.9685296548544424e-05,
+      "loss": 10.299,
+      "step": 250
+    },
+    {
+      "epoch": 0.40173296573454115,
+      "grad_norm": 8.432981491088867,
+      "learning_rate": 4.965903258506806e-05,
+      "loss": 10.3602,
+      "step": 255
+    },
+    {
+      "epoch": 0.40961008270972826,
+      "grad_norm": 7.998964786529541,
+      "learning_rate": 4.9631723661724836e-05,
+      "loss": 10.3991,
+      "step": 260
+    },
+    {
+      "epoch": 0.4174871996849153,
+      "grad_norm": 7.924742698669434,
+      "learning_rate": 4.9603370935767324e-05,
+      "loss": 10.7044,
+      "step": 265
+    },
+    {
+      "epoch": 0.4253643166601024,
+      "grad_norm": 7.88558292388916,
+      "learning_rate": 4.957397560868064e-05,
+      "loss": 10.2436,
+      "step": 270
+    },
+    {
+      "epoch": 0.4332414336352895,
+      "grad_norm": 7.450520038604736,
+      "learning_rate": 4.954353892613154e-05,
+      "loss": 10.2202,
+      "step": 275
+    },
+    {
+      "epoch": 0.44111855061047656,
+      "grad_norm": 8.56113052368164,
+      "learning_rate": 4.951206217791564e-05,
+      "loss": 10.4133,
+      "step": 280
+    },
+    {
+      "epoch": 0.44899566758566367,
+      "grad_norm": 6.7272491455078125,
+      "learning_rate": 4.94795466979027e-05,
+      "loss": 10.1886,
+      "step": 285
+    },
+    {
+      "epoch": 0.45687278456085073,
+      "grad_norm": 8.858597755432129,
+      "learning_rate": 4.94459938639802e-05,
+      "loss": 10.2836,
+      "step": 290
+    },
+    {
+      "epoch": 0.4647499015360378,
+      "grad_norm": 7.176908493041992,
+      "learning_rate": 4.941140509799487e-05,
+      "loss": 10.657,
+      "step": 295
+    },
+    {
+      "epoch": 0.4726270185112249,
+      "grad_norm": 6.454217433929443,
+      "learning_rate": 4.937578186569244e-05,
+      "loss": 9.67,
+      "step": 300
+    },
+    {
+      "epoch": 0.48050413548641197,
+      "grad_norm": 8.159943580627441,
+      "learning_rate": 4.933912567665561e-05,
+      "loss": 10.7037,
+      "step": 305
+    },
+    {
+      "epoch": 0.4883812524615991,
+      "grad_norm": 7.20767879486084,
+      "learning_rate": 4.930143808423995e-05,
+      "loss": 9.7757,
+      "step": 310
+    },
+    {
+      "epoch": 0.49625836943678614,
+      "grad_norm": 8.055448532104492,
+      "learning_rate": 4.926272068550819e-05,
+      "loss": 10.9187,
+      "step": 315
+    },
+    {
+      "epoch": 0.5041354864119733,
+      "grad_norm": 7.440671443939209,
+      "learning_rate": 4.922297512116248e-05,
+      "loss": 10.756,
+      "step": 320
+    },
+    {
+      "epoch": 0.5120126033871603,
+      "grad_norm": 7.851231098175049,
+      "learning_rate": 4.9182203075474884e-05,
+      "loss": 9.9799,
+      "step": 325
+    },
+    {
+      "epoch": 0.5198897203623474,
+      "grad_norm": 7.782782077789307,
+      "learning_rate": 4.914040627621599e-05,
+      "loss": 10.0362,
+      "step": 330
+    },
+    {
+      "epoch": 0.5277668373375345,
+      "grad_norm": 7.82265043258667,
+      "learning_rate": 4.909758649458172e-05,
+      "loss": 10.2922,
+      "step": 335
+    },
+    {
+      "epoch": 0.5356439543127215,
+      "grad_norm": 7.081122875213623,
+      "learning_rate": 4.905374554511825e-05,
+      "loss": 9.93,
+      "step": 340
+    },
+    {
+      "epoch": 0.5435210712879086,
+      "grad_norm": 7.180002212524414,
+      "learning_rate": 4.9008885285645134e-05,
+      "loss": 10.1384,
+      "step": 345
+    },
+    {
+      "epoch": 0.5513981882630957,
+      "grad_norm": 8.38278579711914,
+      "learning_rate": 4.8963007617176535e-05,
+      "loss": 10.1262,
+      "step": 350
+    },
+    {
+      "epoch": 0.5592753052382828,
+      "grad_norm": 7.965811729431152,
+      "learning_rate": 4.8916114483840746e-05,
+      "loss": 10.2287,
+      "step": 355
+    },
+    {
+      "epoch": 0.5671524222134698,
+      "grad_norm": 6.411648750305176,
+      "learning_rate": 4.886820787279773e-05,
+      "loss": 9.9416,
+      "step": 360
+    },
+    {
+      "epoch": 0.575029539188657,
+      "grad_norm": 6.27194356918335,
+      "learning_rate": 4.8819289814154935e-05,
+      "loss": 9.9728,
+      "step": 365
+    },
+    {
+      "epoch": 0.5829066561638441,
+      "grad_norm": 5.867300987243652,
+      "learning_rate": 4.876936238088127e-05,
+      "loss": 10.0274,
+      "step": 370
+    },
+    {
+      "epoch": 0.5907837731390311,
+      "grad_norm": 6.166661739349365,
+      "learning_rate": 4.871842768871928e-05,
+      "loss": 10.2888,
+      "step": 375
+    },
+    {
+      "epoch": 0.5986608901142182,
+      "grad_norm": 6.584941387176514,
+      "learning_rate": 4.866648789609545e-05,
+      "loss": 9.8767,
+      "step": 380
+    },
+    {
+      "epoch": 0.6065380070894053,
+      "grad_norm": 6.197508811950684,
+      "learning_rate": 4.8613545204028764e-05,
+      "loss": 10.2098,
+      "step": 385
+    },
+    {
+      "epoch": 0.6144151240645923,
+      "grad_norm": 7.481927394866943,
+      "learning_rate": 4.855960185603742e-05,
+      "loss": 9.8902,
+      "step": 390
+    },
+    {
+      "epoch": 0.6222922410397794,
+      "grad_norm": 7.165607452392578,
+      "learning_rate": 4.850466013804377e-05,
+      "loss": 10.2372,
+      "step": 395
+    },
+    {
+      "epoch": 0.6301693580149665,
+      "grad_norm": 6.94336462020874,
+      "learning_rate": 4.844872237827745e-05,
+      "loss": 9.9813,
+      "step": 400
+    },
+    {
+      "epoch": 0.6380464749901537,
+      "grad_norm": 6.348086833953857,
+      "learning_rate": 4.8391790947176696e-05,
+      "loss": 10.0216,
+      "step": 405
+    },
+    {
+      "epoch": 0.6459235919653407,
+      "grad_norm": 6.404892444610596,
+      "learning_rate": 4.833386825728794e-05,
+      "loss": 10.108,
+      "step": 410
+    },
+    {
+      "epoch": 0.6538007089405278,
+      "grad_norm": 7.904120445251465,
+      "learning_rate": 4.8274956763163506e-05,
+      "loss": 10.0952,
+      "step": 415
+    },
+    {
+      "epoch": 0.6616778259157149,
+      "grad_norm": 6.545925617218018,
+      "learning_rate": 4.821505896125769e-05,
+      "loss": 9.8825,
+      "step": 420
+    },
+    {
+      "epoch": 0.6695549428909019,
+      "grad_norm": 6.014336109161377,
+      "learning_rate": 4.815417738982085e-05,
+      "loss": 10.0608,
+      "step": 425
+    },
+    {
+      "epoch": 0.677432059866089,
+      "grad_norm": 7.037391185760498,
+      "learning_rate": 4.809231462879196e-05,
+      "loss": 9.6694,
+      "step": 430
+    },
+    {
+      "epoch": 0.6853091768412761,
+      "grad_norm": 6.266141414642334,
+      "learning_rate": 4.802947329968921e-05,
+      "loss": 9.5179,
+      "step": 435
+    },
+    {
+      "epoch": 0.6931862938164631,
+      "grad_norm": 6.4194111824035645,
+      "learning_rate": 4.7965656065498944e-05,
+      "loss": 9.9608,
+      "step": 440
+    },
+    {
+      "epoch": 0.7010634107916502,
+      "grad_norm": 6.709506511688232,
+      "learning_rate": 4.7900865630562786e-05,
+      "loss": 9.8622,
+      "step": 445
+    },
+    {
+      "epoch": 0.7089405277668374,
+      "grad_norm": 5.767233848571777,
+      "learning_rate": 4.783510474046308e-05,
+      "loss": 9.3161,
+      "step": 450
+    },
+    {
+      "epoch": 0.7168176447420245,
+      "grad_norm": 5.416118621826172,
+      "learning_rate": 4.77683761819065e-05,
+      "loss": 9.719,
+      "step": 455
+    },
+    {
+      "epoch": 0.7246947617172115,
+      "grad_norm": 6.180813789367676,
+      "learning_rate": 4.7700682782606015e-05,
+      "loss": 9.9941,
+      "step": 460
+    },
+    {
+      "epoch": 0.7325718786923986,
+      "grad_norm": 5.783968448638916,
+      "learning_rate": 4.763202741116098e-05,
+      "loss": 10.4704,
+      "step": 465
+    },
+    {
+      "epoch": 0.7404489956675857,
+      "grad_norm": 5.99013614654541,
+      "learning_rate": 4.756241297693566e-05,
+      "loss": 9.3854,
+      "step": 470
+    },
+    {
+      "epoch": 0.7483261126427727,
+      "grad_norm": 6.445564270019531,
+      "learning_rate": 4.749184242993587e-05,
+      "loss": 9.835,
+      "step": 475
+    },
+    {
+      "epoch": 0.7562032296179598,
+      "grad_norm": 5.977458477020264,
+      "learning_rate": 4.742031876068402e-05,
+      "loss": 10.2198,
+      "step": 480
+    },
+    {
+      "epoch": 0.7640803465931469,
+      "grad_norm": 7.140058517456055,
+      "learning_rate": 4.734784500009237e-05,
+      "loss": 9.456,
+      "step": 485
+    },
+    {
+      "epoch": 0.771957463568334,
+      "grad_norm": 6.452182769775391,
+      "learning_rate": 4.727442421933456e-05,
+      "loss": 9.904,
+      "step": 490
+    },
+    {
+      "epoch": 0.7798345805435211,
+      "grad_norm": 5.870535373687744,
+      "learning_rate": 4.720005952971551e-05,
+      "loss": 9.9197,
+      "step": 495
+    },
+    {
+      "epoch": 0.7877116975187082,
+      "grad_norm": 6.265655040740967,
+      "learning_rate": 4.7124754082539546e-05,
+      "loss": 9.8827,
+      "step": 500
+    },
+    {
+      "epoch": 0.7955888144938952,
+      "grad_norm": 6.607460975646973,
+      "learning_rate": 4.704851106897689e-05,
+      "loss": 9.6845,
+      "step": 505
+    },
+    {
+      "epoch": 0.8034659314690823,
+      "grad_norm": 7.221487045288086,
+      "learning_rate": 4.6971333719928354e-05,
+      "loss": 9.9776,
+      "step": 510
+    },
+    {
+      "epoch": 0.8113430484442694,
+      "grad_norm": 6.316247463226318,
+      "learning_rate": 4.6893225305888556e-05,
+      "loss": 9.2224,
+      "step": 515
+    },
+    {
+      "epoch": 0.8192201654194565,
+      "grad_norm": 5.804502010345459,
+      "learning_rate": 4.68141891368072e-05,
+      "loss": 9.8942,
+      "step": 520
+    },
+    {
+      "epoch": 0.8270972823946435,
+      "grad_norm": 6.048508644104004,
+      "learning_rate": 4.6734228561948864e-05,
+      "loss": 9.9273,
+      "step": 525
+    },
+    {
+      "epoch": 0.8349743993698306,
+      "grad_norm": 5.967385768890381,
+      "learning_rate": 4.6653346969751096e-05,
+      "loss": 9.8039,
+      "step": 530
+    },
+    {
+      "epoch": 0.8428515163450178,
+      "grad_norm": 5.828855514526367,
+      "learning_rate": 4.65715477876808e-05,
+      "loss": 9.5876,
+      "step": 535
+    },
+    {
+      "epoch": 0.8507286333202048,
+      "grad_norm": 6.0214314460754395,
+      "learning_rate": 4.6488834482088965e-05,
+      "loss": 9.478,
+      "step": 540
+    },
+    {
+      "epoch": 0.8586057502953919,
+      "grad_norm": 6.222189903259277,
+      "learning_rate": 4.640521055806383e-05,
+      "loss": 9.6082,
+      "step": 545
+    },
+    {
+      "epoch": 0.866482867270579,
+      "grad_norm": 6.3668437004089355,
+      "learning_rate": 4.632067955928231e-05,
+      "loss": 9.515,
+      "step": 550
+    },
+    {
+      "epoch": 0.874359984245766,
+      "grad_norm": 6.134618282318115,
+      "learning_rate": 4.6235245067859825e-05,
+      "loss": 9.7517,
+      "step": 555
+    },
+    {
+      "epoch": 0.8822371012209531,
+      "grad_norm": 6.289186000823975,
+      "learning_rate": 4.614891070419855e-05,
+      "loss": 9.7513,
+      "step": 560
+    },
+    {
+      "epoch": 0.8901142181961402,
+      "grad_norm": 6.570099830627441,
+      "learning_rate": 4.606168012683394e-05,
+      "loss": 9.5023,
+      "step": 565
+    },
+    {
+      "epoch": 0.8979913351713273,
+      "grad_norm": 6.501761436462402,
+      "learning_rate": 4.5973557032279726e-05,
+      "loss": 9.7911,
+      "step": 570
+    },
+    {
+      "epoch": 0.9058684521465143,
+      "grad_norm": 6.731903553009033,
+      "learning_rate": 4.588454515487125e-05,
+      "loss": 9.2721,
+      "step": 575
+    },
+    {
+      "epoch": 0.9137455691217015,
+      "grad_norm": 6.2406086921691895,
+      "learning_rate": 4.5794648266607254e-05,
+      "loss": 9.2431,
+      "step": 580
+    },
+    {
+      "epoch": 0.9216226860968886,
+      "grad_norm": 8.897172927856445,
+      "learning_rate": 4.570387017699e-05,
+      "loss": 9.4321,
+      "step": 585
+    },
+    {
+      "epoch": 0.9294998030720756,
+      "grad_norm": 6.036581039428711,
+      "learning_rate": 4.561221473286384e-05,
+      "loss": 10.3099,
+      "step": 590
+    },
+    {
+      "epoch": 0.9373769200472627,
+      "grad_norm": 5.789550304412842,
+      "learning_rate": 4.551968581825222e-05,
+      "loss": 10.217,
+      "step": 595
+    },
+    {
+      "epoch": 0.9452540370224498,
+      "grad_norm": 6.0984110832214355,
+      "learning_rate": 4.5426287354193095e-05,
+      "loss": 9.667,
+      "step": 600
+    },
+    {
+      "epoch": 0.9531311539976368,
+      "grad_norm": 6.172824382781982,
+      "learning_rate": 4.533202329857271e-05,
+      "loss": 9.7751,
+      "step": 605
+    },
+    {
+      "epoch": 0.9610082709728239,
+      "grad_norm": 6.124137878417969,
+      "learning_rate": 4.523689764595797e-05,
+      "loss": 9.4172,
+      "step": 610
+    },
+    {
+      "epoch": 0.968885387948011,
+      "grad_norm": 5.909232139587402,
+      "learning_rate": 4.5140914427427094e-05,
+      "loss": 9.2955,
+      "step": 615
+    },
+    {
+      "epoch": 0.9767625049231982,
+      "grad_norm": 5.878952503204346,
+      "learning_rate": 4.504407771039883e-05,
+      "loss": 9.4788,
+      "step": 620
+    },
+    {
+      "epoch": 0.9846396218983852,
+      "grad_norm": 7.975442886352539,
+      "learning_rate": 4.494639159846006e-05,
+      "loss": 9.15,
+      "step": 625
+    },
+    {
+      "epoch": 0.9925167388735723,
+      "grad_norm": 6.191237449645996,
+      "learning_rate": 4.484786023119194e-05,
+      "loss": 10.031,
+      "step": 630
+    },
+    {
+      "epoch": 1.0,
+      "grad_norm": 5.575296401977539,
+      "learning_rate": 4.474848778399446e-05,
+      "loss": 9.1086,
+      "step": 635
+    },
+    {
+      "epoch": 1.007877116975187,
+      "grad_norm": 5.996377944946289,
+      "learning_rate": 4.464827846790952e-05,
+      "loss": 9.3717,
+      "step": 640
+    },
+    {
+      "epoch": 1.0157542339503742,
+      "grad_norm": 5.924043655395508,
+      "learning_rate": 4.454723652944246e-05,
+      "loss": 9.4512,
+      "step": 645
+    },
+    {
+      "epoch": 1.0236313509255612,
+      "grad_norm": 6.695152282714844,
+      "learning_rate": 4.4445366250382106e-05,
+      "loss": 8.4621,
+      "step": 650
+    },
+    {
+      "epoch": 1.0315084679007482,
+      "grad_norm": 7.838866233825684,
+      "learning_rate": 4.434267194761936e-05,
+      "loss": 9.3423,
+      "step": 655
+    },
+    {
+      "epoch": 1.0393855848759355,
+      "grad_norm": 5.655851364135742,
+      "learning_rate": 4.423915797296425e-05,
+      "loss": 9.2772,
+      "step": 660
+    },
+    {
+      "epoch": 1.0472627018511225,
+      "grad_norm": 6.232006072998047,
+      "learning_rate": 4.413482871296148e-05,
+      "loss": 9.1979,
+      "step": 665
+    },
+    {
+      "epoch": 1.0551398188263095,
+      "grad_norm": 5.975604057312012,
+      "learning_rate": 4.402968858870459e-05,
+      "loss": 9.4245,
+      "step": 670
+    },
+    {
+      "epoch": 1.0630169358014967,
+      "grad_norm": 8.961718559265137,
+      "learning_rate": 4.39237420556486e-05,
+      "loss": 9.2747,
+      "step": 675
+    },
+    {
+      "epoch": 1.0708940527766837,
+      "grad_norm": 6.198498249053955,
+      "learning_rate": 4.381699360342116e-05,
+      "loss": 9.1819,
+      "step": 680
+    },
+    {
+      "epoch": 1.078771169751871,
+      "grad_norm": 5.1888861656188965,
+      "learning_rate": 4.370944775563239e-05,
+      "loss": 8.2554,
+      "step": 685
+    },
+    {
+      "epoch": 1.086648286727058,
+      "grad_norm": 5.546014308929443,
+      "learning_rate": 4.360110906968305e-05,
+      "loss": 9.2945,
+      "step": 690
+    },
+    {
+      "epoch": 1.094525403702245,
+      "grad_norm": 5.926428318023682,
+      "learning_rate": 4.3491982136571554e-05,
+      "loss": 9.384,
+      "step": 695
+    },
+    {
+      "epoch": 1.1024025206774322,
+      "grad_norm": 6.317972660064697,
+      "learning_rate": 4.3382071580699314e-05,
+      "loss": 8.7085,
+      "step": 700
+    },
+    {
+      "epoch": 1.1102796376526192,
+      "grad_norm": 5.735546588897705,
+      "learning_rate": 4.327138205967484e-05,
+      "loss": 9.1479,
+      "step": 705
+    },
+    {
+      "epoch": 1.1181567546278062,
+      "grad_norm": 5.673744201660156,
+      "learning_rate": 4.3159918264116334e-05,
+      "loss": 9.5043,
+      "step": 710
+    },
+    {
+      "epoch": 1.1260338716029934,
+      "grad_norm": 5.20525598526001,
+      "learning_rate": 4.3047684917452936e-05,
+      "loss": 9.1163,
+      "step": 715
+    },
+    {
+      "epoch": 1.1339109885781804,
+      "grad_norm": 6.533771514892578,
+      "learning_rate": 4.293468677572454e-05,
+      "loss": 8.9555,
+      "step": 720
+    },
+    {
+      "epoch": 1.1417881055533674,
+      "grad_norm": 5.279290676116943,
+      "learning_rate": 4.282092862738029e-05,
+      "loss": 8.6573,
+      "step": 725
+    },
+    {
+      "epoch": 1.1496652225285546,
+      "grad_norm": 5.301630020141602,
+      "learning_rate": 4.270641529307562e-05,
+      "loss": 8.9388,
+      "step": 730
+    },
+    {
+      "epoch": 1.1575423395037416,
+      "grad_norm": 7.306978225708008,
+      "learning_rate": 4.2591151625467986e-05,
+      "loss": 9.1701,
+      "step": 735
+    },
+    {
+      "epoch": 1.1654194564789286,
+      "grad_norm": 5.707004070281982,
+      "learning_rate": 4.2475142509011254e-05,
+      "loss": 9.3798,
+      "step": 740
+    },
+    {
+      "epoch": 1.1732965734541159,
+      "grad_norm": 7.129739284515381,
+      "learning_rate": 4.235839285974868e-05,
+      "loss": 9.1854,
+      "step": 745
+    },
+    {
+      "epoch": 1.1811736904293029,
+      "grad_norm": 9.110360145568848,
+      "learning_rate": 4.2240907625104606e-05,
+      "loss": 8.9261,
+      "step": 750
+    },
+    {
+      "epoch": 1.1890508074044899,
+      "grad_norm": 6.3835554122924805,
+      "learning_rate": 4.2122691783674786e-05,
+      "loss": 8.9869,
+      "step": 755
+    },
+    {
+      "epoch": 1.196927924379677,
+      "grad_norm": 5.634960651397705,
+      "learning_rate": 4.200375034501543e-05,
+      "loss": 8.6859,
+      "step": 760
+    },
+    {
+      "epoch": 1.204805041354864,
+      "grad_norm": 5.98690938949585,
+      "learning_rate": 4.188408834943093e-05,
+      "loss": 8.9667,
+      "step": 765
+    },
+    {
+      "epoch": 1.212682158330051,
+      "grad_norm": 5.745028972625732,
+      "learning_rate": 4.176371086776023e-05,
+      "loss": 8.3641,
+      "step": 770
+    },
+    {
+      "epoch": 1.2205592753052383,
+      "grad_norm": 5.545426845550537,
+      "learning_rate": 4.1642623001161976e-05,
+      "loss": 9.2566,
+      "step": 775
+    },
+    {
+      "epoch": 1.2284363922804253,
+      "grad_norm": 6.151079177856445,
+      "learning_rate": 4.152082988089834e-05,
+      "loss": 9.1436,
+      "step": 780
+    },
+    {
+      "epoch": 1.2363135092556123,
+      "grad_norm": 5.903685092926025,
+      "learning_rate": 4.139833666811756e-05,
+      "loss": 9.0897,
+      "step": 785
+    },
+    {
+      "epoch": 1.2441906262307996,
+      "grad_norm": 5.456060886383057,
+      "learning_rate": 4.127514855363524e-05,
+      "loss": 9.2121,
+      "step": 790
+    },
+    {
+      "epoch": 1.2520677432059866,
+      "grad_norm": 5.680197238922119,
+      "learning_rate": 4.1151270757714405e-05,
+      "loss": 9.1453,
+      "step": 795
+    },
+    {
+      "epoch": 1.2599448601811738,
+      "grad_norm": 6.367788791656494,
+      "learning_rate": 4.102670852984425e-05,
+      "loss": 9.0336,
+      "step": 800
+    },
+    {
+      "epoch": 1.2678219771563608,
+      "grad_norm": 5.573480606079102,
+      "learning_rate": 4.090146714851769e-05,
+      "loss": 8.6367,
+      "step": 805
+    },
+    {
+      "epoch": 1.2756990941315478,
+      "grad_norm": 5.5287861824035645,
+      "learning_rate": 4.077555192100772e-05,
+      "loss": 8.8872,
+      "step": 810
+    },
+    {
+      "epoch": 1.2835762111067348,
+      "grad_norm": 5.406901836395264,
+      "learning_rate": 4.0648968183142424e-05,
+      "loss": 8.9955,
+      "step": 815
+    },
+    {
+      "epoch": 1.291453328081922,
+      "grad_norm": 5.834075927734375,
+      "learning_rate": 4.052172129907897e-05,
+      "loss": 8.9981,
+      "step": 820
+    },
+    {
+      "epoch": 1.299330445057109,
+      "grad_norm": 6.222182273864746,
+      "learning_rate": 4.039381666107621e-05,
+      "loss": 9.2987,
+      "step": 825
+    },
+    {
+      "epoch": 1.3072075620322963,
+      "grad_norm": 5.359075546264648,
+      "learning_rate": 4.026525968926624e-05,
+      "loss": 8.97,
+      "step": 830
+    },
+    {
+      "epoch": 1.3150846790074833,
+      "grad_norm": 7.37460994720459,
+      "learning_rate": 4.013605583142464e-05,
+      "loss": 8.8287,
+      "step": 835
+    },
+    {
+      "epoch": 1.3229617959826703,
+      "grad_norm": 7.033757209777832,
+      "learning_rate": 4.000621056273971e-05,
+      "loss": 8.9608,
+      "step": 840
+    },
+    {
+      "epoch": 1.3308389129578575,
+      "grad_norm": 5.98358154296875,
+      "learning_rate": 3.987572938558035e-05,
+      "loss": 8.7479,
+      "step": 845
+    },
+    {
+      "epoch": 1.3387160299330445,
+      "grad_norm": 8.184402465820312,
+      "learning_rate": 3.974461782926299e-05,
+      "loss": 8.6626,
+      "step": 850
+    },
+    {
+      "epoch": 1.3465931469082315,
+      "grad_norm": 9.955259323120117,
+      "learning_rate": 3.96128814498172e-05,
+      "loss": 8.926,
+      "step": 855
+    },
+    {
+      "epoch": 1.3544702638834187,
+      "grad_norm": 6.086458683013916,
+      "learning_rate": 3.9480525829750295e-05,
+      "loss": 8.6291,
+      "step": 860
+    },
+    {
+      "epoch": 1.3623473808586057,
+      "grad_norm": 6.604151725769043,
+      "learning_rate": 3.9347556577810724e-05,
+      "loss": 8.8916,
+      "step": 865
+    },
+    {
+      "epoch": 1.3702244978337927,
+      "grad_norm": 6.385220050811768,
+      "learning_rate": 3.921397932875044e-05,
+      "loss": 8.6173,
+      "step": 870
+    },
+    {
+      "epoch": 1.37810161480898,
+      "grad_norm": 5.808775424957275,
+      "learning_rate": 3.9079799743086076e-05,
+      "loss": 8.8478,
+      "step": 875
+    },
+    {
+      "epoch": 1.385978731784167,
+      "grad_norm": 5.057977676391602,
+      "learning_rate": 3.8945023506859125e-05,
+      "loss": 8.8043,
+      "step": 880
+    },
+    {
+      "epoch": 1.3938558487593542,
+      "grad_norm": 5.664560794830322,
+      "learning_rate": 3.880965633139493e-05,
+      "loss": 8.517,
+      "step": 885
+    },
+    {
+      "epoch": 1.4017329657345412,
+      "grad_norm": 6.176719665527344,
+      "learning_rate": 3.867370395306068e-05,
+      "loss": 9.4023,
+      "step": 890
+    },
+    {
+      "epoch": 1.4096100827097282,
+      "grad_norm": 5.604133605957031,
+      "learning_rate": 3.8537172133022334e-05,
+      "loss": 8.819,
+      "step": 895
+    },
+    {
+      "epoch": 1.4174871996849152,
+      "grad_norm": 6.113736152648926,
+      "learning_rate": 3.840006665700049e-05,
+      "loss": 8.5058,
+      "step": 900
+    },
+    {
+      "epoch": 1.4253643166601024,
+      "grad_norm": 5.689216136932373,
+      "learning_rate": 3.8262393335025175e-05,
+      "loss": 9.0603,
+      "step": 905
+    },
+    {
+      "epoch": 1.4332414336352894,
+      "grad_norm": 5.811161994934082,
+      "learning_rate": 3.812415800118968e-05,
+      "loss": 8.7633,
+      "step": 910
+    },
+    {
+      "epoch": 1.4411185506104767,
+      "grad_norm": 6.930866241455078,
+      "learning_rate": 3.7985366513403275e-05,
+      "loss": 8.8962,
+      "step": 915
+    },
+    {
+      "epoch": 1.4489956675856637,
+      "grad_norm": 4.998627662658691,
+      "learning_rate": 3.784602475314303e-05,
+      "loss": 8.8133,
+      "step": 920
+    },
+    {
+      "epoch": 1.4568727845608507,
+      "grad_norm": 5.640780925750732,
+      "learning_rate": 3.7706138625204565e-05,
+      "loss": 8.5275,
+      "step": 925
+    },
+    {
+      "epoch": 1.4647499015360377,
+      "grad_norm": 6.085617542266846,
+      "learning_rate": 3.756571405745178e-05,
+      "loss": 8.2861,
+      "step": 930
+    },
+    {
+      "epoch": 1.472627018511225,
+      "grad_norm": 6.3180108070373535,
+      "learning_rate": 3.7424757000565734e-05,
+      "loss": 8.8205,
+      "step": 935
+    },
+    {
+      "epoch": 1.480504135486412,
+      "grad_norm": 7.119158744812012,
+      "learning_rate": 3.7283273427792407e-05,
+      "loss": 8.8586,
+      "step": 940
+    },
+    {
+      "epoch": 1.4883812524615991,
+      "grad_norm": 7.279252529144287,
+      "learning_rate": 3.714126933468959e-05,
+      "loss": 8.2839,
+      "step": 945
+    },
+    {
+      "epoch": 1.4962583694367861,
+      "grad_norm": 5.607680320739746,
+      "learning_rate": 3.6998750738872876e-05,
+      "loss": 9.2545,
+      "step": 950
+    },
+    {
+      "epoch": 1.5041354864119731,
+      "grad_norm": 5.1762590408325195,
+      "learning_rate": 3.6855723679760545e-05,
+      "loss": 9.3281,
+      "step": 955
+    },
+    {
+      "epoch": 1.5120126033871601,
+      "grad_norm": 5.414069652557373,
+      "learning_rate": 3.671219421831775e-05,
+      "loss": 8.5181,
+      "step": 960
+    },
+    {
+      "epoch": 1.5198897203623474,
+      "grad_norm": 6.324887275695801,
+      "learning_rate": 3.656816843679959e-05,
+      "loss": 9.1248,
+      "step": 965
+    },
+    {
+      "epoch": 1.5277668373375346,
+      "grad_norm": 6.40478515625,
+      "learning_rate": 3.64236524384934e-05,
+      "loss": 9.0107,
+      "step": 970
+    },
+    {
+      "epoch": 1.5356439543127216,
+      "grad_norm": 6.342474937438965,
+      "learning_rate": 3.627865234746014e-05,
+      "loss": 8.8478,
+      "step": 975
+    },
+    {
+      "epoch": 1.5435210712879086,
+      "grad_norm": 5.396797180175781,
+      "learning_rate": 3.6133174308274826e-05,
+      "loss": 8.8895,
+      "step": 980
+    },
+    {
+      "epoch": 1.5513981882630956,
+      "grad_norm": 5.2031378746032715,
+      "learning_rate": 3.598722448576619e-05,
+      "loss": 8.6624,
+      "step": 985
+    },
+    {
+      "epoch": 1.5592753052382828,
+      "grad_norm": 6.283775329589844,
+      "learning_rate": 3.584080906475542e-05,
+      "loss": 8.7416,
+      "step": 990
+    },
+    {
+      "epoch": 1.5671524222134698,
+      "grad_norm": 5.148292064666748,
+      "learning_rate": 3.569393424979405e-05,
+      "loss": 8.5302,
+      "step": 995
+    },
+    {
+      "epoch": 1.575029539188657,
+      "grad_norm": 5.293840408325195,
+      "learning_rate": 3.554660626490109e-05,
+      "loss": 8.4594,
+      "step": 1000
+    },
+    {
+      "epoch": 1.582906656163844,
+      "grad_norm": 4.885067462921143,
+      "learning_rate": 3.53988313532992e-05,
+      "loss": 8.5849,
+      "step": 1005
+    },
+    {
+      "epoch": 1.590783773139031,
+      "grad_norm": 5.12977933883667,
+      "learning_rate": 3.52506157771502e-05,
+      "loss": 8.2942,
+      "step": 1010
+    },
+    {
+      "epoch": 1.598660890114218,
+      "grad_norm": 6.016343593597412,
+      "learning_rate": 3.510196581728963e-05,
+      "loss": 8.5103,
+      "step": 1015
+    },
+    {
+      "epoch": 1.6065380070894053,
+      "grad_norm": 6.179079055786133,
+      "learning_rate": 3.495288777296067e-05,
+      "loss": 8.3985,
+      "step": 1020
+    },
+    {
+      "epoch": 1.6144151240645923,
+      "grad_norm": 6.393824100494385,
+      "learning_rate": 3.480338796154712e-05,
+      "loss": 8.5286,
+      "step": 1025
+    },
+    {
+      "epoch": 1.6222922410397795,
+      "grad_norm": 6.632625102996826,
+      "learning_rate": 3.4653472718305745e-05,
+      "loss": 8.713,
+      "step": 1030
+    },
+    {
+      "epoch": 1.6301693580149665,
+      "grad_norm": 5.532236576080322,
+      "learning_rate": 3.450314839609781e-05,
+      "loss": 8.6,
+      "step": 1035
+    },
+    {
+      "epoch": 1.6380464749901535,
+      "grad_norm": 5.10947322845459,
+      "learning_rate": 3.435242136511984e-05,
+      "loss": 8.7229,
+      "step": 1040
+    },
+    {
+      "epoch": 1.6459235919653405,
+      "grad_norm": 6.765851020812988,
+      "learning_rate": 3.420129801263367e-05,
+      "loss": 8.7038,
+      "step": 1045
+    },
+    {
+      "epoch": 1.6538007089405278,
+      "grad_norm": 6.051114082336426,
+      "learning_rate": 3.404978474269583e-05,
+      "loss": 8.9874,
+      "step": 1050
+    },
+    {
+      "epoch": 1.661677825915715,
+      "grad_norm": 5.545468807220459,
+      "learning_rate": 3.389788797588611e-05,
+      "loss": 8.7562,
+      "step": 1055
+    },
+    {
+      "epoch": 1.669554942890902,
+      "grad_norm": 5.190369129180908,
+      "learning_rate": 3.374561414903552e-05,
+      "loss": 8.7141,
+      "step": 1060
+    },
+    {
+      "epoch": 1.677432059866089,
+      "grad_norm": 5.89268684387207,
+      "learning_rate": 3.3592969714953464e-05,
+      "loss": 8.5692,
+      "step": 1065
+    },
+    {
+      "epoch": 1.685309176841276,
+      "grad_norm": 5.843109130859375,
+      "learning_rate": 3.343996114215436e-05,
+      "loss": 8.7083,
+      "step": 1070
+    },
+    {
+      "epoch": 1.693186293816463,
+      "grad_norm": 7.01223087310791,
+      "learning_rate": 3.328659491458348e-05,
+      "loss": 8.9503,
+      "step": 1075
+    },
+    {
+      "epoch": 1.7010634107916502,
+      "grad_norm": 5.1653666496276855,
+      "learning_rate": 3.31328775313422e-05,
+      "loss": 8.6678,
+      "step": 1080
+    },
+    {
+      "epoch": 1.7089405277668375,
+      "grad_norm": 4.834114074707031,
+      "learning_rate": 3.297881550641261e-05,
+      "loss": 9.2416,
+      "step": 1085
+    },
+    {
+      "epoch": 1.7168176447420245,
+      "grad_norm": 5.249499320983887,
+      "learning_rate": 3.282441536838144e-05,
+      "loss": 8.4502,
+      "step": 1090
+    },
+    {
+      "epoch": 1.7246947617172115,
+      "grad_norm": 5.539346694946289,
+      "learning_rate": 3.266968366016342e-05,
+      "loss": 8.7632,
+      "step": 1095
+    },
+    {
+      "epoch": 1.7325718786923985,
+      "grad_norm": 5.473989963531494,
+      "learning_rate": 3.251462693872403e-05,
+      "loss": 8.6672,
+      "step": 1100
+    },
+    {
+      "epoch": 1.7404489956675857,
+      "grad_norm": 5.088748455047607,
+      "learning_rate": 3.2359251774801616e-05,
+      "loss": 8.3875,
+      "step": 1105
+    },
+    {
+      "epoch": 1.7483261126427727,
+      "grad_norm": 5.558929920196533,
+      "learning_rate": 3.2203564752628956e-05,
+      "loss": 8.9813,
+      "step": 1110
+    },
+    {
+      "epoch": 1.75620322961796,
+      "grad_norm": 5.3655548095703125,
+      "learning_rate": 3.204757246965424e-05,
+      "loss": 8.9908,
+      "step": 1115
+    },
+    {
+      "epoch": 1.764080346593147,
+      "grad_norm": 6.439647197723389,
+      "learning_rate": 3.18912815362615e-05,
+      "loss": 8.1854,
+      "step": 1120
+    },
+    {
+      "epoch": 1.771957463568334,
+      "grad_norm": 5.290548324584961,
+      "learning_rate": 3.173469857549048e-05,
+      "loss": 8.9399,
+      "step": 1125
+    },
+    {
+      "epoch": 1.779834580543521,
+      "grad_norm": 5.242992401123047,
+      "learning_rate": 3.157783022275599e-05,
+      "loss": 8.6978,
+      "step": 1130
+    },
+    {
+      "epoch": 1.7877116975187082,
+      "grad_norm": 5.243824481964111,
+      "learning_rate": 3.14206831255667e-05,
+      "loss": 8.5899,
+      "step": 1135
+    },
+    {
+      "epoch": 1.7955888144938952,
+      "grad_norm": 6.161407947540283,
+      "learning_rate": 3.126326394324346e-05,
+      "loss": 8.4425,
+      "step": 1140
+    },
+    {
+      "epoch": 1.8034659314690824,
+      "grad_norm": 5.396100997924805,
+      "learning_rate": 3.110557934663708e-05,
+      "loss": 8.4716,
+      "step": 1145
+    },
+    {
+      "epoch": 1.8113430484442694,
+      "grad_norm": 5.102015972137451,
+      "learning_rate": 3.09476360178457e-05,
+      "loss": 8.396,
+      "step": 1150
+    },
+    {
+      "epoch": 1.8192201654194564,
+      "grad_norm": 5.385177135467529,
+      "learning_rate": 3.078944064993152e-05,
+      "loss": 9.0414,
+      "step": 1155
+    },
+    {
+      "epoch": 1.8270972823946434,
+      "grad_norm": 5.5736165046691895,
+      "learning_rate": 3.063099994663731e-05,
+      "loss": 8.5122,
+      "step": 1160
+    },
+    {
+      "epoch": 1.8349743993698306,
+      "grad_norm": 5.4414286613464355,
+      "learning_rate": 3.0472320622102207e-05,
+      "loss": 8.3838,
+      "step": 1165
+    },
+    {
+      "epoch": 1.8428515163450179,
+      "grad_norm": 5.483443737030029,
+      "learning_rate": 3.0313409400577264e-05,
+      "loss": 8.0776,
+      "step": 1170
+    },
+    {
+      "epoch": 1.8507286333202049,
+      "grad_norm": 5.481230735778809,
+      "learning_rate": 3.0154273016140504e-05,
+      "loss": 8.5493,
+      "step": 1175
+    },
+    {
+      "epoch": 1.8586057502953919,
+      "grad_norm": 5.8650312423706055,
+      "learning_rate": 2.999491821241148e-05,
+      "loss": 8.9861,
+      "step": 1180
+    },
+    {
+      "epoch": 1.8664828672705789,
+      "grad_norm": 5.712615013122559,
+      "learning_rate": 2.9835351742265622e-05,
+      "loss": 8.5673,
+      "step": 1185
+    },
+    {
+      "epoch": 1.8743599842457659,
+      "grad_norm": 5.285681247711182,
+      "learning_rate": 2.967558036754796e-05,
+      "loss": 8.4617,
+      "step": 1190
+    },
+    {
+      "epoch": 1.8822371012209531,
+      "grad_norm": 5.035665035247803,
+      "learning_rate": 2.9515610858786657e-05,
+      "loss": 8.1492,
+      "step": 1195
+    },
+    {
+      "epoch": 1.8901142181961403,
+      "grad_norm": 5.165551662445068,
+      "learning_rate": 2.935544999490607e-05,
+      "loss": 8.599,
+      "step": 1200
+    },
+    {
+      "epoch": 1.8979913351713273,
+      "grad_norm": 5.6296610832214355,
+      "learning_rate": 2.9195104562939474e-05,
+      "loss": 8.6841,
+      "step": 1205
+    },
+    {
+      "epoch": 1.9058684521465143,
+      "grad_norm": 5.5025634765625,
+      "learning_rate": 2.90345813577415e-05,
+      "loss": 8.4796,
+      "step": 1210
+    },
+    {
+      "epoch": 1.9137455691217014,
+      "grad_norm": 5.420698642730713,
+      "learning_rate": 2.8873887181700127e-05,
+      "loss": 8.134,
+      "step": 1215
+    },
+    {
+      "epoch": 1.9216226860968886,
+      "grad_norm": 5.495443344116211,
+      "learning_rate": 2.8713028844448464e-05,
+      "loss": 8.4425,
+      "step": 1220
+    },
+    {
+      "epoch": 1.9294998030720756,
+      "grad_norm": 5.478509902954102,
+      "learning_rate": 2.8552013162576186e-05,
+      "loss": 8.2366,
+      "step": 1225
+    },
+    {
+      "epoch": 1.9373769200472628,
+      "grad_norm": 5.962779521942139,
+      "learning_rate": 2.8390846959340638e-05,
+      "loss": 8.4159,
+      "step": 1230
+    },
+    {
+      "epoch": 1.9452540370224498,
+      "grad_norm": 7.839330673217773,
+      "learning_rate": 2.8229537064377735e-05,
+      "loss": 8.2683,
+      "step": 1235
+    },
+    {
+      "epoch": 1.9531311539976368,
+      "grad_norm": 5.109673023223877,
+      "learning_rate": 2.806809031341251e-05,
+      "loss": 8.6595,
+      "step": 1240
+    },
+    {
+      "epoch": 1.9610082709728238,
+      "grad_norm": 5.204682350158691,
+      "learning_rate": 2.790651354796945e-05,
+      "loss": 8.6676,
+      "step": 1245
+    },
+    {
+      "epoch": 1.968885387948011,
+      "grad_norm": 7.151166915893555,
+      "learning_rate": 2.7744813615082588e-05,
+      "loss": 8.2718,
+      "step": 1250
+    },
+    {
+      "epoch": 1.9767625049231983,
+      "grad_norm": 5.639292240142822,
+      "learning_rate": 2.7582997367005337e-05,
+      "loss": 8.4982,
+      "step": 1255
+    },
+    {
+      "epoch": 1.9846396218983853,
+      "grad_norm": 7.254332065582275,
+      "learning_rate": 2.742107166092015e-05,
+      "loss": 8.3822,
+      "step": 1260
+    },
+    {
+      "epoch": 1.9925167388735723,
+      "grad_norm": 5.9872307777404785,
+      "learning_rate": 2.7259043358647863e-05,
+      "loss": 8.5006,
+      "step": 1265
+    },
+    {
+      "epoch": 2.0015754233950376,
+      "grad_norm": 4.944819450378418,
+      "learning_rate": 2.709691932635701e-05,
+      "loss": 9.5234,
+      "step": 1270
+    },
+    {
+      "epoch": 2.0094525403702246,
+      "grad_norm": 4.903031349182129,
+      "learning_rate": 2.6934706434272794e-05,
+      "loss": 8.1337,
+      "step": 1275
+    },
+    {
+      "epoch": 2.0173296573454116,
+      "grad_norm": 5.022507667541504,
+      "learning_rate": 2.6772411556385966e-05,
+      "loss": 8.5159,
+      "step": 1280
+    },
+    {
+      "epoch": 2.0252067743205986,
+      "grad_norm": 5.7532525062561035,
+      "learning_rate": 2.6610041570161563e-05,
+      "loss": 7.8905,
+      "step": 1285
+    },
+    {
+      "epoch": 2.0330838912957856,
+      "grad_norm": 5.3462653160095215,
+      "learning_rate": 2.6447603356247396e-05,
+      "loss": 8.5373,
+      "step": 1290
+    },
+    {
+      "epoch": 2.0409610082709726,
+      "grad_norm": 5.146194934844971,
+      "learning_rate": 2.628510379818256e-05,
+      "loss": 8.375,
+      "step": 1295
+    },
+    {
+      "epoch": 2.04883812524616,
+      "grad_norm": 4.806943893432617,
+      "learning_rate": 2.6122549782105666e-05,
+      "loss": 8.4479,
+      "step": 1300
+    },
+    {
+      "epoch": 2.056715242221347,
+      "grad_norm": 6.281928062438965,
+      "learning_rate": 2.5959948196463058e-05,
+      "loss": 8.1086,
+      "step": 1305
+    },
+    {
+      "epoch": 2.064592359196534,
+      "grad_norm": 5.9106550216674805,
+      "learning_rate": 2.5797305931716924e-05,
+      "loss": 8.3172,
+      "step": 1310
+    },
+    {
+      "epoch": 2.072469476171721,
+      "grad_norm": 5.928641319274902,
+      "learning_rate": 2.5634629880053256e-05,
+      "loss": 8.3275,
+      "step": 1315
+    },
+    {
+      "epoch": 2.080346593146908,
+      "grad_norm": 5.428690433502197,
+      "learning_rate": 2.5471926935089835e-05,
+      "loss": 8.5239,
+      "step": 1320
+    },
+    {
+      "epoch": 2.0882237101220955,
+      "grad_norm": 5.30228853225708,
+      "learning_rate": 2.5309203991584073e-05,
+      "loss": 8.3246,
+      "step": 1325
+    },
+    {
+      "epoch": 2.0961008270972825,
+      "grad_norm": 5.126616477966309,
+      "learning_rate": 2.514646794514085e-05,
+      "loss": 7.8035,
+      "step": 1330
+    },
+    {
+      "epoch": 2.1039779440724695,
+      "grad_norm": 5.645951747894287,
+      "learning_rate": 2.4983725691920294e-05,
+      "loss": 8.3698,
+      "step": 1335
+    },
+    {
+      "epoch": 2.1118550610476565,
+      "grad_norm": 5.491515636444092,
+      "learning_rate": 2.482098412834556e-05,
+      "loss": 7.826,
+      "step": 1340
+    },
+    {
+      "epoch": 2.1197321780228435,
+      "grad_norm": 5.176724910736084,
+      "learning_rate": 2.4658250150810578e-05,
+      "loss": 8.0034,
+      "step": 1345
+    },
+    {
+      "epoch": 2.1276092949980305,
+      "grad_norm": 4.87959098815918,
+      "learning_rate": 2.449553065538781e-05,
+      "loss": 8.2813,
+      "step": 1350
+    },
+    {
+      "epoch": 2.135486411973218,
+      "grad_norm": 5.171248435974121,
+      "learning_rate": 2.4332832537536e-05,
+      "loss": 8.4077,
+      "step": 1355
+    },
+    {
+      "epoch": 2.143363528948405,
+      "grad_norm": 5.680236339569092,
+      "learning_rate": 2.4170162691808034e-05,
+      "loss": 8.2604,
+      "step": 1360
+    },
+    {
+      "epoch": 2.151240645923592,
+      "grad_norm": 5.374014854431152,
+      "learning_rate": 2.4007528011558683e-05,
+      "loss": 8.0889,
+      "step": 1365
+    },
+    {
+      "epoch": 2.159117762898779,
+      "grad_norm": 5.249903678894043,
+      "learning_rate": 2.3844935388652545e-05,
+      "loss": 8.3831,
+      "step": 1370
+    },
+    {
+      "epoch": 2.166994879873966,
+      "grad_norm": 5.103092670440674,
+      "learning_rate": 2.3682391713171998e-05,
+      "loss": 8.314,
+      "step": 1375
+    },
+    {
+      "epoch": 2.174871996849153,
+      "grad_norm": 5.576798915863037,
+      "learning_rate": 2.351990387312519e-05,
+      "loss": 7.9968,
+      "step": 1380
+    },
+    {
+      "epoch": 2.1827491138243404,
+      "grad_norm": 5.228267192840576,
+      "learning_rate": 2.3357478754154173e-05,
+      "loss": 8.3819,
+      "step": 1385
+    },
+    {
+      "epoch": 2.1906262307995275,
+      "grad_norm": 4.981462478637695,
+      "learning_rate": 2.3195123239243098e-05,
+      "loss": 8.5833,
+      "step": 1390
+    },
+    {
+      "epoch": 2.1985033477747145,
+      "grad_norm": 5.161903381347656,
+      "learning_rate": 2.303284420842658e-05,
+      "loss": 8.0129,
+      "step": 1395
+    },
+    {
+      "epoch": 2.2063804647499015,
+      "grad_norm": 5.56730842590332,
+      "learning_rate": 2.2870648538498085e-05,
+      "loss": 8.2146,
+      "step": 1400
+    },
+    {
+      "epoch": 2.2142575817250885,
+      "grad_norm": 5.310222625732422,
+      "learning_rate": 2.2708543102718542e-05,
+      "loss": 8.3771,
+      "step": 1405
+    },
+    {
+      "epoch": 2.222134698700276,
+      "grad_norm": 5.077228546142578,
+      "learning_rate": 2.254653477052514e-05,
+      "loss": 8.0806,
+      "step": 1410
+    },
+    {
+      "epoch": 2.230011815675463,
+      "grad_norm": 5.488475322723389,
+      "learning_rate": 2.238463040724011e-05,
+      "loss": 7.9597,
+      "step": 1415
+    },
+    {
+      "epoch": 2.23788893265065,
+      "grad_norm": 4.879157543182373,
+      "learning_rate": 2.2222836873779888e-05,
+      "loss": 8.1727,
+      "step": 1420
+    },
+    {
+      "epoch": 2.245766049625837,
+      "grad_norm": 5.107478618621826,
+      "learning_rate": 2.2061161026364353e-05,
+      "loss": 8.0123,
+      "step": 1425
+    },
+    {
+      "epoch": 2.253643166601024,
+      "grad_norm": 5.371547222137451,
+      "learning_rate": 2.1899609716226263e-05,
+      "loss": 8.2489,
+      "step": 1430
+    },
+    {
+      "epoch": 2.261520283576211,
+      "grad_norm": 5.334740161895752,
+      "learning_rate": 2.173818978932095e-05,
+      "loss": 8.536,
+      "step": 1435
+    },
+    {
+      "epoch": 2.2693974005513984,
+      "grad_norm": 5.4585394859313965,
+      "learning_rate": 2.1576908086036195e-05,
+      "loss": 8.2132,
+      "step": 1440
+    },
+    {
+      "epoch": 2.2772745175265854,
+      "grad_norm": 6.0415449142456055,
+      "learning_rate": 2.1415771440902384e-05,
+      "loss": 7.9417,
+      "step": 1445
+    },
+    {
+      "epoch": 2.2851516345017724,
+      "grad_norm": 5.975248336791992,
+      "learning_rate": 2.1254786682302847e-05,
+      "loss": 8.2411,
+      "step": 1450
+    },
+    {
+      "epoch": 2.2930287514769594,
+      "grad_norm": 5.171903133392334,
+      "learning_rate": 2.1093960632184513e-05,
+      "loss": 8.4717,
+      "step": 1455
+    },
+    {
+      "epoch": 2.3009058684521464,
+      "grad_norm": 5.419390678405762,
+      "learning_rate": 2.0933300105768867e-05,
+      "loss": 7.9774,
+      "step": 1460
+    },
+    {
+      "epoch": 2.3087829854273334,
+      "grad_norm": 5.009170055389404,
+      "learning_rate": 2.0772811911263048e-05,
+      "loss": 7.9232,
+      "step": 1465
+    },
+    {
+      "epoch": 2.316660102402521,
+      "grad_norm": 5.920315265655518,
+      "learning_rate": 2.0612502849571425e-05,
+      "loss": 8.3842,
+      "step": 1470
+    },
+    {
+      "epoch": 2.324537219377708,
+      "grad_norm": 5.717631816864014,
+      "learning_rate": 2.0452379714007375e-05,
+      "loss": 8.3991,
+      "step": 1475
+    },
+    {
+      "epoch": 2.332414336352895,
+      "grad_norm": 6.118537425994873,
+      "learning_rate": 2.0292449290005395e-05,
+      "loss": 8.1187,
+      "step": 1480
+    },
+    {
+      "epoch": 2.340291453328082,
+      "grad_norm": 4.826847076416016,
+      "learning_rate": 2.013271835483357e-05,
+      "loss": 7.8728,
+      "step": 1485
+    },
+    {
+      "epoch": 2.348168570303269,
+      "grad_norm": 5.45656681060791,
+      "learning_rate": 1.9973193677306377e-05,
+      "loss": 8.1391,
+      "step": 1490
+    },
+    {
+      "epoch": 2.3560456872784563,
+      "grad_norm": 5.478461265563965,
+      "learning_rate": 1.9813882017497848e-05,
+      "loss": 8.1173,
+      "step": 1495
+    },
+    {
+      "epoch": 2.3639228042536433,
+      "grad_norm": 6.264642715454102,
+      "learning_rate": 1.9654790126455103e-05,
+      "loss": 8.1301,
+      "step": 1500
+    },
+    {
+      "epoch": 2.3717999212288303,
+      "grad_norm": 5.903857231140137,
+      "learning_rate": 1.949592474591224e-05,
+      "loss": 7.9252,
+      "step": 1505
+    },
+    {
+      "epoch": 2.3796770382040173,
+      "grad_norm": 4.8199896812438965,
+      "learning_rate": 1.9337292608004715e-05,
+      "loss": 8.3755,
+      "step": 1510
+    },
+    {
+      "epoch": 2.3875541551792043,
+      "grad_norm": 4.588783264160156,
+      "learning_rate": 1.917890043498397e-05,
+      "loss": 8.1026,
+      "step": 1515
+    },
+    {
+      "epoch": 2.3954312721543913,
+      "grad_norm": 5.213608741760254,
+      "learning_rate": 1.9020754938932624e-05,
+      "loss": 8.1211,
+      "step": 1520
+    },
+    {
+      "epoch": 2.4033083891295783,
+      "grad_norm": 5.581782817840576,
+      "learning_rate": 1.8862862821480025e-05,
+      "loss": 7.7893,
+      "step": 1525
+    },
+    {
+      "epoch": 2.411185506104766,
+      "grad_norm": 5.054122447967529,
+      "learning_rate": 1.8705230773518267e-05,
+      "loss": 8.1924,
+      "step": 1530
+    },
+    {
+      "epoch": 2.419062623079953,
+      "grad_norm": 4.978759288787842,
+      "learning_rate": 1.854786547491864e-05,
+      "loss": 8.1678,
+      "step": 1535
+    },
+    {
+      "epoch": 2.42693974005514,
+      "grad_norm": 5.0260329246521,
+      "learning_rate": 1.8390773594248568e-05,
+      "loss": 8.1418,
+      "step": 1540
+    },
+    {
+      "epoch": 2.434816857030327,
+      "grad_norm": 5.8550262451171875,
+      "learning_rate": 1.823396178848903e-05,
+      "loss": 8.2088,
+      "step": 1545
+    },
+    {
+      "epoch": 2.442693974005514,
+      "grad_norm": 6.323260307312012,
+      "learning_rate": 1.8077436702752448e-05,
+      "loss": 7.9444,
+      "step": 1550
+    },
+    {
+      "epoch": 2.4505710909807012,
+      "grad_norm": 5.396198272705078,
+      "learning_rate": 1.7921204970001065e-05,
+      "loss": 7.8484,
+      "step": 1555
+    },
+    {
+      "epoch": 2.4584482079558883,
+      "grad_norm": 5.220623970031738,
+      "learning_rate": 1.776527321076596e-05,
+      "loss": 8.1014,
+      "step": 1560
+    },
+    {
+      "epoch": 2.4663253249310753,
+      "grad_norm": 16.552457809448242,
+      "learning_rate": 1.7609648032866362e-05,
+      "loss": 8.2018,
+      "step": 1565
+    },
+    {
+      "epoch": 2.4742024419062623,
+      "grad_norm": 5.957991600036621,
+      "learning_rate": 1.7454336031129743e-05,
+      "loss": 7.8216,
+      "step": 1570
+    },
+    {
+      "epoch": 2.4820795588814493,
+      "grad_norm": 5.393924713134766,
+      "learning_rate": 1.729934378711228e-05,
+      "loss": 8.3398,
+      "step": 1575
+    },
+    {
+      "epoch": 2.4899566758566367,
+      "grad_norm": 4.859684944152832,
+      "learning_rate": 1.7144677868820026e-05,
+      "loss": 8.6844,
+      "step": 1580
+    },
+    {
+      "epoch": 2.4978337928318237,
+      "grad_norm": 5.323124885559082,
+      "learning_rate": 1.6990344830430498e-05,
+      "loss": 8.0736,
+      "step": 1585
+    },
+    {
+      "epoch": 2.5057109098070107,
+      "grad_norm": 8.831464767456055,
+      "learning_rate": 1.6836351212014996e-05,
+      "loss": 8.1686,
+      "step": 1590
+    },
+    {
+      "epoch": 2.51201260338716,
+      "grad_norm": 5.187417507171631,
+      "learning_rate": 1.6682703539261442e-05,
+      "loss": 7.6429,
+      "step": 1595
+    },
+    {
+      "epoch": 2.5198897203623476,
+      "grad_norm": 6.1180500984191895,
+      "learning_rate": 1.652940832319785e-05,
+      "loss": 7.9255,
+      "step": 1600
+    },
+    {
+      "epoch": 2.5277668373375346,
+      "grad_norm": 8.949240684509277,
+      "learning_rate": 1.6376472059916354e-05,
+      "loss": 8.0755,
+      "step": 1605
+    },
+    {
+      "epoch": 2.5356439543127216,
+      "grad_norm": 4.905629634857178,
+      "learning_rate": 1.6223901230298062e-05,
+      "loss": 7.8017,
+      "step": 1610
+    },
+    {
+      "epoch": 2.5435210712879086,
+      "grad_norm": 5.435753345489502,
+      "learning_rate": 1.6071702299738268e-05,
+      "loss": 7.7271,
+      "step": 1615
+    },
+    {
+      "epoch": 2.5513981882630956,
+      "grad_norm": 5.005056858062744,
+      "learning_rate": 1.5919881717872564e-05,
+      "loss": 8.3681,
+      "step": 1620
+    },
+    {
+      "epoch": 2.559275305238283,
+      "grad_norm": 5.524580001831055,
+      "learning_rate": 1.5768445918303505e-05,
+      "loss": 7.6268,
+      "step": 1625
+    },
+    {
+      "epoch": 2.5671524222134696,
+      "grad_norm": 5.758496284484863,
+      "learning_rate": 1.5617401318327985e-05,
+      "loss": 7.8986,
+      "step": 1630
+    },
+    {
+      "epoch": 2.575029539188657,
+      "grad_norm": 5.624476909637451,
+      "learning_rate": 1.546675431866526e-05,
+      "loss": 7.437,
+      "step": 1635
+    },
+    {
+      "epoch": 2.582906656163844,
+      "grad_norm": 4.7457594871521,
+      "learning_rate": 1.5316511303185755e-05,
+      "loss": 8.1143,
+      "step": 1640
+    },
+    {
+      "epoch": 2.590783773139031,
+      "grad_norm": 5.465321063995361,
+      "learning_rate": 1.5166678638640521e-05,
+      "loss": 7.4633,
+      "step": 1645
+    },
+    {
+      "epoch": 2.598660890114218,
+      "grad_norm": 5.189293384552002,
+      "learning_rate": 1.5017262674391428e-05,
+      "loss": 7.7685,
+      "step": 1650
+    },
+    {
+      "epoch": 2.606538007089405,
+      "grad_norm": 5.231342792510986,
+      "learning_rate": 1.486826974214209e-05,
+      "loss": 7.6534,
+      "step": 1655
+    },
+    {
+      "epoch": 2.6144151240645925,
+      "grad_norm": 5.773874282836914,
+      "learning_rate": 1.4719706155669587e-05,
+      "loss": 7.7808,
+      "step": 1660
+    },
+    {
+      "epoch": 2.6222922410397795,
+      "grad_norm": 5.353034019470215,
+      "learning_rate": 1.4571578210556896e-05,
+      "loss": 7.7763,
+      "step": 1665
+    },
+    {
+      "epoch": 2.6301693580149665,
+      "grad_norm": 5.128270149230957,
+      "learning_rate": 1.4423892183926092e-05,
+      "loss": 7.8019,
+      "step": 1670
+    },
+    {
+      "epoch": 2.6380464749901535,
+      "grad_norm": 5.596746444702148,
+      "learning_rate": 1.4276654334172332e-05,
+      "loss": 7.3829,
+      "step": 1675
+    },
+    {
+      "epoch": 2.6459235919653405,
+      "grad_norm": 4.946563720703125,
+      "learning_rate": 1.4129870900698739e-05,
+      "loss": 8.053,
+      "step": 1680
+    },
+    {
+      "epoch": 2.653800708940528,
+      "grad_norm": 5.08951473236084,
+      "learning_rate": 1.3983548103651861e-05,
+      "loss": 7.9546,
+      "step": 1685
+    },
+    {
+      "epoch": 2.661677825915715,
+      "grad_norm": 6.104579925537109,
+      "learning_rate": 1.3837692143658182e-05,
+      "loss": 7.8367,
+      "step": 1690
+    },
+    {
+      "epoch": 2.669554942890902,
+      "grad_norm": 5.49988317489624,
+      "learning_rate": 1.369230920156134e-05,
+      "loss": 7.8629,
+      "step": 1695
+    },
+    {
+      "epoch": 2.677432059866089,
+      "grad_norm": 5.227181434631348,
+      "learning_rate": 1.3547405438160205e-05,
+      "loss": 8.1967,
+      "step": 1700
+    },
+    {
+      "epoch": 2.685309176841276,
+      "grad_norm": 6.716334819793701,
+      "learning_rate": 1.340298699394777e-05,
+      "loss": 7.4787,
+      "step": 1705
+    },
+    {
+      "epoch": 2.693186293816463,
+      "grad_norm": 5.50681209564209,
+      "learning_rate": 1.3259059988851e-05,
+      "loss": 7.2509,
+      "step": 1710
+    },
+    {
+      "epoch": 2.70106341079165,
+      "grad_norm": 4.969735145568848,
+      "learning_rate": 1.3115630521971468e-05,
+      "loss": 7.8345,
+      "step": 1715
+    },
+    {
+      "epoch": 2.7089405277668375,
+      "grad_norm": 5.385042190551758,
+      "learning_rate": 1.2972704671326863e-05,
+      "loss": 7.6157,
+      "step": 1720
+    },
+    {
+      "epoch": 2.7168176447420245,
+      "grad_norm": 5.288397789001465,
+      "learning_rate": 1.2830288493593451e-05,
+      "loss": 7.6661,
+      "step": 1725
+    },
+    {
+      "epoch": 2.7246947617172115,
+      "grad_norm": 5.0901994705200195,
+      "learning_rate": 1.2688388023849482e-05,
+      "loss": 7.8334,
+      "step": 1730
+    },
+    {
+      "epoch": 2.7325718786923985,
+      "grad_norm": 5.180675983428955,
+      "learning_rate": 1.2547009275319313e-05,
+      "loss": 7.3121,
+      "step": 1735
+    },
+    {
+      "epoch": 2.7404489956675855,
+      "grad_norm": 7.612630367279053,
+      "learning_rate": 1.2406158239118677e-05,
+      "loss": 7.5115,
+      "step": 1740
+    },
+    {
+      "epoch": 2.748326112642773,
+      "grad_norm": 9.157015800476074,
+      "learning_rate": 1.2265840884000792e-05,
+      "loss": 7.7749,
+      "step": 1745
+    },
+    {
+      "epoch": 2.75620322961796,
+      "grad_norm": 5.494700908660889,
+      "learning_rate": 1.2126063156103437e-05,
+      "loss": 7.8249,
+      "step": 1750
+    },
+    {
+      "epoch": 2.764080346593147,
+      "grad_norm": 4.8552985191345215,
+      "learning_rate": 1.1986830978696908e-05,
+      "loss": 7.3518,
+      "step": 1755
+    },
+    {
+      "epoch": 2.771957463568334,
+      "grad_norm": 5.416152477264404,
+      "learning_rate": 1.1848150251933102e-05,
+      "loss": 7.7477,
+      "step": 1760
+    },
+    {
+      "epoch": 2.779834580543521,
+      "grad_norm": 4.609814643859863,
+      "learning_rate": 1.1710026852595446e-05,
+      "loss": 7.7182,
+      "step": 1765
+    },
+    {
+      "epoch": 2.7877116975187084,
+      "grad_norm": 4.770631790161133,
+      "learning_rate": 1.157246663384984e-05,
+      "loss": 7.5309,
+      "step": 1770
+    },
+    {
+      "epoch": 2.795588814493895,
+      "grad_norm": 5.455694675445557,
+      "learning_rate": 1.1435475424996644e-05,
+      "loss": 7.4731,
+      "step": 1775
+    },
+    {
+      "epoch": 2.8034659314690824,
+      "grad_norm": 5.270656585693359,
+      "learning_rate": 1.1299059031223706e-05,
+      "loss": 7.6206,
+      "step": 1780
+    },
+    {
+      "epoch": 2.8113430484442694,
+      "grad_norm": 4.653499126434326,
+      "learning_rate": 1.1163223233360243e-05,
+      "loss": 7.9029,
+      "step": 1785
+    },
+    {
+      "epoch": 2.8192201654194564,
+      "grad_norm": 5.981932163238525,
+      "learning_rate": 1.1027973787631956e-05,
+      "loss": 7.6331,
+      "step": 1790
+    },
+    {
+      "epoch": 2.8270972823946434,
+      "grad_norm": 6.355517864227295,
+      "learning_rate": 1.0893316425417097e-05,
+      "loss": 7.6632,
+      "step": 1795
+    },
+    {
+      "epoch": 2.8349743993698304,
+      "grad_norm": 7.381013870239258,
+      "learning_rate": 1.0759256853003578e-05,
+      "loss": 7.5561,
+      "step": 1800
+    },
+    {
+      "epoch": 2.842851516345018,
+      "grad_norm": 6.400808811187744,
+      "learning_rate": 1.062580075134712e-05,
+      "loss": 7.1496,
+      "step": 1805
+    },
+    {
+      "epoch": 2.850728633320205,
+      "grad_norm": 6.144219875335693,
+      "learning_rate": 1.0492953775830603e-05,
+      "loss": 7.6644,
+      "step": 1810
+    },
+    {
+      "epoch": 2.858605750295392,
+      "grad_norm": 4.925227642059326,
+      "learning_rate": 1.0360721556024348e-05,
+      "loss": 7.9462,
+      "step": 1815
+    },
+    {
+      "epoch": 2.866482867270579,
+      "grad_norm": 6.179694175720215,
+      "learning_rate": 1.0229109695447553e-05,
+      "loss": 7.6671,
+      "step": 1820
+    },
+    {
+      "epoch": 2.874359984245766,
+      "grad_norm": 5.008157730102539,
+      "learning_rate": 1.0098123771330853e-05,
+      "loss": 7.9779,
+      "step": 1825
+    },
+    {
+      "epoch": 2.8822371012209533,
+      "grad_norm": 5.26678466796875,
+      "learning_rate": 9.967769334380029e-06,
+      "loss": 7.4016,
+      "step": 1830
+    },
+    {
+      "epoch": 2.8901142181961403,
+      "grad_norm": 5.3563151359558105,
+      "learning_rate": 9.83805190854068e-06,
+      "loss": 7.6818,
+      "step": 1835
+    },
+    {
+      "epoch": 2.8979913351713273,
+      "grad_norm": 4.609752178192139,
+      "learning_rate": 9.708976990764212e-06,
+      "loss": 7.826,
+      "step": 1840
+    },
+    {
+      "epoch": 2.9058684521465143,
+      "grad_norm": 4.820992946624756,
+      "learning_rate": 9.580550050774906e-06,
+      "loss": 7.7669,
+      "step": 1845
+    },
+    {
+      "epoch": 2.9137455691217014,
+      "grad_norm": 5.011669635772705,
+      "learning_rate": 9.452776530838096e-06,
+      "loss": 7.8166,
+      "step": 1850
+    },
+    {
+      "epoch": 2.921622686096889,
+      "grad_norm": 5.413022041320801,
+      "learning_rate": 9.325661845529533e-06,
+      "loss": 7.3276,
+      "step": 1855
+    },
+    {
+      "epoch": 2.9294998030720754,
+      "grad_norm": 5.638709545135498,
+      "learning_rate": 9.199211381505993e-06,
+      "loss": 7.6476,
+      "step": 1860
+    },
+    {
+      "epoch": 2.937376920047263,
+      "grad_norm": 6.355219841003418,
+      "learning_rate": 9.073430497276969e-06,
+      "loss": 7.5497,
+      "step": 1865
+    },
+    {
+      "epoch": 2.94525403702245,
+      "grad_norm": 4.824697494506836,
+      "learning_rate": 8.94832452297759e-06,
+      "loss": 7.4419,
+      "step": 1870
+    },
+    {
+      "epoch": 2.953131153997637,
+      "grad_norm": 10.201951026916504,
+      "learning_rate": 8.823898760142759e-06,
+      "loss": 7.8972,
+      "step": 1875
+    },
+    {
+      "epoch": 2.961008270972824,
+      "grad_norm": 5.421176433563232,
+      "learning_rate": 8.700158481482543e-06,
+      "loss": 7.5122,
+      "step": 1880
+    },
+    {
+      "epoch": 2.968885387948011,
+      "grad_norm": 4.769680500030518,
+      "learning_rate": 8.577108930658636e-06,
+      "loss": 7.4867,
+      "step": 1885
+    },
+    {
+      "epoch": 2.9767625049231983,
+      "grad_norm": 4.7165727615356445,
+      "learning_rate": 8.454755322062228e-06,
+      "loss": 7.9242,
+      "step": 1890
+    },
+    {
+      "epoch": 2.9846396218983853,
+      "grad_norm": 6.179513931274414,
+      "learning_rate": 8.333102840593015e-06,
+      "loss": 7.585,
+      "step": 1895
+    },
+    {
+      "epoch": 2.9925167388735723,
+      "grad_norm": 5.120509624481201,
+      "learning_rate": 8.212156641439495e-06,
+      "loss": 7.6742,
+      "step": 1900
+    },
+    {
+      "epoch": 3.0015754233950376,
+      "grad_norm": 5.389956951141357,
+      "learning_rate": 8.091921849860466e-06,
+      "loss": 8.6246,
+      "step": 1905
+    },
+    {
+      "epoch": 3.007877116975187,
+      "grad_norm": 5.342672348022461,
+      "learning_rate": 7.972403560967895e-06,
+      "loss": 8.6445,
+      "step": 1910
+    },
+    {
+      "epoch": 3.015754233950374,
+      "grad_norm": 5.166170597076416,
+      "learning_rate": 7.853606839510976e-06,
+      "loss": 8.1686,
+      "step": 1915
+    },
+    {
+      "epoch": 3.0236313509255615,
+      "grad_norm": 4.708956241607666,
+      "learning_rate": 7.735536719661495e-06,
+      "loss": 8.1996,
+      "step": 1920
+    },
+    {
+      "epoch": 3.0315084679007485,
+      "grad_norm": 5.225474834442139,
+      "learning_rate": 7.618198204800509e-06,
+      "loss": 8.5347,
+      "step": 1925
+    },
+    {
+      "epoch": 3.0393855848759355,
+      "grad_norm": 4.633377552032471,
+      "learning_rate": 7.5015962673063375e-06,
+      "loss": 8.4879,
+      "step": 1930
+    },
+    {
+      "epoch": 3.0472627018511225,
+      "grad_norm": 6.010617256164551,
+      "learning_rate": 7.3857358483438385e-06,
+      "loss": 8.3143,
+      "step": 1935
+    },
+    {
+      "epoch": 3.0551398188263095,
+      "grad_norm": 5.735445022583008,
+      "learning_rate": 7.270621857655002e-06,
+      "loss": 8.2262,
+      "step": 1940
+    },
+    {
+      "epoch": 3.0630169358014965,
+      "grad_norm": 5.311731815338135,
+      "learning_rate": 7.156259173350926e-06,
+      "loss": 8.3875,
+      "step": 1945
+    },
+    {
+      "epoch": 3.070894052776684,
+      "grad_norm": 4.930365562438965,
+      "learning_rate": 7.042652641705083e-06,
+      "loss": 8.2483,
+      "step": 1950
+    },
+    {
+      "epoch": 3.078771169751871,
+      "grad_norm": 4.509944915771484,
+      "learning_rate": 6.9298070769479476e-06,
+      "loss": 8.1581,
+      "step": 1955
+    },
+    {
+      "epoch": 3.086648286727058,
+      "grad_norm": 6.027124404907227,
+      "learning_rate": 6.817727261062973e-06,
+      "loss": 8.4293,
+      "step": 1960
+    },
+    {
+      "epoch": 3.094525403702245,
+      "grad_norm": 4.761337757110596,
+      "learning_rate": 6.706417943584023e-06,
+      "loss": 8.0434,
+      "step": 1965
+    },
+    {
+      "epoch": 3.102402520677432,
+      "grad_norm": 4.544243335723877,
+      "learning_rate": 6.5958838413940016e-06,
+      "loss": 8.3465,
+      "step": 1970
+    },
+    {
+      "epoch": 3.110279637652619,
+      "grad_norm": 4.9786882400512695,
+      "learning_rate": 6.486129638525021e-06,
+      "loss": 8.0994,
+      "step": 1975
+    },
+    {
+      "epoch": 3.1181567546278064,
+      "grad_norm": 5.5463972091674805,
+      "learning_rate": 6.377159985959921e-06,
+      "loss": 7.9211,
+      "step": 1980
+    },
+    {
+      "epoch": 3.1260338716029934,
+      "grad_norm": 4.427308082580566,
+      "learning_rate": 6.2689795014351595e-06,
+      "loss": 8.2482,
+      "step": 1985
+    },
+    {
+      "epoch": 3.1339109885781804,
+      "grad_norm": 4.5052642822265625,
+      "learning_rate": 6.161592769245114e-06,
+      "loss": 8.1749,
+      "step": 1990
+    },
+    {
+      "epoch": 3.1417881055533674,
+      "grad_norm": 5.253861904144287,
+      "learning_rate": 6.055004340047849e-06,
+      "loss": 8.2287,
+      "step": 1995
+    },
+    {
+      "epoch": 3.1496652225285544,
+      "grad_norm": 8.412028312683105,
+      "learning_rate": 5.949218730672259e-06,
+      "loss": 7.489,
+      "step": 2000
+    },
+    {
+      "epoch": 3.157542339503742,
+      "grad_norm": 6.002096652984619,
+      "learning_rate": 5.844240423926659e-06,
+      "loss": 7.7578,
+      "step": 2005
+    },
+    {
+      "epoch": 3.165419456478929,
+      "grad_norm": 5.330885410308838,
+      "learning_rate": 5.740073868408799e-06,
+      "loss": 7.6986,
+      "step": 2010
+    },
+    {
+      "epoch": 3.173296573454116,
+      "grad_norm": 6.712745666503906,
+      "learning_rate": 5.636723478317429e-06,
+      "loss": 7.765,
+      "step": 2015
+    },
+    {
+      "epoch": 3.181173690429303,
+      "grad_norm": 5.10537052154541,
+      "learning_rate": 5.5341936332651336e-06,
+      "loss": 7.9934,
+      "step": 2020
+    },
+    {
+      "epoch": 3.18905080740449,
+      "grad_norm": 5.30924129486084,
+      "learning_rate": 5.432488678092807e-06,
+      "loss": 7.8984,
+      "step": 2025
+    },
+    {
+      "epoch": 3.196927924379677,
+      "grad_norm": 6.369449138641357,
+      "learning_rate": 5.331612922685522e-06,
+      "loss": 7.7312,
+      "step": 2030
+    },
+    {
+      "epoch": 3.2048050413548643,
+      "grad_norm": 4.61285924911499,
+      "learning_rate": 5.231570641789899e-06,
+      "loss": 7.9155,
+      "step": 2035
+    },
+    {
+      "epoch": 3.2126821583300513,
+      "grad_norm": 4.817512512207031,
+      "learning_rate": 5.13236607483292e-06,
+      "loss": 7.6799,
+      "step": 2040
+    },
+    {
+      "epoch": 3.2205592753052383,
+      "grad_norm": 5.9151082038879395,
+      "learning_rate": 5.034003425742334e-06,
+      "loss": 8.2068,
+      "step": 2045
+    },
+    {
+      "epoch": 3.2284363922804253,
+      "grad_norm": 4.645730972290039,
+      "learning_rate": 4.9364868627684775e-06,
+      "loss": 7.9157,
+      "step": 2050
+    },
+    {
+      "epoch": 3.2363135092556123,
+      "grad_norm": 4.885422706604004,
+      "learning_rate": 4.839820518307628e-06,
+      "loss": 7.7623,
+      "step": 2055
+    },
+    {
+      "epoch": 3.2441906262307993,
+      "grad_norm": 5.870239734649658,
+      "learning_rate": 4.7440084887269035e-06,
+      "loss": 8.0045,
+      "step": 2060
+    },
+    {
+      "epoch": 3.252067743205987,
+      "grad_norm": 4.6275153160095215,
+      "learning_rate": 4.649054834190686e-06,
+      "loss": 8.0402,
+      "step": 2065
+    },
+    {
+      "epoch": 3.259944860181174,
+      "grad_norm": 5.3590779304504395,
+      "learning_rate": 4.554963578488561e-06,
+      "loss": 7.8615,
+      "step": 2070
+    },
+    {
+      "epoch": 3.267821977156361,
+      "grad_norm": 4.911188125610352,
+      "learning_rate": 4.461738708864768e-06,
+      "loss": 7.8886,
+      "step": 2075
+    },
+    {
+      "epoch": 3.275699094131548,
+      "grad_norm": 5.0650954246521,
+      "learning_rate": 4.369384175849287e-06,
+      "loss": 8.1302,
+      "step": 2080
+    },
+    {
+      "epoch": 3.283576211106735,
+      "grad_norm": 5.238993167877197,
+      "learning_rate": 4.277903893090407e-06,
+      "loss": 7.5243,
+      "step": 2085
+    },
+    {
+      "epoch": 3.2914533280819223,
+      "grad_norm": 4.651856899261475,
+      "learning_rate": 4.187301737188859e-06,
+      "loss": 7.3692,
+      "step": 2090
+    },
+    {
+      "epoch": 3.2993304450571093,
+      "grad_norm": 4.900179386138916,
+      "learning_rate": 4.0975815475335855e-06,
+      "loss": 7.0708,
+      "step": 2095
+    },
+    {
+      "epoch": 3.3072075620322963,
+      "grad_norm": 7.008111953735352,
+      "learning_rate": 4.0087471261390094e-06,
+      "loss": 7.5399,
+      "step": 2100
+    },
+    {
+      "epoch": 3.3150846790074833,
+      "grad_norm": 4.802496910095215,
+      "learning_rate": 3.920802237483912e-06,
+      "loss": 7.7337,
+      "step": 2105
+    },
+    {
+      "epoch": 3.3229617959826703,
+      "grad_norm": 5.664666175842285,
+      "learning_rate": 3.833750608351924e-06,
+      "loss": 7.3805,
+      "step": 2110
+    },
+    {
+      "epoch": 3.3308389129578573,
+      "grad_norm": 5.779867649078369,
+      "learning_rate": 3.747595927673611e-06,
+      "loss": 7.4321,
+      "step": 2115
+    },
+    {
+      "epoch": 3.3387160299330443,
+      "grad_norm": 4.908834457397461,
+      "learning_rate": 3.6623418463701293e-06,
+      "loss": 7.8338,
+      "step": 2120
+    },
+    {
+      "epoch": 3.3465931469082317,
+      "grad_norm": 5.189285755157471,
+      "learning_rate": 3.5779919771985025e-06,
+      "loss": 7.2317,
+      "step": 2125
+    },
+    {
+      "epoch": 3.3544702638834187,
+      "grad_norm": 4.881292819976807,
+      "learning_rate": 3.494549894598556e-06,
+      "loss": 7.1418,
+      "step": 2130
+    },
+    {
+      "epoch": 3.3623473808586057,
+      "grad_norm": 4.643113613128662,
+      "learning_rate": 3.4120191345414415e-06,
+      "loss": 7.9551,
+      "step": 2135
+    },
+    {
+      "epoch": 3.3702244978337927,
+      "grad_norm": 6.662186622619629,
+      "learning_rate": 3.3304031943797527e-06,
+      "loss": 7.9614,
+      "step": 2140
+    },
+    {
+      "epoch": 3.3781016148089797,
+      "grad_norm": 4.886229038238525,
+      "learning_rate": 3.249705532699379e-06,
+      "loss": 7.7298,
+      "step": 2145
+    },
+    {
+      "epoch": 3.385978731784167,
+      "grad_norm": 6.290960788726807,
+      "learning_rate": 3.169929569172919e-06,
+      "loss": 7.4452,
+      "step": 2150
+    },
+    {
+      "epoch": 3.393855848759354,
+      "grad_norm": 5.830533981323242,
+      "learning_rate": 3.0910786844147473e-06,
+      "loss": 7.6276,
+      "step": 2155
+    },
+    {
+      "epoch": 3.401732965734541,
+      "grad_norm": 4.647792339324951,
+      "learning_rate": 3.013156219837776e-06,
+      "loss": 7.7677,
+      "step": 2160
+    },
+    {
+      "epoch": 3.409610082709728,
+      "grad_norm": 4.803062915802002,
+      "learning_rate": 2.936165477511868e-06,
+      "loss": 7.9931,
+      "step": 2165
+    },
+    {
+      "epoch": 3.417487199684915,
+      "grad_norm": 5.041831970214844,
+      "learning_rate": 2.8601097200238987e-06,
+      "loss": 7.7018,
+      "step": 2170
+    },
+    {
+      "epoch": 3.425364316660102,
+      "grad_norm": 5.187193393707275,
+      "learning_rate": 2.7849921703394732e-06,
+      "loss": 7.4835,
+      "step": 2175
+    },
+    {
+      "epoch": 3.4332414336352897,
+      "grad_norm": 5.092511177062988,
+      "learning_rate": 2.7108160116663893e-06,
+      "loss": 7.4102,
+      "step": 2180
+    },
+    {
+      "epoch": 3.4411185506104767,
+      "grad_norm": 4.9641594886779785,
+      "learning_rate": 2.637584387319733e-06,
+      "loss": 8.2661,
+      "step": 2185
+    },
+    {
+      "epoch": 3.4489956675856637,
+      "grad_norm": 4.644840240478516,
+      "learning_rate": 2.565300400588647e-06,
+      "loss": 7.5782,
+      "step": 2190
+    },
+    {
+      "epoch": 3.4568727845608507,
+      "grad_norm": 4.720969200134277,
+      "learning_rate": 2.4939671146048727e-06,
+      "loss": 7.7751,
+      "step": 2195
+    },
+    {
+      "epoch": 3.4647499015360377,
+      "grad_norm": 4.5188889503479,
+      "learning_rate": 2.4235875522128997e-06,
+      "loss": 7.6633,
+      "step": 2200
+    },
+    {
+      "epoch": 3.4726270185112247,
+      "grad_norm": 4.929895877838135,
+      "learning_rate": 2.3541646958419094e-06,
+      "loss": 8.0803,
+      "step": 2205
+    },
+    {
+      "epoch": 3.480504135486412,
+      "grad_norm": 4.896505832672119,
+      "learning_rate": 2.2857014873793577e-06,
+      "loss": 7.508,
+      "step": 2210
+    },
+    {
+      "epoch": 3.488381252461599,
+      "grad_norm": 5.505984783172607,
+      "learning_rate": 2.218200828046324e-06,
+      "loss": 7.1282,
+      "step": 2215
+    },
+    {
+      "epoch": 3.496258369436786,
+      "grad_norm": 5.277995586395264,
+      "learning_rate": 2.151665578274581e-06,
+      "loss": 7.9669,
+      "step": 2220
+    },
+    {
+      "epoch": 3.504135486411973,
+      "grad_norm": 7.832403182983398,
+      "learning_rate": 2.0860985575853366e-06,
+      "loss": 8.0894,
+      "step": 2225
+    },
+    {
+      "epoch": 3.51201260338716,
+      "grad_norm": 5.209780216217041,
+      "learning_rate": 2.021502544469808e-06,
+      "loss": 7.4866,
+      "step": 2230
+    },
+    {
+      "epoch": 3.5198897203623476,
+      "grad_norm": 5.2757697105407715,
+      "learning_rate": 1.9578802762714427e-06,
+      "loss": 7.7545,
+      "step": 2235
+    },
+    {
+      "epoch": 3.5277668373375346,
+      "grad_norm": 4.831712245941162,
+      "learning_rate": 1.8952344490699243e-06,
+      "loss": 7.9045,
+      "step": 2240
+    },
+    {
+      "epoch": 3.5356439543127216,
+      "grad_norm": 4.441675662994385,
+      "learning_rate": 1.833567717566939e-06,
+      "loss": 7.379,
+      "step": 2245
+    },
+    {
+      "epoch": 3.5435210712879086,
+      "grad_norm": 5.417379856109619,
+      "learning_rate": 1.7728826949736598e-06,
+      "loss": 7.6097,
+      "step": 2250
+    },
+    {
+      "epoch": 3.5513981882630956,
+      "grad_norm": 4.504559516906738,
+      "learning_rate": 1.7131819529000226e-06,
+      "loss": 7.2856,
+      "step": 2255
+    },
+    {
+      "epoch": 3.559275305238283,
+      "grad_norm": 4.984797954559326,
+      "learning_rate": 1.6544680212457337e-06,
+      "loss": 7.194,
+      "step": 2260
+    },
+    {
+      "epoch": 3.5671524222134696,
+      "grad_norm": 4.818814754486084,
+      "learning_rate": 1.5967433880930889e-06,
+      "loss": 7.051,
+      "step": 2265
+    },
+    {
+      "epoch": 3.575029539188657,
+      "grad_norm": 4.872919082641602,
+      "learning_rate": 1.540010499601513e-06,
+      "loss": 7.1664,
+      "step": 2270
+    },
+    {
+      "epoch": 3.582906656163844,
+      "grad_norm": 6.4378862380981445,
+      "learning_rate": 1.4842717599039047e-06,
+      "loss": 7.3849,
+      "step": 2275
+    },
+    {
+      "epoch": 3.590783773139031,
+      "grad_norm": 5.795922756195068,
+      "learning_rate": 1.4295295310047668e-06,
+      "loss": 7.4473,
+      "step": 2280
+    },
+    {
+      "epoch": 3.598660890114218,
+      "grad_norm": 4.69597053527832,
+      "learning_rate": 1.375786132680118e-06,
+      "loss": 7.3833,
+      "step": 2285
+    },
+    {
+      "epoch": 3.606538007089405,
+      "grad_norm": 5.1728901863098145,
+      "learning_rate": 1.3230438423791625e-06,
+      "loss": 7.0266,
+      "step": 2290
+    },
+    {
+      "epoch": 3.6144151240645925,
+      "grad_norm": 4.9159440994262695,
+      "learning_rate": 1.2713048951278144e-06,
+      "loss": 7.3836,
+      "step": 2295
+    },
+    {
+      "epoch": 3.6222922410397795,
+      "grad_norm": 5.896880149841309,
+      "learning_rate": 1.2205714834339521e-06,
+      "loss": 7.517,
+      "step": 2300
+    },
+    {
+      "epoch": 3.6301693580149665,
+      "grad_norm": 4.930107593536377,
+      "learning_rate": 1.1708457571945386e-06,
+      "loss": 7.6201,
+      "step": 2305
+    },
+    {
+      "epoch": 3.6380464749901535,
+      "grad_norm": 4.929466724395752,
+      "learning_rate": 1.1221298236044953e-06,
+      "loss": 7.285,
+      "step": 2310
+    },
+    {
+      "epoch": 3.6459235919653405,
+      "grad_norm": 5.38197135925293,
+      "learning_rate": 1.0744257470674147e-06,
+      "loss": 7.2539,
+      "step": 2315
+    },
+    {
+      "epoch": 3.653800708940528,
+      "grad_norm": 5.5960469245910645,
+      "learning_rate": 1.0277355491080804e-06,
+      "loss": 7.4543,
+      "step": 2320
+    },
+    {
+      "epoch": 3.661677825915715,
+      "grad_norm": 5.1494245529174805,
+      "learning_rate": 9.820612082867969e-07,
+      "loss": 7.4204,
+      "step": 2325
+    },
+    {
+      "epoch": 3.669554942890902,
+      "grad_norm": 5.0791015625,
+      "learning_rate": 9.374046601155539e-07,
+      "loss": 7.2228,
+      "step": 2330
+    },
+    {
+      "epoch": 3.677432059866089,
+      "grad_norm": 4.948309421539307,
+      "learning_rate": 8.937677969759861e-07,
+      "loss": 7.5265,
+      "step": 2335
+    },
+    {
+      "epoch": 3.685309176841276,
+      "grad_norm": 5.31011438369751,
+      "learning_rate": 8.511524680392096e-07,
+      "loss": 7.5065,
+      "step": 2340
+    },
+    {
+      "epoch": 3.693186293816463,
+      "grad_norm": 4.728499889373779,
+      "learning_rate": 8.09560479187449e-07,
+      "loss": 7.6349,
+      "step": 2345
+    },
+    {
+      "epoch": 3.70106341079165,
+      "grad_norm": 4.810965538024902,
+      "learning_rate": 7.689935929374892e-07,
+      "loss": 7.2619,
+      "step": 2350
+    },
+    {
+      "epoch": 3.7089405277668375,
+      "grad_norm": 5.238432884216309,
+      "learning_rate": 7.294535283660281e-07,
+      "loss": 7.1741,
+      "step": 2355
+    },
+    {
+      "epoch": 3.7168176447420245,
+      "grad_norm": 4.824338436126709,
+      "learning_rate": 6.909419610367896e-07,
+      "loss": 7.5995,
+      "step": 2360
+    },
+    {
+      "epoch": 3.7246947617172115,
+      "grad_norm": 5.660040855407715,
+      "learning_rate": 6.534605229295415e-07,
+      "loss": 7.513,
+      "step": 2365
+    },
+    {
+      "epoch": 3.7325718786923985,
+      "grad_norm": 5.301452159881592,
+      "learning_rate": 6.170108023709348e-07,
+      "loss": 7.3567,
+      "step": 2370
+    },
+    {
+      "epoch": 3.7404489956675855,
+      "grad_norm": 5.379169940948486,
+      "learning_rate": 5.815943439671901e-07,
+      "loss": 6.9589,
+      "step": 2375
+    },
+    {
+      "epoch": 3.748326112642773,
+      "grad_norm": 5.965606212615967,
+      "learning_rate": 5.472126485386475e-07,
+      "loss": 7.6638,
+      "step": 2380
+    },
+    {
+      "epoch": 3.75620322961796,
+      "grad_norm": 4.746719837188721,
+      "learning_rate": 5.138671730561623e-07,
+      "loss": 7.3243,
+      "step": 2385
+    },
+    {
+      "epoch": 3.764080346593147,
+      "grad_norm": 4.829946994781494,
+      "learning_rate": 4.815593305793703e-07,
+      "loss": 7.6804,
+      "step": 2390
+    },
+    {
+      "epoch": 3.771957463568334,
+      "grad_norm": 6.1915669441223145,
+      "learning_rate": 4.502904901968141e-07,
+      "loss": 7.2188,
+      "step": 2395
+    },
+    {
+      "epoch": 3.779834580543521,
+      "grad_norm": 4.6367998123168945,
+      "learning_rate": 4.200619769679004e-07,
+      "loss": 6.9212,
+      "step": 2400
+    },
+    {
+      "epoch": 3.7877116975187084,
+      "grad_norm": 5.0970611572265625,
+      "learning_rate": 3.908750718667753e-07,
+      "loss": 7.3765,
+      "step": 2405
+    },
+    {
+      "epoch": 3.795588814493895,
+      "grad_norm": 5.42421817779541,
+      "learning_rate": 3.6273101172802914e-07,
+      "loss": 7.402,
+      "step": 2410
+    },
+    {
+      "epoch": 3.8034659314690824,
+      "grad_norm": 5.175075054168701,
+      "learning_rate": 3.3563098919428827e-07,
+      "loss": 7.053,
+      "step": 2415
+    },
+    {
+      "epoch": 3.8113430484442694,
+      "grad_norm": 4.7831854820251465,
+      "learning_rate": 3.095761526656693e-07,
+      "loss": 7.9933,
+      "step": 2420
+    },
+    {
+      "epoch": 3.8192201654194564,
+      "grad_norm": 6.312967300415039,
+      "learning_rate": 2.8456760625111823e-07,
+      "loss": 7.4616,
+      "step": 2425
+    },
+    {
+      "epoch": 3.8270972823946434,
+      "grad_norm": 5.104128360748291,
+      "learning_rate": 2.6060640972162543e-07,
+      "loss": 7.5123,
+      "step": 2430
+    },
+    {
+      "epoch": 3.8349743993698304,
+      "grad_norm": 5.383700847625732,
+      "learning_rate": 2.3769357846531171e-07,
+      "loss": 6.8885,
+      "step": 2435
+    },
+    {
+      "epoch": 3.842851516345018,
+      "grad_norm": 4.860473155975342,
+      "learning_rate": 2.1583008344439603e-07,
+      "loss": 7.0768,
+      "step": 2440
+    },
+    {
+      "epoch": 3.850728633320205,
+      "grad_norm": 5.095125675201416,
+      "learning_rate": 1.9501685115406176e-07,
+      "loss": 7.6519,
+      "step": 2445
+    },
+    {
+      "epoch": 3.858605750295392,
+      "grad_norm": 5.066483497619629,
+      "learning_rate": 1.7525476358318526e-07,
+      "loss": 7.3912,
+      "step": 2450
+    },
+    {
+      "epoch": 3.866482867270579,
+      "grad_norm": 6.104281425476074,
+      "learning_rate": 1.5654465817696584e-07,
+      "loss": 6.9311,
+      "step": 2455
+    },
+    {
+      "epoch": 3.874359984245766,
+      "grad_norm": 4.735925674438477,
+      "learning_rate": 1.3888732780142632e-07,
+      "loss": 7.2688,
+      "step": 2460
+    },
+    {
+      "epoch": 3.8822371012209533,
+      "grad_norm": 4.991415500640869,
+      "learning_rate": 1.2228352070983719e-07,
+      "loss": 6.8569,
+      "step": 2465
+    },
+    {
+      "epoch": 3.8901142181961403,
+      "grad_norm": 5.310063362121582,
+      "learning_rate": 1.0673394051098074e-07,
+      "loss": 7.4524,
+      "step": 2470
+    },
+    {
+      "epoch": 3.8979913351713273,
+      "grad_norm": 6.486708641052246,
+      "learning_rate": 9.223924613935842e-08,
+      "loss": 7.5994,
+      "step": 2475
+    },
+    {
+      "epoch": 3.9058684521465143,
+      "grad_norm": 4.825411319732666,
+      "learning_rate": 7.880005182725469e-08,
+      "loss": 7.0521,
+      "step": 2480
+    },
+    {
+      "epoch": 3.9137455691217014,
+      "grad_norm": 4.599216461181641,
+      "learning_rate": 6.641692707870795e-08,
+      "loss": 7.1754,
+      "step": 2485
+    },
+    {
+      "epoch": 3.921622686096889,
+      "grad_norm": 4.610249042510986,
+      "learning_rate": 5.509039664538251e-08,
+      "loss": 7.6455,
+      "step": 2490
+    },
+    {
+      "epoch": 3.9294998030720754,
+      "grad_norm": 5.008790016174316,
+      "learning_rate": 4.482094050433372e-08,
+      "loss": 7.7298,
+      "step": 2495
+    },
+    {
+      "epoch": 3.937376920047263,
+      "grad_norm": 5.213418006896973,
+      "learning_rate": 3.5608993837651924e-08,
+      "loss": 7.2896,
+      "step": 2500
+    },
+    {
+      "epoch": 3.94525403702245,
+      "grad_norm": 4.882436275482178,
+      "learning_rate": 2.745494701404394e-08,
+      "loss": 7.3404,
+      "step": 2505
+    },
+    {
+      "epoch": 3.953131153997637,
+      "grad_norm": 4.545849800109863,
+      "learning_rate": 2.0359145572276804e-08,
+      "loss": 7.5327,
+      "step": 2510
+    },
+    {
+      "epoch": 3.961008270972824,
+      "grad_norm": 5.91042423248291,
+      "learning_rate": 1.4321890206533961e-08,
+      "loss": 7.468,
+      "step": 2515
+    },
+    {
+      "epoch": 3.968885387948011,
+      "grad_norm": 5.8342766761779785,
+      "learning_rate": 9.34343675368099e-09,
+      "loss": 7.6057,
+      "step": 2520
+    },
+    {
+      "epoch": 3.9767625049231983,
+      "grad_norm": 5.667447566986084,
+      "learning_rate": 5.423996182415958e-09,
+      "loss": 7.2114,
+      "step": 2525
+    },
+    {
+      "epoch": 3.9846396218983853,
+      "grad_norm": 5.622943878173828,
+      "learning_rate": 2.5637345843348982e-09,
+      "loss": 7.2359,
+      "step": 2530
+    },
+    {
+      "epoch": 3.9925167388735723,
+      "grad_norm": 4.587869644165039,
+      "learning_rate": 7.627731668985449e-10,
+      "loss": 7.1989,
+      "step": 2535
+    },
+    {
+      "epoch": 3.9988184324537217,
+      "step": 2539,
+      "total_flos": 4.3565600076592054e+19,
+      "train_loss": 1.895302517482642,
+      "train_runtime": 61070.1581,
+      "train_samples_per_second": 2.661,
+      "train_steps_per_second": 0.042
+    }
+  ],
+  "logging_steps": 5,
+  "max_steps": 2540,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 4,
+  "save_steps": 159,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 4.3565600076592054e+19,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}

SDAR-4B-Math-R1-CoT-SFT/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46943a038bbd8afe345ee35cc3e2d8972bcd53d59b88da7fd7b677b1c4414324
+size 7825