Upload folder using huggingface_hub

README.md

@@ -0,0 +1,69 @@
+---
+library_name: transformers
+tags:
+- alignment-handbook
+- generated_from_trainer
+datasets:
+- HuggingFaceTB/Magpie-Pro-300K-Filtered-H4
+- HuggingFaceTB/self-oss-instruct-sc2-H4
+- HuggingFaceTB/OpenHermes-2.5-H4
+- HuggingFaceTB/everyday-conversations-llama3.1-2k
+- HuggingFaceTB/instruct-data-basics-smollm-H4
+model-index:
+- name: monet-vd-1.4B-100BT-chat-hf
+  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. -->
+
+# monet-vd-1.4B-100BT-chat-hf
+
+This model is a fine-tuned version of [monet-vd-1.4B-100BT-hf](https://huggingface.co/MonetLLM/monet-vd-1.4B-100BT-hf) on the HuggingFaceTB/Magpie-Pro-300K-Filtered-H4, the HuggingFaceTB/self-oss-instruct-sc2-H4, the HuggingFaceTB/OpenHermes-2.5-H4, the HuggingFaceTB/everyday-conversations-llama3.1-2k and the HuggingFaceTB/instruct-data-basics-smollm-H4 datasets.
+It achieves the following results on the evaluation set:
+- Loss: 1.1664
+
+## 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: 0.0003
+- train_batch_size: 4
+- eval_batch_size: 4
+- seed: 42
+- distributed_type: multi-GPU
+- num_devices: 4
+- gradient_accumulation_steps: 8
+- total_train_batch_size: 128
+- total_eval_batch_size: 16
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: cosine
+- lr_scheduler_warmup_ratio: 0.1
+- num_epochs: 1
+
+### Training results
+
+| Training Loss | Epoch  | Step | Validation Loss |
+|:-------------:|:------:|:----:|:---------------:|
+| 0.8032        | 0.9988 | 502  | 1.1664          |
+
+
+### Framework versions
+
+- Transformers 4.44.2
+- Pytorch 2.4.1
+- Datasets 2.21.0
+- Tokenizers 0.19.1

all_results.json

@@ -0,0 +1,14 @@
+{
+    "epoch": 0.9987565282268093,
+    "eval_loss": 1.1664291620254517,
+    "eval_runtime": 1405.1859,
+    "eval_samples": 82730,
+    "eval_samples_per_second": 16.588,
+    "eval_steps_per_second": 1.037,
+    "total_flos": 234414617395200.0,
+    "train_loss": 0.8937448220423968,
+    "train_runtime": 13768.4166,
+    "train_samples": 186330,
+    "train_samples_per_second": 4.672,
+    "train_steps_per_second": 0.036
+}

config.json

@@ -0,0 +1,37 @@
+{
+  "architectures": ["MonetForCausalLM"],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "modeling_monet.MonetConfig",
+    "AutoModelForCausalLM": "modeling_monet.MonetForCausalLM"
+  },
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_act": "relu2",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": null,
+  "max_position_embeddings": 2048,
+  "mlp_bias": null,
+  "model_type": "monet",
+  "moe_decompose": "vertical",
+  "moe_dim": 16,
+  "moe_experts": 512,
+  "moe_groups": 4,
+  "moe_heads": 8,
+  "moe_topk": 8,
+  "num_attention_heads": 16,
+  "num_hidden_layers": 24,
+  "num_key_value_heads": 16,
+  "output_router_probs": false,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-6,
+  "rope_scaling": null,
+  "rope_theta": 10000.0,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.44.2",
+  "use_cache": true,
+  "vocab_size": 32000
+}

eval_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 0.9987565282268093,
+    "eval_loss": 1.1664291620254517,
+    "eval_runtime": 1405.1859,
+    "eval_samples": 82730,
+    "eval_samples_per_second": 16.588,
+    "eval_steps_per_second": 1.037
+}

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "transformers_version": "4.44.2"
+}

interpretation/embeddings-0.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:74554360419b088909efae4ae09e38735b002c9cae03f32228ff86145d3f8136
+size 1073743029

interpretation/embeddings-1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:86dcae4c950c033905c99d0443b514cd2760d2e53e36620023a9f67e8ca5eb90
+size 1073743029

interpretation/embeddings-2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5d87a75cbf5748380f810f673474602027a6fd858a2d435739976270e9b574a
+size 1073743029

interpretation/embeddings-3.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c76e17add23fe8a1cce5827e5b53cd59c46cc1771e932fd635bc60b3961540e
+size 1073743029

interpretation/embeddings-4.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a6373ddea12b56129c67b26c10db4cdf96e1ea1f3e9c4b6c1efb9754ada634b
+size 1073743029

interpretation/embeddings-5.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c8c8de2fde422690d361cf7a59abc5f863442ec223e62ec7029f16db3f59d48
+size 1073743029

interpretation/examples-0.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:71750857dbf288e52c6de4faf6e9808ae787f3db93b09b81263e0e22adafd615
+size 99785078

interpretation/examples-1.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:236966ff9a822d27f38a408a13dd2b7fa948fccdaefa25c4f7746d90750a8b55
+size 186844896

interpretation/examples-2.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ba38c6b4fc4c97794704c33646aeec55643c4f9bba0b86378882b0fad1b866a
+size 151744447

interpretation/examples-3.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:18151accfdec95f236eb3ccbe0b4418b7f30ec027dffc38a2df139a6de076bba
+size 196289141

interpretation/examples-4.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5587d6bcf1562be1bbe006b522849542ba7c126d31ab67435e0f975ac5d58f4a
+size 244767190

interpretation/examples-5.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:450696857c0c06a630f70b512fd047f1436f9d775842bfceba4d33f5c502d101
+size 247488639

interpretation/inputs.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c6d9e5feaca0db26368a82f6151c276638d910f2efaf5f313d3822ad1379f9c
+size 157615191

interpretation/routings-0.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a014eb363ab2fae5d2d08c351b8f0dc81fa4c848216f797ab85dbe50556484b
+size 2715152823

interpretation/routings-1.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a46dfea287041d4121c3a68e3b0109b629bf31fce2f66e296075ba25fefb7ec
+size 2713393723

interpretation/routings-2.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:62e6267a5ed8852d20259789276ad3bb5eb1e14b273ddd740de3296bce97b164
+size 1681898562

interpretation/routings-3.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd7fdf3a3ab31346bdffa781b1d9b80a510015bd1b0afbeb164a3ed27a358b5e
+size 1622189972

interpretation/routings-4.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:22f6cbee6cc421193e91fe3b0e1ce6d8c5fb4abcdf9551ca48b7069d3b70f954
+size 2007108077

interpretation/routings-5.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28446ce8d4f29127499ff503d01ed15a9bf2f71d20ddd7095c81470036203733
+size 2680546231

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab36abbc22160db50fef82a4d24a03151172abcfa6a4a4256b475475d7524f99
+size 2930363080

modeling_monet.py

@@ -0,0 +1,663 @@
+# fmt: off
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+import torch
+import torch.utils.checkpoint
+from scipy.stats import norm
+from torch import nn
+from torch.nn import CrossEntropyLoss
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache, StaticCache
+from transformers.modeling_attn_mask_utils import AttentionMaskConverter
+from transformers.modeling_utils import PreTrainedModel
+from transformers.models.llama.configuration_llama import LlamaConfig
+from transformers.models.llama.modeling_llama import (
+    LLAMA_ATTENTION_CLASSES,
+    LlamaRMSNorm,
+)
+from transformers.utils import ModelOutput, logging
+
+logger = logging.get_logger(__name__)
+
+
+@dataclass
+class MonetModelOutputWithPast(ModelOutput):
+    last_hidden_state: torch.FloatTensor = None
+    past_key_values: tuple[tuple[torch.FloatTensor]] | None = None
+    hidden_states: tuple[torch.FloatTensor, ...] | None = None
+    attentions: tuple[torch.FloatTensor, ...] | None = None
+    router_probs: tuple[tuple[torch.FloatTensor, ...], ...] | None = None
+
+
+@dataclass
+class MonetCausalLMOutputWithPast(ModelOutput):
+    loss: torch.FloatTensor | None = None
+    aux_loss: torch.FloatTensor | None = None
+    logits: torch.FloatTensor = None
+    past_key_values: tuple[tuple[torch.FloatTensor]] | None = None
+    hidden_states: tuple[torch.FloatTensor, ...] | None = None
+    attentions: tuple[torch.FloatTensor, ...] | None = None
+    router_probs: tuple[tuple[torch.FloatTensor, ...], ...] | None = None
+
+
+class MonetConfig(LlamaConfig):
+    model_type = "monet"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=32000,
+        hidden_size=4096,
+        intermediate_size=None,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=None,
+        hidden_act="relu2",
+        max_position_embeddings=2048,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=1,
+        eos_token_id=2,
+        pretraining_tp=1,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        attention_dropout=0.0,
+        mlp_bias=None,
+        moe_dim=8,
+        moe_heads=8,
+        moe_experts=512,
+        moe_topk=32,
+        moe_groups=4,
+        moe_decompose="vertical",
+        output_router_probs=False,
+        **kwargs,
+    ):
+        self.moe_dim = moe_dim
+        self.moe_heads = moe_heads
+        self.moe_experts = moe_experts
+        self.moe_topk = moe_topk
+        self.moe_groups = moe_groups
+        self.moe_decompose = moe_decompose
+        self.output_router_probs = output_router_probs
+
+        super().__init__(
+            vocab_size=vocab_size,
+            hidden_size=hidden_size,
+            intermediate_size=intermediate_size,
+            num_hidden_layers=num_hidden_layers,
+            num_attention_heads=num_attention_heads,
+            num_key_value_heads=num_key_value_heads,
+            hidden_act=hidden_act,
+            max_position_embeddings=max_position_embeddings,
+            initializer_range=initializer_range,
+            rms_norm_eps=rms_norm_eps,
+            use_cache=use_cache,
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            pretraining_tp=pretraining_tp,
+            tie_word_embeddings=tie_word_embeddings,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            attention_bias=attention_bias,
+            attention_dropout=attention_dropout,
+            mlp_bias=mlp_bias,
+            **kwargs,
+        )
+
+
+class MonetRouter(nn.Module):
+    def __init__(self, config: MonetConfig):
+        super().__init__()
+        self.config = config
+        flatten_shape = config.moe_heads * config.moe_experts
+
+        self.w1 = nn.Linear(config.hidden_size, flatten_shape, bias=False)
+        self.w2 = nn.Linear(config.hidden_size, flatten_shape, bias=False)
+        self.norm1 = nn.BatchNorm1d(config.moe_heads, affine=False)
+        self.norm2 = nn.BatchNorm1d(config.moe_heads, affine=False)
+
+    def forward(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+        g1z = self.w1(x).unflatten(-1, (self.config.moe_heads, -1)).float()
+        g2z = self.w2(x).unflatten(-1, (self.config.moe_heads, -1)).float()
+
+        g1n = self.norm1(g1z.transpose(2, 3).flatten(0, -2))
+        g2n = self.norm2(g2z.transpose(2, 3).flatten(0, -2))
+        g1n = g1n.view(g1z.size(0), g1z.size(1), g1z.size(3), -1).transpose(2, 3)
+        g2n = g2n.view(g2z.size(0), g2z.size(1), g2z.size(3), -1).transpose(2, 3)
+
+        sigma = float(norm.ppf(1 - self.config.moe_topk / self.config.moe_experts))
+        g1s = g1n.amax(-1, keepdim=True).clamp_max_(sigma)
+        g2s = g2n.amax(-1, keepdim=True).clamp_max_(sigma)
+
+        g1 = nn.functional.softmax(torch.where(g1n >= g1s, g1z, -1e10), dim=-1)
+        g2 = nn.functional.softmax(torch.where(g2n >= g2s, g2z, -1e10), dim=-1)
+        return g1, g2
+
+
+class MonetMoVDE(nn.Module):
+    def __init__(self, config: MonetConfig):
+        super().__init__()
+        self.config = config
+        self.act_fn = ACT2FN[config.hidden_act]
+        flatten_shape = config.moe_experts * config.moe_dim // 2
+
+        self.u1 = nn.Linear(config.hidden_size, flatten_shape)
+        self.u2 = nn.Linear(config.hidden_size, flatten_shape)
+
+        self.v11 = nn.Linear(flatten_shape, config.hidden_size // 2, bias=False)
+        self.v12 = nn.Linear(flatten_shape, config.hidden_size // 2, bias=False)
+        self.v21 = nn.Linear(flatten_shape, config.hidden_size // 2, bias=False)
+        self.v22 = nn.Linear(flatten_shape, config.hidden_size // 2, bias=False)
+
+        self.b1 = nn.Parameter(torch.zeros(config.moe_experts, config.hidden_size // 2))
+        self.b2 = nn.Parameter(torch.zeros(config.moe_experts, config.hidden_size // 2))
+
+    def forward(
+        self, x: torch.Tensor, g1: torch.Tensor, g2: torch.Tensor
+    ) -> torch.Tensor:
+        g1, g2 = g1.type_as(x), g2.type_as(x)
+        x1 = self.act_fn(self.u1(x).unflatten(-1, (self.config.moe_experts, -1)))
+        x2 = self.act_fn(self.u2(x).unflatten(-1, (self.config.moe_experts, -1)))
+
+        x11 = self.v11(torch.einsum("btim,bthi->btim", x1, g1).flatten(-2))
+        x12 = self.v12(torch.einsum("btjm,bthj,bthi->btim", x2, g2, g1).flatten(-2))
+        x13 = torch.einsum("bthi,id->btd", g1, self.b1.type_as(x))
+
+        x21 = self.v21(torch.einsum("btim,bthi,bthj->btjm", x1, g1, g2).flatten(-2))
+        x22 = self.v22(torch.einsum("btjm,bthj->btjm", x2, g2).flatten(-2))
+        x23 = torch.einsum("bthj,jd->btd", g2, self.b2.type_as(x))
+
+        return torch.cat((x11 + x12 + x13, x21 + x22 + x23), dim=-1)
+
+
+class MonetMoHDE(nn.Module):
+    def __init__(self, config: MonetConfig):
+        super().__init__()
+        self.config = config
+        self.act_fn = ACT2FN[config.hidden_act]
+        flatten_shape = config.moe_experts * config.moe_dim
+
+        self.u = nn.Linear(config.hidden_size, flatten_shape)
+        self.v = nn.Linear(flatten_shape, config.hidden_size, bias=False)
+        self.b = nn.Parameter(torch.zeros(config.moe_experts, config.hidden_size))
+
+    def forward(
+        self, x: torch.Tensor, g1: torch.Tensor, g2: torch.Tensor
+    ) -> torch.Tensor:
+        g1, g2 = g1.type_as(x), g2.type_as(x)
+        x = self.act_fn(self.u(x).unflatten(-1, (self.config.moe_experts, -1)))
+        x = self.v(torch.einsum("btim,bthi,bthj->btjm", x, g1, g2).flatten(-2))
+        return x + torch.einsum("bthj,jd->btd", g2, self.b)
+
+
+class MonetDecoderLayer(nn.Module):
+    def __init__(self, config: MonetConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = LLAMA_ATTENTION_CLASSES[config._attn_implementation](
+            config=config, layer_idx=layer_idx
+        )
+        self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = LlamaRMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+        if config.moe_decompose == "vertical":
+            self.moe = MonetMoVDE(config)
+        elif config.moe_decompose == "horizontal":
+            self.moe = MonetMoHDE(config)
+        if layer_idx % config.moe_groups == 0:
+            self.router = MonetRouter(config).requires_grad_(False)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_value: Cache | None = None,
+        previous_router_probs: tuple[torch.Tensor, torch.Tensor] | None = None,
+        output_attentions: bool | None = False,
+        use_cache: bool | None = False,
+        cache_position: torch.LongTensor | None = None,
+        **kwargs,
+    ) -> tuple[torch.FloatTensor, ...]:
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = 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,
+            cache_position=cache_position,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        g1, g2 = (
+            self.router(hidden_states)
+            if hasattr(self, "router")
+            else previous_router_probs
+        )
+        hidden_states = self.moe(hidden_states, g1, g2)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs + ((g1, g2) if hasattr(self, "router") else None,)
+
+
+class MonetPreTrainedModel(PreTrainedModel):
+    config_class = MonetConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["MonetDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_cache_class = True
+    _supports_quantized_cache = True
+    _supports_static_cache = 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_()
+
+
+class MonetModel(MonetPreTrainedModel):
+    def __init__(self, config: MonetConfig):
+        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)  # noqa
+        self.layers = nn.ModuleList([MonetDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)])  # noqa
+        self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        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
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | list[torch.FloatTensor] | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        use_cache: bool | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        output_router_probs: bool | None = None,
+        return_dict: bool | None = None,
+        cache_position: torch.LongTensor | None = None,
+    ) -> tuple[torch.Tensor, ...] | MonetModelOutputWithPast:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions  # noqa
+        output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states  # noqa
+        output_router_probs = output_router_probs if output_router_probs is not None else self.config.output_router_probs  # noqa
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict  # noqa
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one")  # noqa
+
+        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`.")  # noqa
+            use_cache = False
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        return_legacy_cache = False
+        if use_cache and not isinstance(past_key_values, Cache):  # kept for BC (non `Cache` `past_key_values` inputs)  # noqa
+            return_legacy_cache = True
+            past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+            logger.warning_once(
+                "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. "  # noqa
+                "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)"  # noqa
+            )
+
+        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)  # noqa
+        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)  # noqa
+
+        # embed positions
+        hidden_states = inputs_embeds
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_router_probs = () if output_router_probs else None
+        previous_router_probs, next_decoder_cache = None, None
+
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    previous_router_probs,
+                    output_attentions,
+                    use_cache,
+                    cache_position,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    previous_router_probs=previous_router_probs,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                )
+
+            hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache = layer_outputs[2 if output_attentions else 1]
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+            if output_router_probs:
+                all_router_probs += (layer_outputs[-1],)
+            previous_router_probs = (
+                layer_outputs[-1]
+                if layer_outputs[-1] is not None
+                else previous_router_probs
+            )
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if return_legacy_cache:
+            next_cache = next_cache.to_legacy_cache()
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_router_probs] if v is not None)  # noqa
+        return MonetModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            router_probs=all_router_probs,
+        )
+
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool,
+    ):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0  # noqa
+        using_static_cache = isinstance(past_key_values, StaticCache)
+
+        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:  # noqa
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                is_training=self.training,
+            ):
+                return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        min_dtype = torch.finfo(dtype).min
+        sequence_length = input_tensor.shape[1]
+        if using_static_cache:
+            target_length = past_key_values.get_max_length()
+        else:
+            target_length = (
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
+            )
+
+        if attention_mask is not None and attention_mask.dim() == 4:
+            if attention_mask.max() != 0:
+                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")  # noqa
+            causal_mask = attention_mask
+        else:
+            causal_mask = torch.full(
+                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device  # noqa
+            )
+            if sequence_length != 1:
+                causal_mask = torch.triu(causal_mask, diagonal=1)
+            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)  # noqa
+            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)  # noqa
+            if attention_mask is not None:
+                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit  # noqa
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]  # noqa
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(padding_mask, min_dtype)  # noqa
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type == "cuda"
+            and not output_attentions
+        ):
+            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)  # noqa
+
+        return causal_mask
+
+
+class MonetForCausalLM(MonetPreTrainedModel):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = MonetModel(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 forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | list[torch.FloatTensor] | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        labels: torch.LongTensor | None = None,
+        use_cache: bool | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        output_router_probs: bool | None = None,
+        return_dict: bool | None = None,
+        cache_position: torch.LongTensor | None = None,
+    ) -> tuple[torch.Tensor, ...] | MonetCausalLMOutputWithPast:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions  # noqa
+        output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states  # noqa
+        output_router_probs = output_router_probs if output_router_probs is not None else self.config.output_router_probs  # noqa
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict  # noqa
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            output_router_probs=output_router_probs,
+            return_dict=return_dict,
+            cache_position=cache_position,
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+        logits = logits.float()
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return MonetCausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            router_probs=outputs.router_probs,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        cache_position=None,
+        use_cache=True,
+        **kwargs,
+    ):
+        past_length = 0
+        if past_key_values is not None:
+            past_length = cache_position[0] if cache_position is not None else past_key_values.get_seq_length()  # noqa
+            max_cache_length = (
+                torch.tensor(past_key_values.get_max_length(), device=input_ids.device)
+                if past_key_values.get_max_length() is not None
+                else None
+            )
+            cache_length = past_length if max_cache_length is None else torch.min(max_cache_length, past_length)  # noqa
+
+            # Keep only the unprocessed tokens:
+            if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:  # noqa
+                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
+            # input_ids based on the past_length.
+            elif past_length < input_ids.shape[1]:
+                input_ids = input_ids[:, past_length:]
+
+            if (
+                max_cache_length is not None
+                and attention_mask is not None
+                and cache_length + input_ids.shape[1] > max_cache_length
+            ):
+                attention_mask = attention_mask[:, -max_cache_length:]
+
+        position_ids = kwargs.get("position_ids", None)
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -input_ids.shape[1] :]
+
+        if inputs_embeds is not None and past_length == 0:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids.contiguous()}
+
+        input_length = position_ids.shape[-1] if position_ids is not None else input_ids.shape[-1]  # noqa
+        if cache_position is None:
+            cache_position = torch.arange(past_length, past_length + input_length, device=input_ids.device)  # noqa
+        elif use_cache:
+            cache_position = cache_position[-input_length:]
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "cache_position": cache_position,
+                "past_key_values": past_key_values,
+                "use_cache": use_cache,
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),  # noqa
+            )
+        return reordered_past

special_tokens_map.json

@@ -0,0 +1,24 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "</s>",
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347
+size 499723

tokenizer_config.json

@@ -0,0 +1,44 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "add_prefix_space": true,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<s>",
+  "chat_template": "{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if loop.index0 == 0 and system_message != false %}{% set content = '<<SYS>>\\n' + system_message + '\\n<</SYS>>\\n\\n' + message['content'] %}{% else %}{% set content = message['content'] %}{% endif %}{% if message['role'] == 'user' %}{{ bos_token + '[INST] ' + content.strip() + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ ' '  + content.strip() + ' ' + eos_token }}{% endif %}{% endfor %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "legacy": false,
+  "model_max_length": 2048,
+  "pad_token": "</s>",
+  "padding_side": "right",
+  "sp_model_kwargs": {},
+  "spaces_between_special_tokens": false,
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": "<unk>",
+  "use_default_system_prompt": false
+}

train_results.json

@@ -0,0 +1,9 @@
+{
+    "epoch": 0.9987565282268093,
+    "total_flos": 234414617395200.0,
+    "train_loss": 0.8937448220423968,
+    "train_runtime": 13768.4166,
+    "train_samples": 186330,
+    "train_samples_per_second": 4.672,
+    "train_steps_per_second": 0.036
+}

trainer_state.json

@@ -0,0 +1,757 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.9987565282268093,
+  "eval_steps": 500,
+  "global_step": 502,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.001989554837105198,
+      "grad_norm": 2.019071375807948,
+      "learning_rate": 5.88235294117647e-06,
+      "loss": 1.4931,
+      "step": 1
+    },
+    {
+      "epoch": 0.009947774185525988,
+      "grad_norm": 1.5423217337624162,
+      "learning_rate": 2.941176470588235e-05,
+      "loss": 1.4424,
+      "step": 5
+    },
+    {
+      "epoch": 0.019895548371051976,
+      "grad_norm": 0.8308012142463063,
+      "learning_rate": 5.88235294117647e-05,
+      "loss": 1.2728,
+      "step": 10
+    },
+    {
+      "epoch": 0.029843322556577966,
+      "grad_norm": 0.5145126949446309,
+      "learning_rate": 8.823529411764705e-05,
+      "loss": 1.1614,
+      "step": 15
+    },
+    {
+      "epoch": 0.03979109674210395,
+      "grad_norm": 0.36488357130003074,
+      "learning_rate": 0.0001176470588235294,
+      "loss": 1.1138,
+      "step": 20
+    },
+    {
+      "epoch": 0.04973887092762994,
+      "grad_norm": 0.45152250150726514,
+      "learning_rate": 0.00014705882352941175,
+      "loss": 1.07,
+      "step": 25
+    },
+    {
+      "epoch": 0.05968664511315593,
+      "grad_norm": 0.3640886200970852,
+      "learning_rate": 0.0001764705882352941,
+      "loss": 1.0509,
+      "step": 30
+    },
+    {
+      "epoch": 0.06963441929868192,
+      "grad_norm": 1.7718021462555353,
+      "learning_rate": 0.00020588235294117645,
+      "loss": 1.0168,
+      "step": 35
+    },
+    {
+      "epoch": 0.0795821934842079,
+      "grad_norm": 0.3768307338988579,
+      "learning_rate": 0.0002352941176470588,
+      "loss": 1.0086,
+      "step": 40
+    },
+    {
+      "epoch": 0.0895299676697339,
+      "grad_norm": 0.33093766782578965,
+      "learning_rate": 0.00026470588235294115,
+      "loss": 1.0008,
+      "step": 45
+    },
+    {
+      "epoch": 0.09947774185525989,
+      "grad_norm": 0.37314906796958397,
+      "learning_rate": 0.0002941176470588235,
+      "loss": 0.9975,
+      "step": 50
+    },
+    {
+      "epoch": 0.10942551604078588,
+      "grad_norm": 0.39939561360809905,
+      "learning_rate": 0.00029994177629874796,
+      "loss": 0.9884,
+      "step": 55
+    },
+    {
+      "epoch": 0.11937329022631187,
+      "grad_norm": 0.3919443406321306,
+      "learning_rate": 0.00029970531997706437,
+      "loss": 0.9843,
+      "step": 60
+    },
+    {
+      "epoch": 0.12932106441183785,
+      "grad_norm": 0.3915947062044229,
+      "learning_rate": 0.00029928727864250395,
+      "loss": 0.9913,
+      "step": 65
+    },
+    {
+      "epoch": 0.13926883859736383,
+      "grad_norm": 0.3534055778790222,
+      "learning_rate": 0.00029868815935814996,
+      "loss": 0.9893,
+      "step": 70
+    },
+    {
+      "epoch": 0.14921661278288983,
+      "grad_norm": 0.36735315351449394,
+      "learning_rate": 0.0002979086888255182,
+      "loss": 0.9775,
+      "step": 75
+    },
+    {
+      "epoch": 0.1591643869684158,
+      "grad_norm": 0.3088341190569696,
+      "learning_rate": 0.00029694981250310496,
+      "loss": 0.981,
+      "step": 80
+    },
+    {
+      "epoch": 0.16911216115394181,
+      "grad_norm": 0.325839880040256,
+      "learning_rate": 0.0002958126934595933,
+      "loss": 0.9659,
+      "step": 85
+    },
+    {
+      "epoch": 0.1790599353394678,
+      "grad_norm": 1.4788476638739554,
+      "learning_rate": 0.0002944987109631094,
+      "loss": 0.9681,
+      "step": 90
+    },
+    {
+      "epoch": 0.1890077095249938,
+      "grad_norm": 0.3505592753718136,
+      "learning_rate": 0.00029300945880823956,
+      "loss": 0.9653,
+      "step": 95
+    },
+    {
+      "epoch": 0.19895548371051977,
+      "grad_norm": 0.28451648461527196,
+      "learning_rate": 0.0002913467433828382,
+      "loss": 0.9511,
+      "step": 100
+    },
+    {
+      "epoch": 0.20890325789604575,
+      "grad_norm": 0.3138636289880952,
+      "learning_rate": 0.00028951258147696967,
+      "loss": 0.9572,
+      "step": 105
+    },
+    {
+      "epoch": 0.21885103208157175,
+      "grad_norm": 0.2732249403521582,
+      "learning_rate": 0.00028750919783664407,
+      "loss": 0.9617,
+      "step": 110
+    },
+    {
+      "epoch": 0.22879880626709773,
+      "grad_norm": 0.311679721602736,
+      "learning_rate": 0.000285339022465312,
+      "loss": 0.9484,
+      "step": 115
+    },
+    {
+      "epoch": 0.23874658045262373,
+      "grad_norm": 0.2676711284506792,
+      "learning_rate": 0.00028300468767639305,
+      "loss": 0.9397,
+      "step": 120
+    },
+    {
+      "epoch": 0.2486943546381497,
+      "grad_norm": 0.25074023315013144,
+      "learning_rate": 0.00028050902490041194,
+      "loss": 0.9457,
+      "step": 125
+    },
+    {
+      "epoch": 0.2586421288236757,
+      "grad_norm": 0.27129765694428115,
+      "learning_rate": 0.00027785506125061604,
+      "loss": 0.9268,
+      "step": 130
+    },
+    {
+      "epoch": 0.2685899030092017,
+      "grad_norm": 2.105526076305897,
+      "learning_rate": 0.00027504601585123963,
+      "loss": 0.9459,
+      "step": 135
+    },
+    {
+      "epoch": 0.27853767719472766,
+      "grad_norm": 0.27202583577769746,
+      "learning_rate": 0.00027208529593286804,
+      "loss": 0.9395,
+      "step": 140
+    },
+    {
+      "epoch": 0.28848545138025367,
+      "grad_norm": 0.26240068306380243,
+      "learning_rate": 0.00026897649269963866,
+      "loss": 0.9166,
+      "step": 145
+    },
+    {
+      "epoch": 0.29843322556577967,
+      "grad_norm": 0.2631437322326041,
+      "learning_rate": 0.00026572337697329144,
+      "loss": 0.92,
+      "step": 150
+    },
+    {
+      "epoch": 0.3083809997513056,
+      "grad_norm": 0.2780992526939389,
+      "learning_rate": 0.00026232989461935164,
+      "loss": 0.929,
+      "step": 155
+    },
+    {
+      "epoch": 0.3183287739368316,
+      "grad_norm": 0.2644307921832001,
+      "learning_rate": 0.000258800161760994,
+      "loss": 0.9119,
+      "step": 160
+    },
+    {
+      "epoch": 0.3282765481223576,
+      "grad_norm": 0.27009906970951136,
+      "learning_rate": 0.0002551384597863925,
+      "loss": 0.9141,
+      "step": 165
+    },
+    {
+      "epoch": 0.33822432230788363,
+      "grad_norm": 0.24907417369935828,
+      "learning_rate": 0.0002513492301556124,
+      "loss": 0.9045,
+      "step": 170
+    },
+    {
+      "epoch": 0.3481720964934096,
+      "grad_norm": 0.4637150153647203,
+      "learning_rate": 0.0002474370690133423,
+      "loss": 0.9185,
+      "step": 175
+    },
+    {
+      "epoch": 0.3581198706789356,
+      "grad_norm": 0.28404376319877433,
+      "learning_rate": 0.00024340672161400278,
+      "loss": 0.9224,
+      "step": 180
+    },
+    {
+      "epoch": 0.3680676448644616,
+      "grad_norm": 0.2633604383062445,
+      "learning_rate": 0.00023926307656599145,
+      "loss": 0.9049,
+      "step": 185
+    },
+    {
+      "epoch": 0.3780154190499876,
+      "grad_norm": 0.3089691505145177,
+      "learning_rate": 0.00023501115990204728,
+      "loss": 0.906,
+      "step": 190
+    },
+    {
+      "epoch": 0.38796319323551354,
+      "grad_norm": 0.2712113108379062,
+      "learning_rate": 0.00023065612898292607,
+      "loss": 0.9033,
+      "step": 195
+    },
+    {
+      "epoch": 0.39791096742103954,
+      "grad_norm": 0.2759532883918752,
+      "learning_rate": 0.00022620326624178135,
+      "loss": 0.9047,
+      "step": 200
+    },
+    {
+      "epoch": 0.40785874160656554,
+      "grad_norm": 0.25413106263769636,
+      "learning_rate": 0.0002216579727768394,
+      "loss": 0.8884,
+      "step": 205
+    },
+    {
+      "epoch": 0.4178065157920915,
+      "grad_norm": 0.2679789855086274,
+      "learning_rate": 0.00021702576180013906,
+      "loss": 0.892,
+      "step": 210
+    },
+    {
+      "epoch": 0.4277542899776175,
+      "grad_norm": 0.2531713028754476,
+      "learning_rate": 0.00021231225195028297,
+      "loss": 0.8907,
+      "step": 215
+    },
+    {
+      "epoch": 0.4377020641631435,
+      "grad_norm": 0.24842966918028864,
+      "learning_rate": 0.00020752316047731214,
+      "loss": 0.882,
+      "step": 220
+    },
+    {
+      "epoch": 0.4476498383486695,
+      "grad_norm": 0.23591143252036872,
+      "learning_rate": 0.00020266429630796956,
+      "loss": 0.8846,
+      "step": 225
+    },
+    {
+      "epoch": 0.45759761253419545,
+      "grad_norm": 0.23767648270009806,
+      "learning_rate": 0.00019774155299976477,
+      "loss": 0.8793,
+      "step": 230
+    },
+    {
+      "epoch": 0.46754538671972146,
+      "grad_norm": 0.2271591056583853,
+      "learning_rate": 0.00019276090159238524,
+      "loss": 0.8741,
+      "step": 235
+    },
+    {
+      "epoch": 0.47749316090524746,
+      "grad_norm": 0.22901636532179012,
+      "learning_rate": 0.000187728383365126,
+      "loss": 0.8837,
+      "step": 240
+    },
+    {
+      "epoch": 0.48744093509077346,
+      "grad_norm": 0.22668623781094616,
+      "learning_rate": 0.0001826501025091223,
+      "loss": 0.8735,
+      "step": 245
+    },
+    {
+      "epoch": 0.4973887092762994,
+      "grad_norm": 0.23947671322760095,
+      "learning_rate": 0.00017753221872327318,
+      "loss": 0.8692,
+      "step": 250
+    },
+    {
+      "epoch": 0.5073364834618255,
+      "grad_norm": 0.26156533719751,
+      "learning_rate": 0.00017238093974283674,
+      "loss": 0.8625,
+      "step": 255
+    },
+    {
+      "epoch": 0.5172842576473514,
+      "grad_norm": 0.25671509792902836,
+      "learning_rate": 0.00016720251380976007,
+      "loss": 0.8604,
+      "step": 260
+    },
+    {
+      "epoch": 0.5272320318328774,
+      "grad_norm": 0.24704965941321674,
+      "learning_rate": 0.00016200322209387663,
+      "loss": 0.8626,
+      "step": 265
+    },
+    {
+      "epoch": 0.5371798060184034,
+      "grad_norm": 0.2514007967545614,
+      "learning_rate": 0.00015678937107416343,
+      "loss": 0.8528,
+      "step": 270
+    },
+    {
+      "epoch": 0.5471275802039294,
+      "grad_norm": 0.24028475499604857,
+      "learning_rate": 0.00015156728488929967,
+      "loss": 0.8574,
+      "step": 275
+    },
+    {
+      "epoch": 0.5570753543894553,
+      "grad_norm": 0.23213673166180135,
+      "learning_rate": 0.0001463432976668051,
+      "loss": 0.86,
+      "step": 280
+    },
+    {
+      "epoch": 0.5670231285749814,
+      "grad_norm": 0.23102662796096035,
+      "learning_rate": 0.00014112374584006253,
+      "loss": 0.8617,
+      "step": 285
+    },
+    {
+      "epoch": 0.5769709027605073,
+      "grad_norm": 0.23635078987821154,
+      "learning_rate": 0.00013591496046254278,
+      "loss": 0.8468,
+      "step": 290
+    },
+    {
+      "epoch": 0.5869186769460333,
+      "grad_norm": 0.2473298815208931,
+      "learning_rate": 0.00013072325952855624,
+      "loss": 0.8465,
+      "step": 295
+    },
+    {
+      "epoch": 0.5968664511315593,
+      "grad_norm": 0.22199589321301555,
+      "learning_rate": 0.00012555494030984393,
+      "loss": 0.8474,
+      "step": 300
+    },
+    {
+      "epoch": 0.6068142253170853,
+      "grad_norm": 0.224149793992071,
+      "learning_rate": 0.00012041627171730368,
+      "loss": 0.8523,
+      "step": 305
+    },
+    {
+      "epoch": 0.6167619995026112,
+      "grad_norm": 0.2216854964579631,
+      "learning_rate": 0.00011531348669711734,
+      "loss": 0.8296,
+      "step": 310
+    },
+    {
+      "epoch": 0.6267097736881373,
+      "grad_norm": 0.25823608221836053,
+      "learning_rate": 0.00011025277467050076,
+      "loss": 0.8275,
+      "step": 315
+    },
+    {
+      "epoch": 0.6366575478736632,
+      "grad_norm": 0.25511921593962283,
+      "learning_rate": 0.00010524027402624775,
+      "loss": 0.8379,
+      "step": 320
+    },
+    {
+      "epoch": 0.6466053220591893,
+      "grad_norm": 0.2169176240841302,
+      "learning_rate": 0.00010028206467517357,
+      "loss": 0.842,
+      "step": 325
+    },
+    {
+      "epoch": 0.6565530962447153,
+      "grad_norm": 0.23684621611339568,
+      "learning_rate": 9.538416067548939e-05,
+      "loss": 0.8363,
+      "step": 330
+    },
+    {
+      "epoch": 0.6665008704302412,
+      "grad_norm": 0.21588755800082085,
+      "learning_rate": 9.055250293805247e-05,
+      "loss": 0.8257,
+      "step": 335
+    },
+    {
+      "epoch": 0.6764486446157673,
+      "grad_norm": 0.22987884680681675,
+      "learning_rate": 8.579295202034084e-05,
+      "loss": 0.8434,
+      "step": 340
+    },
+    {
+      "epoch": 0.6863964188012932,
+      "grad_norm": 0.23044188787166803,
+      "learning_rate": 8.111128101789177e-05,
+      "loss": 0.8368,
+      "step": 345
+    },
+    {
+      "epoch": 0.6963441929868192,
+      "grad_norm": 0.20871504598447846,
+      "learning_rate": 7.651316856182797e-05,
+      "loss": 0.8235,
+      "step": 350
+    },
+    {
+      "epoch": 0.7062919671723452,
+      "grad_norm": 0.21764586591268964,
+      "learning_rate": 7.200419193096416e-05,
+      "loss": 0.8366,
+      "step": 355
+    },
+    {
+      "epoch": 0.7162397413578712,
+      "grad_norm": 0.20093905811705248,
+      "learning_rate": 6.758982028684842e-05,
+      "loss": 0.8212,
+      "step": 360
+    },
+    {
+      "epoch": 0.7261875155433971,
+      "grad_norm": 0.20201151370456955,
+      "learning_rate": 6.327540803994507e-05,
+      "loss": 0.8132,
+      "step": 365
+    },
+    {
+      "epoch": 0.7361352897289232,
+      "grad_norm": 0.20205806825962228,
+      "learning_rate": 5.9066188355004337e-05,
+      "loss": 0.8115,
+      "step": 370
+    },
+    {
+      "epoch": 0.7460830639144491,
+      "grad_norm": 0.4492212407917222,
+      "learning_rate": 5.4967266803496726e-05,
+      "loss": 0.8178,
+      "step": 375
+    },
+    {
+      "epoch": 0.7560308380999752,
+      "grad_norm": 0.20165879598033634,
+      "learning_rate": 5.0983615170812656e-05,
+      "loss": 0.8202,
+      "step": 380
+    },
+    {
+      "epoch": 0.7659786122855011,
+      "grad_norm": 0.20731013148125454,
+      "learning_rate": 4.7120065425736744e-05,
+      "loss": 0.8224,
+      "step": 385
+    },
+    {
+      "epoch": 0.7759263864710271,
+      "grad_norm": 0.19813282037840257,
+      "learning_rate": 4.3381303859513076e-05,
+      "loss": 0.8031,
+      "step": 390
+    },
+    {
+      "epoch": 0.7858741606565531,
+      "grad_norm": 0.19914236210141723,
+      "learning_rate": 3.977186540161016e-05,
+      "loss": 0.8146,
+      "step": 395
+    },
+    {
+      "epoch": 0.7958219348420791,
+      "grad_norm": 0.20390757189890973,
+      "learning_rate": 3.629612811907965e-05,
+      "loss": 0.8132,
+      "step": 400
+    },
+    {
+      "epoch": 0.805769709027605,
+      "grad_norm": 0.19393901917008016,
+      "learning_rate": 3.295830790618167e-05,
+      "loss": 0.8142,
+      "step": 405
+    },
+    {
+      "epoch": 0.8157174832131311,
+      "grad_norm": 0.1907323400363912,
+      "learning_rate": 2.976245337071748e-05,
+      "loss": 0.8129,
+      "step": 410
+    },
+    {
+      "epoch": 0.825665257398657,
+      "grad_norm": 0.199622342411806,
+      "learning_rate": 2.671244092327191e-05,
+      "loss": 0.7951,
+      "step": 415
+    },
+    {
+      "epoch": 0.835613031584183,
+      "grad_norm": 0.20412611674653627,
+      "learning_rate": 2.38119700753228e-05,
+      "loss": 0.8143,
+      "step": 420
+    },
+    {
+      "epoch": 0.845560805769709,
+      "grad_norm": 0.19936335181674378,
+      "learning_rate": 2.106455895191985e-05,
+      "loss": 0.802,
+      "step": 425
+    },
+    {
+      "epoch": 0.855508579955235,
+      "grad_norm": 0.19669632646100263,
+      "learning_rate": 1.847354002437588e-05,
+      "loss": 0.7948,
+      "step": 430
+    },
+    {
+      "epoch": 0.865456354140761,
+      "grad_norm": 0.19474265071015343,
+      "learning_rate": 1.6042056068147402e-05,
+      "loss": 0.8078,
+      "step": 435
+    },
+    {
+      "epoch": 0.875404128326287,
+      "grad_norm": 0.19134910562287546,
+      "learning_rate": 1.3773056350806022e-05,
+      "loss": 0.8067,
+      "step": 440
+    },
+    {
+      "epoch": 0.885351902511813,
+      "grad_norm": 0.18803806717884775,
+      "learning_rate": 1.1669293054725392e-05,
+      "loss": 0.7952,
+      "step": 445
+    },
+    {
+      "epoch": 0.895299676697339,
+      "grad_norm": 0.18932335388504165,
+      "learning_rate": 9.7333179388228e-06,
+      "loss": 0.8102,
+      "step": 450
+    },
+    {
+      "epoch": 0.905247450882865,
+      "grad_norm": 0.19773216218571474,
+      "learning_rate": 7.967479243403913e-06,
+      "loss": 0.8015,
+      "step": 455
+    },
+    {
+      "epoch": 0.9151952250683909,
+      "grad_norm": 0.18837812375636145,
+      "learning_rate": 6.373918841865727e-06,
+      "loss": 0.7997,
+      "step": 460
+    },
+    {
+      "epoch": 0.925142999253917,
+      "grad_norm": 0.18803298571004357,
+      "learning_rate": 4.954569642711964e-06,
+      "loss": 0.8068,
+      "step": 465
+    },
+    {
+      "epoch": 0.9350907734394429,
+      "grad_norm": 0.1810703612957102,
+      "learning_rate": 3.711153245032361e-06,
+      "loss": 0.7992,
+      "step": 470
+    },
+    {
+      "epoch": 0.9450385476249689,
+      "grad_norm": 0.18601132727527311,
+      "learning_rate": 2.645177850289787e-06,
+      "loss": 0.8039,
+      "step": 475
+    },
+    {
+      "epoch": 0.9549863218104949,
+      "grad_norm": 0.18789475936900635,
+      "learning_rate": 1.7579364329477375e-06,
+      "loss": 0.8024,
+      "step": 480
+    },
+    {
+      "epoch": 0.9649340959960209,
+      "grad_norm": 0.7153792725698854,
+      "learning_rate": 1.0505051721574398e-06,
+      "loss": 0.8047,
+      "step": 485
+    },
+    {
+      "epoch": 0.9748818701815469,
+      "grad_norm": 0.19016384162520752,
+      "learning_rate": 5.23742146406858e-07,
+      "loss": 0.8004,
+      "step": 490
+    },
+    {
+      "epoch": 0.9848296443670729,
+      "grad_norm": 0.19392554346569527,
+      "learning_rate": 1.7828629271456894e-07,
+      "loss": 0.7991,
+      "step": 495
+    },
+    {
+      "epoch": 0.9947774185525988,
+      "grad_norm": 0.19065289345162353,
+      "learning_rate": 1.4556631631429393e-08,
+      "loss": 0.8032,
+      "step": 500
+    },
+    {
+      "epoch": 0.9987565282268093,
+      "eval_loss": 1.1664291620254517,
+      "eval_runtime": 1405.6452,
+      "eval_samples_per_second": 16.582,
+      "eval_steps_per_second": 1.037,
+      "step": 502
+    },
+    {
+      "epoch": 0.9987565282268093,
+      "step": 502,
+      "total_flos": 234414617395200.0,
+      "train_loss": 0.8937448220423968,
+      "train_runtime": 13768.4166,
+      "train_samples_per_second": 4.672,
+      "train_steps_per_second": 0.036
+    }
+  ],
+  "logging_steps": 5,
+  "max_steps": 502,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 1,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": false,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 234414617395200.0,
+  "train_batch_size": 4,
+  "trial_name": null,
+  "trial_params": null
+}