End of training

README.md

@@ -0,0 +1,116 @@
+---
+tags:
+- generated_from_trainer
+model-index:
+- name: myBit-Llama2-jp-127M-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. -->
+
+# myBit-Llama2-jp-127M-4
+
+This model is a fine-tuned version of [](https://huggingface.co/) on an unknown dataset.
+It achieves the following results on the evaluation set:
+- Loss: 3.0920
+
+## 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: 0.0024
+- train_batch_size: 96
+- eval_batch_size: 96
+- seed: 42
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: polynomial
+- lr_scheduler_warmup_steps: 5000
+- num_epochs: 1
+
+### Training results
+
+| Training Loss | Epoch | Step   | Validation Loss |
+|:-------------:|:-----:|:------:|:---------------:|
+| 4.6932        | 0.02  | 2000   | 3.3504          |
+| 3.252         | 0.03  | 4000   | 3.1987          |
+| 3.1379        | 0.05  | 6000   | 3.0873          |
+| 3.0466        | 0.06  | 8000   | 3.0233          |
+| 2.9925        | 0.08  | 10000  | 2.9819          |
+| 2.9553        | 0.1   | 12000  | 2.9471          |
+| 2.9292        | 0.11  | 14000  | 2.9278          |
+| 2.9158        | 0.13  | 16000  | 2.9159          |
+| 2.907         | 0.15  | 18000  | 2.9084          |
+| 2.9018        | 0.16  | 20000  | 2.9015          |
+| 2.8945        | 0.18  | 22000  | 2.8971          |
+| 2.8901        | 0.19  | 24000  | 2.9014          |
+| 2.8906        | 0.21  | 26000  | 2.8980          |
+| 2.8943        | 0.23  | 28000  | 2.9010          |
+| 2.8985        | 0.24  | 30000  | 2.9165          |
+| 3.0191        | 0.26  | 32000  | 3.3484          |
+| 3.5616        | 0.28  | 34000  | 3.4516          |
+| 3.2849        | 0.29  | 36000  | 3.0454          |
+| 3.2425        | 0.31  | 38000  | 3.7183          |
+| 3.655         | 0.32  | 40000  | 3.8947          |
+| 3.3151        | 0.34  | 42000  | 3.6150          |
+| 3.3482        | 0.36  | 44000  | 3.1714          |
+| 3.1433        | 0.37  | 46000  | 3.1073          |
+| 3.0462        | 0.39  | 48000  | 2.9786          |
+| 3.0889        | 0.41  | 50000  | 3.3002          |
+| 3.4652        | 0.42  | 52000  | 3.3920          |
+| 3.3726        | 0.44  | 54000  | 3.1293          |
+| 3.2314        | 0.45  | 56000  | 3.3841          |
+| 3.5303        | 0.47  | 58000  | 3.3865          |
+| 3.2828        | 0.49  | 60000  | 3.2591          |
+| 3.0219        | 0.5   | 62000  | 2.9431          |
+| 3.0714        | 0.52  | 64000  | 3.2328          |
+| 3.1354        | 0.54  | 66000  | 3.0794          |
+| 3.2194        | 0.55  | 68000  | 3.1326          |
+| 3.394         | 0.57  | 70000  | 3.5974          |
+| 3.2692        | 0.58  | 72000  | 3.1522          |
+| 3.1513        | 0.6   | 74000  | 3.1398          |
+| 3.2473        | 0.62  | 76000  | 3.1921          |
+| 3.1717        | 0.63  | 78000  | 3.1827          |
+| 3.211         | 0.65  | 80000  | 3.0845          |
+| 2.9955        | 0.67  | 82000  | 3.0229          |
+| 3.3145        | 0.68  | 84000  | 3.3382          |
+| 3.0703        | 0.7   | 86000  | 3.5395          |
+| 3.234         | 0.71  | 88000  | 2.9486          |
+| 3.1077        | 0.73  | 90000  | 2.9488          |
+| 3.1097        | 0.75  | 92000  | 2.9597          |
+| 2.8979        | 0.76  | 94000  | 3.0215          |
+| 3.236         | 0.78  | 96000  | 3.1758          |
+| 3.1365        | 0.8   | 98000  | 3.4841          |
+| 3.1954        | 0.81  | 100000 | 2.9520          |
+| 3.2054        | 0.83  | 102000 | 3.6384          |
+| 3.2957        | 0.84  | 104000 | 2.9212          |
+| 2.9358        | 0.86  | 106000 | 3.0166          |
+| 3.221         | 0.88  | 108000 | 3.3753          |
+| 3.2241        | 0.89  | 110000 | 3.0858          |
+| 3.1497        | 0.91  | 112000 | 2.9252          |
+| 3.198         | 0.93  | 114000 | 3.8514          |
+| 3.1427        | 0.94  | 116000 | 4.1130          |
+| 3.2371        | 0.96  | 118000 | 2.8639          |
+| 3.2576        | 0.97  | 120000 | 2.9192          |
+| 3.3229        | 0.99  | 122000 | 3.0920          |
+
+
+### Framework versions
+
+- Transformers 4.38.2
+- Pytorch 2.3.0+cu121
+- Datasets 2.20.0
+- Tokenizers 0.15.2

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "transformers_version": "4.38.2"
+}

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:788c820a2b381e8ff7123dbad9d370af4c617d6e6cb769dc7f2eefdf18a0acfd
+oid sha256:0342f4324869091d5f08b1fdc85932b8dd0c4d690da3e0747329ddb83fc78d3c
 size 510960712

modeling_bit_llama.py

@@ -0,0 +1,134 @@
+import warnings
+from typing import Optional, Tuple
+from transformers.models.llama.modeling_llama import (
+    LlamaConfig, 
+    LlamaModel,
+    LlamaForCausalLM,
+    LlamaAttention,
+    LlamaFlashAttention2,
+    LlamaSdpaAttention,
+    LlamaMLP,
+    LlamaDecoderLayer,
+)
+from mybitnet.bitnet import BitLinear
+import torch
+from torch import nn
+
+class BitLlamaConfig(LlamaConfig):
+    model_type = "bit_llama"
+
+    def __init__(self, bits=8, **kwargs):
+        super().__init__(**kwargs)
+        self.bits = bits
+
+class BitLlamaMLP(LlamaMLP):
+    def __init__(self, config):
+        super().__init__(config)
+        self.gate_proj = BitLinear(self.hidden_size, self.intermediate_size, bias=False, bits=config.bits, flg_before_linear=False)
+        self.up_proj = BitLinear(self.hidden_size, self.intermediate_size, bias=False, bits=config.bits, flg_before_linear=True)
+        self.down_proj = BitLinear(self.intermediate_size, self.hidden_size, bias=False, bits=config.bits, flg_before_linear=True)
+        
+class BitLlamaAttention(LlamaAttention):
+    def __init__(self, config: BitLlamaConfig, layer_idx: Optional[int] = None):
+        super().__init__(config)
+        self.q_proj = BitLinear(self.hidden_size, self.num_heads * self.head_dim, bias=False, bits=config.bits, flg_before_linear=True)
+        self.k_proj = BitLinear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False, bits=config.bits, flg_before_linear=True)
+        self.v_proj = BitLinear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False, bits=config.bits, flg_before_linear=True)
+        self.o_proj = BitLinear(self.hidden_size, self.hidden_size, bias=False, bits=config.bits, flg_before_linear=True)
+
+class BitLlamaFlashAttention2(LlamaFlashAttention2):
+    def __init__(self, config: BitLlamaConfig, layer_idx: Optional[int] = None):
+        super().__init__(config, layer_idx)
+        self.q_proj = BitLinear(self.hidden_size, self.num_heads * self.head_dim, bias=False, bits=config.bits, flg_before_linear=True)
+        self.k_proj = BitLinear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False, bits=config.bits, flg_before_linear=True)
+        self.v_proj = BitLinear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False, bits=config.bits, flg_before_linear=True)
+        self.o_proj = BitLinear(self.hidden_size, self.hidden_size, bias=False, bits=config.bits, flg_before_linear=True)
+
+class BitLlamaSdpaAttention(LlamaSdpaAttention):
+    def __init__(self, config: BitLlamaConfig, layer_idx: Optional[int] = None):
+        super().__init__(config, layer_idx)
+        self.q_proj = BitLinear(self.hidden_size, self.num_heads * self.head_dim, bias=False, bits=config.bits, flg_before_linear=True)
+        self.k_proj = BitLinear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False, bits=config.bits, flg_before_linear=True)
+        self.v_proj = BitLinear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False, bits=config.bits, flg_before_linear=True)
+        self.o_proj = BitLinear(self.hidden_size, self.hidden_size, bias=False, bits=config.bits, flg_before_linear=True)
+
+BITLLAMA_ATTENTION_CLASSES = {
+    "eager": BitLlamaAttention,
+    "flash_attention_2": BitLlamaFlashAttention2,
+    "sdpa": BitLlamaSdpaAttention,
+}
+
+class BitLlamaDecoderLayer(LlamaDecoderLayer):
+    def __init__(self, config: BitLlamaConfig, layer_idx: int):
+        super().__init__(config, layer_idx)
+        self.self_attn = BITLLAMA_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
+        self.mlp = BitLlamaMLP(config)
+        del self.input_layernorm
+        del self.post_attention_layernorm
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        """
+        refers: https://github.com/huggingface/transformers/blob/c5f0288bc7d76f65996586f79f69fba8867a0e67/src/transformers/models/llama/modeling_llama.py#L693
+        """
+        if "padding_mask" in kwargs:
+            warnings.warn(
+                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+            )
+
+        residual = 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,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+class BitLlamaModel(LlamaModel):
+    config_class = BitLlamaConfig
+
+    def __init__(self, config: BitLlamaConfig):
+        super().__init__(config)
+        self.layers = nn.ModuleList(
+            [BitLlamaDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+
+class BitLlamaForCausalLM(LlamaForCausalLM):
+    config_class = BitLlamaConfig
+
+    def __init__(self, config: BitLlamaConfig):
+        super().__init__(config)
+        self.model = BitLlamaModel(config)
+        self.lm_head = BitLinear(config.hidden_size, config.vocab_size, bias=False, bits=config.bits, flg_before_linear=True)
+