End of training

README.md

@@ -0,0 +1,75 @@
+---
+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: 2.9790
+
+## 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 |
+|:-------------:|:-----:|:-----:|:---------------:|
+| 4.8696        | 0.05  | 2000  | 3.8588          |
+| 3.7027        | 0.1   | 4000  | 3.6106          |
+| 3.5648        | 0.15  | 6000  | 3.5014          |
+| 3.448         | 0.2   | 8000  | 3.4153          |
+| 3.3884        | 0.25  | 10000 | 3.3650          |
+| 3.3462        | 0.29  | 12000 | 3.3280          |
+| 3.3155        | 0.34  | 14000 | 3.3053          |
+| 3.2932        | 0.39  | 16000 | 3.2891          |
+| 3.2762        | 0.44  | 18000 | 3.2673          |
+| 3.2594        | 0.49  | 20000 | 3.2533          |
+| 3.2432        | 0.54  | 22000 | 3.2398          |
+| 3.2286        | 0.59  | 24000 | 3.2186          |
+| 3.2083        | 0.64  | 26000 | 3.1957          |
+| 3.1867        | 0.69  | 28000 | 3.1769          |
+| 3.1676        | 0.74  | 30000 | 3.1568          |
+| 3.14          | 0.79  | 32000 | 3.1286          |
+| 3.114         | 0.83  | 34000 | 3.1006          |
+| 3.0848        | 0.88  | 36000 | 3.0696          |
+| 3.0511        | 0.93  | 38000 | 3.0301          |
+| 3.005         | 0.98  | 40000 | 2.9790          |
+
+
+### Framework versions
+
+- Transformers 4.38.2
+- Pytorch 2.2.1+cu121
+- Datasets 2.18.0
+- 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:a9c5c5f345079769e7d5009917b9ff49d4ced4d001e44923525a0bf70b88d5fa
+oid sha256:0f1ee1502ac904b2fb98516525a5b48da3aeec927df5b75cf426c788eeabf0f0
 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)
+