End of training

README.md

@@ -0,0 +1,91 @@
+---
+library_name: transformers
+tags:
+- generated_from_trainer
+model-index:
+- name: myBit-Llama2-jp-127M-2B4TLike-aozora-sort-3epc
+  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-2B4TLike-aozora-sort-3epc
+
+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: 3.2302
+
+## 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: 24
+- eval_batch_size: 24
+- seed: 42
+- gradient_accumulation_steps: 4
+- total_train_batch_size: 96
+- optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.95) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
+- lr_scheduler_type: linear
+- lr_scheduler_warmup_steps: 100
+- num_epochs: 3
+
+### Training results
+
+| Training Loss | Epoch  | Step | Validation Loss |
+|:-------------:|:------:|:----:|:---------------:|
+| 6.9258        | 0.0883 | 100  | 5.2708          |
+| 4.7937        | 0.1765 | 200  | 4.4660          |
+| 4.2565        | 0.2648 | 300  | 4.1755          |
+| 3.9951        | 0.3530 | 400  | 4.0060          |
+| 3.8438        | 0.4413 | 500  | 3.8854          |
+| 3.7223        | 0.5296 | 600  | 3.7829          |
+| 3.6523        | 0.6178 | 700  | 3.7125          |
+| 3.5985        | 0.7061 | 800  | 3.6535          |
+| 3.5666        | 0.7944 | 900  | 3.6039          |
+| 3.5519        | 0.8826 | 1000 | 3.5693          |
+| 3.5365        | 0.9709 | 1100 | 3.5404          |
+| 3.6085        | 1.0591 | 1200 | 3.5638          |
+| 3.4953        | 1.1474 | 1300 | 3.4983          |
+| 3.425         | 1.2357 | 1400 | 3.4737          |
+| 3.3693        | 1.3239 | 1500 | 3.4579          |
+| 3.3396        | 1.4122 | 1600 | 3.4431          |
+| 3.3187        | 1.5004 | 1700 | 3.4259          |
+| 3.3013        | 1.5887 | 1800 | 3.4121          |
+| 3.3036        | 1.6770 | 1900 | 3.4004          |
+| 3.2947        | 1.7652 | 2000 | 3.3808          |
+| 3.3041        | 1.8535 | 2100 | 3.3653          |
+| 3.304         | 1.9417 | 2200 | 3.3541          |
+| 3.3582        | 2.0300 | 2300 | 3.4233          |
+| 3.3097        | 2.1183 | 2400 | 3.3351          |
+| 3.2426        | 2.2065 | 2500 | 3.3234          |
+| 3.2034        | 2.2948 | 2600 | 3.3149          |
+| 3.1675        | 2.3831 | 2700 | 3.3033          |
+| 3.1611        | 2.4713 | 2800 | 3.2953          |
+| 3.1344        | 2.5596 | 2900 | 3.2832          |
+| 3.1391        | 2.6478 | 3000 | 3.2729          |
+| 3.1324        | 2.7361 | 3100 | 3.2572          |
+| 3.1355        | 2.8244 | 3200 | 3.2440          |
+| 3.1417        | 2.9126 | 3300 | 3.2302          |
+
+
+### Framework versions
+
+- Transformers 4.47.1
+- Pytorch 2.6.0+cu124
+- Datasets 3.5.1
+- Tokenizers 0.21.1

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "pad_token_id": 2,
+  "transformers_version": "4.47.1"
+}

modeling_bit_llama.py

@@ -0,0 +1,169 @@
+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, BitLinear158b
+import torch
+from torch import nn
+
+class BitLlamaConfig(LlamaConfig):
+    model_type = "bit_llama"
+
+    def __init__(self, bitnet_type="1.58b", bits=8, **kwargs):
+        super().__init__(**kwargs)
+        self.bitnet_type = bitnet_type
+        if self.bitnet_type not in ["1.58b", "1b"]:
+            raise ValueError("bitnet_type must be either '1.58b' or '1b'.")
+        self.bits = bits
+
+
+class BitLlamaMLP(LlamaMLP):
+    def __init__(self, config):
+        super().__init__(config)
+        if config.bitnet_type=="1b":
+            self.gate_proj = BitLinear(self.hidden_size, self.intermediate_size, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits, flg_before_linear=False)
+            self.up_proj = BitLinear(self.hidden_size, self.intermediate_size, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits, flg_before_linear=True)
+            self.down_proj = BitLinear(self.intermediate_size, self.hidden_size, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits, flg_before_linear=True)
+        elif config.bitnet_type=="1.58b":
+            self.gate_proj = BitLinear158b(self.hidden_size, self.intermediate_size, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits)
+            self.up_proj = BitLinear158b(self.hidden_size, self.intermediate_size, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits)
+            self.down_proj = BitLinear158b(self.intermediate_size, self.hidden_size, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits)
+        else:
+            raise ValueError("bitnet_type must be either '1.58b' or '1b'.")
+        
+class BitLlamaAttention(LlamaAttention):
+    def __init__(self, config: BitLlamaConfig, layer_idx: Optional[int] = None):
+        super().__init__(config, layer_idx)  # Set `layer_idx` to avoid `self.layer_idx` to be `None`
+        if config.bitnet_type=="1b":
+            self.q_proj = BitLinear(self.hidden_size, self.num_heads * self.head_dim, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits, flg_before_linear=True)
+            self.k_proj = BitLinear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits, flg_before_linear=True)
+            self.v_proj = BitLinear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits, flg_before_linear=True)
+            self.o_proj = BitLinear(self.hidden_size, self.hidden_size, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits, flg_before_linear=True)
+        elif config.bitnet_type=="1.58b":
+            self.q_proj = BitLinear158b(self.hidden_size, self.num_heads * self.head_dim, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits)
+            self.k_proj = BitLinear158b(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits)
+            self.v_proj = BitLinear158b(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits)
+            self.o_proj = BitLinear158b(self.hidden_size, self.hidden_size, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits)
+        else:
+            raise ValueError("bitnet_type must be either '1.58b' or '1b'.")
+
+class BitLlamaFlashAttention2(LlamaFlashAttention2):
+    def __init__(self, config: BitLlamaConfig, layer_idx: Optional[int] = None):
+        super().__init__(config, layer_idx)
+        if config.bitnet_type=="1b":
+            self.q_proj = BitLinear(self.hidden_size, self.num_heads * self.head_dim, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits, flg_before_linear=True)
+            self.k_proj = BitLinear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits, flg_before_linear=True)
+            self.v_proj = BitLinear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits, flg_before_linear=True)
+            self.o_proj = BitLinear(self.hidden_size, self.hidden_size, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits, flg_before_linear=True)
+        elif config.bitnet_type=="1.58b":
+            self.q_proj = BitLinear158b(self.hidden_size, self.num_heads * self.head_dim, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits)
+            self.k_proj = BitLinear158b(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits)
+            self.v_proj = BitLinear158b(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits)
+            self.o_proj = BitLinear158b(self.hidden_size, self.hidden_size, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits)
+        else:
+            raise ValueError("bitnet_type must be either '1.58b' or '1b'.")
+
+class BitLlamaSdpaAttention(LlamaSdpaAttention):
+    def __init__(self, config: BitLlamaConfig, layer_idx: Optional[int] = None):
+        super().__init__(config, layer_idx)
+        if config.bitnet_type=="1b":
+            self.q_proj = BitLinear(self.hidden_size, self.num_heads * self.head_dim, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits, flg_before_linear=True)
+            self.k_proj = BitLinear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits, flg_before_linear=True)
+            self.v_proj = BitLinear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits, flg_before_linear=True)
+            self.o_proj = BitLinear(self.hidden_size, self.hidden_size, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits, flg_before_linear=True)
+        elif config.bitnet_type=="1.58b":
+            self.q_proj = BitLinear158b(self.hidden_size, self.num_heads * self.head_dim, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits)
+            self.k_proj = BitLinear158b(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits)
+            self.v_proj = BitLinear158b(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits)
+            self.o_proj = BitLinear158b(self.hidden_size, self.hidden_size, bias=False, rms_norm_eps=config.rms_norm_eps, bits=config.bits)
+        else:
+            raise ValueError("bitnet_type must be either '1.58b' or '1b'.")
+
+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)
+