End of training

README.md

@@ -0,0 +1,75 @@
+---
+tags:
+- generated_from_trainer
+model-index:
+- name: myBit-Llama2-jp-127M-3
+  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-3
+
+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: 3.8396
+
+## 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: 0.00024
+- 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 |
+|:-------------:|:-----:|:-----:|:---------------:|
+| 6.9048        | 0.05  | 2000  | 4.7602          |
+| 4.4421        | 0.1   | 4000  | 4.2117          |
+| 4.0625        | 0.15  | 6000  | 3.9227          |
+| 3.807         | 0.2   | 8000  | 3.7181          |
+| 3.6547        | 0.25  | 10000 | 3.5929          |
+| 3.5296        | 0.29  | 12000 | 3.4812          |
+| 3.4492        | 0.34  | 14000 | 3.4236          |
+| 3.4065        | 0.39  | 16000 | 3.3923          |
+| 3.3816        | 0.44  | 18000 | 3.3778          |
+| 3.3815        | 0.49  | 20000 | 3.3907          |
+| 3.431         | 0.54  | 22000 | 3.4870          |
+| 3.5507        | 0.59  | 24000 | 3.5969          |
+| 3.6557        | 0.64  | 26000 | 3.6918          |
+| 3.715         | 0.69  | 28000 | 3.7377          |
+| 3.7646        | 0.74  | 30000 | 3.7620          |
+| 3.8005        | 0.79  | 32000 | 3.8221          |
+| 3.8288        | 0.83  | 34000 | 3.8550          |
+| 3.8552        | 0.88  | 36000 | 3.8449          |
+| 3.8591        | 0.93  | 38000 | 3.8483          |
+| 3.8452        | 0.98  | 40000 | 3.8396          |
+
+
+### 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 sha256:33f98e759c9066855e2181121fc9aca54c9ec4f4bff7cb84d2a454aa08b608a2
+oid sha256:6dc02a25486dbef4062dacf02ce49d5f447547e728955a207c2b5ff9721b63d3
 size 511344824

modeling_bit_llama.py

@@ -0,0 +1,79 @@
+from typing import Optional
+from transformers.models.llama.modeling_llama import (
+    LlamaConfig, 
+    LlamaModel,
+    LlamaForCausalLM,
+    LlamaAttention,
+    LlamaFlashAttention2,
+    LlamaSdpaAttention,
+    LlamaMLP,
+    LlamaDecoderLayer,
+)
+from mybitnet.bitnet import BitLinear
+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=True)
+        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=False)
+        
+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)
+
+class BitLlamaModel(LlamaModel):
+    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)
+