End of training

README.md

@@ -0,0 +1,56 @@
+---
+tags:
+- generated_from_trainer
+model-index:
+- name: bitllama-Llama2
+  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. -->
+
+# bitllama-Llama2
+
+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: 3.7504
+
+## 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=(0.9,0.95) and epsilon=1e-08
+- lr_scheduler_type: linear
+- lr_scheduler_warmup_steps: 5000
+- num_epochs: 1
+
+### Training results
+
+| Training Loss | Epoch | Step | Validation Loss |
+|:-------------:|:-----:|:----:|:---------------:|
+| 4.6791        | 0.74  | 2000 | 3.7504          |
+
+
+### Framework versions
+
+- Transformers 4.38.2
+- Pytorch 2.8.0+cu126
+- Datasets 4.0.0
+- Tokenizers 0.15.2

bitllama_layers.py

@@ -0,0 +1,163 @@
+import warnings 
+from typing import Optional,Tuple
+from transformers.models.llama.modeling_llama import (
+    LlamaConfig,
+    LlamaModel,
+    LlamaForCausalLM,
+    LlamaAttention,
+    LlamaFlashAttention2,
+    LlamaSdpaAttention,
+    LlamaMLP,
+    LlamaDecoderLayer
+)
+from 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:BitLlamaConfig):
+        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)
+        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)
+            self.o_proj=BitLinear158b(self.hidden_size,self.hidden_size,bias=False,rms_norm_eps=config.rms_norm_eps)
+        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 BitLlamaSpdaAttention(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_config=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_esp=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":BitLlamaSpdaAttention,
+}     
+
+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]]]:
+        
+        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
+
+        hidden_states,self_attn_weight,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
+
+        residual=hidden_states
+        hidden_states=self.mlp(hidden_states)
+        hidden_states=residual+hidden_states
+
+        outputs=(hidden_states,)
+
+        if output_attentions:
+            outputs+=(self_attn_weight,)
+        
+        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)

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:cc186dae20a199ce00a1d07bc4da59dba3743cfb308d034721a3a2b606153e68
+oid sha256:c269b518dbf4a5c57251f3ef15b172074e8a98174de85da677c454cf3b4fd152
 size 510960712