Training in progress, step 100

README.md

@@ -13,7 +13,12 @@ should probably proofread and complete it, then remove this comment. -->
 
 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
+- eval_loss: 14.1081
+- eval_runtime: 0.0886
+- eval_samples_per_second: 237.106
+- eval_steps_per_second: 11.291
+- epoch: 50.0
+- step: 400
 
 ## Model description
 
@@ -32,44 +37,17 @@ More information needed
 ### Training hyperparameters
 
 The following hyperparameters were used during training:
-- learning_rate: 0.00024
-- train_batch_size: 96
-- eval_batch_size: 96
+- learning_rate: 0.001
+- train_batch_size: 32
+- eval_batch_size: 32
 - 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          |
-
+- num_epochs: 100
 
 ### Framework versions
 
-- Transformers 4.38.2
+- Transformers 4.39.1
 - Pytorch 2.2.1+cu121
 - Datasets 2.18.0
 - Tokenizers 0.15.2

generation_config.json

@@ -2,5 +2,5 @@
   "_from_model_config": true,
   "bos_token_id": 1,
   "eos_token_id": 2,
-  "transformers_version": "4.38.2"
+  "transformers_version": "4.39.1"
 }

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f2993fec007900dcd9d7c675dc75369a67f5269c908bfef209667ce5be0910d4
+oid sha256:55c1433d65751299063935edf3737aecb3b4f6176af451137f1c51308b56c0aa
 size 510960712

modeling_bit_llama.py

@@ -1,4 +1,5 @@
-from typing import Optional
+import warnings
+from typing import Optional, Tuple
 from transformers.models.llama.modeling_llama import (
     LlamaConfig, 
     LlamaModel,
@@ -10,6 +11,7 @@ from transformers.models.llama.modeling_llama import (
     LlamaDecoderLayer,
 )
 from mybitnet.bitnet import BitLinear
+import torch
 from torch import nn
 
 class BitLlamaConfig(LlamaConfig):
@@ -22,9 +24,9 @@ class BitLlamaConfig(LlamaConfig):
 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.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=False)
+        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):
@@ -61,8 +63,61 @@ class BitLlamaDecoderLayer(LlamaDecoderLayer):
         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(

training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:374bbfaab702b8ea66e7026aa109e4f27c10eb644bac7ca5847e09d38b90158b
+oid sha256:f68b6d8610db40c9080edd428332a28b58687a273c0815ea5842db873611ca2a
 size 4920