modify: update modeling code

configuration_baichuan.py

@@ -1,3 +1,4 @@
+# Copyright (c) 2023, Baichuan Intelligent Technology. All rights reserved.
 
 from transformers.configuration_utils import PretrainedConfig
 
@@ -21,6 +22,7 @@ class BaichuanConfig(PretrainedConfig):
         bos_token_id=1,
         eos_token_id=2,
         tie_word_embeddings=False,
+        gradient_checkpointing=False,
         **kwargs,
     ):
         self.vocab_size = vocab_size
@@ -33,6 +35,7 @@ class BaichuanConfig(PretrainedConfig):
         self.initializer_range = initializer_range
         self.rms_norm_eps = rms_norm_eps
         self.use_cache = use_cache
+        self.gradient_checkpointing = gradient_checkpointing,
         super().__init__(
             pad_token_id=pad_token_id,
             bos_token_id=bos_token_id,

modeling_baichuan.py

@@ -1,3 +1,5 @@
+# Copyright (c) 2023, Baichuan Intelligent Technology. All rights reserved.
+
 import math
 from typing import List, Optional, Tuple, Union
 
@@ -7,30 +9,31 @@ from transformers import PreTrainedModel
 from transformers.activations import ACT2FN
 from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
 from transformers.utils import logging
+from transformers.generation.utils import GenerationConfig
 
 from .configuration_baichuan import BaichuanConfig
 
 logger = logging.get_logger(__name__)
 
-def _get_slopes(n):
-    def _get_slopes_power_of_2(n):
+def _get_interleave(n):
+    def _get_interleave_power_of_2(n):
         start = (2 ** (-2 ** -(math.log2(n) - 3)))
         ratio = start
         return [start * ratio ** i for i in range(n)]
 
     if math.log2(n).is_integer():
-        return _get_slopes_power_of_2(n)
+        return _get_interleave_power_of_2(n)
     else:
         closest_power_of_2 = 2 ** math.floor(math.log2(n))
-        return _get_slopes_power_of_2(closest_power_of_2) + \
-               _get_slopes(2 * closest_power_of_2)[0::2][:n - closest_power_of_2]
+        return _get_interleave_power_of_2(closest_power_of_2) + \
+               _get_interleave(2 * closest_power_of_2)[0::2][:n - closest_power_of_2]
 
 def _fill_with_neg_inf(t):
     """FP16-compatible function that fills a tensor with -inf."""
     return t.float().fill_(float("-inf")).type_as(t)
 
 def _gen_alibi_mask(n_head, max_pos):
-    slopes = torch.Tensor(_get_slopes(n_head))
+    slopes = torch.Tensor(_get_interleave(n_head))
     alibi = slopes.unsqueeze(1).unsqueeze(1) * torch.arange(max_pos).unsqueeze(0).unsqueeze(0).expand(
         n_head, -1, -1)
     alibi = alibi.view(n_head, 1, max_pos)
@@ -87,8 +90,7 @@ class BaichuanAttention(torch.nn.Module):
 
         if (self.head_dim * self.num_heads) != self.hidden_size:
             raise ValueError(
-                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
-                f" and `num_heads`: {self.num_heads})."
+                f"hidden_size {self.hidden_size} is not divisible by num_heads {self.num_heads}"
             )
         self.W_pack = torch.nn.Linear(self.hidden_size, 3 * self.hidden_size, bias=False)
         self.o_proj = torch.nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
@@ -100,7 +102,6 @@ class BaichuanAttention(torch.nn.Module):
             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: bool = False,
             use_cache: bool = False,
@@ -127,12 +128,6 @@ class BaichuanAttention(torch.nn.Module):
 
         attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
 
-        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
-            raise ValueError(
-                f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
-                f" {attn_weights.size()}"
-            )
-
         if attention_mask is not None:
             if attn_weights.size(-2) == 1:
                 attention_mask = attention_mask[:, -1:, :]
@@ -142,12 +137,6 @@ class BaichuanAttention(torch.nn.Module):
         attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1)
         attn_output = torch.matmul(attn_weights, value_states)
 
-        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
-            raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
-                f" {attn_output.size()}"
-            )
-
         attn_output = attn_output.transpose(1, 2)
         attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
         attn_output = self.o_proj(attn_output)
@@ -175,7 +164,6 @@ class BaichuanLayer(torch.nn.Module):
             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,
@@ -189,7 +177,6 @@ class BaichuanLayer(torch.nn.Module):
         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,
@@ -244,7 +231,7 @@ class BaichuanModel(BaichuanPreTrainedModel):
         self.layers = torch.nn.ModuleList([BaichuanLayer(config) for _ in range(config.num_hidden_layers)])
         self.norm = RMSNorm(config.hidden_size, epsilon=config.rms_norm_eps)
 
-        self.gradient_checkpointing = False
+        self.gradient_checkpointing = config.gradient_checkpointing
         self.post_init()
         self.max_cache_pos = config.model_max_length
         self.first_run = True    
@@ -253,7 +240,7 @@ class BaichuanModel(BaichuanPreTrainedModel):
         if self.first_run:
             self.first_run = False
             self.register_buffer("future_mask", _gen_alibi_mask(self.n_head, self.max_cache_pos).to(tensor), persistent=False)
-        if (seq_length_with_past > self.max_cache_pos):
+        if seq_length_with_past > self.max_cache_pos:
             self.max_cache_pos = seq_length_with_past
             self.register_buffer("future_mask", _gen_alibi_mask(self.n_head, self.max_cache_pos).to(tensor), persistent=False)
         mask = self.future_mask[:self.n_head, :seq_length_with_past, :seq_length_with_past] 
@@ -262,8 +249,6 @@ class BaichuanModel(BaichuanPreTrainedModel):
     def forward(
             self,
             input_ids: torch.LongTensor = None,
-            attention_mask: Optional[torch.Tensor] = None,
-            position_ids: Optional[torch.LongTensor] = None,
             past_key_values: Optional[List[torch.FloatTensor]] = None,
             inputs_embeds: Optional[torch.FloatTensor] = None,
             use_cache: Optional[bool] = False,
@@ -273,34 +258,24 @@ class BaichuanModel(BaichuanPreTrainedModel):
     ) -> Union[Tuple, BaseModelOutputWithPast]:
 
 
-        # retrieve input_ids and inputs_embeds
         if input_ids is not None and inputs_embeds is not None:
-            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+            raise ValueError("You cannot provide both input_ids and inputs_embeds simultaneously")
         elif input_ids is not None:
             batch_size, seq_length = input_ids.shape
         elif inputs_embeds is not None:
             batch_size, seq_length, _ = inputs_embeds.shape
         else:
-            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+            raise ValueError("You need to provide input_ids or inputs_embeds")
 
         seq_length_with_past = seq_length
-        past_key_values_length = 0 
 
         if past_key_values is not None:
             past_key_values_length = past_key_values[0][0].shape[2]
             seq_length_with_past = seq_length_with_past + past_key_values_length
 
-        if position_ids is None:
-            device = input_ids.device if input_ids is not None else inputs_embeds.device
-            position_ids = torch.arange(
-                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
-            )   
-            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
-        else:
-            position_ids = position_ids.view(-1, seq_length).long()
-
         if inputs_embeds is None:
             inputs_embeds = self.embed_tokens(input_ids)
+
         # embed positions
         attention_mask = self.get_alibi_mask(inputs_embeds, seq_length_with_past)
 
@@ -337,14 +312,12 @@ class BaichuanModel(BaichuanPreTrainedModel):
                     create_custom_forward(decoder_layer),
                     hidden_states,
                     attention_mask,
-                    position_ids,
                     None,
                 )
             else:
                 layer_outputs = decoder_layer(
                     hidden_states,
                     attention_mask=attention_mask,
-                    position_ids=position_ids,
                     past_key_value=past_key_value,
                     output_attentions=output_attentions,
                     use_cache=use_cache,
@@ -387,8 +360,6 @@ class BaichuanForCausalLM(BaichuanPreTrainedModel):
     def forward(
             self,
             input_ids: torch.LongTensor = None,
-            attention_mask: Optional[torch.Tensor] = None,
-            position_ids: Optional[torch.LongTensor] = None,
             past_key_values: Optional[List[torch.FloatTensor]] = None,
             inputs_embeds: Optional[torch.FloatTensor] = None,
             labels: Optional[torch.LongTensor] = None,
@@ -396,14 +367,13 @@ class BaichuanForCausalLM(BaichuanPreTrainedModel):
             output_attentions: Optional[bool] = False,
             output_hidden_states: Optional[bool] = False,
             return_dict: Optional[bool] = True,
+            **kwargs
     ) -> Union[Tuple, CausalLMOutputWithPast]:
 
 
         # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
         outputs = self.model(
             input_ids=input_ids,
-            attention_mask=attention_mask,
-            position_ids=position_ids,
             past_key_values=past_key_values,
             inputs_embeds=inputs_embeds,
             use_cache=use_cache,
@@ -446,14 +416,6 @@ class BaichuanForCausalLM(BaichuanPreTrainedModel):
         if past_key_values:
             input_ids = input_ids[:, -1:]
 
-        position_ids = kwargs.get("position_ids", None)
-        if attention_mask is not None and position_ids is None:
-            # create position_ids on the fly for batch generation
-            position_ids = attention_mask.long().cumsum(-1) - 1 
-            position_ids.masked_fill_(attention_mask == 0, 1)
-            if past_key_values:
-                position_ids = position_ids[:, -1].unsqueeze(-1)
-
         # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
         if inputs_embeds is not None and past_key_values is None:
             model_inputs = {"inputs_embeds": inputs_embeds}
@@ -462,10 +424,8 @@ class BaichuanForCausalLM(BaichuanPreTrainedModel):
 
         model_inputs.update(
             {   
-                "position_ids": position_ids,
                 "past_key_values": past_key_values,
                 "use_cache": kwargs.get("use_cache"),
-                "attention_mask": attention_mask,
             }   
         )   
         return model_inputs
@@ -477,12 +437,13 @@ class BaichuanForCausalLM(BaichuanPreTrainedModel):
             for layer_past in past_key_values
         )
 
+
     def quantize(self, bits: int):
         try:
             from .quantizer import QLinear
         except ImportError:
             raise ImportError(
-                f"Error: Needs QLinear to run quantize."
+                f"Needs QLinear to run quantize."
             )
         
         for layer in self.model.layers:
@@ -512,3 +473,57 @@ class BaichuanForCausalLM(BaichuanPreTrainedModel):
                 bias = None,
             )
         return self 
+
+    def _build_chat_input(self, tokenizer, messages: List[dict], max_new_tokens: int=0):
+        max_new_tokens = max_new_tokens or self.generation_config.max_new_tokens
+        max_input_tokens = self.config.model_max_length - max_new_tokens
+        max_input_tokens = max(self.config.model_max_length // 2, max_input_tokens)
+        total_input, round_input = [], []
+        for i, message in enumerate(messages[::-1]):
+            content_tokens = tokenizer.encode(message['content'])
+            if message['role'] == 'user':
+                round_input = [self.generation_config.user_token_id] + content_tokens + round_input
+                if total_input and len(total_input) + len(round_input) > max_input_tokens:
+                    break
+                else:
+                    total_input = round_input + total_input
+                    if len(total_input) >= max_input_tokens:
+                        break
+                    else:
+                        round_input = []
+            elif message['role'] == 'assistant':
+                round_input = [
+                    self.generation_config.assistant_token_id
+                ] + content_tokens + [
+                    self.generation_config.eos_token_id
+                ] + round_input
+            else:
+                raise ValueError(f"message role not supported yet: {message['role']}")
+        total_input = total_input[-max_input_tokens:]  # truncate left
+        total_input.append(self.generation_config.assistant_token_id)
+        total_input = torch.LongTensor([total_input]).to(self.device)
+        return total_input
+
+    @torch.no_grad()
+    def chat(self, tokenizer, messages: List[dict], stream=False,
+             generation_config: Optional[GenerationConfig]=None):
+        generation_config = generation_config or self.generation_config
+        input_ids = self._build_chat_input(tokenizer, messages, generation_config.max_new_tokens)
+        if stream:
+            from transformers_stream_generator.main import NewGenerationMixin, StreamGenerationConfig
+            self.__class__.generate = NewGenerationMixin.generate
+            self.__class__.sample_stream = NewGenerationMixin.sample_stream
+            stream_config = StreamGenerationConfig(**generation_config.to_dict(), do_stream=True)
+
+            def stream_generator():
+                outputs = []
+                for token in self.generate(input_ids, generation_config=stream_config):
+                    outputs.append(token.item())
+                    yield tokenizer.decode(outputs, skip_special_tokens=True)
+
+            return stream_generator()
+        else:
+            self.__class__.generate = PreTrainedModel.generate  # disable stream
+            outputs = self.generate(input_ids, generation_config=generation_config)
+            response = tokenizer.decode(outputs[0][len(input_ids[0]):], skip_special_tokens=True)
+            return response

quantizer.py

@@ -1,24 +1,123 @@
+# Copyright (c) 2023, Baichuan Intelligent Technology. All rights reserved.
+
 import torch
+from typing import List
+import bz2
+import base64
+import ctypes
+from transformers.utils import logging
+logger = logging.get_logger(__name__)
+
+try:
+    from cpm_kernels.kernels.base import LazyKernelCModule, KernelFunction, round_up
+
+    class Kernel:
+        def __init__(self, code: bytes, function_names: List[str]):
+            self.code = code
+            self._function_names = function_names
+            self._cmodule = LazyKernelCModule(self.code)
+
+            for name in self._function_names:
+                setattr(self, name, KernelFunction(self._cmodule, name))
+    quantization_code = "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"
+    kernels = Kernel(
+        bz2.decompress(base64.b64decode(quantization_code)),
+        [
+            "int4_to_fp16",
+            "fp16_to_int4",
+            "int8_to_fp16",
+            "fp16_to_int8",
+            "int4_to_bf16",
+            "bf16_to_int4",
+            "int8_to_bf16",
+            "bf16_to_int8",
+        ],
+    )
+except Exception as exception:
+    kernels = None
+    logger.warning("Failed to load kernels:" + str(exception))
+
+def quant4(weight: torch.Tensor, scale: torch.Tensor):
+    stream = torch.cuda.current_stream()
+    num_row = weight.size(0)
+    num_chan_fp16 = weight.size(1)
+    # 4bit
+    num_chan_int = num_chan_fp16 // 8
+    qweight = torch.zeros((num_row, num_chan_int), dtype=torch.int32, device=weight.device)
+    intweight = torch.empty(num_row, num_chan_fp16, dtype = torch.int32)
+    intweight = torch.clip(torch.round(weight.to(scale.dtype) / scale[:, None]),-16, 15).to(dtype=torch.int32)
+
+    for j in range(num_chan_int):
+        qweight[:, j] = ((intweight[:, j*8+7] & 0x0f) << 28) \
+            | ((intweight[:, j*8+6] & 0x0f) << 24) \
+            | ((intweight[:, j*8+5] & 0x0f) << 20) \
+            | ((intweight[:, j*8+4] & 0x0f) << 16) \
+            | ((intweight[:, j*8+3] & 0x0f) << 12) \
+            | ((intweight[:, j*8+2] & 0x0f) << 8) \
+            | ((intweight[:, j*8+1] & 0x0f) << 4) \
+            | ((intweight[:, j*8] & 0x0f))
+    return qweight
+
+def dequant4(qweight: torch.Tensor, scale: torch.Tensor, input: torch.Tensor):
+    stream = torch.cuda.current_stream()
+    num_row = qweight.size(0)
+    num_chan_int = qweight.size(1)
+    # 4bit
+    num_chan_fp16 = num_chan_int * 8
+    
+    out = torch.empty((num_row, num_chan_fp16), dtype=input.dtype, device=qweight.device)
+
+    blockDim = (128, 1, 1)
+    gridDim = ((num_chan_int + blockDim[0] - 1) // blockDim[0], num_row, 1)
+    if input.dtype == torch.bfloat16:
+    	kernels.int4_to_bf16(
+            gridDim,
+       	    blockDim,
+            0,
+            stream,
+            [ctypes.c_void_p(out.data_ptr()), ctypes.c_void_p(qweight.data_ptr()), 
+            ctypes.c_void_p(scale.data_ptr()), ctypes.c_int32(num_row), ctypes.c_int32(num_chan_int), ctypes.c_int32(num_chan_fp16)],
+        )
+    elif input.dtype == torch.float16:
+        kernels.int4_to_fp16(
+            gridDim,
+            blockDim,
+            0,
+            stream,
+            [ctypes.c_void_p(out.data_ptr()), ctypes.c_void_p(qweight.data_ptr()),
+            ctypes.c_void_p(scale.data_ptr()), ctypes.c_int32(num_row), ctypes.c_int32(num_chan_int), ctypes.c_int32(num_chan_fp16)],
+        )
+    return out
 
 class QLinear(torch.nn.Module):
     def __init__(self, bits: int, weight: torch.Tensor, bias=None):
         super().__init__()
         self.quant_bits = bits
-        if self.quant_bits != 8:
-            raise ValueError(
-                f'Only supprt int8 quant in current version'
-            )
         self.scale = weight.abs().max(dim=-1).values / ((2 ** (bits - 1)) - 1)
-        self.weight = torch.round(weight / self.scale[:, None]).to(torch.int8)
-        self.weight = self.weight.T
+        self.scale = self.scale.to(torch.float32)
+        if self.quant_bits == 4:
+            self.weight = quant4(weight, self.scale)
+        elif self.quant_bits == 8:
+            self.weight = torch.round(weight.to(self.scale.dtype) / self.scale[:, None]).to(torch.int8)
+        if self.quant_bits == 8:
+            self.weight = self.weight.T
         self.bias = None
 
     def forward(self, input):
+        if self.quant_bits == 4:
+            assert(input.dtype == torch.bfloat16 or input.dtype == torch.float16)            
+
         if self.weight.device != input.device:
             self.weight = self.weight.to(input.device)
             self.scale = self.scale.to(input.device)
         
-        output = torch.matmul(input, self.weight.to(input.dtype)) * self.scale.to(input.dtype)[None,None, :]
+        if self.quant_bits == 4:
+            self.scale = self.scale.to(input.dtype)
+            rweight = dequant4(self.weight, self.scale, input).T
+            output = torch.matmul(input, rweight)
+        elif self.quant_bits == 8:
+            rweight = self.weight.to(input.dtype) * self.scale.to(input.dtype)
+            output = torch.matmul(input, rweight)
         if self.bias is not None:
             output = output + self.bias
         return output

requirements.txt

@@ -0,0 +1,6 @@
+accelerate
+colorama
+cpm_kernels
+sentencepiece
+streamlit
+transformers_stream_generator

tokenization_baichuan.py

@@ -1,9 +1,10 @@
+# Copyright (c) 2023, Baichuan Intelligent Technology. All rights reserved.
+
 import os
 from shutil import copyfile
 from typing import Any, Dict, List, Optional, Tuple
 
 import sentencepiece as spm
-
 from transformers.tokenization_utils import AddedToken, PreTrainedTokenizer
 from transformers.utils import logging