glaiveai/glaive-function-calling-v2-small

1"""A simple, flexible implementation of a GPT model.2 3Inspired by https://github.com/karpathy/minGPT/blob/master/mingpt/model.py4"""5import math6import warnings7from typing import List, Optional, Tuple, Union8import torch9import torch.nn as nn10import torch.nn.functional as F11from transformers import PreTrainedModel, PreTrainedTokenizer, PreTrainedTokenizerFast12from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast13from .attention import attn_bias_shape, build_attn_bias14from .blocks import MPTBlock15from .norm import NORM_CLASS_REGISTRY16from .configuration_mpt import MPTConfig17from .adapt_tokenizer import AutoTokenizerForMOD, adapt_tokenizer_for_denoising18from .hf_prefixlm_converter import add_bidirectional_mask_if_missing, convert_hf_causal_lm_to_prefix_lm19from .meta_init_context import init_empty_weights20from .param_init_fns import MODEL_INIT_REGISTRY, generic_param_init_fn_21Tokenizer = Union[PreTrainedTokenizer, PreTrainedTokenizerFast]22 23class MPTPreTrainedModel(PreTrainedModel):24    config_class = MPTConfig25    base_model_prefix = 'model'26    _no_split_modules=["MPTBlock"]27 28class MPTModel(MPTPreTrainedModel):29 30    def __init__(self, config: MPTConfig):31        config._validate_config()32        super().__init__(config)33        self.attn_impl = config.attn_config['attn_impl']34        self.prefix_lm = config.attn_config['prefix_lm']35        self.attn_uses_sequence_id = config.attn_config['attn_uses_sequence_id']36        self.alibi = config.attn_config['alibi']37        self.alibi_bias_max = config.attn_config['alibi_bias_max']38        if config.norm_type.lower() not in NORM_CLASS_REGISTRY.keys():39            norm_options = ' | '.join(NORM_CLASS_REGISTRY.keys())40            raise NotImplementedError(f'Requested norm type ({config.norm_type}) is not implemented within this repo (Options: {norm_options}).')41        norm_class = NORM_CLASS_REGISTRY[config.norm_type.lower()]42        self.embedding_fraction = config.embedding_fraction43        self.wte = nn.Embedding(config.vocab_size, config.d_model, device=config.init_device)44        if not self.alibi:45            self.wpe = nn.Embedding(config.max_seq_len, config.d_model, device=config.init_device)46        self.emb_drop = nn.Dropout(config.emb_pdrop)47        self.blocks = nn.ModuleList([MPTBlock(device=config.init_device, **config.to_dict()) for _ in range(config.n_layers)])48        self.norm_f = norm_class(config.d_model, device=config.init_device)49        if config.init_device != 'meta':50            print(f'You are using config.init_device={config.init_device!r}, but you can also use config.init_device="meta" with Composer + FSDP for fast initialization.')51            self.apply(self.param_init_fn)52        self.is_causal = not self.prefix_lm53        self._attn_bias_initialized = False54        self.attn_bias = None55        self.attn_bias_shape = attn_bias_shape(self.attn_impl, config.n_heads, config.max_seq_len, self.alibi, prefix_lm=self.prefix_lm, causal=self.is_causal, use_sequence_id=self.attn_uses_sequence_id)56        if config.no_bias:57            for module in self.modules():58                if hasattr(module, 'bias') and isinstance(module.bias, nn.Parameter):59                    if config.verbose:60                        warnings.warn(f'Removing bias ({module.bias}) from {module}.')61                    module.register_parameter('bias', None)62        if config.verbose and config.verbose > 2:63            print(self)64        if 'verbose' not in self.config.init_config:65            self.config.init_config['verbose'] = self.config.verbose66        if self.config.init_config['verbose'] > 1:67            init_fn_name = self.config.init_config['name']68            warnings.warn(f'Using {init_fn_name} initialization.')69 70    def get_input_embeddings(self):71        return self.wte72 73    def set_input_embeddings(self, value):74        self.wte = value75 76    @torch.no_grad()77    def _attn_bias(self, device, dtype, attention_mask: Optional[torch.ByteTensor]=None, prefix_mask: Optional[torch.ByteTensor]=None, sequence_id: Optional[torch.LongTensor]=None):78        if not self._attn_bias_initialized:79            if self.attn_bias_shape:80                self.attn_bias = torch.zeros(self.attn_bias_shape, device=device, dtype=dtype)81                self.attn_bias = build_attn_bias(self.attn_impl, self.attn_bias, self.config.n_heads, self.config.max_seq_len, causal=self.is_causal, alibi=self.alibi, alibi_bias_max=self.alibi_bias_max)82            self._attn_bias_initialized = True83        if self.attn_impl == 'flash':84            return (self.attn_bias, attention_mask)85        if self.attn_bias is not None:86            self.attn_bias = self.attn_bias.to(dtype=dtype, device=device)87        attn_bias = self.attn_bias88        if self.prefix_lm:89            assert isinstance(attn_bias, torch.Tensor)90            assert isinstance(prefix_mask, torch.Tensor)91            attn_bias = self._apply_prefix_mask(attn_bias, prefix_mask)92        if self.attn_uses_sequence_id and sequence_id is not None:93            assert isinstance(attn_bias, torch.Tensor)94            attn_bias = self._apply_sequence_id(attn_bias, sequence_id)95        if attention_mask is not None:96            s_k = attention_mask.shape[-1]97            if attn_bias is None:98                attn_bias = torch.zeros((1, 1, 1, s_k), device=device, dtype=dtype)99            else:100                attn_bias = attn_bias[:, :, :, -s_k:]101            if prefix_mask is not None and attention_mask.shape != prefix_mask.shape:102                raise ValueError(f'attention_mask shape={attention_mask.shape} ' + f'and prefix_mask shape={prefix_mask.shape} are not equal.')103            min_val = torch.finfo(attn_bias.dtype).min104            attn_bias = attn_bias.masked_fill(~attention_mask.view(-1, 1, 1, s_k), min_val)105        return (attn_bias, None)106 107    def _apply_prefix_mask(self, attn_bias: torch.Tensor, prefix_mask: torch.Tensor):108        (s_k, s_q) = attn_bias.shape[-2:]109        if s_k != self.config.max_seq_len or s_q != self.config.max_seq_len:110            raise ValueError('attn_bias does not match the expected shape. ' + f'The last two dimensions should both be {self.config.max_length} ' + f'but are {s_k} and {s_q}.')111        seq_len = prefix_mask.shape[-1]112        if seq_len > self.config.max_seq_len:113            raise ValueError(f'prefix_mask sequence length cannot exceed max_seq_len={self.config.max_seq_len}')114        attn_bias = attn_bias[..., :seq_len, :seq_len]115        causal = torch.tril(torch.ones((seq_len, seq_len), dtype=torch.bool, device=prefix_mask.device)).view(1, 1, seq_len, seq_len)116        prefix = prefix_mask.view(-1, 1, 1, seq_len)117        cannot_attend = ~torch.logical_or(causal, prefix.bool())118        min_val = torch.finfo(attn_bias.dtype).min119        attn_bias = attn_bias.masked_fill(cannot_attend, min_val)120        return attn_bias121 122    def _apply_sequence_id(self, attn_bias: torch.Tensor, sequence_id: torch.LongTensor):123        seq_len = sequence_id.shape[-1]124        if seq_len > self.config.max_seq_len:125            raise ValueError(f'sequence_id sequence length cannot exceed max_seq_len={self.config.max_seq_len}')126        attn_bias = attn_bias[..., :seq_len, :seq_len]127        cannot_attend = torch.logical_not(torch.eq(sequence_id.view(-1, seq_len, 1), sequence_id.view(-1, 1, seq_len))).unsqueeze(1)128        min_val = torch.finfo(attn_bias.dtype).min129        attn_bias = attn_bias.masked_fill(cannot_attend, min_val)130        return attn_bias131 132    def forward(self, input_ids: torch.LongTensor, past_key_values: Optional[List[Tuple[torch.FloatTensor]]]=None, attention_mask: Optional[torch.ByteTensor]=None, prefix_mask: Optional[torch.ByteTensor]=None, sequence_id: Optional[torch.LongTensor]=None, return_dict: Optional[bool]=None, output_attentions: Optional[bool]=None, output_hidden_states: Optional[bool]=None, use_cache: Optional[bool]=None):133        return_dict = return_dict if return_dict is not None else self.config.return_dict134        use_cache = use_cache if use_cache is not None else self.config.use_cache135        if attention_mask is not None:136            attention_mask = attention_mask.bool()137        if prefix_mask is not None:138            prefix_mask = prefix_mask.bool()139        if not return_dict:140            raise NotImplementedError('return_dict False is not implemented yet for MPT')141        if output_attentions:142            raise NotImplementedError('output_attentions is not implemented yet for MPT')143        if attention_mask is not None and attention_mask[:, 0].sum() != attention_mask.shape[0] and self.training:144            raise NotImplementedError('MPT does not support training with left padding.')145        if self.prefix_lm and prefix_mask is None:146            raise ValueError('prefix_mask is a required argument when MPT is configured with prefix_lm=True.')147        if self.training:148            if self.attn_uses_sequence_id and sequence_id is None:149                raise ValueError('sequence_id is a required argument when MPT is configured with attn_uses_sequence_id=True ' + 'and the model is in train mode.')150            elif self.attn_uses_sequence_id is False and sequence_id is not None:151                warnings.warn('MPT received non-None input for `sequence_id` but is configured with attn_uses_sequence_id=False. ' + 'This input will be ignored. If you want the model to use `sequence_id`, set attn_uses_sequence_id to True.')152        S = input_ids.size(1)153        assert S <= self.config.max_seq_len, f'Cannot forward input with seq_len={S}, this model only supports seq_len<={self.config.max_seq_len}'154        tok_emb = self.wte(input_ids)155        if self.alibi:156            x = tok_emb157        else:158            past_position = 0159            if past_key_values is not None:160                if len(past_key_values) != self.config.n_layers:161                    raise ValueError(f'past_key_values must provide a past_key_value for each attention ' + f'layer in the network (len(past_key_values)={len(past_key_values)!r}; self.config.n_layers={self.config.n_layers!r}).')162                past_position = past_key_values[0][0].size(1)163            if S + past_position > self.config.max_seq_len:164                raise ValueError(f'Cannot forward input with past sequence length {past_position} and current sequence length {S + 1}, this model only supports total sequence length <= {self.config.max_seq_len}.')165            pos = torch.arange(past_position, S + past_position, dtype=torch.long, device=input_ids.device).unsqueeze(0)166            if attention_mask is not None:167                pos = torch.clamp(pos - torch.cumsum((~attention_mask).to(torch.int32), dim=1)[:, past_position:], min=0)168            pos_emb = self.wpe(pos)169            x = tok_emb + pos_emb170        if self.embedding_fraction == 1:171            x = self.emb_drop(x)172        else:173            x_shrunk = x * self.embedding_fraction + x.detach() * (1 - self.embedding_fraction)174            assert isinstance(self.emb_drop, nn.Module)175            x = self.emb_drop(x_shrunk)176        (attn_bias, attention_mask) = self._attn_bias(device=x.device, dtype=x.dtype, attention_mask=attention_mask, prefix_mask=prefix_mask, sequence_id=sequence_id)177        if use_cache and past_key_values is None:178            past_key_values = [() for _ in range(self.config.n_layers)]179        all_hidden_states = () if output_hidden_states else None180        for (b_idx, block) in enumerate(self.blocks):181            if output_hidden_states:182                assert all_hidden_states is not None183                all_hidden_states = all_hidden_states + (x,)184            past_key_value = past_key_values[b_idx] if past_key_values is not None else None185            (x, past_key_value) = block(x, past_key_value=past_key_value, attn_bias=attn_bias, attention_mask=attention_mask, is_causal=self.is_causal)186            if past_key_values is not None:187                past_key_values[b_idx] = past_key_value188        x = self.norm_f(x)189        return BaseModelOutputWithPast(last_hidden_state=x, past_key_values=past_key_values, hidden_states=all_hidden_states)190 191    def param_init_fn(self, module):192        init_fn_name = self.config.init_config['name']193        MODEL_INIT_REGISTRY[init_fn_name](module=module, n_layers=self.config.n_layers, d_model=self.config.d_model, **self.config.init_config)194 195    def fsdp_wrap_fn(self, module):196        return isinstance(module, MPTBlock)197 198    def activation_checkpointing_fn(self, module):199        return isinstance(module, MPTBlock)200 201class MPTForCausalLM(MPTPreTrainedModel):202 203    def __init__(self, config: MPTConfig):204        super().__init__(config)205        if not config.tie_word_embeddings:206            raise ValueError('MPTForCausalLM only supports tied word embeddings')207        self.transformer = MPTModel(config)208        self.logit_scale = None209        if config.logit_scale is not None:210            logit_scale = config.logit_scale211            if isinstance(logit_scale, str):212                if logit_scale == 'inv_sqrt_d_model':213                    logit_scale = 1 / math.sqrt(config.d_model)214                else:215                    raise ValueError(f"logit_scale={logit_scale!r} is not recognized as an option; use numeric value or 'inv_sqrt_d_model'.")216            self.logit_scale = logit_scale217 218    def get_input_embeddings(self):219        return self.transformer.wte220 221    def set_input_embeddings(self, value):222        self.transformer.wte = value223 224    def get_output_embeddings(self):225        return self.transformer.wte226 227    def set_output_embeddings(self, new_embeddings):228        self.transformer.wte = new_embeddings229 230    def set_decoder(self, decoder):231        self.transformer = decoder232 233    def get_decoder(self):234        return self.transformer235 236    def forward(self, input_ids: torch.LongTensor, past_key_values: Optional[List[Tuple[torch.FloatTensor]]]=None, attention_mask: Optional[torch.ByteTensor]=None, prefix_mask: Optional[torch.ByteTensor]=None, sequence_id: Optional[torch.LongTensor]=None, labels: Optional[torch.LongTensor]=None, return_dict: Optional[bool]=None, output_attentions: Optional[bool]=None, output_hidden_states: Optional[bool]=None, use_cache: Optional[bool]=None):237        return_dict = return_dict if return_dict is not None else self.config.return_dict238        use_cache = use_cache if use_cache is not None else self.config.use_cache239        outputs = self.transformer(input_ids=input_ids, past_key_values=past_key_values, attention_mask=attention_mask, prefix_mask=prefix_mask, sequence_id=sequence_id, return_dict=return_dict, output_attentions=output_attentions, output_hidden_states=output_hidden_states, use_cache=use_cache)240        logits = F.linear(outputs.last_hidden_state, self.transformer.wte.weight)241        if self.logit_scale is not None:242            if self.logit_scale == 0:243                warnings.warn(f'Multiplying logits by self.logit_scale={self.logit_scale!r}. This will produce uniform (uninformative) outputs.')244            logits *= self.logit_scale245        loss = None246        if labels is not None:247            labels = torch.roll(labels, shifts=-1)248            labels[:, -1] = -100249            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), labels.to(logits.device).view(-1))250        return CausalLMOutputWithPast(loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states)251 252    def param_init_fn(self, module):253        init_fn_name = self.config.init_config['name']254        MODEL_INIT_REGISTRY[init_fn_name](module=module, n_layers=self.config.n_layers, d_model=self.config.d_model, **self.config.init_config)255 256    def fsdp_wrap_fn(self, module):257        return isinstance(module, MPTBlock)258 259    def activation_checkpointing_fn(self, module):260        return isinstance(module, MPTBlock)261 262    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):263        if inputs_embeds is not None:264            raise NotImplementedError('inputs_embeds is not implemented for MPT yet')265        attention_mask = kwargs['attention_mask'].bool()266        if attention_mask[:, -1].sum() != attention_mask.shape[0]:267            raise NotImplementedError('MPT does not support generation with right padding.')268        if self.transformer.attn_uses_sequence_id and self.training:269            sequence_id = torch.zeros_like(input_ids[:1])270        else:271            sequence_id = None272        if past_key_values is not None:273            input_ids = input_ids[:, -1].unsqueeze(-1)274        if self.transformer.prefix_lm:275            prefix_mask = torch.ones_like(attention_mask)276            if kwargs.get('use_cache') == False:277                raise NotImplementedError('MPT with prefix_lm=True does not support use_cache=False.')278        else:279            prefix_mask = None280        return {'input_ids': input_ids, 'attention_mask': attention_mask, 'prefix_mask': prefix_mask, 'sequence_id': sequence_id, 'past_key_values': past_key_values, 'use_cache': kwargs.get('use_cache', True)}281 282    @staticmethod283    def _reorder_cache(past_key_values, beam_idx):284        """Used by HuggingFace generate when using beam search with kv-caching.285 286        See https://github.com/huggingface/transformers/blob/3ec7a47664ebe40c40f4b722f6bb1cd30c3821ec/src/transformers/models/gpt2/modeling_gpt2.py#L1122-L1133287        for an example in transformers.288        """289        reordered_past = []290        for layer_past in past_key_values:291            reordered_past += [tuple((past_state.index_select(0, beam_idx) for past_state in layer_past))]292        return reordered_past