lance_ai_model.py

import math
import torch
import torch.nn as nn
from torch.nn import functional as F
from transformers import PreTrainedModel, PretrainedConfig, GenerationMixin
from transformers.modeling_outputs import CausalLMOutputWithPast
from transformers.models.auto.configuration_auto import CONFIG_MAPPING
from transformers.models.auto.modeling_auto import MODEL_FOR_CAUSAL_LM_MAPPING

class LanceAIConfig(PretrainedConfig):
    model_type = "lance_ai"

    def __init__(
        self,
        vocab_size=151936,
        hidden_size=896,
        intermediate_size=4864,
        num_hidden_layers=24,
        num_attention_heads=14,
        num_key_value_heads=2,
        hidden_act="silu",
        max_position_embeddings=32768,
        initializer_range=0.02,
        rms_norm_eps=1e-6,
        use_cache=True,
        tie_word_embeddings=True,
        rope_theta=1000000.0,
        attention_dropout=0.0,
        pad_token_id=None,
        bos_token_id=None,
        eos_token_id=None,
        head_dim=64,
        **kwargs
    ):
        super().__init__(**kwargs)
        self.vocab_size = vocab_size
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.num_key_value_heads = num_key_value_heads
        self.hidden_act = hidden_act
        self.max_position_embeddings = max_position_embeddings
        self.initializer_range = initializer_range
        self.rms_norm_eps = rms_norm_eps
        self.use_cache = use_cache
        self.tie_word_embeddings = tie_word_embeddings
        self.rope_theta = rope_theta
        self.attention_dropout = attention_dropout
        self.head_dim = head_dim
        self.pad_token_id = pad_token_id
        self.bos_token_id = bos_token_id
        self.eos_token_id = eos_token_id

class LanceAIRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        input_dtype = hidden_states.dtype
        hidden_states = hidden_states.to(torch.float32)
        variance = hidden_states.pow(2).mean(-1, keepdim=True)
        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
        return self.weight * hidden_states.to(input_dtype)

class LanceAIRotaryEmbedding(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.max_seq_len_cached = config.max_position_embeddings
        base = config.rope_theta
        dim = config.head_dim
        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.int64).float() / dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)

    @torch.no_grad()
    def forward(self, x, position_ids):
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
        position_ids_expanded = position_ids[:, None, :].float()
        freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
        emb = torch.cat((freqs, freqs), dim=-1)
        cos = emb.cos()
        sin = emb.sin()
        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)

def rotate_half(x):
    x1 = x[..., :x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2:]
    return torch.cat((-x2, x1), dim=-1)

def apply_rotary_pos_emb(q, k, cos, sin):
    cos = cos.unsqueeze(1)
    sin = sin.unsqueeze(1)
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed

def repeat_kv(hidden_states, n_rep):
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)

class LanceAIAttention(nn.Module):
    def __init__(self, config, layer_idx):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.head_dim = config.head_dim
        self.num_heads = config.num_attention_heads
        self.num_kv_heads = config.num_key_value_heads
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
        self.scaling = self.head_dim ** -0.5
        self.attention_dropout = config.attention_dropout
        self.is_causal = True

        self.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=True)
        self.k_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=True)
        self.v_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=True)
        self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False)

    def forward(self, hidden_states, attention_mask=None, position_embeddings=None, past_key_values=None, use_cache=False):
        batch_size, seq_len, _ = hidden_states.shape
        query_states = self.q_proj(hidden_states).view(batch_size, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = self.k_proj(hidden_states).view(batch_size, seq_len, self.num_kv_heads, self.head_dim).transpose(1, 2)
        value_states = self.v_proj(hidden_states).view(batch_size, seq_len, self.num_kv_heads, self.head_dim).transpose(1, 2)

        cos, sin = position_embeddings
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_values is not None:
            if past_key_values[0] is not None:
                key_states = torch.cat([past_key_values[0], key_states], dim=2)
            if past_key_values[1] is not None:
                value_states = torch.cat([past_key_values[1], value_states], dim=2)

        past = (key_states, value_states) if use_cache else None

        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)

        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) * self.scaling

        if attention_mask is not None:
            attn_weights = attn_weights + attention_mask

        attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
        attn_weights = F.dropout(attn_weights, p=self.attention_dropout, training=self.training)
        attn_output = torch.matmul(attn_weights, value_states)
        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.reshape(batch_size, seq_len, -1)
        attn_output = self.o_proj(attn_output)

        return attn_output, past

class LanceAIMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
        self.up_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)

    def forward(self, x):
        return self.down_proj(F.silu(self.gate_proj(x)) * self.up_proj(x))

class LanceAIDecoderLayer(nn.Module):
    def __init__(self, config, layer_idx):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = LanceAIAttention(config, layer_idx)
        self.mlp = LanceAIMLP(config)
        self.input_layernorm = LanceAIRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = LanceAIRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(self, hidden_states, attention_mask=None, position_embeddings=None, past_key_values=None, use_cache=False):
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        hidden_states, past_kv = self.self_attn(
            hidden_states,
            attention_mask=attention_mask,
            position_embeddings=position_embeddings,
            past_key_values=past_key_values,
            use_cache=use_cache,
        )
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        return hidden_states, past_kv

class LanceAIPreTrainedModel(PreTrainedModel):
    config_class = LanceAIConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["LanceAIDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_sdpa = True

class LanceAIModel(LanceAIPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
        self.layers = nn.ModuleList(
            [LanceAIDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = LanceAIRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = LanceAIRotaryEmbedding(config)

        self.gradient_checkpointing = False
        self.gradient_checkpointing_func = None
        self.post_init()

    def _set_gradient_checkpointing(self, enable=True, gradient_checkpointing_func=None):
        self.gradient_checkpointing = enable
        if gradient_checkpointing_func is not None:
            self.gradient_checkpointing_func = gradient_checkpointing_func

    def _past_seq_length(self, past_key_values):
        """Get sequence length from past cache (tuple or DynamicCache)"""
        if past_key_values is None:
            return 0
        if hasattr(past_key_values, 'get_seq_length'):
            return past_key_values.get_seq_length()
        if len(past_key_values) > 0 and past_key_values[0] is not None:
            if past_key_values[0][0] is not None:
                return past_key_values[0][0].shape[2]
        return 0

    def _convert_past(self, past_key_values, use_cache):
        """Convert DynamicCache to our tuple format if needed"""
        if past_key_values is None or not use_cache:
            return past_key_values, use_cache
        if isinstance(past_key_values, list):
            return past_key_values, use_cache
        # DynamicCache -> list of (k, v) tuples
        if hasattr(past_key_values, 'get_seq_length'):
            converted = []
            for i in range(len(self.layers)):
                if i < len(past_key_values):
                    kv = past_key_values[i]
                    converted.append((kv[0], kv[1]))
                else:
                    converted.append((None, None))
            return converted, use_cache
        return past_key_values, use_cache

    def forward(self, input_ids=None, attention_mask=None, position_ids=None, past_key_values=None, inputs_embeds=None, use_cache=None):
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("Specify exactly one of input_ids or inputs_embeds")
        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)
        batch_size, seq_len = inputs_embeds.shape[:2]

        past_key_values, use_cache = self._convert_past(past_key_values, use_cache)

        if position_ids is None:
            past_seen_tokens = self._past_seq_length(past_key_values)
            position_ids = torch.arange(seq_len, device=inputs_embeds.device) + past_seen_tokens
            position_ids = position_ids.unsqueeze(0).expand(batch_size, -1)

        position_embeddings = self.rotary_emb(inputs_embeds, position_ids)

        if attention_mask is not None:
            causal_mask = self._make_causal_mask(inputs_embeds, past_key_values)
            attention_mask = attention_mask[:, None, None, :]
            attention_mask = (1.0 - attention_mask) * torch.finfo(inputs_embeds.dtype).min
            attention_mask = attention_mask + causal_mask
        else:
            attention_mask = self._make_causal_mask(inputs_embeds, past_key_values)

        hidden_states = inputs_embeds
        new_past = [] if use_cache else None

        for i, layer in enumerate(self.layers):
            layer_past = past_key_values[i] if past_key_values is not None else None

            if self.gradient_checkpointing and self.training:
                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        return module(*inputs, past_key_values=None, use_cache=False)
                    return custom_forward
                hidden_states, _ = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(layer),
                    hidden_states,
                    attention_mask,
                    position_embeddings,
                    use_reentrant=True,
                )
            else:
                hidden_states, layer_past = layer(
                    hidden_states,
                    attention_mask=attention_mask,
                    position_embeddings=position_embeddings,
                    past_key_values=layer_past,
                    use_cache=use_cache,
                )

            if use_cache:
                new_past.append(layer_past)

        hidden_states = self.norm(hidden_states)
        return hidden_states, new_past

    def _make_causal_mask(self, inputs_embeds, past_key_values=None):
        batch_size, seq_len, _ = inputs_embeds.shape
        past_len = self._past_seq_length(past_key_values)
        total_len = past_len + seq_len
        mask = torch.full((seq_len, total_len), float('-inf'), device=inputs_embeds.device)
        mask = torch.triu(mask, diagonal=1 + past_len)
        return mask[None, None, :, :]

class LanceAI(LanceAIPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}

    def __init__(self, config):
        super().__init__(config)
        self.model = LanceAIModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

        self.post_init()

    def _set_gradient_checkpointing(self, enable=True, gradient_checkpointing_func=None):
        self.model._set_gradient_checkpointing(enable, gradient_checkpointing_func)

    def forward(self, input_ids=None, attention_mask=None, position_ids=None, past_key_values=None, inputs_embeds=None, labels=None, use_cache=None, **kwargs):
        hidden_states, past = 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,
        )

        logits = self.lm_head(hidden_states)

        loss = None
        if labels is not None:
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            loss = F.cross_entropy(
                shift_logits.view(-1, shift_logits.size(-1)),
                shift_labels.view(-1),
                ignore_index=-100,
            )

        return CausalLMOutputWithPast(loss=loss, logits=logits, past_key_values=past)

    def _past_seq_len(self, past_key_values):
        if past_key_values is None:
            return 0
        if hasattr(past_key_values, 'get_seq_length'):
            return past_key_values.get_seq_length()
        if isinstance(past_key_values, list) and len(past_key_values) > 0:
            if past_key_values[0] is not None:
                k, v = past_key_values[0]
                if k is not None:
                    return k.shape[2]
        return 0

    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, attention_mask=None, **kwargs):
        if self._past_seq_len(past_key_values) > 0:
            input_ids = input_ids[:, -1:]
        return {
            "input_ids": input_ids,
            "attention_mask": attention_mask,
            "past_key_values": past_key_values,
            "use_cache": True,
        }

    def _reorder_cache(self, past_key_values, beam_idx):
        reordered = []
        for layer_past in past_key_values:
            layer_k, layer_v = layer_past
            reordered.append((layer_k.index_select(0, beam_idx), layer_v.index_select(0, beam_idx)))
        return reordered

CONFIG_MAPPING.register("lance_ai", LanceAIConfig)
MODEL_FOR_CAUSAL_LM_MAPPING.register(LanceAIConfig, LanceAI)
LanceAIConfig.register_for_auto_class("AutoConfig")
LanceAI.register_for_auto_class("AutoModelForCausalLM")