update config

config.json

@@ -1,10 +1,10 @@
 {
   "architectures": [
-    "StepAudioForCausalLM"
+    "Step1ForCausalLM"
   ],
   "auto_map": {
-    "AutoConfig": "configuration_stepaudio.StepAudioConfig",
-    "AutoModelForCausalLM": "modeling_stepaudio.StepAudioForCausalLM"
+    "AutoConfig": "configuration_step1.Step1Config",
+    "AutoModelForCausalLM": "modeling_step1.Step1ForCausalLM"
   },
   "model_type": "step_audio",
   "bos_token_id": 1,

configuration_step1.py

@@ -0,0 +1,41 @@
+from typing import Optional, List, Any, Dict
+from transformers.configuration_utils import PretrainedConfig
+
+
+
+class Step1Config(PretrainedConfig):
+    model_type = "step1"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    
+    def __init__(
+        self,
+        hidden_size: int = 5120,
+        intermediate_size: int = 13312,
+        num_attention_heads: int = 40,
+        num_attention_groups: int = 8,
+        num_hidden_layers: int = 48,
+        max_seq_len: int = 4096,
+        vocab_size: int = 65536,
+        rms_norm_eps: float = 1e-5,
+        bos_token_id: int = 1,
+        eos_token_id: int = 3,
+        pad_token_id: int = 0,
+        **kwargs,
+    ) -> None:
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_attention_heads = num_attention_heads
+        self.num_attention_groups = num_attention_groups
+        self.num_hidden_layers = num_hidden_layers
+        self.max_seq_len = max_seq_len
+        self.vocab_size = vocab_size
+        self.rms_norm_eps = rms_norm_eps
+        super().__init__(
+            bos_token_id=bos_token_id,
+            pad_token_id=pad_token_id,
+            eos_token_id=eos_token_id,
+            **kwargs
+        )
+
+
+__all__ = ["Step1Config"]

modeling_step1.py

@@ -0,0 +1,392 @@
+import math
+from typing import Optional, Tuple, Union, List
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from transformers.generation import GenerationMixin
+
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+from .configuration_step1 import Step1Config
+from transformers.cache_utils import Cache, DynamicCache
+from einops import rearrange
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+)
+
+logger = logging.get_logger(__name__)
+
+
+def build_alibi_cache(block_size, n_heads, dtype, device):
+    # get slopes
+    n = 2 ** math.floor(math.log2(n_heads))  # nearest 2**n to n_heads
+    m0 = 2.0 ** (-8.0 / n)
+    # 2^(-8/n), 2^(-8*2/n), 2^(-8*3/n), ...
+    slopes = torch.pow(m0, torch.arange(1, n + 1))
+    if n < n_heads:
+        m1 = 2.0 ** (-4.0 / n)
+        # 2^(-8/(2n)), 2^(-8*3/(2n)), 2^(-8*5/(2n)), ...
+        mm = torch.pow(m1, torch.arange(1, 1 + 2 * (n_heads - n), 2))
+        slopes = torch.cat([slopes, mm])
+    slopes = slopes.to(device)
+
+    tril = torch.tril(torch.ones(1, 1, block_size, block_size, device=device))
+
+    bias_rows = torch.arange(block_size, device=device).view(1, -1)
+    bias_cols = torch.arange(block_size, device=device).view(-1, 1)
+    bias = -torch.sqrt(bias_cols - bias_rows)
+    bias = bias.view(1, block_size, block_size) * slopes.view(-1, 1, 1)
+    bias = bias.masked_fill(tril == 0, float("-inf"))
+
+    return bias.type(dtype)
+
+
+class StepRMSNorm(torch.nn.Module):
+    def __init__(self, hidden_size, eps=1e-5):
+        super().__init__()
+        self.weight = torch.nn.Parameter(torch.ones(hidden_size))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor):
+        var = x.float().pow(2).mean(-1, keepdim=True)
+        x = x * torch.rsqrt(var + self.eps).to(x.dtype)
+        x = x * self.weight
+        return x
+
+
+class StepAttention(torch.nn.Module):
+    def __init__(self, hidden_size, num_heads, num_groups, layer_idx: int):
+        super().__init__()
+
+        self.num_heads = num_heads
+        self.num_groups = num_groups
+        self.hidden_size = hidden_size
+        self.head_dim = hidden_size // num_heads
+
+        self.q_proj = torch.nn.Linear(hidden_size, hidden_size, bias=False)
+        self.k_proj = torch.nn.Linear(
+            hidden_size, num_groups * self.head_dim, bias=False
+        )
+        self.v_proj = torch.nn.Linear(
+            hidden_size, num_groups * self.head_dim, bias=False
+        )
+        self.o_proj = torch.nn.Linear(hidden_size, hidden_size, bias=False)
+
+        self.layer_idx = layer_idx
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        past_key_value: Optional[Cache] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+    ):
+
+        q: torch.Tensor = self.q_proj(x)
+        k: torch.Tensor = self.k_proj(x)
+        v: torch.Tensor = self.v_proj(x)
+        if past_key_value is not None:
+            cache_kwargs = {"cache_position": cache_position}
+            k, v = past_key_value.update(k, v, self.layer_idx, cache_kwargs)
+
+        q = rearrange(q, "b s (h d) -> b s h d", h=self.num_heads)
+        k = rearrange(k, "b s (g d) -> b s g d", g=self.num_groups)
+        v = rearrange(v, "b s (g d) -> b s g d", g=self.num_groups)
+
+        k = k.repeat_interleave(self.num_heads // self.num_groups, dim=-2)
+        v = v.repeat_interleave(self.num_heads // self.num_groups, dim=-2)
+
+        attention_mask = build_alibi_cache(
+            k.size(1), self.num_heads, dtype=q.dtype, device=q.device
+        )[:, :, -q.size(1) :, :].contiguous()
+
+        q = q.transpose(1, 2)
+        k = k.transpose(1, 2)
+        v = v.transpose(1, 2)
+
+        o: torch.Tensor = torch.nn.functional.scaled_dot_product_attention(
+            q, k, v, attn_mask=attention_mask
+        )
+        o = o.transpose(1, 2).flatten(-2, -1)
+
+        o = self.o_proj(o)
+        return o
+
+
+class StepMLP(torch.nn.Module):
+    def __init__(self, hidden_size, intermediate_size):
+        super().__init__()
+        self.gate_proj = torch.nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.up_proj = torch.nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.down_proj = torch.nn.Linear(intermediate_size, hidden_size, bias=False)
+
+    def forward(self, x):
+        gate = self.gate_proj(x)
+        up = self.up_proj(x)
+        x = torch.nn.functional.silu(gate) * up
+        x = self.down_proj(x)
+        return x
+
+
+class StepLayer(torch.nn.Module):
+    def __init__(self, config: Step1Config, layer_idx: int):
+        super().__init__()
+        self.layer_idx = layer_idx
+        self.self_attn = StepAttention(
+            hidden_size=config.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_groups=config.num_attention_groups,
+            layer_idx=layer_idx,
+        )
+        self.mlp = StepMLP(
+            hidden_size=config.hidden_size,
+            intermediate_size=config.intermediate_size,
+        )
+        self.input_layernorm = StepRMSNorm(
+            hidden_size=config.hidden_size, eps=config.rms_norm_eps
+        )
+        self.post_attention_layernorm = StepRMSNorm(
+            hidden_size=config.hidden_size, eps=config.rms_norm_eps
+        )
+
+    def forward(
+        self,
+        x,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+    ):
+        def f(x):
+            x = self.input_layernorm(x)
+            x = self.self_attn(x, past_key_value, attention_mask, cache_position)
+            return x
+
+        x = x + f(x)
+
+        def f(x):
+            x = self.post_attention_layernorm(x)
+            x = self.mlp(x)
+            return x
+
+        x = x + f(x)
+
+        return x
+
+
+class StepPreTrainedModel(PreTrainedModel):
+    config_class = Step1Config
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["StepLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_cache_class = True
+    _supports_static_cache = True
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+class Step1Model(StepPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`LlamaDecoderLayer`]
+
+    Args:
+        config: Step1Config
+    """
+
+    def __init__(self, config: Step1Config):
+        super().__init__(config)
+        self.config = config
+        self.embed_tokens = torch.nn.Embedding(config.vocab_size, config.hidden_size)
+
+        self.layers = torch.nn.Sequential(
+            *[
+                StepLayer(config, layer_idx)
+                for layer_idx in range(config.num_hidden_layers)
+            ]
+        )
+
+        self.norm = StepRMSNorm(
+            hidden_size=config.hidden_size, eps=config.rms_norm_eps
+        )
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = False
+        output_hidden_states = False
+
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError(
+                "You must specify exactly one of input_ids or inputs_embeds"
+            )
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if cache_position is None:
+            past_seen_tokens = (
+                past_key_values.get_seq_length() if past_key_values is not None else 0
+            )
+            cache_position = torch.arange(
+                past_seen_tokens,
+                past_seen_tokens + inputs_embeds.shape[1],
+                device=inputs_embeds.device,
+            )
+
+        causal_mask = attention_mask
+
+        hidden_states = inputs_embeds
+
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+            layer_outputs = decoder_layer(
+                hidden_states,
+                attention_mask=causal_mask,
+                past_key_value=past_key_values,
+                cache_position=cache_position,
+            )
+
+            hidden_states = layer_outputs
+
+        hidden_states = self.norm(hidden_states)
+
+        output = BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values if use_cache else None,
+            hidden_states=hidden_states,
+            attentions=None,
+        )
+        return output if return_dict else output.to_tuple()
+
+
+class Step1ForCausalLM(StepPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = Step1Model(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    # def get_output_embeddings(self):
+    #     return self.lm_head
+
+    # def set_output_embeddings(self, new_embeddings):
+    #     self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        num_logits_to_keep: int = 0,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        # output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_attentions = False
+        output_hidden_states = False
+        # output_hidden_states = (
+        #     output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        # )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+        )
+
+        hidden_states = outputs[0]
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+
+        logits = self.lm_head(hidden_states)
+
+        # logits = torch.matmul(hidden_states, lm_stat)
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs
+            )
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )