| import os |
| import torch |
| import torch.nn as nn |
|
|
| from diffulex.attention import Attention |
| from diffulex.layer.layernorm import RMSNorm |
| from diffulex.layer.activation import SiluAndMul |
| from diffulex.layer.rotary_embedding import get_rope |
| from diffulex.model.auto_model import AutoModelForDiffusionLM |
| from diffulex.layer.linear import RowParallelLinear, ColumnParallelLinear |
| from diffulex.layer.embed_head import VocabParallelEmbedding, ParallelLMHead |
| from diffulex.model.config.fast_dllm_v2.configuration_fast_dllm_v2 import ( |
| FastdLLMV2Config, |
| ) |
| from diffulex.distributed.parallel_state import fetch_parallel_state |
|
|
|
|
| if os.environ.get("TRITON_INTERPRET", None) == "1": |
| torch._dynamo.reset() |
| torch._dynamo.config.suppress_errors = True |
| torch.backends.optimized_mode = False |
|
|
|
|
| class FastdLLMV2RMSNorm(RMSNorm): |
| def __init__(self, hidden_size, eps=1e-6): |
| super().__init__(hidden_size, eps) |
|
|
|
|
| class FastdLLMV2Attention(nn.Module): |
| """FastdLLM V2 attention mechanism.""" |
|
|
| def __init__( |
| self, |
| hidden_size: int, |
| num_heads: int, |
| num_kv_heads: int, |
| max_position: int = 32768, |
| head_dim: int | None = None, |
| rms_norm_eps: float = 1e-6, |
| qkv_bias: bool = True, |
| rope_theta: float = 10000, |
| rope_scaling: tuple | None = None, |
| attn_impl: str = "triton", |
| ) -> None: |
| super().__init__() |
| parallel_state = fetch_parallel_state() |
| tp_size = parallel_state.get_tp_world_size() |
| self.total_num_heads = num_heads |
| assert self.total_num_heads % tp_size == 0 |
| self.num_heads = self.total_num_heads // tp_size |
| self.total_num_kv_heads = num_kv_heads |
| assert self.total_num_kv_heads % tp_size == 0 |
| self.num_kv_heads = self.total_num_kv_heads // tp_size |
| self.head_dim = head_dim or hidden_size // self.total_num_heads |
| self.q_size = self.num_heads * self.head_dim |
| self.kv_size = self.num_kv_heads * self.head_dim |
| self.scaling = self.head_dim**-0.5 |
|
|
| self.q_proj = ColumnParallelLinear( |
| hidden_size, |
| self.total_num_heads * self.head_dim, |
| bias=qkv_bias, |
| ) |
| self.k_proj = ColumnParallelLinear( |
| hidden_size, |
| self.total_num_kv_heads * self.head_dim, |
| bias=qkv_bias, |
| ) |
| self.v_proj = ColumnParallelLinear( |
| hidden_size, |
| self.total_num_kv_heads * self.head_dim, |
| bias=qkv_bias, |
| ) |
| self.o_proj = RowParallelLinear( |
| self.total_num_heads * self.head_dim, |
| hidden_size, |
| bias=False, |
| ) |
| self.rotary_emb = get_rope( |
| self.head_dim, |
| rotary_dim=self.head_dim, |
| max_position=max_position, |
| base=rope_theta, |
| rope_scaling=rope_scaling, |
| ) |
| self.attn = Attention( |
| self.num_heads, |
| self.head_dim, |
| self.scaling, |
| self.num_kv_heads, |
| attn_impl=attn_impl, |
| ) |
|
|
| def forward( |
| self, |
| positions: torch.Tensor, |
| hidden_states: torch.Tensor, |
| mask: torch.Tensor | None = None, |
| ) -> torch.Tensor: |
| q = self.q_proj(hidden_states) |
| k = self.k_proj(hidden_states) |
| v = self.v_proj(hidden_states) |
|
|
| q, k = self.rotary_emb(positions, q, k) |
| o = self.attn(q, k, v, mask) |
| output = self.o_proj(o) |
| return output |
|
|
|
|
| class FastdLLMV2MLP(nn.Module): |
| """FastdLLM V2 MLP with SiLU activation.""" |
|
|
| def __init__( |
| self, |
| hidden_size: int, |
| intermediate_size: int, |
| hidden_act: str, |
| ) -> None: |
| super().__init__() |
| self.gate_proj = ColumnParallelLinear( |
| hidden_size, |
| intermediate_size, |
| bias=False, |
| ) |
| self.up_proj = ColumnParallelLinear( |
| hidden_size, |
| intermediate_size, |
| bias=False, |
| ) |
| self.down_proj = RowParallelLinear( |
| intermediate_size, |
| hidden_size, |
| bias=False, |
| ) |
| assert hidden_act == "silu" |
| self.act_fn = SiluAndMul() |
|
|
| def forward(self, x): |
| gate = self.gate_proj(x) |
| up = self.up_proj(x) |
| x = self.act_fn(torch.cat([gate, up], dim=-1)) |
| x = self.down_proj(x) |
| return x |
|
|
|
|
| class FastdLLMV2DecoderLayer(nn.Module): |
| """FastdLLM V2 transformer decoder layer.""" |
|
|
| def __init__( |
| self, |
| config: FastdLLMV2Config, |
| ) -> None: |
| super().__init__() |
| self.self_attn = FastdLLMV2Attention( |
| hidden_size=config.hidden_size, |
| num_heads=config.num_attention_heads, |
| num_kv_heads=config.num_key_value_heads, |
| max_position=config.max_position_embeddings, |
| rms_norm_eps=config.rms_norm_eps, |
| qkv_bias=True, |
| head_dim=getattr(config, "head_dim", None), |
| rope_theta=getattr(config, "rope_theta", 10000), |
| rope_scaling=getattr(config, "rope_scaling", None), |
| attn_impl=getattr(config, "attn_impl", "triton"), |
| ) |
| self.mlp = FastdLLMV2MLP( |
| hidden_size=config.hidden_size, |
| intermediate_size=config.intermediate_size, |
| hidden_act=config.hidden_act, |
| ) |
| self.input_layernorm = FastdLLMV2RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| self.post_attention_layernorm = FastdLLMV2RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
|
|
| def forward( |
| self, |
| positions: torch.Tensor, |
| hidden_states: torch.Tensor, |
| residual: torch.Tensor | None, |
| mask: torch.Tensor | None = None, |
| ) -> tuple[torch.Tensor, torch.Tensor]: |
| if residual is None: |
| residual = hidden_states |
| hidden_states = self.input_layernorm(hidden_states) |
| else: |
| hidden_states, residual = self.input_layernorm(hidden_states, residual) |
| hidden_states = self.self_attn(positions, hidden_states, mask) |
| hidden_states, residual = self.post_attention_layernorm(hidden_states, residual) |
| hidden_states = self.mlp(hidden_states) |
| return hidden_states, residual |
|
|
|
|
| class FastdLLMV2Model(nn.Module): |
| """FastdLLM V2 model for diffusion language modeling.""" |
|
|
| def __init__( |
| self, |
| config: FastdLLMV2Config, |
| ) -> None: |
| super().__init__() |
| self.embed_tokens = VocabParallelEmbedding(config.vocab_size, config.hidden_size) |
| self.layers = nn.ModuleList([FastdLLMV2DecoderLayer(config) for _ in range(config.num_hidden_layers)]) |
| self.norm = FastdLLMV2RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
|
|
| def forward( |
| self, |
| input_ids: torch.Tensor, |
| positions: torch.Tensor, |
| mask: torch.Tensor | None = None, |
| ) -> torch.Tensor: |
| hidden_states = self.embed_tokens(input_ids) |
| residual = None |
| for _, layer in enumerate(self.layers): |
| hidden_states, residual = layer(positions, hidden_states, residual, mask) |
| hidden_states, _ = self.norm(hidden_states, residual) |
| return hidden_states |
|
|
|
|
| @AutoModelForDiffusionLM.register("fast_dllm_v2") |
| class FastdLLMV2ForDiffusionLM(nn.Module): |
| """FastdLLM V2 model for diffusion language modeling with LM head.""" |
|
|
| packed_modules_mapping = {} |
|
|
| def __init__( |
| self, |
| config: FastdLLMV2Config, |
| ) -> None: |
| super().__init__() |
| self.model = FastdLLMV2Model(config) |
| self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size) |
| if getattr(config, "tie_word_embeddings", False): |
| self.lm_head.weight.data = self.model.embed_tokens.weight.data |
|
|
| def forward( |
| self, |
| input_ids: torch.Tensor, |
| positions: torch.Tensor, |
| mask: torch.Tensor | None = None, |
| ) -> torch.Tensor: |
| hidden_states = self.model(input_ids, positions, mask) |
| return hidden_states |
|
|
| def compute_logits( |
| self, |
| hidden_states: torch.Tensor, |
| ) -> torch.Tensor: |
| logits = self.lm_head(hidden_states) |
| return logits |
|
|
