""" Implementation for Deepseek architecture. """ import dataclasses from functools import partial from typing import Any, Dict, Optional # noqa: UP035 from tvm import tirx from tvm.relax.frontend import nn from tvm.relax.frontend.nn import Tensor, op from mlc_llm import op as op_ext from mlc_llm.model.model_utils import index_last_token from mlc_llm.nn import PagedKVCache, RopeMode from mlc_llm.nn.expert import MixtralExperts from mlc_llm.support import logging from mlc_llm.support import tensor_parallel as tp from mlc_llm.support.config import ConfigBase from mlc_llm.support.style import bold logger = logging.getLogger(__name__) @dataclasses.dataclass class DeepseekConfig(ConfigBase): """Configuration of the Deepseek model.""" vocab_size: int hidden_size: int intermediate_size: int moe_intermediate_size: int num_hidden_layers: int num_attention_heads: int num_key_value_heads: int n_shared_experts: int n_routed_experts: int moe_layer_freq: int first_k_dense_replace: int hidden_act: str norm_topk_prob: bool attention_bias: bool rms_norm_eps: float use_cache: bool bos_token_id: int eos_token_id: int tie_word_embeddings: bool = False rope_theta: int = 10000 context_window_size: int = 0 prefill_chunk_size: int = 0 tensor_parallel_shards: int = 1 head_dim: int = 0 max_batch_size: int = 1 num_experts_per_tok: int = 0 kwargs: Dict[str, Any] = dataclasses.field(default_factory=dict) # noqa: UP006 def __post_init__(self): if self.context_window_size == 0: for name in ["max_position_embeddings", "max_sequence_length"]: if name in self.kwargs: self.context_window_size = self.kwargs.pop(name) logger.info( "%s not found in config.json. Falling back to %s (%d)", bold("context_window_size"), bold(name), self.context_window_size, ) break else: raise ValueError( "Unable to determine the maximum sequence length, because none of " "`context_window_size`, `max_position_embeddings` or `max_sequence_length` is " "provided in `config.json`." ) if self.head_dim == 0: self.head_dim = self.hidden_size // self.num_attention_heads assert self.head_dim * self.num_attention_heads == self.hidden_size if self.prefill_chunk_size == 0: logger.info( "%s defaults to %d", bold("prefill_chunk_size"), min(self.context_window_size, 8192), ) self.prefill_chunk_size = min(self.context_window_size, 8192) elif self.prefill_chunk_size > self.context_window_size: logger.info( "Overriding %s from %d to %d", bold("prefill_chunk_size"), self.prefill_chunk_size, min(self.context_window_size, 8192), ) self.prefill_chunk_size = min(self.context_window_size, 8192) class DeepseekAttention(nn.Module): def __init__(self, config: DeepseekConfig): super().__init__() # Make sure to call the parent class constructor self.hidden_size = config.hidden_size self.rope_theta = config.rope_theta self.tensor_parallel_shards = config.tensor_parallel_shards if config.num_attention_heads % config.tensor_parallel_shards != 0: raise ValueError( f"Cannot split {config.num_attention_heads} attention heads " f"evenly to {config.tensor_parallel_shards} GPUs." ) self.attention_bias = config.attention_bias self.num_heads = config.num_attention_heads // self.tensor_parallel_shards self.num_key_value_heads = config.num_key_value_heads // self.tensor_parallel_shards self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.head_dim = config.head_dim self.max_position_embeddings = config.context_window_size self.wqkv_pack = nn.Linear( in_features=self.hidden_size, out_features=(self.num_heads + 2 * self.num_key_value_heads) * self.head_dim, bias=self.attention_bias, ) self.o_proj = nn.Linear( self.num_heads * self.head_dim, self.hidden_size, bias=self.attention_bias ) def forward(self, hidden_states: Tensor, paged_kv_cache: PagedKVCache, layer_id: int): d, h_q, h_kv = self.head_dim, self.num_heads, self.num_key_value_heads b, s, _ = hidden_states.shape qkv = self.wqkv_pack(hidden_states) qkv = op.reshape(qkv, (b, s, h_q + h_kv + h_kv, d)) output = op.reshape( paged_kv_cache.attention_with_fused_qkv( layer_id, qkv, self.num_heads, sm_scale=self.head_dim**-0.5 ), (b, s, h_q * d), ) attn_output = self.o_proj(output) return attn_output ACT2FN = { "gelu": partial(nn.gelu, approximate=False), "relu": nn.relu, "silu": nn.silu, "swish": nn.silu, "gelu_new": partial(nn.gelu, approximate=True), } class DeepseekMLP(nn.Module): def __init__(self, config: DeepseekConfig, intermediate_size=None): self.hidden_size = config.hidden_size if config.intermediate_size % config.tensor_parallel_shards != 0: raise ValueError( f"Cannot split MLP intermediate size {config.intermediate_size} " f"evenly to {config.tensor_parallel_shards} GPUs." ) self.intermediate_size = ( config.intermediate_size if intermediate_size is None else intermediate_size ) // config.tensor_parallel_shards self.gate_up_proj = nn.Linear(self.hidden_size, 2 * self.intermediate_size, bias=False) self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False) self.act_fn = ACT2FN[config.hidden_act] def forward(self, x: Tensor): concat_x1_x2 = self.gate_up_proj(x) x1, x2 = op.split(concat_x1_x2, 2, axis=-1) return self.down_proj(op.silu(x1) * x2) class DeepseekMoE(nn.Module): def __init__(self, config: DeepseekConfig): self.num_local_experts = config.n_routed_experts self.num_experts_per_tok = config.num_experts_per_tok self.gate = nn.Linear(config.hidden_size, config.n_routed_experts, bias=False) self.norm_topk_prob = config.norm_topk_prob self.moe_intermediate_size = config.moe_intermediate_size // config.tensor_parallel_shards self.moe_gate_up_proj = MixtralExperts( self.num_local_experts, in_features=config.hidden_size, out_features=2 * self.moe_intermediate_size, tensor_parallel_shards=config.tensor_parallel_shards, ) self.moe_down_proj = MixtralExperts( self.num_local_experts, in_features=self.moe_intermediate_size, out_features=config.hidden_size, tensor_parallel_shards=config.tensor_parallel_shards, ) self.dtype = "float32" if config.n_shared_experts is not None: intermediate_size = self.moe_intermediate_size * config.n_shared_experts self.shared_experts = DeepseekMLP(config, intermediate_size=intermediate_size) def forward(self, x: Tensor): def _expert_forward(x: Tensor, indptr: Tensor): x1_x3 = self.moe_gate_up_proj(x, indptr) x1, x3 = op.split(x1_x3, indices_or_sections=2, axis=-1) x = self.moe_down_proj(op.silu(x1) * x3, indptr) return x experts_per_tok = self.num_experts_per_tok num_experts = self.num_local_experts b, s, h = x.shape num_tokens = b * s x = op.reshape(x, (num_tokens, h)) gate = self.gate(x) # (b * s, num_routed_experts) expert_weights, expert_indices = op_ext.moe_misc.gating_softmax_topk( gate, experts_per_tok, norm_topk_prob=self.norm_topk_prob ) if num_tokens == 1: # x: [num_tokens * experts_per_tok, hidden_size] moe_hidden_states = _expert_forward(x, expert_indices) else: # cumsum: [num_tokens * local_experts] cumsum = op_ext.moe_misc.moe_cumsum(expert_indices, num_experts) # indices: [num_tokens * experts_per_tok] reverse_indices, token_indices = op_ext.moe_misc.get_indices(cumsum, expert_indices) # indptr: [num_local_experts + 1] indptr = op_ext.moe_misc.get_indptr( cumsum, num_experts, num_tokens, inclusive=False, out_dtype="int32" ) # x: [num_tokens * experts_per_tok, hidden_size] moe_hidden_states = op.take(x, token_indices, axis=0) moe_hidden_states = _expert_forward(moe_hidden_states, indptr) moe_hidden_states = op_ext.moe_misc.scatter_output(moe_hidden_states, reverse_indices) # moe_hidden_states: [num_tokens, experts_per_tok, hidden_size] expert_weights = expert_weights.reshape(num_tokens, experts_per_tok, 1) moe_hidden_states = ( moe_hidden_states.reshape(num_tokens, experts_per_tok, h) * expert_weights ) # moe_hidden_states: [num_tokens, hidden_size] moe_hidden_states = op_ext.moe_misc.moe_sum(moe_hidden_states, dim=1) shared_expert_hidden_states = self.shared_experts(x) final_hidden_states = moe_hidden_states + shared_expert_hidden_states final_hidden_states = op.reshape(final_hidden_states, (b, s, h)) return final_hidden_states class DeepseekDecoderLayer(nn.Module): def __init__(self, config: DeepseekConfig, layer_idx: int): self.hidden_size = config.hidden_size self.num_hidden_layers = config.num_hidden_layers self.self_attn = DeepseekAttention(config) self.num_experts = config.n_routed_experts self.mlp = ( DeepseekMoE(config) if ( config.n_routed_experts is not None and layer_idx >= config.first_k_dense_replace and layer_idx % config.moe_layer_freq == 0 ) else DeepseekMLP(config) ) self.input_layernorm = nn.RMSNorm(config.hidden_size, -1, config.rms_norm_eps, bias=False) self.post_attention_layernorm = nn.RMSNorm( config.hidden_size, -1, config.rms_norm_eps, bias=False ) def _set_tp(): def _set(layer, hint): layer.attrs["shard_strategy"] = hint hd = config.head_dim q = self.self_attn.num_heads * hd k = self.self_attn.num_key_value_heads * hd v = self.self_attn.num_key_value_heads * hd if ( config.n_routed_experts is not None and layer_idx >= config.first_k_dense_replace and layer_idx % config.moe_layer_freq == 0 ): i = self.mlp.moe_intermediate_size else: i = self.mlp.intermediate_size _set( self.self_attn.wqkv_pack.weight, tp.ShardSingleDim("_shard_qkv_weight", dim=0, segs=[q, k, v]), ) _set(self.self_attn.o_proj.weight, tp.ShardSingleDim("_shard_o", dim=1)) if ( config.n_routed_experts is not None and layer_idx >= config.first_k_dense_replace and layer_idx % config.moe_layer_freq == 0 ): _set( self.mlp.moe_gate_up_proj.weight, tp.ShardSingleDim("_shard_mlp_up", segs=[i, i], dim=1), ) _set( self.mlp.moe_down_proj.weight, tp.ShardSingleDim("_shard_mlp_down", dim=2), ) else: _set( self.mlp.gate_up_proj.weight, tp.ShardSingleDim("_shard_mlp_up", segs=[i, i], dim=0), ) _set( self.mlp.down_proj.weight, tp.ShardSingleDim("_shard_mlp_down", dim=1), ) self.tensor_parallel_shards = config.tensor_parallel_shards _set_tp() def forward(self, hidden_states: Tensor, paged_kv_cache: PagedKVCache, layer_id: int): out = self.input_layernorm(hidden_states) out = self.self_attn(out, paged_kv_cache, layer_id) hidden_states = self._apply_residual(out, residual=hidden_states) out = self.post_attention_layernorm(hidden_states) out = self.mlp(out) hidden_states = self._apply_residual(out, residual=hidden_states) return hidden_states def _apply_residual(self, out, residual): if self.tensor_parallel_shards > 1: return op.ccl_allreduce(out, "sum") + residual return out + residual class DeepseekModel(nn.Module): def __init__(self, config: DeepseekConfig): assert config.hidden_size % config.num_attention_heads == 0 self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size) self.layers = nn.ModuleList( [ DeepseekDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers) ] ) self.norm = nn.RMSNorm(config.hidden_size, -1, config.rms_norm_eps, bias=False) def forward(self, inputs: Tensor, paged_kv_cache: PagedKVCache): hidden_states = inputs for layer_id, layer in enumerate(self.layers): hidden_states = layer(hidden_states, paged_kv_cache, layer_id) hidden_states = self.norm(hidden_states) return hidden_states class DeepseekForCausalLM(nn.Module): def __init__(self, config: DeepseekConfig): self.model = DeepseekModel(config) self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) self.vocab_size = config.vocab_size self.num_hidden_layers = config.num_hidden_layers self.hidden_size = config.hidden_size self.num_attention_heads = config.num_attention_heads self.num_key_value_heads = config.num_key_value_heads self.tensor_parallel_shards = config.tensor_parallel_shards self.head_dim = config.head_dim self.vocab_size = config.vocab_size self.rope_theta = config.rope_theta self.dtype = "float32" def to(self, dtype: Optional[str] = None): super().to(dtype=dtype) if dtype is not None: self.dtype = dtype def batch_forward( self, input_embeds: Tensor, paged_kv_cache: PagedKVCache, logit_positions: Optional[Tensor] = None, ): op_ext.configure() hidden_states = self.model(input_embeds, paged_kv_cache) if logit_positions is not None: hidden_states = op.take(hidden_states, logit_positions, axis=1) logits = self.lm_head(hidden_states) if logits.dtype != "float32": logits = logits.astype("float32") return logits def embed(self, input_ids: Tensor): if self.tensor_parallel_shards > 1: input_ids = op.ccl_broadcast_from_worker0(input_ids) return self.model.embed_tokens(input_ids) def prefill(self, input_embed: Tensor, paged_kv_cache: PagedKVCache): op_ext.configure() hidden_states = self.model(input_embed, paged_kv_cache) hidden_states = index_last_token(hidden_states) logits = self.lm_head(hidden_states) if logits.dtype != "float32": logits = logits.astype("float32") return logits, paged_kv_cache def decode(self, input_embed: Tensor, paged_kv_cache: PagedKVCache): op_ext.configure() hidden_states = self.model(input_embed, paged_kv_cache) logits = self.lm_head(hidden_states) if logits.dtype != "float32": logits = logits.astype("float32") return logits, paged_kv_cache def batch_prefill( self, input_embeds: Tensor, logit_positions: Tensor, paged_kv_cache: PagedKVCache, ): if self.tensor_parallel_shards > 1: logit_positions = op.ccl_broadcast_from_worker0(logit_positions) logits = self.batch_forward(input_embeds, paged_kv_cache, logit_positions) return logits, paged_kv_cache def batch_decode(self, input_embeds: Tensor, paged_kv_cache: PagedKVCache): logits = self.batch_forward(input_embeds, paged_kv_cache) return logits, paged_kv_cache def batch_verify(self, input_embeds: Tensor, paged_kv_cache: PagedKVCache): logits = self.batch_forward(input_embeds, paged_kv_cache) return logits, paged_kv_cache def create_paged_kv_cache( self, max_batch_size: tirx.Var, max_total_seq_len: tirx.Var, prefill_chunk_size: tirx.Var, page_size: tirx.Var, support_sliding_window: tirx.Var, ) -> PagedKVCache: return PagedKVCache.create_generic( attn_kind="mha", max_batch_size=max_batch_size, max_total_seq_len=max_total_seq_len, prefill_chunk_size=prefill_chunk_size, page_size=page_size, support_sliding_window=support_sliding_window, num_hidden_layers=self.num_hidden_layers, num_attention_heads=self.num_attention_heads // self.tensor_parallel_shards, num_key_value_heads=self.num_key_value_heads // self.tensor_parallel_shards, qk_head_dim=self.head_dim, v_head_dim=self.head_dim, rope_mode=RopeMode.NORMAL, rope_scale=1, rope_theta=self.rope_theta, dtype=self.dtype, ) def get_default_spec(self): mod_spec = { "embed": { "input_ids": nn.spec.Tensor(["seq_len"], "int32"), "$": { "param_mode": "packed", "effect_mode": "none", }, }, "prefill": { "input_embed": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype), "paged_kv_cache": nn.spec.Object(object_type=PagedKVCache), "$": { "param_mode": "packed", "effect_mode": "none", }, }, "decode": { "input_embed": nn.spec.Tensor([1, 1, self.hidden_size], self.dtype), "paged_kv_cache": nn.spec.Object(object_type=PagedKVCache), "$": { "param_mode": "packed", "effect_mode": "none", }, }, "batch_prefill": { "input_embeds": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype), "logit_positions": nn.spec.Tensor(["batch_size"], "int32"), "paged_kv_cache": nn.spec.Object(object_type=PagedKVCache), "$": { "param_mode": "packed", "effect_mode": "none", }, }, "batch_decode": { "input_embeds": nn.spec.Tensor(["batch_size", 1, self.hidden_size], self.dtype), "paged_kv_cache": nn.spec.Object(object_type=PagedKVCache), "$": { "param_mode": "packed", "effect_mode": "none", }, }, "batch_verify": { "input_embeds": nn.spec.Tensor([1, "seq_len", self.hidden_size], self.dtype), "paged_kv_cache": nn.spec.Object(object_type=PagedKVCache), "$": { "param_mode": "packed", "effect_mode": "none", }, }, "create_paged_kv_cache": { "max_batch_size": int, "max_total_seq_len": int, "prefill_chunk_size": int, "page_size": int, "support_sliding_window": int, "$": { "param_mode": "none", "effect_mode": "none", }, }, } return nn.spec.ModuleSpec.from_raw(mod_spec, self)