# Copyright 2023-2024 SGLang Team # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== import logging from typing import Any, Dict, Iterable, Optional, Tuple, Union import torch import torch.nn.functional as F from torch import nn from sglang.srt.batch_overlap.two_batch_overlap import model_forward_maybe_tbo from sglang.srt.distributed import ( get_moe_expert_parallel_world_size, get_pp_group, get_tensor_model_parallel_world_size, tensor_model_parallel_all_reduce, ) from sglang.srt.eplb.expert_distribution import get_global_expert_distribution_recorder from sglang.srt.eplb.expert_location import ModelConfigForExpertLocation from sglang.srt.eplb.expert_location_dispatch import ExpertLocationDispatchInfo from sglang.srt.layers.activation import SiluAndMul from sglang.srt.layers.communicator import ( LayerCommunicator, LayerScatterModes, ScatterMode, enable_moe_dense_fully_dp, ) from sglang.srt.layers.dp_attention import ( get_attention_tp_rank, get_attention_tp_size, is_dp_attention_enabled, ) from sglang.srt.layers.layernorm import RMSNorm from sglang.srt.layers.linear import ( MergedColumnParallelLinear, QKVParallelLinear, RowParallelLinear, ) from sglang.srt.layers.logits_processor import LogitsProcessor from sglang.srt.layers.moe import ( get_moe_a2a_backend, get_moe_runner_backend, should_use_flashinfer_cutlass_moe_fp4_allgather, ) from sglang.srt.layers.moe.ep_moe.layer import DeepEPMoE, get_moe_impl_class from sglang.srt.layers.moe.topk import TopK, TopKOutputFormat from sglang.srt.layers.quantization.base_config import QuantizationConfig from sglang.srt.layers.radix_attention import RadixAttention from sglang.srt.layers.rotary_embedding import get_rope from sglang.srt.layers.utils import PPMissingLayer, get_layer_id from sglang.srt.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding, ) from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors from sglang.srt.model_loader.weight_utils import ( default_weight_loader, kv_cache_scales_loader, ) from sglang.srt.server_args import get_global_server_args from sglang.srt.utils import ( LazyValue, add_prefix, is_non_idle_and_non_empty, make_layers, ) MiMoV2FlashConfig = None logger = logging.getLogger(__name__) class MiMoV2MLP(nn.Module): def __init__( self, hidden_size: int, intermediate_size: int, hidden_act: str, quant_config: Optional[QuantizationConfig] = None, reduce_results: bool = True, prefix: str = "", tp_rank: Optional[int] = None, tp_size: Optional[int] = None, ) -> None: super().__init__() self.tp_size = tp_size self.gate_up_proj = MergedColumnParallelLinear( hidden_size, [intermediate_size] * 2, bias=False, quant_config=quant_config, prefix=add_prefix("gate_up_proj", prefix), tp_rank=tp_rank, tp_size=tp_size, ) self.down_proj = RowParallelLinear( intermediate_size, hidden_size, bias=False, quant_config=quant_config, reduce_results=reduce_results, prefix=add_prefix("down_proj", prefix), tp_rank=tp_rank, tp_size=tp_size, ) if hidden_act != "silu": raise ValueError( f"Unsupported activation: {hidden_act}. " "Only silu is supported for now." ) self.act_fn = SiluAndMul() def forward( self, x, forward_batch: ForwardBatch = None, should_allreduce_fusion: bool = False, use_reduce_scatter: bool = False, ): if (self.tp_size == 1) and x.shape[0] == 0: return x gate_up, _ = self.gate_up_proj(x) x = self.act_fn(gate_up) x, _ = self.down_proj( x, skip_all_reduce=should_allreduce_fusion or use_reduce_scatter ) return x class MoEGate(nn.Module): def __init__( self, config, quant_config, prefix: str = "", is_nextn: bool = False, ): super().__init__() self.is_nextn = is_nextn self.dtype = torch.float32 self.weight = nn.Parameter( torch.empty((config.n_routed_experts, config.hidden_size), dtype=self.dtype) ) if config.topk_method == "noaux_tc": correction_bias_dtype = ( torch.bfloat16 if quant_config is not None and quant_config.get_name() == "modelopt_fp4" and get_moe_runner_backend().is_flashinfer_trtllm() else self.dtype ) self.e_score_correction_bias = nn.Parameter( torch.empty((config.n_routed_experts), dtype=correction_bias_dtype) ) else: self.e_score_correction_bias = None def forward(self, hidden_states): logits = F.linear(hidden_states.to(self.dtype), self.weight, None) return logits class MiMoV2MoE(nn.Module): def __init__( self, config: MiMoV2FlashConfig, layer_id: int, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", is_nextn: bool = False, ): super().__init__() self.tp_size = get_tensor_model_parallel_world_size() self.config = config self.layer_id = layer_id if self.tp_size > config.n_routed_experts: raise ValueError( f"Tensor parallel size {self.tp_size} is greater than " f"the number of experts {config.n_routed_experts}." ) if config.hidden_act != "silu": raise ValueError( f"Unsupported activation: {config.hidden_act}. " "Only silu is supported for now." ) self.gate = MoEGate( config=config, quant_config=quant_config, prefix=add_prefix("gate", prefix), is_nextn=is_nextn, ) experts_type = get_moe_impl_class(quant_config) self.experts = experts_type( num_experts=config.n_routed_experts + get_global_server_args().ep_num_redundant_experts, top_k=config.num_experts_per_tok, hidden_size=config.hidden_size, intermediate_size=config.moe_intermediate_size, layer_id=self.layer_id, quant_config=quant_config, routed_scaling_factor=1.0, prefix=add_prefix("experts", prefix), ) self.topk = TopK( top_k=config.num_experts_per_tok, renormalize=config.norm_topk_prob, use_grouped_topk=True, num_expert_group=config.n_group, topk_group=config.topk_group, correction_bias=self.gate.e_score_correction_bias, quant_config=quant_config, routed_scaling_factor=1.0, apply_routed_scaling_factor_on_output=self.experts.should_fuse_routed_scaling_factor_in_topk, # Some Fp4 MoE backends require the output format to be bypassed but the MTP layers are unquantized # and requires the output format to be standard. We use quant_config to determine the output format. output_format=TopKOutputFormat.STANDARD if quant_config is None else None, ) # todo : implement tbo forward needed if get_moe_a2a_backend().is_deepep() or get_moe_a2a_backend().is_mooncake(): # TODO: we will support tp < ep in the future self.ep_size = get_moe_expert_parallel_world_size() self.num_experts = ( config.n_routed_experts + get_global_server_args().ep_num_redundant_experts ) self.renormalize = config.norm_topk_prob self.topk_group = config.topk_group self.num_expert_group = config.n_group self.correction_bias = ( self.gate.e_score_correction_bias.data if self.gate.e_score_correction_bias is not None else None ) self._enable_a2a_moe = ( get_moe_a2a_backend().is_deepep() or get_moe_a2a_backend().is_mooncake() ) def get_moe_weights(self): return [ x.data for name, x in self.experts.named_parameters() if name not in ["correction_bias"] ] def forward( self, hidden_states: torch.Tensor, forward_batch: Optional[ForwardBatch] = None, should_allreduce_fusion: bool = False, use_reduce_scatter: bool = False, ) -> torch.Tensor: if not self._enable_a2a_moe: return self.forward_normal( hidden_states, should_allreduce_fusion, use_reduce_scatter, ) else: return self.forward_deepep(hidden_states, forward_batch) def forward_normal( self, hidden_states: torch.Tensor, should_allreduce_fusion: bool = False, use_reduce_scatter: bool = False, ) -> torch.Tensor: if hidden_states.shape[0] > 0: # router_logits: (num_tokens, n_experts) router_logits = self.gate(hidden_states) topk_output = self.topk(hidden_states, router_logits) else: topk_output = self.topk.empty_topk_output(hidden_states.device) final_hidden_states = self.experts(hidden_states, topk_output) if ( self.tp_size > 1 and not should_allreduce_fusion and not use_reduce_scatter and not should_use_flashinfer_cutlass_moe_fp4_allgather() ): final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states) return final_hidden_states def forward_deepep( self, hidden_states: torch.Tensor, forward_batch: ForwardBatch ) -> torch.Tensor: if hidden_states.shape[0] > 0: # router_logits: (num_tokens, n_experts) router_logits = self.gate(hidden_states) topk_output = self.topk( hidden_states, router_logits, num_token_non_padded=forward_batch.num_token_non_padded, expert_location_dispatch_info=ExpertLocationDispatchInfo.init_new( layer_id=self.layer_id, ), ) else: topk_output = self.topk.empty_topk_output(hidden_states.device) final_hidden_states = self.experts( hidden_states=hidden_states, topk_output=topk_output ) return final_hidden_states def op_gate(self, state): if is_non_idle_and_non_empty( state.forward_batch.forward_mode, state.hidden_states_mlp_input ): # router_logits: (num_tokens, n_experts) state.router_logits = self.gate(state.hidden_states_mlp_input) else: state.router_logits = None def op_select_experts(self, state): router_logits = state.pop("router_logits") hidden_states = state.hidden_states_mlp_input if router_logits is not None: with get_global_expert_distribution_recorder().with_current_layer( self.layer_id ): state.topk_output = self.topk( hidden_states=hidden_states, router_logits=router_logits, num_token_non_padded=state.forward_batch.num_token_non_padded, expert_location_dispatch_info=ExpertLocationDispatchInfo.init_new( layer_id=self.layer_id, ), ) else: state.topk_output = self.topk.empty_topk_output(hidden_states.device) def op_dispatch_a(self, state): if self.ep_size > 1: self.experts.dispatcher.dispatch_a( hidden_states=state.pop("hidden_states_mlp_input"), topk_output=state.pop("topk_output"), tbo_subbatch_index=state.get("tbo_subbatch_index"), ) def op_dispatch_b(self, state): if self.ep_size > 1: with get_global_expert_distribution_recorder().with_current_layer( self.layer_id ): state.dispatch_output = self.experts.dispatcher.dispatch_b( tbo_subbatch_index=state.get("tbo_subbatch_index"), ) def op_experts(self, state): state.combine_input = self.experts.run_moe_core( dispatch_output=state.dispatch_output, ) def op_combine_a(self, state): if self.ep_size > 1: self.experts.dispatcher.combine_a( combine_input=state.pop("combine_input"), tbo_subbatch_index=state.get("tbo_subbatch_index"), ) state.pop("dispatch_output") def op_combine_b(self, state): if self.ep_size > 1: state.hidden_states_after_combine = self.experts.dispatcher.combine_b( tbo_subbatch_index=state.get("tbo_subbatch_index"), ) def op_output(self, state): state.hidden_states_mlp_output = state.pop("hidden_states_after_combine") class MiMoV2Attention(nn.Module): def __init__( self, hidden_size: int, num_heads: int, num_kv_heads: int, head_dim: Optional[int] = None, v_head_dim: Optional[int] = None, v_scale: Optional[float] = None, sliding_window_size: int = -1, # if is -1 ,normal attention,else ,window attention attention_bias: bool = False, attention_sink_bias: bool = False, layer_id: int = 0, rope_theta: float = 1000000, rope_scaling: Optional[Dict[str, Any]] = None, max_position_embeddings: int = 32768, quant_config: Optional[QuantizationConfig] = None, partial_rotary_factor: float = 1.0, prefix: str = "", ) -> None: super().__init__() self.hidden_size = hidden_size attn_tp_rank = get_attention_tp_rank() attn_tp_size = get_attention_tp_size() self.total_num_heads = num_heads assert self.total_num_heads % attn_tp_size == 0 self.num_heads = self.total_num_heads // attn_tp_size self.total_num_kv_heads = num_kv_heads if self.total_num_kv_heads >= attn_tp_size: # Number of KV heads is greater than TP size, so we partition # the KV heads across multiple tensor parallel GPUs. assert self.total_num_kv_heads % attn_tp_size == 0 else: # Number of KV heads is less than TP size, so we replicate # the KV heads across multiple tensor parallel GPUs. assert attn_tp_size % self.total_num_kv_heads == 0 self.num_kv_heads = max(1, self.total_num_kv_heads // attn_tp_size) self.head_dim = head_dim self.v_head_dim = v_head_dim if v_head_dim is not None else head_dim self.q_size = self.num_heads * self.head_dim self.k_size = self.num_kv_heads * self.head_dim self.v_size = self.num_kv_heads * self.v_head_dim self.v_scale = v_scale self.scaling = self.head_dim**-0.5 self.qkv_proj = QKVParallelLinear( hidden_size, self.head_dim, self.total_num_heads, self.total_num_kv_heads, v_head_size=self.v_head_dim, bias=attention_bias, quant_config=quant_config, tp_rank=attn_tp_rank, tp_size=attn_tp_size, prefix=add_prefix("qkv_proj", prefix), skip_block_quant_check=True, ) self.o_proj = RowParallelLinear( self.total_num_heads * self.v_head_dim, hidden_size, bias=False, quant_config=quant_config, tp_rank=attn_tp_rank, tp_size=attn_tp_size, reduce_results=False, prefix=add_prefix("o_proj", prefix), ) self.rotary_emb = get_rope( self.head_dim, rotary_dim=self.head_dim, max_position=max_position_embeddings, base=rope_theta, rope_scaling=rope_scaling, partial_rotary_factor=partial_rotary_factor, ) self.attn = RadixAttention( self.num_heads, self.head_dim, self.scaling, num_kv_heads=self.num_kv_heads, layer_id=layer_id, v_head_dim=self.v_head_dim, sliding_window_size=sliding_window_size, # if is -1 ,normal attention,else ,window attention quant_config=quant_config, prefix=add_prefix("attn", prefix), ) self.attention_sink_bias = ( torch.nn.Parameter(torch.empty(self.num_heads), requires_grad=False) if attention_sink_bias else None ) def op_prepare(self, state): state.attn_intermediate_state = self.forward_prepare( positions=state.positions, hidden_states=state.pop("hidden_states_after_comm_pre_attn"), forward_batch=state.forward_batch, ) def op_core(self, state): state.hidden_states_after_attn = self.forward_core( state.pop("attn_intermediate_state") ) def forward_prepare( self, positions: torch.Tensor, hidden_states: torch.Tensor, forward_batch: ForwardBatch, ): if hidden_states.shape[0] == 0: return hidden_states, forward_batch, None qkv, _ = self.qkv_proj(hidden_states) q, k, v = qkv.split([self.q_size, self.k_size, self.v_size], dim=-1) q, k = self.rotary_emb(positions, q, k) if self.v_scale is not None: v = v * self.v_scale inner_state = q, k, v, forward_batch return None, forward_batch, inner_state def forward_core(self, intermediate_state): hidden_states, forward_batch, inner_state = intermediate_state if inner_state is None: return hidden_states attn_output = self.attn( *inner_state, sinks=self.attention_sink_bias, ) output, _ = self.o_proj(attn_output) return output def forward( self, positions: torch.Tensor, hidden_states: torch.Tensor, forward_batch: ForwardBatch, ) -> torch.Tensor: qkv, _ = self.qkv_proj(hidden_states) q, k, v = qkv.split([self.q_size, self.k_size, self.v_size], dim=-1) # [t, h, dr] q, k = self.rotary_emb(positions, q, k) # [t, h, d] if self.v_scale is not None: v = v * self.v_scale attn_output = self.attn(q, k, v, forward_batch, sinks=self.attention_sink_bias) output, _ = self.o_proj(attn_output) return output class MiMoV2DecoderLayer(nn.Module): def __init__( self, config: MiMoV2FlashConfig, layer_id: int = 0, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ) -> None: super().__init__() self.config = config self.hidden_size = config.hidden_size self.layer_id = layer_id rope_theta = getattr(config, "rope_theta", 1000000) rope_scaling = getattr(config, "rope_scaling", None) max_position_embeddings = getattr(config, "max_position_embeddings", 32768) if self.is_swa_layer(): self.self_attn = MiMoV2Attention( hidden_size=self.hidden_size, num_heads=config.swa_num_attention_heads, num_kv_heads=config.swa_num_key_value_heads, head_dim=config.swa_head_dim, v_head_dim=getattr(config, "swa_v_head_dim", None), v_scale=getattr(config, "attention_value_scale", None), sliding_window_size=config.sliding_window_size, attention_bias=config.attention_bias, attention_sink_bias=getattr( config, "add_swa_attention_sink_bias", False ), layer_id=layer_id, rope_theta=getattr(config, "swa_rope_theta", rope_theta), rope_scaling=rope_scaling, max_position_embeddings=max_position_embeddings, quant_config=quant_config, partial_rotary_factor=getattr(config, "partial_rotary_factor", 1.0), prefix=add_prefix("self_attn", prefix), ) else: self.self_attn = MiMoV2Attention( hidden_size=self.hidden_size, num_heads=self.config.num_attention_heads, num_kv_heads=config.num_key_value_heads, head_dim=config.head_dim, v_head_dim=getattr(config, "v_head_dim", None), v_scale=getattr(config, "attention_value_scale", None), sliding_window_size=-1, # normal attention attention_bias=config.attention_bias, attention_sink_bias=getattr( config, "add_full_attention_sink_bias", False ), layer_id=layer_id, rope_theta=rope_theta, rope_scaling=rope_scaling, max_position_embeddings=max_position_embeddings, quant_config=quant_config, partial_rotary_factor=getattr(config, "partial_rotary_factor", 1.0), prefix=add_prefix("self_attn", prefix), ) self.is_layer_sparse = self.is_moe_layer(layer_id) is_previous_layer_sparse = self.is_moe_layer(layer_id - 1) is_next_layer_sparse = self.is_moe_layer(layer_id + 1) if self.is_layer_sparse: self.mlp = MiMoV2MoE( config=config, quant_config=quant_config, prefix=add_prefix("mlp", prefix), layer_id=layer_id, ) else: if enable_moe_dense_fully_dp(): mlp_tp_rank, mlp_tp_size = 0, 1 else: mlp_tp_rank, mlp_tp_size = None, None self.mlp = MiMoV2MLP( hidden_size=self.hidden_size, intermediate_size=config.intermediate_size, hidden_act=config.hidden_act, quant_config=quant_config, prefix=add_prefix("mlp", prefix), tp_rank=mlp_tp_rank, tp_size=mlp_tp_size, ) self.input_layernorm = RMSNorm(config.hidden_size, eps=config.layernorm_epsilon) self.post_attention_layernorm = RMSNorm( config.hidden_size, eps=config.layernorm_epsilon ) self.layer_scatter_modes = LayerScatterModes.init_new( layer_id=layer_id, num_layers=config.num_hidden_layers, is_layer_sparse=self.is_layer_sparse, is_previous_layer_sparse=is_previous_layer_sparse, is_next_layer_sparse=is_next_layer_sparse, ) self.layer_communicator = LayerCommunicator( layer_scatter_modes=self.layer_scatter_modes, input_layernorm=self.input_layernorm, post_attention_layernorm=self.post_attention_layernorm, allow_reduce_scatter=True, is_last_layer=(self.layer_id == self.config.num_hidden_layers - 1), ) def forward( self, positions: torch.Tensor, hidden_states: torch.Tensor, forward_batch: ForwardBatch, residual: Optional[torch.Tensor], ) -> Tuple[torch.Tensor, torch.Tensor]: # Self Attention hidden_states, residual = self.layer_communicator.prepare_attn( hidden_states, residual, forward_batch ) if hidden_states.shape[0] != 0: hidden_states = self.self_attn( positions=positions, hidden_states=hidden_states, forward_batch=forward_batch, ) hidden_states, residual = self.layer_communicator.prepare_mlp( hidden_states, residual, forward_batch ) should_allreduce_fusion = ( self.layer_communicator.should_fuse_mlp_allreduce_with_next_layer( forward_batch ) ) # For DP with padding, reduce scatter can be used instead of all-reduce. use_reduce_scatter = self.layer_communicator.should_use_reduce_scatter( forward_batch ) hidden_states = self.mlp( hidden_states, forward_batch, should_allreduce_fusion, use_reduce_scatter ) if should_allreduce_fusion: hidden_states._sglang_needs_allreduce_fusion = True else: hidden_states, residual = self.layer_communicator.postprocess_layer( hidden_states, residual, forward_batch ) return hidden_states, residual def is_moe_layer(self, layer_idx: int) -> bool: return ( hasattr(self.config, "moe_layer_freq") and 0 <= layer_idx < len(self.config.moe_layer_freq) and not isinstance(self.config.moe_layer_freq, int) and self.config.moe_layer_freq[layer_idx] ) def is_swa_layer(self) -> bool: return self.config.hybrid_layer_pattern[self.layer_id] == 1 def op_comm_prepare_attn( self, state, positions: torch.Tensor, hidden_states: torch.Tensor, forward_batch: ForwardBatch, residual: Optional[torch.Tensor], tbo_subbatch_index: Optional[int] = None, ): state.hidden_states_after_comm_pre_attn, state.residual_after_input_ln = ( self.layer_communicator.prepare_attn(hidden_states, residual, forward_batch) ) state.update( dict( forward_batch=forward_batch, positions=positions, tbo_subbatch_index=tbo_subbatch_index, ) ) def op_comm_prepare_mlp(self, state): state.hidden_states_mlp_input, state.residual_after_comm_pre_mlp = ( self.layer_communicator.prepare_mlp( state.pop("hidden_states_after_attn"), state.pop("residual_after_input_ln"), state.forward_batch, ) ) def op_mlp(self, state): hidden_states = state.pop("hidden_states_mlp_input") state.hidden_states_mlp_output = self.mlp(hidden_states, state.forward_batch) def op_comm_postprocess_layer(self, state): hidden_states, residual = self.layer_communicator.postprocess_layer( state.pop("hidden_states_mlp_output"), state.pop("residual_after_comm_pre_mlp"), state.forward_batch, ) output = dict( positions=state.positions, hidden_states=hidden_states, residual=residual, forward_batch=state.forward_batch, tbo_subbatch_index=state.tbo_subbatch_index, ) state.clear( expect_keys={ "positions", "forward_batch", "tbo_subbatch_index", } ) return output class MiMoV2Model(nn.Module): def __init__( self, config: MiMoV2FlashConfig, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", decoder_layer_type: type[nn.Module] = MiMoV2DecoderLayer, ) -> None: super().__init__() self.config = config self.padding_idx = config.pad_token_id self.vocab_size = config.vocab_size self.pp_group = get_pp_group() if self.pp_group.is_first_rank: self.embed_tokens = VocabParallelEmbedding( config.vocab_size, config.hidden_size, quant_config=quant_config, use_attn_tp_group=is_dp_attention_enabled(), prefix=add_prefix("embed_tokens", prefix), ) else: self.embed_tokens = PPMissingLayer() # Use the provided decoder layer type or default to MiMoV2DecoderLayer decoder_layer_type = decoder_layer_type or MiMoV2DecoderLayer self.layers, self.start_layer, self.end_layer = make_layers( config.num_hidden_layers, layer_fn=lambda idx, prefix: decoder_layer_type( layer_id=idx, config=config, quant_config=quant_config, prefix=prefix, ), pp_rank=self.pp_group.rank_in_group, pp_size=self.pp_group.world_size, prefix=add_prefix("layers", prefix), ) if self.pp_group.is_last_rank: self.norm = RMSNorm(config.hidden_size, eps=config.layernorm_epsilon) else: self.norm = PPMissingLayer(return_tuple=True) def get_input_embedding(self, input_ids: torch.Tensor) -> torch.Tensor: if hasattr(self.config, "scale_emb"): return self.get_input_embeddings()(input_ids) * self.config.scale_emb else: return self.get_input_embeddings()(input_ids) def get_input_embeddings(self) -> nn.Embedding: return self.embed_tokens def forward( self, input_ids: torch.Tensor, positions: torch.Tensor, forward_batch: ForwardBatch, input_embeds: torch.Tensor = None, pp_proxy_tensors: Optional[PPProxyTensors] = None, ) -> Union[torch.Tensor, PPProxyTensors]: if self.pp_group.is_first_rank: if input_embeds is None: hidden_states = self.embed_tokens(input_ids) else: hidden_states = input_embeds residual = None else: assert pp_proxy_tensors is not None hidden_states = pp_proxy_tensors["hidden_states"] residual = pp_proxy_tensors["residual"] if forward_batch.can_run_tbo: tbo_start_layer = self.start_layer tbo_end_layer = self.end_layer # skip first layer for TBO when starting from layer 0 if self.start_layer == 0: layer = self.layers[0] hidden_states, residual = layer( positions, hidden_states, forward_batch, residual ) tbo_start_layer = tbo_start_layer + 1 hidden_states, residual = model_forward_maybe_tbo( layers=self.layers[tbo_start_layer:tbo_end_layer], enable_tbo=True, input_data_scatter_mode=( ScatterMode.model_input_output() if tbo_start_layer == self.start_layer else self.layers[ tbo_start_layer - 1 ].layer_scatter_modes.layer_output_mode ), positions=positions, forward_batch=forward_batch, hidden_states=hidden_states, residual=residual, ) else: for i in range(self.start_layer, self.end_layer): layer = self.layers[i] hidden_states, residual = layer( positions, hidden_states, forward_batch, residual, ) hidden_states_before_norm = None if not self.pp_group.is_last_rank: return PPProxyTensors( { "hidden_states": hidden_states, "residual": residual, } ) else: if hidden_states.shape[0] > 0: if forward_batch.return_hidden_states_before_norm: hidden_states_before_norm = ( hidden_states if residual is None else hidden_states + residual ) if residual is None: hidden_states = self.norm(hidden_states) else: hidden_states, _ = self.norm(hidden_states, residual) return hidden_states, hidden_states_before_norm # If this function is called, it should always initialize KV cache scale # factors (or else raise an exception). Thus, handled exceptions should # make sure to leave KV cache scale factors in a known good (dummy) state def load_kv_cache_scales(self, quantization_param_path: str) -> None: attn_tp_rank = get_attention_tp_rank() attn_tp_size = get_attention_tp_size() for layer_idx, scaling_factor in kv_cache_scales_loader( quantization_param_path, attn_tp_rank, attn_tp_size, self.config.num_hidden_layers, self.config.__class__.model_type, ): if not isinstance(self.layers[layer_idx], nn.Identity): layer_self_attn = self.layers[layer_idx].self_attn if hasattr(layer_self_attn.attn, "k_scale"): layer_self_attn.attn.k_scale = scaling_factor layer_self_attn.attn.v_scale = scaling_factor else: raise RuntimeError( "Self attention has no KV cache scaling " "factor attribute!" ) class MiMoV2FlashForCausalLM(nn.Module): # BitandBytes specific attributes default_bitsandbytes_target_modules = [ ".gate_proj.", ".down_proj.", ".up_proj.", ".q_proj.", ".k_proj.", ".v_proj.", ".o_proj.", ] bitsandbytes_stacked_params_mapping = { # shard_name, weight_name, index "q_proj": ("qkv_proj", 0), "k_proj": ("qkv_proj", 1), "v_proj": ("qkv_proj", 2), "gate_proj": ("gate_up_proj", 0), "up_proj": ("gate_up_proj", 1), } def __init__( self, config: MiMoV2FlashConfig, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ) -> None: super().__init__() self.pp_group = get_pp_group() self.config = config self.quant_config = quant_config self.model = MiMoV2Model( config, quant_config=quant_config, prefix=add_prefix("model", prefix) ) if self.pp_group.is_last_rank: self.lm_head = ParallelLMHead( config.vocab_size, config.hidden_size, quant_config=quant_config, prefix=add_prefix("lm_head", prefix), use_attn_tp_group=get_global_server_args().enable_dp_lm_head, ) else: # ranks other than the last rank will have a placeholder layer self.lm_head = PPMissingLayer() self.logits_processor = LogitsProcessor(config) self._routed_experts_weights_of_layer = LazyValue( lambda: { layer_id: layer.mlp.get_moe_weights() for layer_id, layer in enumerate(self.model.layers) if isinstance(layer.mlp, MiMoV2MoE) } ) @property def routed_experts_weights_of_layer(self): return self._routed_experts_weights_of_layer.value def get_input_embedding(self, input_ids: torch.Tensor) -> torch.Tensor: return self.model.get_input_embedding(input_ids) def get_input_embeddings(self) -> nn.Embedding: return self.model.embed_tokens @torch.no_grad() def forward( self, input_ids: torch.Tensor, positions: torch.Tensor, forward_batch: ForwardBatch, input_embeds: torch.Tensor = None, pp_proxy_tensors: Optional[PPProxyTensors] = None, ) -> torch.Tensor: hidden_states, hidden_states_before_norm = self.model( input_ids, positions, forward_batch, input_embeds, pp_proxy_tensors=pp_proxy_tensors, ) if self.pp_group.is_last_rank: return self.logits_processor( input_ids, hidden_states, self.lm_head, forward_batch, hidden_states_before_norm=hidden_states_before_norm, ) else: return hidden_states @property def start_layer(self): return self.model.start_layer @property def end_layer(self): return self.model.end_layer def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]): stacked_params_mapping = [ # (param_name, shard_name, shard_id) ("qkv_proj", "q_proj", "q"), ("qkv_proj", "k_proj", "k"), ("qkv_proj", "v_proj", "v"), ("gate_up_proj", "gate_proj", 0), ("gate_up_proj", "up_proj", 1), ] # (param_name, weight_name, expert_id, shard_id) expert_params_mapping = DeepEPMoE.make_expert_params_mapping( ckpt_gate_proj_name="gate_proj", ckpt_down_proj_name="down_proj", ckpt_up_proj_name="up_proj", num_experts=self.config.n_routed_experts, ) params_dict = dict(self.named_parameters()) for name, loaded_weight in weights: layer_id = get_layer_id(name) if ( layer_id is not None and hasattr(self.model, "start_layer") and ( layer_id < self.model.start_layer or layer_id >= self.model.end_layer ) ): continue if "rotary_emb.inv_freq" in name or "projector" in name: continue if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name: # Models trained using ColossalAI may include these tensors in # the checkpoint. Skip them. continue if self.config.tie_word_embeddings and "lm_head.weight" in name: if self.pp_group.world_size > 1 and self.pp_group.is_last_rank: # Handle pp weight tying here # find the embed_tokens.weight in the weights embed_token_weights = next( filter(lambda x: x[0] == "model.embed_tokens.weight", weights) )[1] loaded_weight = embed_token_weights else: continue # TODO: skip mtp weights for now, need to implement mtp if "mtp" in name: continue for param_name, weight_name, shard_id in stacked_params_mapping: if weight_name not in name: continue if ("mlp.experts." in name) and name not in params_dict: continue name = name.replace(weight_name, param_name) # Skip loading extra bias for GPTQ models. if name.endswith(".bias") and name not in params_dict: continue param = params_dict[name] weight_loader = param.weight_loader weight_loader(param, loaded_weight, shard_id) break else: for mapping in expert_params_mapping: param_name, weight_name, expert_id, shard_id = mapping if weight_name not in name: continue name = name.replace(weight_name, param_name) param = params_dict[name] weight_loader = param.weight_loader weight_loader( param, loaded_weight, name, shard_id=shard_id, expert_id=expert_id, ) break else: # Skip loading extra bias for GPTQ models. if name.endswith(".bias") and name not in params_dict: continue if name in params_dict.keys(): param = params_dict[name] if "attention_sink_bias" in name: start = get_attention_tp_rank() * param.numel() param.data.copy_( loaded_weight[start : start + param.numel()] ) else: weight_loader = getattr( param, "weight_loader", default_weight_loader ) weight_loader(param, loaded_weight) else: logger.warning(f"Parameter {name} not found in params_dict") def get_embed_and_head(self): return self.model.embed_tokens.weight, self.lm_head.weight def set_embed_and_head(self, embed, head): del self.model.embed_tokens.weight del self.lm_head.weight self.model.embed_tokens.weight = embed self.lm_head.weight = head torch.cuda.empty_cache() torch.cuda.synchronize() def load_kv_cache_scales(self, quantization_param_path: str) -> None: self.model.load_kv_cache_scales(quantization_param_path) @classmethod def get_model_config_for_expert_location(cls, config): return ModelConfigForExpertLocation( num_layers=config.num_hidden_layers, num_logical_experts=getattr(config, "n_routed_experts", 1), num_groups=getattr(config, "n_group", None), ) EntryClass = MiMoV2FlashForCausalLM