from typing import Callable, List, Optional, Tuple import torch import torch.nn as nn from sglang.srt.configs.mamba_utils import ( Mamba2CacheParams, extra_groups_for_head_shards, ) from sglang.srt.distributed import ( divide, get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size, ) from sglang.srt.layers.attention.mamba.mamba2_metadata import Mamba2Metadata from sglang.srt.layers.attention.mamba.mixer2_rms_norm_gated import Mixer2RMSNormGated from sglang.srt.layers.attention.mamba.ops import ( mamba_chunk_scan_combined, selective_state_update, ) from sglang.srt.layers.linear import ( ColumnParallelLinear, MergedColumnParallelLinear, RowParallelLinear, ) from sglang.srt.layers.quantization.base_config import QuantizationConfig from sglang.srt.mem_cache.memory_pool import MambaPool from sglang.srt.model_loader.weight_utils import ( composed_weight_loader, sharded_weight_loader, ) from sglang.srt.utils import is_cpu, is_cuda, is_npu, set_weight_attrs if is_cuda(): from sglang.srt.layers.attention.mamba.causal_conv1d import ( causal_conv1d_fn, causal_conv1d_update, ) from sglang.srt.layers.attention.mamba.causal_conv1d_triton import ( causal_conv1d_fn as causal_conv1d_fn_triton, ) from sglang.srt.layers.attention.mamba.causal_conv1d_triton import ( causal_conv1d_update as causal_conv1d_update_triton, ) elif is_npu(): from sgl_kernel_npu.mamba.causal_conv1d import ( causal_conv1d_fn_npu as causal_conv1d_fn, ) from sgl_kernel_npu.mamba.causal_conv1d import ( causal_conv1d_update_npu as causal_conv1d_update, ) LoaderFunction = Callable[[torch.Tensor, torch.Tensor], None] def mamba_v2_sharded_weight_loader( shard_spec: List[Tuple[int, int, float]], tp_size: int, tp_rank: int, ) -> LoaderFunction: """Create a weight loader for mamba v2. This ensures that the projections are correctly sharded so that they can be split into x, B, C. It also ensures the the all the groups corresponding to a head shard is placed together with it. """ def loader(param: torch.Tensor, loaded_weight: torch.Tensor) -> None: # - track boundary of (sharded) param, and loaded_weight, respectively boundary, loaded_boundary = 0, 0 # Calculate padding size for CPU when TP odd size if is_cpu(): full_dim_sum = 0 full_dim_list = [] weight_full_dim_list = [] for full_dim, _, _ in shard_spec: full_dim_sum = full_dim_sum + full_dim full_dim_list.append(full_dim) for full_dim in full_dim_list: weight_full_dim_list.append( int(full_dim / full_dim_sum * loaded_weight.size(0)) ) # - iterate over the shard specs for full_dim, extra, duplicate_groups in shard_spec: # - full dim is the model dim (before TP). # - extra > 0, means there is expected overall increase # of dimensions. This is so because of replication. # - ratio is used map the tp_rank to the actual shard # rank. This is useful when there is replication of # groups to accompany head shards. # - size of the loaded shard shard_size = full_dim // tp_size # - compute the rank into the loaded shard. # - if there is replication, different TP shards will # take from the same rank. # NOTE: currently we only support duplication # in the case where num_groups == 1 rank = 0 if duplicate_groups else tp_rank # - leftmost boundary index into loaded weight. loaded_skip = rank * shard_size loaded_start_idx = loaded_boundary + loaded_skip # - take these many dims from the loaded weight. take = min(shard_size, full_dim - extra - loaded_skip) # CPU logic of padding size for qwen3-next # TODO : make this common for all mamba. if is_cpu() and loaded_weight.size(0) % tp_size != 0: import copy loaded_weight_ = copy.deepcopy(loaded_weight) q, k, v = torch.split( loaded_weight_, weight_full_dim_list, dim=0, ) pad_qk = torch.zeros( full_dim_list[0] - weight_full_dim_list[0], loaded_weight.size(1), loaded_weight.size(2), ).to(loaded_weight.dtype) pad_v = torch.zeros( full_dim_list[2] - weight_full_dim_list[2], loaded_weight.size(1), loaded_weight.size(2), ).to(loaded_weight.dtype) q = torch.cat((q, pad_qk), dim=0) k = torch.cat((k, pad_qk), dim=0) v = torch.cat((v, pad_v), dim=0) loaded_weight_qk = torch.cat((q, k), dim=0) loaded_weight = torch.cat((loaded_weight_qk, v), dim=0) # - always shard on dim 0 # - the ignore is for a mundane mypy error as it does not # seem to handle slices well. # https://github.com/python/mypy/issues/2410 param.data[ boundary : (boundary + take), ... # type: ignore[misc] ] = loaded_weight[ loaded_start_idx : (loaded_start_idx + take) # type: ignore[misc] ] # type: ignore[misc] # move indexing boundaries boundary += shard_size loaded_boundary += full_dim - extra return loader class MambaMixer2(torch.nn.Module): """ Compute ∆, A, B, C, and D the state space parameters and compute the `contextualized_states`. A, D are input independent (see Mamba paper [1] Section 3.5.2 "Interpretation of A" for why A isn't selective) ∆, B, C are input-dependent (this is a key difference between Mamba and the linear time invariant S4, and is why Mamba is called **selective** state spaces) """ def __init__( self, cache_params: Mamba2CacheParams, hidden_size: int, use_conv_bias: bool, use_bias: bool, n_groups: int = 1, rms_norm_eps: float = 1e-5, activation: str = "silu", use_rms_norm: bool = True, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ): super().__init__() # For TP, the sharding plan is as follows: # - for the conv modules, since # conv_dim = intermediate_size * 2 * n_groups * ssm_state_size, # we shard intermediate_size and n_groups # - since intermediate_size = n_heads * head_dim, sharding on # intermediate_size is achieved by sharding on n_heads. # - IF, world_size divides groups, then sharding # (n_groups / world_size, n_heads / world_size) # also maintains the invariant n_heads % n_groups == 0 # - HOWEVER IF, world_size DOES NOT divide groups, then we need # to allocate extra space in the shard, such that groups # may be replicated to follow the head shard. # - NOTE: currently for the world size DOES NOT divide groups # case, we only support the case when n_groups == 1 self.tp_size = get_tensor_model_parallel_world_size() self.tp_rank = get_tensor_model_parallel_rank() self.num_heads = num_heads = cache_params.shape.num_heads self.head_dim = cache_params.shape.head_dim assert ( num_heads % self.tp_size == 0 ), "Tensor parallel world size must divide num heads." assert (n_groups % self.tp_size) == 0 or n_groups == 1, ( "If tensor parallel world size does not divide num_groups, " "then num_groups must equal 1." ) assert ( (n_groups % self.tp_size == 0) or self.tp_size == 1 or quant_config is None ), ( "Tensor parallel currently supported for quantized models only " "if tensor parallel world size divides num groups." ) self.ssm_state_size = cache_params.shape.ssm_state_size self.activation = activation conv_kernel_size = cache_params.shape.conv_kernel self.intermediate_size = intermediate_size = ( cache_params.shape.intermediate_size ) self.n_groups = n_groups if n_groups % self.tp_size != 0: # - for TP we shard conv_dim by sharding on n_groups, # - but if n_groups cannot divide tp_size, we need to # extend some extra groups groups = extra_groups_for_head_shards(n_groups, self.tp_size) self.n_groups = n_groups + groups self.groups_ssm_state_size = self.n_groups * self.ssm_state_size self.conv_dim = cache_params.shape.conv_dim if n_groups % self.tp_size == 0: self.conv1d = MergedColumnParallelLinear( input_size=conv_kernel_size, output_sizes=[ intermediate_size, self.groups_ssm_state_size, self.groups_ssm_state_size, ], bias=use_conv_bias, quant_config=None, prefix=f"{prefix}.conv1d", ) self.in_proj = MergedColumnParallelLinear( input_size=hidden_size, output_sizes=[ intermediate_size, intermediate_size, self.groups_ssm_state_size, self.groups_ssm_state_size, self.num_heads, ], bias=use_bias, quant_config=quant_config, prefix=f"{prefix}.in_proj", ) else: # This is the n_groups == 1 case, # where we need to duplicate groups if TP>1. self.conv1d = ColumnParallelLinear( input_size=conv_kernel_size, output_size=self.conv_dim, bias=use_conv_bias, quant_config=None, prefix=f"{prefix}.conv1d", ) self.in_proj = ColumnParallelLinear( input_size=hidden_size, output_size=intermediate_size + self.conv_dim + self.num_heads, bias=use_bias, quant_config=quant_config, prefix=f"{prefix}.in_proj", ) # - because in_proj is a concatenation of 3 weights, we # need to interleave them before sharding # - use the custom weight loader mamba_v2_sharded_weight_loader # for conv1d.bias, covn1d.weight and in_proj.weight # - need to set these settings, to assign the groups # to the head shards group_shard_settings = ( self.groups_ssm_state_size, # expected model size (self.n_groups - n_groups) * self.ssm_state_size, # extra dims assigned n_groups == 1, # if there was only one group ) intermediate_settings = (intermediate_size, 0, False) head_settings = (self.num_heads, 0, False) # - the weight already has a "weight_loader" attribute # which set_weight_attrs will raise if we do not # delete before trying to override it # - ditto for the other two weights below delattr(self.conv1d.bias, "weight_loader") set_weight_attrs( self.conv1d.bias, { "weight_loader": mamba_v2_sharded_weight_loader( [ intermediate_settings, group_shard_settings, group_shard_settings, ], self.tp_size, self.tp_rank, ) }, ) delattr(self.conv1d.weight, "weight_loader") set_weight_attrs( self.conv1d.weight, { "weight_loader": mamba_v2_sharded_weight_loader( [ intermediate_settings, group_shard_settings, group_shard_settings, ], self.tp_size, self.tp_rank, ) }, ) if quant_config is None: # - quant layers do not have a weight loader delattr(self.in_proj.weight, "weight_loader") set_weight_attrs( self.in_proj.weight, { "weight_loader": mamba_v2_sharded_weight_loader( [ intermediate_settings, # for gate intermediate_settings, group_shard_settings, group_shard_settings, head_settings, # for dt ], self.tp_size, self.tp_rank, ) }, ) # unsqueeze to fit conv1d weights shape into the linear weights shape. # Can't do this in `weight_loader` since it already exists in # `ColumnParallelLinear` and `MergedColumnParallelLinear`, # and `set_weight_attrs` doesn't allow to override it self.conv1d.weight.data = self.conv1d.weight.data.unsqueeze(1) # - these are TPed by heads to reduce the size of the # temporal shape self.A = nn.Parameter( torch.empty( divide(num_heads, self.tp_size), dtype=torch.float32, ) ) self.D = nn.Parameter(torch.ones(num_heads // self.tp_size)) self.dt_bias = nn.Parameter(torch.ones(num_heads // self.tp_size)) self.use_rms_norm = use_rms_norm set_weight_attrs(self.D, {"weight_loader": sharded_weight_loader(0)}) a_weight_loader = composed_weight_loader( sharded_weight_loader(0), lambda x: -torch.exp(x.float()) ) set_weight_attrs(self.A, {"weight_loader": a_weight_loader}) set_weight_attrs(self.dt_bias, {"weight_loader": sharded_weight_loader(0)}) self.out_proj = RowParallelLinear( intermediate_size, hidden_size, bias=use_bias, input_is_parallel=True, quant_config=quant_config, prefix=f"{prefix}.out_proj", ) self.norm = Mixer2RMSNormGated( intermediate_size, n_groups, self.use_rms_norm, eps=rms_norm_eps ) self.prefix = prefix def forward( self, *, hidden_states: torch.Tensor, output: torch.Tensor, layer_cache: MambaPool.State, metadata: Mamba2Metadata, mup_vector: Optional[torch.Tensor] = None, use_triton_causal_conv: bool = False, ): # metadata contains metadata necessary for the mamba2 triton # kernels to operate in continuous batching and in chunked prefill # modes; they are computed at top-level model forward since they # stay the same and reused for all mamba layers in the same iteration state_indices_tensor = metadata.mamba_cache_indices conv_state = layer_cache.conv[0] ssm_state = layer_cache.temporal query_start_loc = metadata.query_start_loc # 1. Gated MLP's linear projection projected_states, _ = self.in_proj(hidden_states) if mup_vector is not None: projected_states = projected_states * mup_vector gate, hidden_states_B_C, dt = torch.split( projected_states, [ self.intermediate_size // self.tp_size, self.conv_dim // self.tp_size, self.num_heads // self.tp_size, ], dim=-1, ) conv_weights = self.conv1d.weight.view( self.conv1d.weight.size(0), self.conv1d.weight.size(2) ) # - get hidden_states, B and C after depthwise convolution. split_hidden_states_B_C_fn = lambda hidden_states_B_C: torch.split( hidden_states_B_C, [ self.intermediate_size // self.tp_size, self.groups_ssm_state_size // self.tp_size, self.groups_ssm_state_size // self.tp_size, ], dim=-1, ) num_prefills = metadata.num_prefills # request count num_decodes = metadata.num_decodes # token count (=request) num_decode_tokens = ( num_decodes * metadata.draft_token_num if metadata.is_target_verify else num_decodes ) num_prefill_tokens = metadata.num_prefill_tokens # token count has_prefill = num_prefills > 0 has_decode = num_decodes > 0 num_actual_tokens = num_prefill_tokens + num_decode_tokens assert num_actual_tokens == projected_states.shape[0] # NOTE: V0 put prefill before decode # Separate prefill and decode by splitting varlen input # Split along token dimension hidden_states_B_C_p, hidden_states_B_C_d = torch.split( hidden_states_B_C, [num_prefill_tokens, num_decode_tokens], dim=0, ) dt_p, dt_d = torch.split( dt, [num_prefill_tokens, num_decode_tokens], dim=0, ) # Split along batch dimension state_indices_tensor_p, state_indices_tensor_d = torch.split( state_indices_tensor, [num_prefills, num_decodes], dim=0, ) query_start_loc_p = query_start_loc[: num_prefills + 1] if has_prefill else None # Preallocate output tensor to avoid memcpy cost for merging prefill # and decode outputs preallocated_ssm_out = torch.empty( [ projected_states.shape[0], (self.num_heads * self.head_dim) // self.tp_size, ], dtype=hidden_states.dtype, device=hidden_states.device, ) preallocated_ssm_out_p, preallocated_ssm_out_d = torch.split( preallocated_ssm_out, [num_prefill_tokens, num_decode_tokens], dim=0, ) # Process prefill requests if has_prefill: mixed_metadata = metadata.mixed_metadata assert mixed_metadata is not None # 2. Convolution sequence transformation # - "cache_indices" updates the conv_state cache in positions # pointed to by "state_indices_tensor" has_initial_states_p = mixed_metadata.has_initial_states prep_initial_states = mixed_metadata.prep_initial_states cache_indices = state_indices_tensor_p x = hidden_states_B_C_p.transpose( 0, 1 ) # this is the form that causal-conv see ccfn = ( causal_conv1d_fn if not use_triton_causal_conv else causal_conv1d_fn_triton ) hidden_states_B_C_p = ccfn( x, conv_weights, self.conv1d.bias, activation=self.activation, conv_states=conv_state, has_initial_state=has_initial_states_p, cache_indices=cache_indices, query_start_loc=query_start_loc_p, seq_lens_cpu=mixed_metadata.extend_seq_lens_cpu, ).transpose(0, 1)[:num_prefill_tokens] hidden_states_p, B_p, C_p = split_hidden_states_B_C_fn(hidden_states_B_C_p) # 3. State Space Model sequence transformation initial_states = None if has_initial_states_p is not None and prep_initial_states: initial_states = torch.where( has_initial_states_p[:, None, None, None], ssm_state[state_indices_tensor_p], 0, ) # NOTE: final output is an in-place update of out tensor varlen_state = mamba_chunk_scan_combined( hidden_states_p.view( 1, num_prefill_tokens, self.num_heads // self.tp_size, self.head_dim ), dt_p.unsqueeze(0), self.A, B_p.view(1, num_prefill_tokens, self.n_groups // self.tp_size, -1), C_p.view(1, num_prefill_tokens, self.n_groups // self.tp_size, -1), chunk_size=mixed_metadata.chunk_size, D=self.D, z=None, dt_bias=self.dt_bias, seq_idx=mixed_metadata.seq_idx, chunk_indices=mixed_metadata.chunk_indices, chunk_offsets=mixed_metadata.chunk_offsets, cu_seqlens=query_start_loc_p, initial_states=initial_states, return_varlen_states=True, return_final_states=False, dt_softplus=True, dt_limit=(0.0, float("inf")), out=preallocated_ssm_out_p.view( 1, num_prefill_tokens, -1, self.head_dim ), state_dtype=ssm_state.dtype, ) # update ssm states # - varlen state is a (num_prefills, nheads, headdim, dstate) tensor ssm_state[state_indices_tensor_p] = varlen_state # Process decode requests if has_decode: is_target_verify = metadata.is_target_verify # 2. Convolution sequence transformation if is_target_verify: assert ( use_triton_causal_conv ), "Speculative decoding requires use_triton_causal_conv=True for intermediate state support" assert isinstance( layer_cache, MambaPool.SpeculativeState ), "layer_cache must be SpeculativeState for speculative decoding" draft_token_num = metadata.draft_token_num self.intermediate_state_indices = torch.arange( num_decodes, dtype=torch.int32, device=state_indices_tensor_d.device ) # Reshape for batch processing hidden_states_B_C_d_reshaped = hidden_states_B_C_d.view( num_decodes, draft_token_num, -1 ).transpose(1, 2) hidden_states_B_C_d_processed = causal_conv1d_update_triton( hidden_states_B_C_d_reshaped, conv_state, conv_weights, self.conv1d.bias, self.activation, conv_state_indices=state_indices_tensor_d[:num_decodes], intermediate_conv_window=layer_cache.intermediate_conv_window[0], intermediate_state_indices=self.intermediate_state_indices, retrieve_next_token=metadata.retrieve_next_token, retrieve_next_sibling=metadata.retrieve_next_sibling, retrieve_parent_token=metadata.retrieve_parent_token, ) hidden_states_B_C_d = hidden_states_B_C_d_processed.transpose( 1, 2 ).view(num_decode_tokens, -1) else: ccu = ( causal_conv1d_update if not use_triton_causal_conv else causal_conv1d_update_triton ) hidden_states_B_C_d = ccu( hidden_states_B_C_d, conv_state, conv_weights, self.conv1d.bias, self.activation, conv_state_indices=state_indices_tensor_d, ) hidden_states_d, B_d, C_d = split_hidden_states_B_C_fn(hidden_states_B_C_d) # 3. State Space Model sequence transformation n_groups = self.n_groups // self.tp_size A_d = ( self.A[:, None, ...][:, :, None] .expand(-1, self.head_dim, self.ssm_state_size) .to(dtype=torch.float32) ) dt_d = dt_d[:, :, None].expand(-1, -1, self.head_dim) dt_bias = self.dt_bias[:, None, ...].expand(-1, self.head_dim) D_d = self.D[:, None, ...].expand(-1, self.head_dim) B_d = B_d.view(-1, n_groups, B_d.shape[1] // n_groups) C_d = C_d.view(-1, n_groups, C_d.shape[1] // n_groups) hidden_states_d = hidden_states_d.view( -1, self.num_heads // self.tp_size, self.head_dim ) if is_target_verify: selective_state_update( ssm_state, hidden_states_d.view( num_decodes, draft_token_num, self.num_heads // self.tp_size, self.head_dim, ), dt_d.view( num_decodes, draft_token_num, self.num_heads // self.tp_size, self.head_dim, ), A_d, B_d.view(num_decodes, draft_token_num, n_groups, -1), C_d.view(num_decodes, draft_token_num, n_groups, -1), D_d, z=None, dt_bias=dt_bias, dt_softplus=True, state_batch_indices=state_indices_tensor_d[:num_decodes], out=preallocated_ssm_out_d.view( num_decodes, draft_token_num, self.num_heads // self.tp_size, self.head_dim, ), disable_state_update=True, intermediate_states_buffer=layer_cache.intermediate_ssm, cache_steps=draft_token_num, retrieve_parent_token=metadata.retrieve_parent_token, intermediate_state_indices=self.intermediate_state_indices, ) else: selective_state_update( ssm_state, hidden_states_d, dt_d, A_d, B_d, C_d, D_d, z=None, dt_bias=dt_bias, dt_softplus=True, state_batch_indices=state_indices_tensor_d, out=preallocated_ssm_out_d.view(num_decodes, -1, self.head_dim), ) # 4. gated MLP # GatedRMSNorm internally applying SiLU to the gate # SiLU is applied internally before normalization, unlike standard # norm usage hidden_states = self.norm(preallocated_ssm_out, gate[:num_actual_tokens]) # 5. Final linear projection output[:num_actual_tokens], _ = self.out_proj(hidden_states) @property def mamba_type(self) -> str: return "mamba2"