sglang 0.5.3rc0__py3-none-any.whl → 0.5.3rc2__py3-none-any.whl

sglang/srt/layers/attention/mamba/mamba.py

@@ -1,6 +1,39 @@
-from typing import Callable, List, Tuple
+from typing import Callable, List, Optional, Tuple, Union
 
 import torch
+import torch.nn as nn
+
+from sglang.srt.configs.model_config import ModelConfig
+from sglang.srt.custom_op import CustomOp
+from sglang.srt.distributed import (
+    get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size,
+    tensor_model_parallel_all_gather,
+    tensor_model_parallel_all_reduce,
+)
+from sglang.srt.distributed.utils import divide
+from sglang.srt.layers.attention.fla.layernorm_gated import layernorm_fn
+from sglang.srt.layers.attention.mamba.causal_conv1d import (
+    causal_conv1d_fn,
+    causal_conv1d_update,
+)
+from sglang.srt.layers.attention.mamba.mamba_utils import MambaStateShapeCalculator
+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.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_loader.weight_utils import (
+    composed_weight_loader,
+    sharded_weight_loader,
+)
+from sglang.srt.utils import set_weight_attrs
 
 LoaderFunction = Callable[[torch.Tensor, torch.Tensor], None]
 
@@ -62,3 +95,535 @@ def mamba_v2_sharded_weight_loader(
             loaded_boundary += full_dim - extra
 
     return loader
+
+
+class Mixer2RMSNormGated(CustomOp):
+
+    def __init__(
+        self,
+        full_hidden_size: int,
+        full_n_groups: int,
+        use_rms_norm: bool = True,
+        eps: float = 1e-6,
+    ):
+        super().__init__()
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.tp_rank = get_tensor_model_parallel_rank()
+        self.full_hidden_size = full_hidden_size
+        self.group_size = full_hidden_size // full_n_groups
+        self.per_rank_hidden_size = full_hidden_size // self.tp_size
+        self.n_groups = full_hidden_size // self.group_size
+
+        self.variance_epsilon = eps
+        self.use_rms_norm = use_rms_norm
+        if self.use_rms_norm:
+            # Register norm weight only if we're actually applying RMSNorm
+            self.weight = nn.Parameter(torch.ones(self.per_rank_hidden_size))
+            set_weight_attrs(self.weight, {"weight_loader": sharded_weight_loader(0)})
+        else:
+            # Avoid checkpoint mismatch by skipping unused parameter
+            self.register_parameter("weight", None)
+        assert (
+            self.full_hidden_size % self.tp_size == 0
+        ), "Tensor parallel world size must divide hidden size."
+
+    def forward_native(
+        self,
+        x: torch.Tensor,
+        gate: torch.Tensor,
+    ):
+        # Three tensor-parallel cases:
+        #   1. n_groups is 1
+        #      In this case we parallelize along the reduction dim.
+        #      Each rank computes a local sum of squares followed by AllReduce
+        #   2. tp_size divides n_groups
+        #      Each rank only reduces within its local group(s).
+        #      No collective ops necessary.
+        #   3. The general case can be pretty complicated so we AllGather
+        #      the input and then redundantly compute the RMSNorm.
+        input_dtype = x.dtype
+        x = x * nn.functional.silu(gate.to(torch.float32))
+        if not self.use_rms_norm:
+            return x.to(input_dtype)
+
+        if self.n_groups == 1:
+            if self.tp_size > 1:
+                # Compute local sum and then reduce to obtain global sum
+                local_sums = x.pow(2).sum(dim=-1, keepdim=True)
+                global_sums = tensor_model_parallel_all_reduce(local_sums)
+                # Calculate the variance
+                count = self.tp_size * x.shape[-1]
+                variance = global_sums / count
+
+            else:
+                variance = x.pow(2).mean(-1, keepdim=True)
+            x = x * torch.rsqrt(variance + self.variance_epsilon)
+        else:
+            redundant_tp: bool = self.n_groups % self.tp_size != 0
+            if redundant_tp:
+                # To handle the general case, redundantly apply the variance
+                x = tensor_model_parallel_all_gather(x, -1)
+
+            *prefix_dims, hidden_dim = x.shape
+            group_count = hidden_dim // self.group_size
+            x_grouped = x.view(*prefix_dims, group_count, self.group_size)
+            variance = x_grouped.pow(2).mean(-1, keepdim=True)
+            x_grouped = x_grouped * torch.rsqrt(variance + self.variance_epsilon)
+            x = x_grouped.view(*prefix_dims, hidden_dim)
+
+            if redundant_tp:
+                start = self.per_rank_hidden_size * self.tp_rank
+                end = start + self.per_rank_hidden_size
+                x = x[..., start:end]
+
+        return self.weight * x.to(input_dtype)
+
+    def forward_cuda(
+        self,
+        x: torch.Tensor,
+        gate: torch.Tensor,
+    ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
+        input_dtype = x.dtype
+        if not self.use_rms_norm:
+            # Keep gate in float32 for numerical stability during silu
+            return x * nn.functional.silu(gate.to(torch.float32)).to(input_dtype)
+
+        if ((self.n_groups % self.tp_size) != 0) or self.n_groups != 1:
+            return self.forward_native(x, gate)
+
+        return layernorm_fn(
+            x,
+            self.weight.data,
+            bias=None,
+            z=gate,
+            eps=self.variance_epsilon,
+            norm_before_gate=False,
+        )
+
+
+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,
+        hidden_size: int,
+        ssm_state_size: int,
+        conv_kernel_size: int,
+        intermediate_size: int,
+        use_conv_bias: bool,
+        use_bias: bool,
+        chunk_size: int,
+        layer_id: int,
+        n_groups: int = 1,
+        num_heads: int = 128,
+        head_dim: int = 64,
+        rms_norm_eps: float = 1e-5,
+        activation: str = "silu",
+        use_rms_norm: bool = True,
+        model_config: Optional[ModelConfig] = None,
+        #  cache_config: Optional[CacheConfig] = None,
+        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()
+
+        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."
+        )
+
+        self.ssm_state_size = ssm_state_size
+        self.conv_kernel_size = conv_kernel_size
+        self.activation = activation
+        self.layer_id = layer_id
+
+        self.intermediate_size = intermediate_size
+        self.head_dim = head_dim
+        self.num_heads = num_heads
+        self.chunk_size = chunk_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 = MambaStateShapeCalculator.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 = intermediate_size + 2 * self.groups_ssm_state_size
+
+        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,
+            prefix=f"{prefix}.in_proj",
+        )
+        if n_groups % self.tp_size != 0:
+            # This is the n_groups == 1 case,
+            # where we need to duplicate groups if TP>1.
+
+            # - 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",
+            reduce_results=False,
+        )
+
+        self.norm = Mixer2RMSNormGated(
+            intermediate_size, n_groups, self.use_rms_norm, eps=rms_norm_eps
+        )
+
+        # The tuple is (conv_state, ssm_state)
+        self.kv_cache = (torch.tensor([]), torch.tensor([]))
+
+        self.model_config = model_config
+        self.prefix = prefix
+
+    def forward_native(
+        self,
+        hidden_states: torch.Tensor,
+        output: torch.Tensor,
+        mup_vector: Optional[torch.Tensor] = None,
+    ):
+        pass
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        output: torch.Tensor,
+        forward_batch: ForwardBatch,
+        mup_vector: Optional[torch.Tensor] = None,
+    ):
+        # attn_backend_list[-1] gives access to MambaAttnBackend
+        mamba_backend = forward_batch.attn_backend.attn_backend_list[-1]
+        attn_metadata = mamba_backend.forward_metadata
+        state_indices_tensor = attn_metadata.mamba_cache_indices
+        chunk_size = self.chunk_size
+
+        conv_state, ssm_state, *rest = mamba_backend.req_to_token_pool.get_mamba_params(
+            self.layer_id
+        )
+
+        assert (
+            ssm_state.size(1) == self.ssm_state_size
+        ), f"dstate must be {self.ssm_state_size}, got {ssm_state.size(1)}"
+
+        query_start_loc = attn_metadata.query_start_loc
+
+        chunk_size = self.chunk_size
+
+        # TODO: properly support this
+        prep_initial_states = False
+
+        # 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,
+        )
+
+        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,
+        )
+
+        # Process prefill requests
+        if forward_batch.forward_mode.is_extend():
+            # 2. Convolution sequence transformation
+            # - "cache_indices" updates the conv_state cache in positions
+            #   pointed to by "state_indices_tensor"
+            num_prefill_tokens = forward_batch.extend_num_tokens or 0
+            has_initial_states = forward_batch.extend_prefix_lens > 0
+            cache_indices = attn_metadata.mamba_cache_indices
+
+            x = hidden_states_B_C.transpose(
+                0, 1
+            )  # this is the form that causal-conv see
+            hidden_states_B_C = causal_conv1d_fn(
+                x,
+                conv_weights,
+                self.conv1d.bias,
+                activation=self.activation,
+                conv_states=conv_state,
+                has_initial_state=has_initial_states,
+                cache_indices=cache_indices,
+                query_start_loc=query_start_loc,
+                seq_lens_cpu=forward_batch.extend_seq_lens_cpu,
+            ).transpose(0, 1)
+
+            hidden_states, B, C = split_hidden_states_B_C_fn(hidden_states_B_C)
+
+            # 3. State Space Model sequence transformation
+            initial_states = None
+
+            if has_initial_states is not None and prep_initial_states:
+                initial_states = torch.where(
+                    has_initial_states[:, None, None, None],
+                    ssm_state[state_indices_tensor],
+                    0,
+                )
+
+            # NOTE: final output is an in-place update of out tensor
+            varlen_state = mamba_chunk_scan_combined(
+                hidden_states.view(
+                    1, num_prefill_tokens, self.num_heads // self.tp_size, self.head_dim
+                ),
+                dt.unsqueeze(0),
+                self.A,
+                B.view(1, num_prefill_tokens, self.n_groups // self.tp_size, -1),
+                C.view(1, num_prefill_tokens, self.n_groups // self.tp_size, -1),
+                chunk_size=chunk_size,
+                D=self.D,
+                z=None,
+                dt_bias=self.dt_bias,
+                cu_seqlens=query_start_loc,
+                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.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] = varlen_state.permute(0, 3, 2, 1)
+        elif forward_batch.forward_mode.is_decode():
+            num_decodes = len(query_start_loc) - 1
+            # 2. Convolution sequence transformation
+            hidden_states_B_C = causal_conv1d_update(
+                hidden_states_B_C,
+                conv_state,
+                conv_weights,
+                self.conv1d.bias,
+                self.activation,
+                conv_state_indices=state_indices_tensor,
+            )
+
+            hidden_states, B, C = split_hidden_states_B_C_fn(hidden_states_B_C)
+
+            # 3. State Space Model sequence transformation
+            n_groups = self.n_groups // self.tp_size
+            A = (
+                self.A[:, None, ...][:, :, None]
+                .expand(-1, self.head_dim, self.ssm_state_size)
+                .to(dtype=torch.float32)
+            )
+            dt = dt[:, :, None].expand(-1, -1, self.head_dim)
+            dt_bias = self.dt_bias[:, None, ...].expand(-1, self.head_dim)
+            D = self.D[:, None, ...].expand(-1, self.head_dim)
+            B = B.view(-1, n_groups, B.shape[1] // n_groups)
+            C = C.view(-1, n_groups, C.shape[1] // n_groups)
+            hidden_states = hidden_states.view(
+                -1, self.num_heads // self.tp_size, self.head_dim
+            )
+
+            # - the hidden is reshaped into (bs, num_heads, head_dim)
+            # - mamba_cache_params.ssm_state's slots will be selected
+            #   using state_indices_tensor_d
+            # NOTE: final output is an in-place update of out tensor
+            selective_state_update(
+                ssm_state.permute(0, 3, 2, 1),
+                hidden_states,
+                dt,
+                A,
+                B,
+                C,
+                D,
+                z=None,
+                dt_bias=dt_bias,
+                dt_softplus=True,
+                state_batch_indices=state_indices_tensor,
+                out=preallocated_ssm_out.view(num_decodes, -1, self.head_dim),
+            )
+        elif forward_batch.forward_mode.is_idle():
+            preallocated_ssm_out = preallocated_ssm_out
+
+        # 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)
+
+        # 5. Final linear projection
+        output[:], _ = self.out_proj(hidden_states)
+
+    @property
+    def mamba_type(self) -> str:
+        return "mamba2"

sglang/srt/layers/attention/mamba/mamba_utils.py

@@ -0,0 +1,81 @@
+# Adapted from: https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/mamba/mamba_utils.py
+from sglang.srt.distributed.utils import divide
+
+
+class MambaStateShapeCalculator:
+
+    @classmethod
+    def linear_attention_state_shape(
+        cls,
+        num_heads: int,
+        tp_size: int,
+        head_dim: int,
+    ) -> tuple[tuple[int, int, int], ...]:
+
+        state_shape = (num_heads // tp_size, head_dim, head_dim)
+        return (state_shape,)
+
+    @classmethod
+    def mamba1_state_shape(
+        cls,
+        tp_world_size: int,
+        intermediate_size: int,
+        state_size: int,
+        conv_kernel: int,
+    ) -> tuple[tuple[int, int], tuple[int, int]]:
+        conv_state_shape = (divide(intermediate_size, tp_world_size), conv_kernel - 1)
+
+        temporal_state_shape = (divide(intermediate_size, tp_world_size), state_size)
+
+        conv_state_shape = conv_state_shape[1], conv_state_shape[0]
+
+        return conv_state_shape, temporal_state_shape
+
+    @classmethod
+    def mamba2_state_shape(
+        cls,
+        tp_world_size: int,
+        intermediate_size: int,
+        n_groups: int,
+        num_heads: int,
+        head_dim: int,
+        state_size: int,
+        conv_kernel: int,
+    ) -> tuple[tuple[int, int], tuple[int, int, int]]:
+        # if n_groups is not divisible by world_size, need to extend the shards
+        # to ensure all groups needed by a head is sharded along with it
+        n_groups = n_groups + cls.extra_groups_for_head_shards(n_groups, tp_world_size)
+        # heads and n_groups are TP-ed
+        conv_dim = intermediate_size + 2 * n_groups * state_size
+
+        # contiguous along 'dim' axis
+        conv_state_shape = (conv_kernel - 1, divide(conv_dim, tp_world_size))
+
+        # These are not TP-ed as they depend on A, dt_bias, D
+        # - they are typically small
+        #   e.g., (h_heads, head_dim, state_size) = (128, 64, 128)
+        temporal_state_shape = (divide(num_heads, tp_world_size), head_dim, state_size)
+        return conv_state_shape, temporal_state_shape
+
+    @classmethod
+    def short_conv_state_shape(
+        cls,
+        tp_world_size: int,
+        intermediate_size: int,
+        conv_kernel: int,
+    ) -> tuple[tuple[int, int]]:
+        conv_dim = divide(intermediate_size, tp_world_size)
+        conv_state_shape = (conv_kernel - 1, conv_dim)
+        return (conv_state_shape,)
+
+    @classmethod
+    def extra_groups_for_head_shards(cls, ngroups: int, tp_size: int):
+        """Compute the increase in group numbers to account for
+        replication in order to accompany the head shards."""
+
+        # in the case ngoups % tp_size == 0, this will be zero
+        if ngroups % tp_size == 0:
+            return 0
+
+        # for n_groups == 1, this is exactly tp_size - n_groups
+        return tp_size - ngroups

sglang/srt/layers/attention/mamba/ops/__init__.py

@@ -0,0 +1,2 @@
+from .mamba_ssm import selective_state_update
+from .ssd_combined import mamba_chunk_scan_combined