vllm-cpu 0.12.0__cp313-cp313-manylinux_2_17_aarch64.whl

vllm/model_executor/layers/mamba/ops/ssd_combined.py

@@ -0,0 +1,230 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Copyright (c) 2024, Tri Dao, Albert Gu.
+# Adapted from https://github.com/state-spaces/mamba/blob/v2.2.4/mamba_ssm/ops/triton/ssd_combined.py
+
+# ruff: noqa: E501
+
+import torch
+from einops import rearrange
+from packaging import version
+
+from vllm.triton_utils import triton
+
+from .ssd_bmm import _bmm_chunk_fwd
+from .ssd_chunk_scan import _chunk_scan_fwd
+from .ssd_chunk_state import _chunk_cumsum_fwd, _chunk_state_fwd
+from .ssd_state_passing import _state_passing_fwd
+
+TRITON_22 = version.parse(triton.__version__) >= version.parse("2.2.0")
+
+
+def is_int_pow_2(n):
+    return isinstance(n, int) and n > 0 and (n & (n - 1)) == 0
+
+
+def _mamba_chunk_scan_combined_fwd(
+    x,
+    dt,
+    A,
+    B,
+    C,
+    chunk_size,
+    out,
+    D=None,
+    z=None,
+    dt_bias=None,
+    initial_states=None,
+    return_intermediate_states=False,
+    seq_idx=None,
+    cu_seqlens=None,
+    cu_chunk_seqlens=None,
+    last_chunk_indices=None,
+    dt_softplus=False,
+    dt_limit=(0.0, float("inf")),
+    state_dtype=None,
+):
+    assert is_int_pow_2(chunk_size), "chunk_size must be integer power of 2"
+    seqlen, nheads, headdim = x.shape
+    _, ngroups, dstate = B.shape
+    assert nheads % ngroups == 0
+    assert B.shape == (seqlen, ngroups, dstate)
+    assert dt.shape == (seqlen, nheads)
+    assert A.shape == (nheads,)
+    assert C.shape == B.shape
+    if z is not None:
+        assert z.shape == x.shape
+    if D is not None:
+        assert D.shape == (nheads, headdim) or D.shape == (nheads,)
+    if seq_idx is not None:
+        assert seq_idx.shape == (cu_chunk_seqlens.shape[0] - 1,)
+    if B.stride(-1) != 1:
+        B = B.contiguous()
+    if C.stride(-1) != 1:
+        C = C.contiguous()
+    if (
+        x.stride(-1) != 1 and x.stride(0) != 1
+    ):  # Either M or K dimension should be contiguous
+        x = x.contiguous()
+    if (
+        z is not None and z.stride(-1) != 1 and z.stride(0) != 1
+    ):  # Either M or K dimension should be contiguous
+        z = z.contiguous()
+    if D is not None and D.stride(-1) != 1:
+        D = D.contiguous()
+    assert cu_seqlens is not None, "Assuming varlen input - must supply cu_seqlens"
+
+    if initial_states is not None:
+        assert initial_states.shape == (len(cu_seqlens) - 1, nheads, headdim, dstate)
+
+    # This function executes 5 sub-functions for computing mamba
+    # - a good resource is the blog https://goombalab.github.io/blog/2024/mamba2-part3-algorithm/
+    #   which has a minimal implementation to understand the below operations
+    # - as explained by the blog, mamba is a special case of causal attention
+    # - the idea is to chunk the attention matrix and compute each
+    #   submatrix separately using different optimizations.
+    # - see the blog and paper for a visualization of the submatrices
+    #   which we refer to in the comments below
+
+    # 1. Compute chunked cumsum of A * dt
+    # - here dt may go through a softplus activation
+    dA_cumsum, dt = _chunk_cumsum_fwd(
+        dt,
+        A,
+        chunk_size,
+        cu_chunk_seqlens,
+        dt_bias=dt_bias,
+        dt_softplus=dt_softplus,
+        dt_limit=dt_limit,
+    )
+
+    # 2. Compute the state for each intra-chunk
+    # (right term of low-rank factorization of off-diagonal blocks; B terms)
+    states = _chunk_state_fwd(
+        B, x, dt, dA_cumsum, cu_chunk_seqlens, states_in_fp32=True
+    )
+
+    # 3. Compute the inter-chunk SSM recurrence; produces correct SSM states at chunk boundaries
+    # (middle term of factorization of off-diag blocks; A terms)
+    # - for handling chunked prefill, this requires i) initial_states and
+    #   ii) seq_idx to be all specified.
+    # - When a new seq_idx is detected, we will stop passing the prev_state
+    #   and switch accordingly to the init_state corresponding to the new seq_idx.
+    states = _state_passing_fwd(
+        rearrange(states, "... p n -> ... (p n)"),
+        dA_cumsum,  # (nheads, nchunks, chunk_size)
+        cu_chunk_seqlens,
+        initial_states=rearrange(initial_states, "... p n -> ... (p n)")
+        if initial_states is not None
+        else None,  # (batch, nheads, headdim*dstate)
+        seq_idx=seq_idx,
+        out_dtype=state_dtype if state_dtype is not None else C.dtype,
+    )
+    states = rearrange(states, "... (p n) -> ... p n", n=dstate)
+
+    # 4. Compute batched matrix multiply for C_j^T B_i terms
+    CB = _bmm_chunk_fwd(C, B, chunk_size, cu_chunk_seqlens, output_dtype=torch.float32)
+
+    # 5. Scan and compute the diagonal blocks, taking into
+    #    account past causal states.
+    # - if initial states are provided, then states information will be
+    #   augmented with initial_states.
+    # - to do this properly, we need to account for example changes in
+    #   the continuous batch, therefore we introduce pseudo chunks, which is
+    #   a chunk that is split up each time an example changes.
+    # - in each (pseudo) chunk, we detect if the previous (pseudo) chunk had
+    #   a seq_idx change, in which case we take states information from
+    #   init_states.
+    _chunk_scan_fwd(
+        CB,
+        x,
+        dt,
+        dA_cumsum,
+        C,
+        states,
+        cu_chunk_seqlens,
+        out,  # in-place update
+        seq_idx,
+        D=D,
+        z=z,
+        initial_states=initial_states,
+    )
+
+    if return_intermediate_states:
+        return states
+    else:
+        return states[last_chunk_indices]
+
+
+def mamba_chunk_scan_combined_varlen(
+    x,
+    dt,
+    A,
+    B,
+    C,
+    chunk_size,
+    cu_seqlens,
+    cu_chunk_seqlens,
+    last_chunk_indices,
+    seq_idx,
+    out,
+    D=None,
+    z=None,
+    dt_bias=None,
+    initial_states=None,
+    dt_softplus=False,
+    dt_limit=(0.0, float("inf")),
+    return_intermediate_states=False,
+    state_dtype=None,
+):
+    """
+    Argument:
+        x: (seqlen, nheads, headdim)
+        dt: (seqlen, nheads)
+        A: (nheads)
+        B: (seqlen, ngroups, dstate)
+        C: (seqlen, ngroups, dstate)
+        chunk_size: int
+        cu_seqlens: (batch + 1,)
+        cu_chunk_seqlens: (nchunks + 1,)
+        last_chunk_indices: (batch,)
+        seq_idx: (nchunks,)
+        out: (seqlen, nheads, headdim) preallocated output tensor
+        D: (nheads, headdim) or (nheads,)
+        z: (seqlen, nheads, headdim)
+        dt_bias: (nheads,)
+        initial_states: (batch, nheads, headdim, dstate)
+        dt_softplus: Whether to apply softplus to dt
+        out: (seqlen, nheads, headdim) preallocated output tensor
+        state_dtype: The data type of the ssm state
+    Return:
+        varlen_states: (batch, nheads, headdim, dstate)
+    """
+
+    assert cu_seqlens is not None, "cu_seqlens must be provided assuming varlen input"
+    assert seq_idx is not None
+
+    varlen_states = _mamba_chunk_scan_combined_fwd(
+        x,
+        dt,
+        A,
+        B,
+        C,
+        chunk_size,
+        out,
+        D=D,
+        z=z,
+        dt_bias=dt_bias,
+        initial_states=initial_states,
+        return_intermediate_states=return_intermediate_states,
+        seq_idx=seq_idx,
+        cu_seqlens=cu_seqlens,
+        cu_chunk_seqlens=cu_chunk_seqlens,
+        last_chunk_indices=last_chunk_indices,
+        dt_softplus=dt_softplus,
+        dt_limit=dt_limit,
+        state_dtype=state_dtype,
+    )
+
+    return varlen_states

vllm/model_executor/layers/mamba/ops/ssd_state_passing.py

@@ -0,0 +1,157 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Copyright (c) 2024, Tri Dao, Albert Gu.
+# Adapted from https://github.com/state-spaces/mamba/blob/v2.2.4/mamba_ssm/ops/triton/ssd_state_passing.py
+
+# ruff: noqa: E501
+
+import torch
+
+from vllm.triton_utils import tl, triton
+
+
+@triton.autotune(
+    configs=[
+        triton.Config({"BLOCK_SIZE": 64}),
+        triton.Config({"BLOCK_SIZE": 128}),
+        triton.Config({"BLOCK_SIZE": 256}),
+        triton.Config({"BLOCK_SIZE": 512}),
+        triton.Config({"BLOCK_SIZE": 1024}),
+        triton.Config({"BLOCK_SIZE": 2048}),
+    ],
+    key=["dim"],
+)
+@triton.jit
+def _state_passing_fwd_kernel(
+    # Pointers to matrices
+    states_ptr,
+    out_ptr,
+    dA_cs_ptr,
+    initstates_ptr,
+    seq_idx_ptr,
+    cu_chunk_seqlens_ptr,
+    # Matrix dimensions
+    dim: tl.constexpr,
+    nchunks,
+    seqlen,
+    chunk_size: tl.constexpr,
+    # Strides
+    stride_states_chunk: tl.int64,
+    stride_states_head: tl.int64,
+    stride_states_dim: tl.constexpr,
+    stride_out_chunk: tl.int64,
+    stride_out_head: tl.int64,
+    stride_out_dim: tl.constexpr,
+    stride_dA_cs_head: tl.int64,
+    stride_dA_cs_chunk: tl.int64,
+    stride_dA_cs_csize: tl.constexpr,
+    stride_initstates_batch: tl.int64,
+    stride_initstates_head: tl.int64,
+    stride_initstates_dim: tl.constexpr,
+    stride_seq_idx_chunk: tl.constexpr,
+    # Meta-parameters
+    HAS_INITSTATES: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+):
+    pid_h = tl.program_id(axis=1)
+    pid_m = tl.program_id(axis=0)
+
+    states_ptr += pid_h * stride_states_head
+    dA_cs_ptr += pid_h * stride_dA_cs_head + (chunk_size - 1) * stride_dA_cs_csize
+    out_ptr += pid_h * stride_out_head
+
+    offs_m = pid_m * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
+    states_ptrs = states_ptr + offs_m * stride_states_dim
+    out_ptrs = out_ptr + offs_m * stride_out_dim
+
+    if HAS_INITSTATES:
+        initstates_ptrs = (
+            initstates_ptr
+            + pid_h * stride_initstates_head
+            + offs_m * stride_initstates_dim
+        )
+
+        states = tl.load(initstates_ptrs, mask=offs_m < dim, other=0.0).to(tl.float32)
+    else:
+        states = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
+
+    prev_seq_idx = 0
+    for c in range(nchunks):
+        new_states = tl.load(states_ptrs, mask=offs_m < dim, other=0.0).to(tl.float32)
+        dA_cs = tl.load(dA_cs_ptr).to(tl.float32)
+        seq_idx = tl.load(seq_idx_ptr + c * stride_seq_idx_chunk)
+        # we have started a new sequence
+        if prev_seq_idx != seq_idx:
+            if HAS_INITSTATES:
+                initstates_ptrs = (
+                    initstates_ptr
+                    + seq_idx * stride_initstates_batch
+                    + pid_h * stride_initstates_head
+                    + offs_m * stride_initstates_dim
+                )
+                states = tl.load(initstates_ptrs, mask=offs_m < dim, other=0.0).to(
+                    tl.float32
+                )
+            else:
+                states = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
+
+        prev_seq_idx = seq_idx
+        states = tl.exp(dA_cs) * states + new_states
+        tl.store(out_ptrs, states, mask=offs_m < dim)
+
+        states_ptrs += stride_states_chunk
+        dA_cs_ptr += stride_dA_cs_chunk
+        out_ptrs += stride_out_chunk
+
+
+def _state_passing_fwd(
+    states,
+    dA_cumsum,
+    cu_chunk_seqlens,
+    seq_idx,
+    initial_states=None,
+    out_dtype=None,
+):
+    nchunks, nheads, dim = states.shape
+    chunk_size = dA_cumsum.shape[-1]
+    assert dA_cumsum.shape == (nheads, nchunks, chunk_size)
+    seqlen = seq_idx.shape[-1]
+    out_dtype = states.dtype if out_dtype is None else out_dtype
+    out = torch.empty((nchunks, nheads, dim), device=states.device, dtype=out_dtype)
+
+    initial_states_strides = (
+        (initial_states.stride(0), initial_states.stride(1), initial_states.stride(2))
+        if initial_states is not None
+        else (0, 0, 0)
+    )
+
+    grid = lambda META: (triton.cdiv(dim, META["BLOCK_SIZE"]), nheads)
+    with torch.cuda.device(states.device.index):
+        _state_passing_fwd_kernel[grid](
+            states_ptr=states,
+            out_ptr=out,
+            dA_cs_ptr=dA_cumsum,
+            initstates_ptr=initial_states,
+            seq_idx_ptr=seq_idx,
+            cu_chunk_seqlens_ptr=cu_chunk_seqlens,
+            dim=dim,
+            nchunks=nchunks,
+            seqlen=seqlen if seq_idx is not None else 0,
+            chunk_size=chunk_size if seq_idx is not None else 0,
+            stride_states_chunk=states.stride(0),
+            stride_states_head=states.stride(1),
+            stride_states_dim=states.stride(2),
+            stride_out_chunk=out.stride(0),
+            stride_out_head=out.stride(1),
+            stride_out_dim=out.stride(2),
+            stride_dA_cs_head=dA_cumsum.stride(0),
+            stride_dA_cs_chunk=dA_cumsum.stride(1),
+            stride_dA_cs_csize=dA_cumsum.stride(2),
+            stride_initstates_batch=initial_states_strides[0],
+            stride_initstates_head=initial_states_strides[1],
+            stride_initstates_dim=initial_states_strides[2],
+            stride_seq_idx_chunk=seq_idx.stride(0),
+            HAS_INITSTATES=initial_states is not None,
+        )
+    return out

vllm/model_executor/layers/mamba/short_conv.py

@@ -0,0 +1,255 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+
+import torch
+
+from vllm.attention.backends.abstract import AttentionMetadata
+from vllm.config import CacheConfig, ModelConfig, get_current_vllm_config
+from vllm.distributed import get_tensor_model_parallel_world_size
+from vllm.forward_context import ForwardContext, get_forward_context
+from vllm.model_executor.custom_op import CustomOp
+from vllm.model_executor.layers.linear import (
+    ColumnParallelLinear,
+    MergedColumnParallelLinear,
+    RowParallelLinear,
+)
+from vllm.model_executor.layers.mamba.abstract import MambaBase
+from vllm.model_executor.layers.mamba.mamba_utils import (
+    MambaStateDtypeCalculator,
+    MambaStateShapeCalculator,
+)
+from vllm.model_executor.layers.mamba.ops.causal_conv1d import (
+    causal_conv1d_fn,
+    causal_conv1d_update,
+)
+from vllm.utils.torch_utils import direct_register_custom_op
+from vllm.v1.attention.backends.short_conv_attn import ShortConvAttentionMetadata
+
+
+@CustomOp.register("short_conv")
+class ShortConv(MambaBase, CustomOp):
+    def __init__(
+        self,
+        config,
+        dim: int,
+        layer_idx: int,
+        model_config: ModelConfig | None = None,
+        cache_config: CacheConfig | None = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.conv_dim = dim
+        self.L_cache = config.conv_L_cache
+        self.bias = config.conv_bias
+
+        self.conv = ColumnParallelLinear(
+            input_size=self.L_cache,
+            output_size=dim,
+            bias=self.bias,
+            prefix=f"{prefix}.conv1d",
+        )
+        # 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 `set_weight_attrs`
+        # doesn't allow to override it
+        self.conv.weight.data = self.conv.weight.data.unsqueeze(1)
+
+        self.in_proj = MergedColumnParallelLinear(
+            input_size=dim,
+            output_sizes=[dim] * 3,
+            bias=self.bias,
+            prefix=f"{prefix}.in_proj",
+        )
+        self.out_proj = RowParallelLinear(
+            input_size=dim,
+            output_size=dim,
+            bias=self.bias,
+            prefix=f"{prefix}.out_proj",
+        )
+
+        compilation_config = get_current_vllm_config().compilation_config
+        if prefix in compilation_config.static_forward_context:
+            raise ValueError(f"Duplicate layer name: {prefix}")
+        compilation_config.static_forward_context[prefix] = self
+        self.kv_cache = (torch.tensor([]),)
+
+        self.model_config = model_config
+        self.cache_config = cache_config
+        self.prefix = prefix
+
+    def forward_native(
+        self,
+        hidden_states: torch.Tensor,
+        output: torch.Tensor,
+    ):
+        return
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        output: torch.Tensor,
+    ):
+        torch.ops.vllm.short_conv(
+            hidden_states,
+            output,
+            self.prefix,
+        )
+
+    def forward_cuda(
+        self,
+        hidden_states: torch.Tensor,
+        output: torch.Tensor,
+    ):
+        forward_context = get_forward_context()
+        # ShortConvAttentionMetadata contains metadata necessary for the
+        # short_conv 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.
+        attn_metadata: AttentionMetadata = forward_context.attn_metadata
+        if attn_metadata is not None:
+            assert isinstance(attn_metadata, dict)
+            attn_metadata = attn_metadata[self.prefix]
+            assert isinstance(attn_metadata, ShortConvAttentionMetadata)
+            self_kv_cache = self.kv_cache[forward_context.virtual_engine]
+            conv_state = self_kv_cache[0].transpose(-1, -2)
+            state_indices_tensor = attn_metadata.state_indices_tensor
+            has_initial_states_p = attn_metadata.has_initial_states_p
+
+        BCx, _ = self.in_proj(hidden_states)
+
+        B, C, x = BCx.chunk(3, dim=-1)
+
+        conv_weights = self.conv.weight.view(
+            self.conv.weight.size(0), self.conv.weight.size(2)
+        )
+
+        if attn_metadata is None:
+            # V1 profile run
+            Bx = (B * x).contiguous()
+            hidden_states = C * Bx
+            contextualized_states, _ = self.out_proj(hidden_states)
+            return contextualized_states
+
+        num_prefills = attn_metadata.num_prefills  # request count
+        num_decodes = attn_metadata.num_decode_tokens  # token count (=request)
+        num_prefill_tokens = attn_metadata.num_prefill_tokens  # token count
+        has_prefill = num_prefills > 0
+        has_decode = num_decodes > 0
+        num_actual_tokens = num_decodes + num_prefill_tokens
+
+        # NOTE: V1 puts decode before prefill
+        # Separate prefill and decode by splitting varlen input
+        # Split along token dimension
+        B_d, B_p = torch.split(
+            B[:num_actual_tokens],
+            [num_decodes, num_prefill_tokens],
+            dim=0,
+        )
+        C_d, C_p = torch.split(
+            C[:num_actual_tokens],
+            [num_decodes, num_prefill_tokens],
+            dim=0,
+        )
+        x_d, x_p = torch.split(
+            x[:num_actual_tokens],
+            [num_decodes, num_prefill_tokens],
+            dim=0,
+        )
+        # Split along batch dimension
+        state_indices_tensor_d, state_indices_tensor_p = torch.split(
+            state_indices_tensor,
+            [num_decodes, num_prefills],
+            dim=0,
+        )
+        query_start_loc_p = (
+            attn_metadata.query_start_loc[-num_prefills - 1 :] - num_decodes
+            if has_prefill
+            else None
+        )
+
+        conv_output_list = []
+
+        if has_prefill:
+            Bx_p = (B_p * x_p).transpose(0, 1)
+            Bx = causal_conv1d_fn(
+                Bx_p,
+                conv_weights,
+                self.conv.bias,
+                activation=None,
+                conv_states=conv_state,
+                has_initial_state=has_initial_states_p,
+                cache_indices=state_indices_tensor_p,
+                metadata=attn_metadata,
+                query_start_loc=query_start_loc_p,
+            ).transpose(0, 1)[:num_prefill_tokens]
+
+            y = C_p * Bx
+            conv_output_list.append(y)
+
+        if has_decode:
+            Bx_d = (B_d * x_d).contiguous()
+            Bx = causal_conv1d_update(
+                Bx_d,
+                conv_state,
+                conv_weights,
+                self.conv.bias,
+                activation=None,
+                conv_state_indices=state_indices_tensor_d,
+            )
+            y = C_d * Bx
+            conv_output_list.insert(0, y)
+
+        # Merge prefill and decode outputs before passing to gated MLP
+        hidden_states = torch.vstack(conv_output_list)
+
+        # Final linear projection
+        output[:num_actual_tokens], _ = self.out_proj(hidden_states)
+
+    def get_state_dtype(self) -> tuple[torch.dtype, ...]:
+        assert self.model_config is not None
+        assert self.cache_config is not None
+        return MambaStateDtypeCalculator.short_conv_state_dtype(
+            self.model_config.dtype,
+            self.cache_config.mamba_cache_dtype,
+        )
+
+    def get_state_shape(self) -> tuple[tuple[int, ...]]:
+        return MambaStateShapeCalculator.short_conv_state_shape(
+            tp_world_size=get_tensor_model_parallel_world_size(),
+            intermediate_size=self.conv_dim,
+            conv_kernel=self.L_cache,
+        )
+
+    @property
+    def mamba_type(self) -> str:
+        return "short_conv"
+
+
+def short_conv(
+    hidden_states: torch.Tensor,
+    output: torch.Tensor,
+    layer_name: str,
+) -> None:
+    forward_context: ForwardContext = get_forward_context()
+    self = forward_context.no_compile_layers[layer_name]
+    self.forward_cuda(hidden_states=hidden_states, output=output)
+
+
+def short_conv_fake(
+    hidden_states: torch.Tensor,
+    output: torch.Tensor,
+    layer_name: str,
+) -> None:
+    return
+
+
+direct_register_custom_op(
+    op_name="short_conv",
+    op_func=short_conv,
+    mutates_args=["output"],
+    fake_impl=short_conv_fake,
+)