moe.py

import torch
import triton
from typing import Dict
import triton.language as tl

try:
    import awq_ext  # with CUDA kernels

    AWQ_INSTALLED = True
except:
    AWQ_INSTALLED = False


class FusedSparseMoeBlock(torch.nn.Module):
    def __init__(
        self,
        top_k,
        gate,
        ws,
        w2s,
    ):
        super().__init__()
        self.gate = gate
        self.top_k = top_k
        self.ws = ws
        self.w2s = w2s

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        batch_size, sequence_length, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)

        # router_logits: (batch * sequence_length, n_experts)
        router_logits = self.gate(hidden_states)

        final_hidden_states = apply_moe_weights(
            self.ws,
            self.w2s,
            hidden_states,
            router_logits,
            self.top_k,
            renormalize=True,
        )

        return final_hidden_states.view(batch_size, sequence_length, hidden_dim)


def apply_moe_weights(
    w1: Dict[str, torch.Tensor],
    w2: Dict[str, torch.Tensor],
    x: torch.Tensor,
    gating_output: torch.Tensor,
    topk: int,
    renormalize: bool,
) -> torch.Tensor:
    FP16_MATMUL_HEURISTIC_CONDITION = x.shape[:-1].numel() >= 1024
    if FP16_MATMUL_HEURISTIC_CONDITION:
        dequant_w1 = awq_ext.dequantize_weights_cuda(
            w1.qweight, w1.scales, w1.qzeros, 0, 0, 0, False
        ).permute(0, 2, 1)
        dequant_w2 = awq_ext.dequantize_weights_cuda(
            w2.qweight, w2.scales, w2.qzeros, 0, 0, 0, False
        ).permute(0, 2, 1)
        return fused_moe(x, dequant_w1, dequant_w2, gating_output, topk, renormalize)

    topk_weights, topk_ids = fused_topk(gating_output, topk, renormalize)
    (sorted_token_ids, expert_ids, num_tokens_post_padded) = moe_align_block_size(
        topk_ids, 16, w1.qweight.shape[0]
    )

    x = x.view(x.shape[0], 1, *x.shape[1:])

    gate_up = awq_ext.grouped_gemm_forward(
        x,
        w1.qweight,
        w1.scales,
        w1.qzeros,
        topk_weights,
        sorted_token_ids,
        expert_ids,
        num_tokens_post_padded,
        False,
        8,
    )

    out = torch.empty(
        (gate_up.shape[:-1] + (gate_up.shape[-1] // 2,)), dtype=x.dtype, device=x.device
    )
    awq_ext.silu_and_mul(out, gate_up)

    out = awq_ext.grouped_gemm_forward(
        out,
        w2.qweight,
        w2.scales,
        w2.qzeros,
        topk_weights,
        sorted_token_ids,
        expert_ids,
        num_tokens_post_padded,
        True,
        8,
    )

    return torch.sum(out, dim=1)


@triton.jit
def fused_moe_kernel(
    # Pointers to matrices
    a_ptr,
    b_ptr,
    c_ptr,
    topk_weights_ptr,
    sorted_token_ids_ptr,
    expert_ids_ptr,
    num_tokens_post_padded_ptr,
    # Matrix dimensions
    N,
    K,
    EM,
    num_valid_tokens,
    # The stride variables represent how much to increase the ptr by when moving by 1
    # element in a particular dimension. E.g. `stride_am` is how much to increase `a_ptr`
    # by to get the element one row down (A has M rows).
    stride_am,
    stride_ak,
    stride_be,
    stride_bk,
    stride_bn,
    stride_cm,
    stride_cn,
    # Meta-parameters
    BLOCK_SIZE_M: tl.constexpr,
    BLOCK_SIZE_N: tl.constexpr,
    BLOCK_SIZE_K: tl.constexpr,
    GROUP_SIZE_M: tl.constexpr,
    MUL_ROUTED_WEIGHT: tl.constexpr,
    top_k: tl.constexpr,
    compute_type: tl.constexpr,
):
    """
    Implements the fused computation for a Mixture of Experts (MOE) using token and expert matrices.

    Key Parameters:
    - A: The input tensor representing tokens with shape (*, K), where '*' can be any shape representing batches and K is the feature dimension of each token.
    - B: The stacked MOE weight tensor with shape (E, N, K), where E is the number of experts, K is the input feature dimension, and N is the output feature dimension.
    - C: The output cache tensor with shape (M, topk, N), where M is the total number of tokens post padding, topk is the number of times each token is repeated,
        and N is the output feature dimension.
    - sorted_token_ids: A tensor containing the sorted indices of tokens, repeated topk times and arranged by the expert index they are assigned to.
    - expert_ids: A tensor containing the indices of the expert for each block. It determines which expert matrix from B should be used for each block in A.
    This kernel performs the multiplication of a token by its corresponding expert matrix as determined by `expert_ids`. The sorting of `sorted_token_ids`
    by expert index and padding ensures divisibility by BLOCK_SIZE_M, which is necessary to maintain consistency in block matrix multiplication across different blocks processed by the same expert.
    """
    # -----------------------------------------------------------
    # Map program ids `pid` to the block of C it should compute.
    # This is done in a grouped ordering to promote L2 data reuse.
    pid = tl.program_id(axis=0)
    num_pid_m = tl.cdiv(EM, BLOCK_SIZE_M)
    num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
    num_pid_in_group = GROUP_SIZE_M * num_pid_n
    group_id = pid // num_pid_in_group
    first_pid_m = group_id * GROUP_SIZE_M
    group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
    pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m)
    pid_n = (pid % num_pid_in_group) // group_size_m

    # ----------------------------------------------------------
    # Create pointers for the first blocks of A and B.
    # We will advance this pointer as we move in the K direction
    # and accumulate
    # `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers
    # `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers
    num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
    if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
        return
    offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
    offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
    token_mask = offs_token < num_valid_tokens

    offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
    offs_k = tl.arange(0, BLOCK_SIZE_K)
    a_ptrs = a_ptr + (
        offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
    )

    off_experts = tl.load(expert_ids_ptr + pid_m)
    b_ptrs = (
        b_ptr
        + off_experts * stride_be
        + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn)
    )

    # -----------------------------------------------------------
    # Iterate to compute a block of the C matrix.
    # We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block
    # of fp32 values for higher accuracy.
    # `accumulator` will be converted back to fp16 after the loop.
    accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)

    for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
        # Load the next block of A and B, generate a mask by checking the K dimension.
        a = tl.load(
            a_ptrs,
            mask=token_mask[:, None] & (offs_k[None, :] < K - k * BLOCK_SIZE_K),
            other=0.0,
        )
        b = tl.load(b_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0)
        # We accumulate along the K dimension.
        accumulator += tl.dot(a, b)
        # Advance the ptrs to the next K block.
        a_ptrs += BLOCK_SIZE_K * stride_ak
        b_ptrs += BLOCK_SIZE_K * stride_bk

    if MUL_ROUTED_WEIGHT:
        moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0)
        accumulator = accumulator * moe_weight[:, None]

    accumulator = accumulator.to(compute_type)
    # -----------------------------------------------------------
    # Write back the block of the output
    offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
    c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :]
    c_mask = token_mask[:, None] & (offs_cn[None, :] < N)
    tl.store(c_ptrs, accumulator, mask=c_mask)


def moe_align_block_size(topk_ids: torch.Tensor, block_size: int, num_experts: int):
    """
    Aligns the token distribution across experts to be compatible with block size for matrix multiplication.

    Parameters:
    - topk_ids: A tensor of shape [total_tokens, top_k] representing the top-k expert indices for each token.
    - block_size: The block size used in block matrix multiplication.
    - num_experts: The total number of experts.

    Returns:
    - sorted_token_ids: A tensor containing the sorted token indices according to their allocated expert.
    - expert_ids: A tensor indicating the assigned expert index for each block.
    - num_tokens_post_padded: The total number of tokens after padding, ensuring divisibility by block_size.

    This function pads the number of tokens that each expert needs to process so that it is divisible by block_size.
    Padding ensures that during block matrix multiplication, the dimensions align correctly.

    Example:
    Given topk_ids = [[2, 3, 4], [1, 2, 4], [1, 3, 4], [1, 2, 3]], block_size = 4, and num_experts = 4:
    - We initially have 12 tokens (after repeating 'top_k' times) and 4 experts, with each expert needing to process 3 tokens.
    - As block_size is 4, we pad 1 token for each expert.
    - First, flatten topk_ids to [2, 3, 4, 1, 2, 4, 1, 3, 4, 1, 2, 3].
    - Then append padding tokens [12, 12, 12, 12] for each block.
    - After sorting by expert index, we obtain token_ids [3, 6, 9, 12, 0, 4, 10, 12, 1, 7, 11, 12, 2, 5, 8, 12].
        Tokens 12 are non-existent (padding) and are ignored in the subsequent matrix multiplication.
    - The padding ensures that the total number of tokens is now divisible by block_size for proper block matrix operations.
    """
    sorted_ids = torch.empty(
        (topk_ids.numel() + num_experts * (block_size - 1),),
        dtype=torch.int32,
        device=topk_ids.device,
    )
    expert_ids = torch.empty(
        (topk_ids.numel() + num_experts,), dtype=torch.int32, device=topk_ids.device
    )
    sorted_ids.fill_(topk_ids.numel())
    num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device)
    awq_ext.moe_alig_block_size(
        topk_ids, num_experts, block_size, sorted_ids, expert_ids, num_tokens_post_pad
    )
    return sorted_ids, expert_ids, num_tokens_post_pad


def invoke_fused_moe_kernel(
    A: torch.Tensor,
    B: torch.Tensor,
    C: torch.Tensor,
    topk_weights: torch.Tensor,
    topk_ids: torch.Tensor,
    sorted_token_ids: torch.Tensor,
    expert_ids: torch.Tensor,
    num_tokens_post_padded: torch.Tensor,
    mul_routed_weight: bool,
    top_k: int,
    config: dict,
):
    assert topk_weights.stride(1) == 1
    assert sorted_token_ids.stride(0) == 1

    grid = lambda META: (
        triton.cdiv(sorted_token_ids.shape[0], META["BLOCK_SIZE_M"])
        * triton.cdiv(B.shape[1], META["BLOCK_SIZE_N"]),
    )

    fused_moe_kernel[grid](
        A,
        B,
        C,
        topk_weights,
        sorted_token_ids,
        expert_ids,
        num_tokens_post_padded,
        B.shape[1],
        B.shape[2],
        sorted_token_ids.shape[0],
        topk_ids.numel(),
        A.stride(0),
        A.stride(1),
        B.stride(0),
        B.stride(2),
        B.stride(1),
        C.stride(1),
        C.stride(2),
        MUL_ROUTED_WEIGHT=mul_routed_weight,
        top_k=top_k,
        compute_type=tl.bfloat16 if A.dtype == torch.bfloat16 else tl.float16,
        **config,
    )


def fused_topk(
    gating_output: torch.Tensor,
    topk: int,
    renormalize: bool,
):
    """Compute top-k indice and weights from gating logits

    Args:
        gating_output (torch.Tensor): The output of the gating operation (before softmax).
        topk (int): The number of top-k experts to select.
        renormalize (bool): If True, renormalize the top-k weights to sum to 1.
    """
    M = gating_output.shape[0]
    if torch.version.hip is not None:
        # The MoE kernels are not yet supported on ROCm.
        routing_weights = torch.softmax(gating_output, dim=-1, dtype=torch.float32)
        topk_weights, topk_ids = torch.topk(routing_weights, topk, dim=-1)
    else:
        topk_weights = torch.empty(
            M, topk, dtype=torch.float32, device=gating_output.device
        )
        topk_ids = torch.empty(M, topk, dtype=torch.int32, device=gating_output.device)
        token_expert_indicies = torch.empty(
            M, topk, dtype=torch.int32, device=gating_output.device
        )
        awq_ext.topk_softmax(
            topk_weights,
            topk_ids,
            token_expert_indicies,
            gating_output.float(),  # TODO(woosuk): Optimize this.
        )
        del token_expert_indicies  # Not used. Will be used in the future.
    if renormalize:
        topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
    return topk_weights, topk_ids


def fused_moe(
    hidden_states: torch.Tensor,
    w1: torch.Tensor,
    w2: torch.Tensor,
    gating_output: torch.Tensor,
    topk: int,
    renormalize: bool,
    inplace: bool = True,
) -> torch.Tensor:
    """
    This function computes a Mixture of Experts (MoE) layer using two sets of weights, w1 and w2, and top-k gating mechanism.

    Parameters:
    - hidden_states (torch.Tensor): The input tensor to the MoE layer.
    - w1 (torch.Tensor): The first set of expert weights.
    - w2 (torch.Tensor): The second set of expert weights.
    - gating_output (torch.Tensor): The output of the gating operation (before softmax).
    - topk (int): The number of top-k experts to select.
    - renormalize (bool): If True, renormalize the top-k weights to sum to 1.
    - inplace (bool): If True, perform the operation in-place. Defaults to False.

    Returns:
    - torch.Tensor: The output tensor after applying the MoE layer.
    """
    # Check constraints.
    assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
    assert hidden_states.shape[1] == w1.shape[2], "Hidden size mismatch"
    assert gating_output.shape[1] == w1.shape[0], "Number of experts mismatch"
    assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
    # assert w1.is_contiguous(), "Expert weights1 must be contiguous"
    # assert w2.is_contiguous(), "Expert weights2 must be contiguous"
    assert hidden_states.dtype in [torch.float32, torch.float16, torch.bfloat16]
    M, _ = hidden_states.shape
    E, N, _ = w1.shape

    topk_weights, topk_ids = fused_topk(gating_output, topk, renormalize)

    config = {
        "BLOCK_SIZE_M": 64,
        "BLOCK_SIZE_N": 64,
        "BLOCK_SIZE_K": 32,
        "GROUP_SIZE_M": 8,
    }

    if topk_ids.numel() <= w1.shape[0]:
        config = {
            "BLOCK_SIZE_M": 16,
            "BLOCK_SIZE_N": 32,
            "BLOCK_SIZE_K": 64,
            "GROUP_SIZE_M": 1,
        }

    intermediate_cache1 = torch.empty(
        (M, topk_ids.shape[1], N),
        device=hidden_states.device,
        dtype=hidden_states.dtype,
    )
    intermediate_cache2 = torch.empty(
        (M * topk_ids.shape[1], N // 2),
        device=hidden_states.device,
        dtype=hidden_states.dtype,
    )
    intermediate_cache3 = torch.empty(
        (M, topk_ids.shape[1], w2.shape[1]),
        device=hidden_states.device,
        dtype=hidden_states.dtype,
    )

    sorted_token_ids, expert_ids, num_tokens_post_padded = moe_align_block_size(
        topk_ids, config["BLOCK_SIZE_M"], E
    )

    invoke_fused_moe_kernel(
        hidden_states,
        w1,
        intermediate_cache1,
        topk_weights,
        topk_ids,
        sorted_token_ids,
        expert_ids,
        num_tokens_post_padded,
        False,
        topk_ids.shape[1],
        config,
    )

    awq_ext.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))

    invoke_fused_moe_kernel(
        intermediate_cache2,
        w2,
        intermediate_cache3,
        topk_weights,
        topk_ids,
        sorted_token_ids,
        expert_ids,
        num_tokens_post_padded,
        True,
        1,
        config,
    )

    if inplace:
        return torch.sum(
            intermediate_cache3.view(*intermediate_cache3.shape),
            dim=1,
            out=hidden_states,
        )
    return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1)