How much acceleration can Mixtral8x7b achieve using Column-Major MoE?

[Jacob0226]

In my tests, using 8 H100 GPUs did not show any acceleration. However, when I ran the script kernels/triton/inference/col_major_moe_gemm/perf_test_moe.py , I did see a 2-3 times speedup. Later, I tested Column-Major MoE and the MoE from vllm on 2 H100 GPUs separately and found about a 25% speedup, but this acceleration was not observed on 8 H100 GPUs. Is this result reasonable?

To reproduce:

# Docker
docker run -it --gpus all --rm --ipc=host  nvcr.io/nvidia/pytorch:23.10-py3

# Download vllm library
pip install vllm==v0.4.0.post1
# Download vllm repo
git clone https://github.com/vllm-project/vllm.git
cd vllm/benchmarks/; git checkout tags/v0.4.0.post1

# Download Mixtral
from huggingface_hub import snapshot_download, login
hf_token = "You should use your own token!" 
login(token=hf_token)
snapshot_download(repo_id="mistralai/Mixtral-8x7B-Instruct-v0.1")

# Run benchmark with vllm MoE
# If you see error about quantization_param_path, just comment it and run again.
python benchmark_throughput.py \
--model=mistralai/Mixtral-8x7B-Instruct-v0.1 --input-len=256 --output-len=64 \
--tensor-parallel-size=8  --num-prompts 400 --worker-use-ray

# Run benchmark with column-major MoE
# Find out the MoE script,
vim /usr/local/lib/python3.10/dist-packages/vllm/model_executor/layers/fused_moe/fused_moe.py
# Copy and paste the script from Column-Major_fused_moe.py
python benchmark_throughput.py \
--model=mistralai/Mixtral-8x7B-Instruct-v0.1 --input-len=256 --output-len=64 \
--tensor-parallel-size=8  --num-prompts 400 --worker-use-ray

I got roughly the same throughput from the above benchmark.

The code for Column-Major_fused_moe.py

"""Column-major Fused MoE kernel."""
import functools
import json
import os
from typing import Any, Dict, Optional, Tuple

import torch
import triton
import triton.language as tl

from vllm._C import ops
from vllm.logger import init_logger
from vllm.utils import is_hip

import time
MEASURE_TIME = False

logger = init_logger(__name__)

@triton.jit()
def col_major(pid,
              m, n,
              block_m: tl.constexpr, block_n: tl.constexpr):

    grid_m = tl.cdiv(m, block_m)
    grid_n = tl.cdiv(n, block_n)

    pid_m = (pid % grid_n)
    pid_n = pid // grid_m

    return pid_m, pid_n

@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,
    stride_weight,
    stride_token_id,
    # Meta-parameters
    block_m: tl.constexpr,
    block_n: tl.constexpr,
    block_k: 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_m, which is necessary to maintain consistency in block matrix
    multiplication across different blocks processed by the same expert.
    """
    
    pid = tl.program_id(axis=0)
    pid_m, pid_n = col_major(pid,
                             EM, N,
                             block_m, block_n,)

    num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
    if pid_m * block_m >= num_tokens_post_padded:
        return

    offs_token_id = pid_m * block_m + tl.arange(0, block_m)
    offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
    token_mask = offs_token < num_valid_tokens

    offs_bn = (pid_n * block_n + tl.arange(0, block_n)) % N
    offs_k = tl.arange(0, block_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_m, block_n]` block
    # of fp32 values for higher accuracy.
    # `accumulator` will be converted back to fp16 after the loop.
    accumulator = tl.zeros((block_m, block_n), dtype=tl.float32)

    for k in range(0, tl.cdiv(K, block_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_k),
                    other=0.0)
        b = tl.load(b_ptrs,
                    mask=offs_k[:, None] < K - k * block_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_k * stride_ak
        b_ptrs += block_k * stride_bk

    if MUL_ROUTED_WEIGHT:
        moe_weight = tl.load(topk_weights_ptr + offs_token * stride_weight,
                             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_n + tl.arange(0, block_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) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    """
    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)
    ops.moe_align_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[str, Any]) -> None:
    assert topk_weights.stride(1) == 1
    assert sorted_token_ids.stride(0) == 1

    EM = sorted_token_ids.shape[0]
    N = B.shape[1]

    grid = lambda META: (triton.cdiv(EM, META['block_m']) *
                         triton.cdiv(N, META['block_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),
        topk_weights.stride(1),       # New argument
        sorted_token_ids.stride(0),   # New argument
        MUL_ROUTED_WEIGHT=mul_routed_weight,
        top_k=top_k,
        compute_type=tl.bfloat16 if A.dtype == torch.bfloat16 else tl.float16,
         **config,
    )


def get_config_file_name(E: int, N: int) -> str:
    device_name = torch.cuda.get_device_name().replace(" ", "_")
    return f"E={E},N={N},device_name={device_name}.json"


@functools.lru_cache
def get_moe_configs(E: int, N: int) -> Optional[Dict[int, Any]]:
    """
    Return optimized configurations for the fused MoE kernel.

    The return value will be a dictionary that maps an irregular grid of
    batch sizes to configurations of the fused_moe kernel. To evaluate the
    kernel on a given batch size bs, the closest batch size in the grid should
    be picked and the associated configuration chosen to invoke the kernel.
    """

    # First look up if an optimized configuration is available in the configs
    # directory
    json_file_name = get_config_file_name(E, N)

    config_file_path = os.path.join(
        os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name)
    if os.path.exists(config_file_path):
        with open(config_file_path) as f:
            logger.info(
                f"Using configuration from {config_file_path} for MoE layer.")
            # If a configuration has been found, return it
            return {int(key): val for key, val in json.load(f).items()}

    # If no optimized configuration is available, we will use the default
    # configuration
    return None


def fused_moe(
    hidden_states: torch.Tensor,
    w1: torch.Tensor,
    w2: torch.Tensor,
    gating_output: torch.Tensor,                      # Not in Pytorch-Labs col_major
    topk: int,                                        # Pytorch-Labs pass topk_weights and topk_ids
    renormalize: bool,                                # Not in Pytorch-Labs col_major
    inplace: bool = False,
    override_config: Optional[Dict[str, Any]] = None, # Not in Pytorch-Labs col_major
) -> 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.
    - override_config (Optional[Dict[str, Any]]): Optional override
        for the kernel configuration.

    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
    # print("hidden_states shape or BatchSize =",hidden_states.shape)
    # print("w1 shape =",w1.shape)
    # print("w2 shape =",w2.shape)
    if MEASURE_TIME:
        start_time = time.perf_counter()

    if is_hip():
        # 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:
        import vllm._moe_C as moe_kernels

        topk_weights = torch.empty(M,
                                   topk,
                                   dtype=torch.float32,
                                   device=hidden_states.device)
        topk_ids = torch.empty(M,
                               topk,
                               dtype=torch.int32,
                               device=hidden_states.device)
        token_expert_indicies = torch.empty(M,
                                            topk,
                                            dtype=torch.int32,
                                            device=hidden_states.device)
        moe_kernels.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)

    config = {
        'block_m': 128,
        'block_n': 128,
        'block_k': 64,
    }
    
    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_m'], E)

    # print(f"First MoE: a.shape = {hidden_states.shape}, b.shape = {w1.shape}, c.shape = {intermediate_cache1.shape},\n"
    #       f"Second MoE:a.shape = {intermediate_cache2.shape}, b.shape = {w2.shape}, c.shape = {intermediate_cache3.shape}")
    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)

    ops.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:
        result = torch.sum(intermediate_cache3.view(*intermediate_cache3.shape),
                         dim=1,
                         out=hidden_states)
    else:
        result torch.sum(intermediate_cache3.view(*intermediate_cache3.shape),
                     dim=1)

    if MEASURE_TIME:
        end_time = time.perf_counter()
        elapsed_time = (end_time - start_time) * 1_000  # Convert to milliseconds
        print(f"Batch size = {M}, elapsed_time = {elapsed_time} ms")

    return result