Skip to content
KernelIndex
Search⌘K

amd-mixture-of-experts · submission 15074

KernelBot user 247 · python · License unknown

Use it

Vendorable · source mirrored · license unknownView source →

No package. Vendor the mirrored source: 1 lines, June 9 Researcher Reciprocity License v1.0.

triton_v03.py
curl "https://kernelindex.com/api/v1/implementations/kernelbot-amd-mixture-of-experts-15074?include=source"
interfacepython
Compatibility
measured onAMD Instinct MI300X
declared hardwareAMD Instinct MI300X
architecturesgfx942
dtypesfp16

Benchmark evidence

2 measurements across 1 GPU, fastest first.

Operation / workload
Hardware
Latency
Rank
Observed
AMD DeepSeek-style MoE layerfp16 · [1, 2048, 7168]
AMD Instinct MI300X
13.5ms
#60 of 118
2025-05-03
AMD DeepSeek-style MoE layerfp16 · [1, 8192, 7168]
AMD Instinct MI300X
35.4ms
#65 of 118
2025-05-03

Reported · How evidence levels are derived →

Source and license

sourceavailable
revision digestsha256:133c6932aa9651d1e36a525debc707f71091372109e4afbfbbf6875b7009b007
license declaredunknown
license concludedunknown
imported2026-08-15

Kernel source

triton_v03.py1 lines
from functools import partial
from typing import Dict, Tuple, Callable
import time

import torch
import torch.nn as nn
import triton
import triton.language as tl

from task import input_t, output_t


@triton.jit
def expert_up_kernel(
    x_ptr,
    gate_ptr,
    up_ptr,
    out_ptr,
    M,
    N: tl.constexpr,
    K: tl.constexpr,
    BLOCK_M: tl.constexpr,
    BLOCK_N: tl.constexpr,
    BLOCK_K: tl.constexpr,
):
    b_pid = tl.program_id(0)
    pid = tl.program_id(1)
    m_blocks = tl.cdiv(M, BLOCK_M)
    pid_m = pid % m_blocks
    pid_n = pid // m_blocks

    m_range = tl.arange(0, BLOCK_M)[:, None] + pid_m * BLOCK_M
    n_range = tl.arange(0, BLOCK_N)[None, :] + pid_n * BLOCK_N
    k1_range = tl.arange(0, BLOCK_K)[None, :]
    k2_range = tl.arange(0, BLOCK_K)[:, None]

    x_offset = x_ptr + m_range * K + b_pid * (M * K)
    gate_offset = gate_ptr + n_range
    up_offset = up_ptr + n_range
    up = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32)
    gate = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32)
    for k in tl.range(0, tl.cdiv(K, BLOCK_K)):
        x = tl.load(x_offset + k1_range, (m_range < M) & (k1_range < K))
        w = tl.load(gate_offset + k2_range * N, (n_range < N) & (k2_range < K))
        gate = tl.dot(x, w, gate)
        w = tl.load(up_offset + k2_range * N, (n_range < N) & (k2_range < K))
        up = tl.dot(x, w, up)
        k1_range += BLOCK_K
        k2_range += BLOCK_K

    out = up * gate * tl.sigmoid(gate)
    out_range = m_range * N + n_range + b_pid * (M * N)
    tl.store(out_ptr + out_range, out.to(tl.float16), (m_range < M) & (n_range < N))


@triton.jit
def expert_down_kernel(
    in_ptr,
    down_ptr,
    out_ptr,
    M,
    N: tl.constexpr,
    K: tl.constexpr,
    BLOCK_M: tl.constexpr,
    BLOCK_N: tl.constexpr,
    BLOCK_K: tl.constexpr,
):
    b_pid = tl.program_id(0)
    pid = tl.program_id(1)
    m_blocks = tl.cdiv(M, BLOCK_M)
    pid_m = pid % m_blocks
    pid_n = pid // m_blocks

    m_range = tl.arange(0, BLOCK_M)[:, None] + pid_m * BLOCK_M
    n_range = tl.arange(0, BLOCK_N)[None, :] + pid_n * BLOCK_N
    k1_range = tl.arange(0, BLOCK_K)[None, :]
    k2_range = tl.arange(0, BLOCK_K)[:, None]

    in_offset = in_ptr + m_range * K + b_pid * (M * K)
    down_offset = down_ptr + n_range
    out = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32)
    for k in tl.range(0, tl.cdiv(K, BLOCK_K)):
        x = tl.load(in_offset + k1_range, (m_range < M) & (k1_range < K))
        w = tl.load(down_offset + k2_range * N, (n_range < N) & (k2_range < K))
        out = tl.dot(x, w, out)
        k1_range += BLOCK_K
        k2_range += BLOCK_K

    out_range = m_range * N + n_range + b_pid * (M * N)
    tl.store(out_ptr + out_range, out.to(tl.float16), (m_range < M) & (n_range < N))


def expert(w_gate, w_up, w_down, x):
    if len(x.shape) == 2:
        B = 1
        M, K = x.shape
    else:
        B, M, K = x.shape

    K, N = w_gate.shape

    out = torch.zeros_like(x)
    tmp = torch.zeros((B, M, N), dtype=x.dtype, device=x.device)

    BLOCK_M = 32
    BLOCK_N = 32
    BLOCK_K = 128

    grid = (B, triton.cdiv(M, BLOCK_M) * triton.cdiv(N, BLOCK_N))    
    expert_up_kernel[grid](
        x,
        w_gate,
        w_up,
        tmp,
        M,
        N,
        K,
        BLOCK_M,
        BLOCK_N,
        BLOCK_K,
        num_stages = 1,
        num_warps = 2,
    )

    BLOCK_N = 128
    BLOCK_K = 32
    grid = (B, triton.cdiv(M, BLOCK_M) * triton.cdiv(K, BLOCK_K))
    expert_down_kernel[grid](
        tmp,
        w_down,
        out,
        M,
        K,
        N,
        BLOCK_M,
        BLOCK_K,
        BLOCK_N,
        num_stages = 1,
        num_warps = 2,
    )    
    del tmp
    return out


class MoEGate(nn.Module):
    def __init__(self, config: Dict):
        super().__init__()
        self.top_k: int = config["n_experts_per_token"]
        self.num_experts: int = config["n_routed_experts"]
        self.d_hidden: int = config["d_hidden"]

        self.W_g = nn.Linear(self.d_hidden, self.num_experts, bias=False)

    def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
        logits = self.W_g(x)
        scores = logits.softmax(dim=-1)
        topk_scores, topk_indices = torch.topk(
            scores, k=self.top_k, dim=-1, sorted=False
        )

        return topk_indices, topk_scores


class MoE(nn.Module):
    def __init__(self, config: Dict, experts: list[Callable], s_expert: Callable):
        super().__init__()
        self.config = config
        self.experts = experts
        self.gating_network = MoEGate(config)
        self.shared_expert = s_expert

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        shared_output = self.shared_expert(x)
        expert_indices, expert_scores = self.gating_network(x)
        batch_size, seq_len, hidden_dim = x.shape
        orig_shape = x.shape
        x_flat = x.view(-1, hidden_dim)
        flat_expert_indices = expert_indices.view(-1)
        flat_expert_weights = expert_scores.view(-1, 1)
        routed_output_flat = self.moe_infer(
            x_flat, flat_expert_indices, flat_expert_weights
        )

        routed_output = routed_output_flat.view(*orig_shape)
        return routed_output + shared_output

    @torch.no_grad()
    def moe_infer(
        self,
        x: torch.Tensor,
        flat_expert_indices: torch.Tensor,
        flat_expert_weights: torch.Tensor,
    ) -> torch.Tensor:
        expert_cache = torch.zeros_like(x)
        idxs = flat_expert_indices.argsort()
        counts = flat_expert_indices.bincount().cpu().numpy()
        tokens_per_expert = counts.cumsum()
        num_per_tok = self.config["n_experts_per_token"]
        token_idxs = idxs // num_per_tok
        for expert_id, end_idx in enumerate(tokens_per_expert):
            start_idx = 0 if expert_id == 0 else tokens_per_expert[expert_id - 1]
            if start_idx == end_idx:
                continue
            exp_token_idxs = token_idxs[start_idx:end_idx]
            expert_tokens = x[exp_token_idxs]
            expert_out = self.experts[expert_id](expert_tokens)
            expert_out.mul_(flat_expert_weights[idxs[start_idx:end_idx]])
            expert_cache.scatter_reduce_(
                0,
                exp_token_idxs.view(-1, 1).repeat(1, x.shape[-1]),
                expert_out,
                reduce="sum",
            )

        return expert_cache


def custom_kernel(data: input_t) -> output_t:
    """
    Submission template for DeepSeek-style Mixture of Experts using PyTorch.

    Args:
        data: Tuple of (input: torch.Tensor, weights: Dict[str, torch.Tensor], config: Dict)
            - input: Input tensor of shape [batch_size, seq_len, hidden_size]
            - weights: Dictionary containing model weights
            - config: Dictionary containing model configuration parameters

    Returns:
        Tuple containing:
            - output: Processed tensor [batch_size, seq_len, d_model]
            - aux_data: Dictionary with auxiliary data
    """
    input_tensor, weights, config = data
    num_experts = config["n_routed_experts"]

    experts = [    
        partial(
            expert,
            weights[f"experts.{i}.0.weight"],
            weights[f"experts.{i}.1.weight"],
            weights[f"experts.{i}.2.weight"],
        )        
        for i in range(num_experts)
    ]
    
    s_expert = partial(
        expert,
        weights["shared_experts.0.weight"],
        weights["shared_experts.1.weight"],
        weights["shared_experts.2.weight"],
    )
    
    moe = MoE(config, experts, s_expert)

    # Fill in the given weights of the model
    moe.gating_network.W_g.weight = nn.Parameter(weights["router.weight"])

    # Run the model
    output = moe(input_tensor)

    return output


Source code from GPU Mode and the KernelBot dataset · June 9 Researcher Reciprocity License v1.0

Best evidence level for this revision: reported

JSON