AICASGC/triton_linear_gelu.py

import triton
import triton.language as tl

@triton.jit
def _fc1_bias_gelu_kernel(
    X_ptr, W_ptr, B_ptr, Y_ptr,
    M: tl.constexpr, N: tl.constexpr, K: tl.constexpr,
    stride_xm, stride_xk,
    stride_wn, stride_wk,   # W is [N, K]
    stride_ym, stride_yn,
    BLOCK_SIZE_M: tl.constexpr, BLOCK_SIZE_N: tl.constexpr, BLOCK_SIZE_K: tl.constexpr,
    GROUP_SIZE_M: tl.constexpr
):
    """Kernel for computing the matmul C = A x B.
    A has shape (M, K), B has shape (K, N) and C has shape (M, N)
    """
    # -----------------------------------------------------------
    # Map program ids `pid` to the block of C it should compute.
    # This is done in a grouped ordering to promote L2 data reuse.
    # See above `L2 Cache Optimizations` section for details.
    pid = tl.program_id(axis=0)
    num_pid_m = tl.cdiv(M, 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

    # -----------------------------------------------------------
    # Add some integer bound assumptions.
    # This helps to guide integer analysis in the backend to optimize
    # load/store offset address calculation
    tl.assume(pid_m >= 0)
    tl.assume(pid_n >= 0)
    tl.assume(stride_xm > 0)
    tl.assume(stride_xk > 0)
    tl.assume(stride_wn > 0)
    tl.assume(stride_wk > 0)
    tl.assume(stride_ym > 0)
    tl.assume(stride_yn > 0)

    # ----------------------------------------------------------
    # Create pointers for the first blocks of A and B.
    # We will advance this pointer as we move in the K direction
    # and accumulate
    # `X_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers
    # `w_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers
    # See above `Pointer Arithmetic` section for details
    offs_xm = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M
    offs_wn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
    offs_k = tl.arange(0, BLOCK_SIZE_K)
    X_ptrs = X_ptr + (offs_xm[:, None] * stride_xm + offs_k[None, :] * stride_xk)
    W_ptrs = W_ptr + (offs_k[:, None] * stride_wk + offs_wn[None, :] * stride_wn)

    # -----------------------------------------------------------
    # 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.
        # If it is out of bounds, set it to 0.
        a = tl.load(X_ptrs, mask=offs_k[None, :] < K - k * BLOCK_SIZE_K, other=0.0)
        b = tl.load(W_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0)
        # We accumulate along the K dimension.
        accumulator = tl.dot(a, b, accumulator)
        # Advance the ptrs to the next K block.
        X_ptrs += BLOCK_SIZE_K * stride_xk
        W_ptrs += BLOCK_SIZE_K * stride_wk

    # Bias add (broadcast over M)
    b = tl.load(B_ptr + offs_wn, mask=offs_wn < N, other=0.0).to(tl.float32)
    accumulator = accumulator + b[None, :]

    # GELU(tanh) epilogue (fp32 compute)
    # You can fuse arbitrary activation functions here
    # while the accumulator is still in FP32!
    c0 = 0.7978845608028654
    c1 = 0.044715
    y = c0 * (accumulator + c1 * accumulator * accumulator * accumulator)
    # exp_2y = tl.exp(2.0 * y)
    # tanh_y = (exp_2y - 1.0) / (exp_2y + 1.0)
    ay = tl.abs(y)
    e = tl.exp(-2.0 * ay)
    t = (1.0 - e) / (1.0 + e)
    t = tl.where(y >= 0, t, -t)
    accumulator = 0.5 * accumulator * (1.0 + t)

    y = accumulator.to(tl.float16)

    # -----------------------------------------------------------
    # Write back the block of the output matrix C with masks.
    offs_ym = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
    offs_yn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
    Y_ptrs = Y_ptr + stride_ym * offs_ym[:, None] + stride_yn * offs_yn[None, :]
    Y_mask = (offs_ym[:, None] < M) & (offs_yn[None, :] < N)
    tl.store(Y_ptrs, y, mask=Y_mask)


def fc1_bias_gelu_triton(x, w, b,
                         BLOCK_SIZE_M=128, BLOCK_SIZE_N=128, BLOCK_SIZE_K=32, GROUP_SIZE_M=8, num_warps=8):
    """
    x: [M, K]
    w: [N, K]  (PyTorch Linear weight layout)
    b: [N]
    returns y: [M, N]
    """
    import torch
    assert x.is_cuda and w.is_cuda and b.is_cuda
    M, K = x.shape
    N, K2 = w.shape
    assert K == K2 and b.shape[0] == N

    y = torch.empty((M, N), device=x.device, dtype=x.dtype)

    grid = (triton.cdiv(M, BLOCK_SIZE_M) * triton.cdiv(N, BLOCK_SIZE_N),)

    _fc1_bias_gelu_kernel[grid](
        x, w, b, y,
        M=M, N=N, K=K,
        stride_xm=x.stride(0), stride_xk=x.stride(1),
        stride_wn=w.stride(0), stride_wk=w.stride(1),
        stride_ym=y.stride(0), stride_yn=y.stride(1),
        BLOCK_SIZE_M=BLOCK_SIZE_M, BLOCK_SIZE_N=BLOCK_SIZE_N, BLOCK_SIZE_K=BLOCK_SIZE_K,
        GROUP_SIZE_M=GROUP_SIZE_M,
        num_warps=num_warps
    )
    return y

import functools
from torch import nn

def path_linear_forward(self, hidden_state):
    return self.linear_fc2(fc1_bias_gelu_triton(hidden_state, self.linear_fc1.weight, self.linear_fc1.bias))