"""FP8 GEMM with Triton — heuristic + version-based repad."""
import torch
import torch.nn as nn
import triton
import triton.language as tl

OP_TYPE = "gemm"
SUPPORTED_PRECISIONS = ["fp8_e4m3"]
HARDWARE_REQUIRED = ["RTX_PRO_6000", "H100", "B200"]
E4M3_MAX = 448.0

# Pad K to multiple of 128 for efficient MMA.
K_BLOCK_ALIGN = 128


@triton.jit
def fp8_gemm_kernel(
    a_ptr, b_ptr, c_ptr,
    M, N, K,
    stride_am, stride_ak,
    stride_bn, stride_bk,
    scale_ptr,
    stride_cm, stride_cn,
    BLOCK_M: tl.constexpr,
    BLOCK_N: tl.constexpr,
    BLOCK_K: tl.constexpr,
    GROUP_M: tl.constexpr,
):
    pid = tl.program_id(0)
    grid_m = tl.cdiv(M, BLOCK_M)
    grid_n = tl.cdiv(N, BLOCK_N)
    width = GROUP_M * grid_n
    group_id = pid // width
    group_size = min(grid_m - group_id * GROUP_M, GROUP_M)
    pid_m = group_id * GROUP_M + (pid % group_size)
    pid_n = (pid % width) // group_size
    
    offs_m = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
    offs_n = pid_n * BLOCK_N + tl.arange(0, BLOCK_N)
    offs_k = tl.arange(0, BLOCK_K)
    
    a_ptrs = a_ptr + (offs_m[:, None] * stride_am + offs_k[None, :] * stride_ak)
    b_ptrs = b_ptr + (offs_n[None, :] * stride_bn + offs_k[:, None] * stride_bk)
    
    acc = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32)
    for k in range(0, K, BLOCK_K):
        a = tl.load(a_ptrs)
        b = tl.load(b_ptrs)
        acc = tl.dot(a, b, acc)
        a_ptrs += BLOCK_K * stride_ak
        b_ptrs += BLOCK_K * stride_bk
    
    scale = tl.load(scale_ptr + offs_n)
    acc = acc * scale[None, :]
    
    c_ptrs = c_ptr + offs_m[:, None] * stride_cm + offs_n[None, :] * stride_cn
    mask = (offs_m[:, None] < M) & (offs_n[None, :] < N)
    tl.store(c_ptrs, acc.to(tl.bfloat16), mask=mask)


def fp8_gemm_dispatch(x, w, weight_scale, M, N, K_pad):
    if M <= 32:
        BM, BN, BK, GM, NW, NS = 32, 64, 256, 4, 4, 2
    else:
        BM, BN, BK, GM, NW, NS = 128, 256, 128, 4, 8, 3
    
    c = torch.empty((M, N), device=x.device, dtype=torch.bfloat16)
    grid = (triton.cdiv(M, BM) * triton.cdiv(N, BN),)
    fp8_gemm_kernel[grid](
        x, w, c, M, N, K_pad,
        x.stride(0), x.stride(1),
        w.stride(0), w.stride(1),
        weight_scale,
        c.stride(0), c.stride(1),
        BLOCK_M=BM, BLOCK_N=BN, BLOCK_K=BK, GROUP_M=GM,
        num_warps=NW, num_stages=NS,
    )
    return c


class Model(nn.Module):
    def __init__(self, M: int, N: int, K: int):
        super().__init__()
        self.M, self.N, self.K = M, N, K
        w = torch.empty(N, K, dtype=torch.bfloat16)
        nn.init.normal_(w, std=0.02)
        s = (w.float().abs().amax(dim=1, keepdim=True) / E4M3_MAX).clamp(min=1e-12)
        w_fp8 = (w.float() / s).to(torch.float8_e4m3fn)
        self.register_buffer("weight", w_fp8)
        self.register_buffer("weight_scale", s.squeeze(1).to(torch.float32))
        K_pad = ((K + K_BLOCK_ALIGN - 1) // K_BLOCK_ALIGN) * K_BLOCK_ALIGN
        self._K_pad = K_pad
        if K_pad != K:
            w_pad = torch.zeros((N, K_pad), dtype=w_fp8.dtype)
            w_pad[:, :K].copy_(w_fp8)
            self.register_buffer("_w_padded", w_pad, persistent=False)
            self._w_version = -1
        else:
            self._w_padded = None
            self._w_version = -1

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        K = self.K
        K_pad = self._K_pad
        M, N = self.M, self.N
        
        if K_pad != K:
            if self.weight._version != self._w_version:
                self._w_padded[:, :K].copy_(self.weight)
                self._w_version = self.weight._version
            x_pad = getattr(self, '_x_pad', None)
            if x_pad is None or x_pad.device != x.device:
                x_pad = torch.zeros((M, K_pad), device=x.device, dtype=x.dtype)
                self._x_pad = x_pad
            x_pad[:, :K].copy_(x)
            x_use = x_pad
            w_use = self._w_padded
        else:
            x_use = x
            w_use = self.weight
        
        return fp8_gemm_dispatch(x_use, w_use, self.weight_scale, M, N, K_pad)


M = 4096
N = 4096
K = 4096


def get_inputs():
    x = (torch.rand(M, K) * 8 - 4).to(torch.float8_e4m3fn)
    return [x]


def get_init_inputs():
    return [M, N, K]