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

@triton.jit
def matmul_kernel(
    a_ptr, b_ptr, c_ptr, scale_ptr,
    M, N, K,
    stride_am, stride_ak,
    stride_bn, stride_bk,
    stride_cm, stride_cn,
    stride_scale,
    BLOCK_M: tl.constexpr, BLOCK_N: tl.constexpr, BLOCK_K: tl.constexpr,
):
    pid = tl.program_id(axis=0)
    num_pid_m = tl.cdiv(M, BLOCK_M)
    num_pid_n = tl.cdiv(N, BLOCK_N)
    pid_m = pid // num_pid_n
    pid_n = pid % num_pid_n

    offs_am = (pid_m * BLOCK_M + tl.arange(0, BLOCK_M)) % M
    offs_bn = (pid_n * BLOCK_N + tl.arange(0, BLOCK_N)) % N
    offs_k = tl.arange(0, BLOCK_K)

    a_ptrs = a_ptr + offs_am[:, None] * stride_am + offs_k[None, :] * stride_ak
    b_ptrs = b_ptr + offs_bn[None, :] * stride_bn + offs_k[:, None] * stride_bk

    accumulator = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32)
    for k in range(0, tl.cdiv(K, BLOCK_K)):
        a = tl.load(a_ptrs, mask=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)
        
        accumulator = tl.dot(a, b, accumulator)
        a_ptrs += BLOCK_K * stride_ak
        b_ptrs += BLOCK_K * stride_bk

    # Load scales
    scale_ptrs = scale_ptr + offs_bn * stride_scale
    scale = tl.load(scale_ptrs, mask=offs_bn < N, other=1.0)
    accumulator = accumulator * scale[None, :]

    offs_cm = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
    offs_cn = pid_n * BLOCK_N + tl.arange(0, BLOCK_N)
    c_ptrs = c_ptr + offs_cm[:, None] * stride_cm + offs_cn[None, :] * stride_cn
    c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N)
    tl.store(c_ptrs, accumulator.to(tl.bfloat16), mask=c_mask)

class Model(nn.Module):
    def __init__(self, M: int, N: int, K: int):
        super().__init__()
        self.M, self.N, self.K = M, N, K
        # Setup temporary buffers matching reference
        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))

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        # Run our Triton GEMM
        M, K = x.shape
        N = self.weight.shape[0]
        c = torch.empty((M, N), device=x.device, dtype=torch.bfloat16)
        
        BLOCK_M = 64
        BLOCK_N = 64
        BLOCK_K = 64
        
        grid = lambda META: (triton.cdiv(M, META['BLOCK_M']) * triton.cdiv(N, META['BLOCK_N']),)
        
        matmul_kernel[grid](
            x, self.weight, c, self.weight_scale,
            M, N, K,
            x.stride(0), x.stride(1),
            self.weight.stride(0), self.weight.stride(1),
            c.stride(0), c.stride(1),
            self.weight_scale.stride(0),
            BLOCK_M=BLOCK_M, BLOCK_N=BLOCK_N, BLOCK_K=BLOCK_K,
        )
        return c

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]