"""Diagnostic: print TMA partition tensor shapes for multi-tile K/V."""
import torch, cutlass, cutlass.cute as cute, cutlass.utils as utils
from cutlass.cute.nvgpu import cpasync, tcgen05
from cutlass import Float32, BFloat16, Int32
import cutlass.torch as ct
import math

HEAD_DIM = 64
n = 256

q = torch.randn(128, HEAD_DIM, 1, dtype=torch.bfloat16, device='cuda')
k = torch.randn(n, HEAD_DIM, 1, dtype=torch.bfloat16, device='cuda')
v = torch.randn(n, HEAD_DIM, dtype=torch.bfloat16, device='cuda')
v_kernel = v.unsqueeze(-1)

mQ = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q))
mK = ct.from_dlpack(k).mark_layout_dynamic(leading_dim=ct.get_leading_dim(k))
mV = ct.from_dlpack(v_kernel).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v_kernel))

# Hardcode major modes since LayoutEnum.from_tensor needs JIT context
qk_mma = utils.sm100.make_trivial_tiled_mma(BFloat16, BFloat16, cute.nvgpu.OperandMajorMode.K, cute.nvgpu.OperandMajorMode.K, Float32, tcgen05.CtaGroup.ONE, (128,128), tcgen05.OperandSource.SMEM)
pv_mma = utils.sm100.make_trivial_tiled_mma(BFloat16, BFloat16, cute.nvgpu.OperandMajorMode.K, cute.nvgpu.OperandMajorMode.MN, Float32, tcgen05.CtaGroup.ONE, (128,HEAD_DIM), tcgen05.OperandSource.TMEM)

qk_ik = cute.size(qk_mma.shape_mnk, mode=[2])
qk_mma_tiler = (128, 128, qk_ik * 4)
pv_ik = cute.size(pv_mma.shape_mnk, mode=[2])
pv_mma_tiler = (128, HEAD_DIM, pv_ik * (128 // pv_ik))
cluster_layout_vmnk = cute.tiled_divide(cute.make_layout((1,1,1)), (qk_mma.thr_id.shape,))

print(f'qk_mma_tiler: {qk_mma_tiler}')
print(f'pv_mma_tiler: {pv_mma_tiler}')

kv_stage = 2
k_smem_s = utils.sm100.make_smem_layout_b(qk_mma, qk_mma_tiler, BFloat16, kv_stage)
v_smem_s = utils.sm100.make_smem_layout_b(pv_mma, pv_mma_tiler, BFloat16, kv_stage)

k_s = cute.slice_(k_smem_s,(None,None,None,0))
v_s = cute.slice_(v_smem_s,(None,None,None,0))

print(f'k_s shape: {cute.shape(k_s)}')
print(f'v_s shape: {cute.shape(v_s)}')

tma_k, mK_tma = cute.nvgpu.make_tiled_tma_atom_B(
    utils.sm100.cluster_shape_to_tma_atom_B(cluster_layout_vmnk.shape, qk_mma.thr_id),
    mK, k_s, qk_mma_tiler, qk_mma, cluster_layout_vmnk.shape
)

gK = cute.local_tile(mK_tma, cute.slice_(qk_mma_tiler,(0,None,None)),(None,None,None))
print(f'mK_tma shape: {cute.shape(mK_tma)}')
print(f'gK shape: {cute.shape(gK)}')
print(f'n_kv_tiles: {cute.size(gK, mode=[3])}')

qk_thr = qk_mma.get_slice(0)
tCgK = qk_thr.partition_B(gK)
print(f'tCgK shape: {cute.shape(tCgK)}')

sK = cute.make_tensor(BFloat16, k_s)
b_lay = cute.make_layout(cute.slice_(cluster_layout_vmnk,(0,None,0,0)).shape)

tBsK, tBgK = cpasync.tma_partition(tma_k, 0, b_lay, cute.group_modes(sK,0,3), cute.group_modes(tCgK,0,3))

print(f'tBsK shape: {cute.shape(tBsK)}')
print(f'tBgK shape: {cute.shape(tBgK)}')
print(f'tBsK layout: {tBsK.layout}')
print(f'tBgK layout: {tBgK.layout}')

# Test slices
for desc, sl in [
    ("(None,0,None,0)", (None,0,None,0)),  # Current (broken)
    ("(None,None,0,0)", (None,None,0,0)),  # CUTLASS reference style
    ("(0,None,None,0)", (0,None,None,0)),  # Alternative
]:
    try:
        result = tBgK[sl]
        print(f'tBgK after {desc} shape: {cute.shape(result)}')
    except Exception as e:
        print(f'tBgK after {desc}: ERROR {type(e).__name__}: {e}')

# Also check V
tma_v, mV_tma = cute.nvgpu.make_tiled_tma_atom_B(
    utils.sm100.cluster_shape_to_tma_atom_B(cluster_layout_vmnk.shape, pv_mma.thr_id),
    mV, v_s, pv_mma_tiler, pv_mma, cluster_layout_vmnk.shape
)
gV = cute.local_tile(mV_tma, cute.slice_(pv_mma_tiler,(0,None,None)),(None,None,None))
pv_thr = pv_mma.get_slice(0)
tCgV = pv_thr.partition_B(gV)
sV = cute.make_tensor(BFloat16, v_s)
tVsV, tVgV = cpasync.tma_partition(tma_v, 0, b_lay, cute.group_modes(sV,0,3), cute.group_modes(tCgV,0,3))
print(f'tVgV shape: {cute.shape(tVgV)}')
for desc, sl in [
    ("(None,0,None,0)", (None,0,None,0)),
    ("(None,None,0,0)", (None,None,0,0)),
]:
    try:
        result = tVgV[sl]
        print(f'tVgV after {desc} shape: {cute.shape(result)}')
    except Exception as e:
        print(f'tVgV after {desc}: ERROR {type(e).__name__}: {e}')