[BUG] Int8 Matmul Wrong Answer (on 4090) + Compiler Internal Error (on certain shapes)

[Triang-jyed-driung]

Required prerequisites

• I have read the documentation https://tilelang.com.
• I have searched the Issue Tracker that this hasn't already been reported. (comment there if it has.)

What version of TileLang are you using?

0.1.9

System information

addict                       2.4.0
aiohappyeyeballs             2.6.1
aiohttp                      3.13.5
aiosignal                    1.4.0
anaconda-anon-usage          0.7.6
anaconda-auth                0.14.2
anaconda-cli-base            0.8.2
annotated-types              0.6.0
anyio                        4.12.1
apache-tvm-ffi               0.1.11
archspec                     0.2.5
attrs                        26.1.0
beautifulsoup4               4.14.3
boltons                      25.0.0
brotlicffi                   1.2.0.0
certifi                      2026.4.22
cffi                         1.17.1
charset-normalizer           3.4.4
click                        8.2.1
cloudpickle                  3.1.2
conda                        26.3.2
conda-anaconda-telemetry     0.3.0
conda-anaconda-tos           0.2.2
conda-content-trust          0.3.1
conda-libmamba-solver        26.4.0
conda-package-handling       2.4.0
conda_package_streaming      0.12.0
contourpy                    1.3.3
cryptography                 45.0.7
cuda-bindings                13.2.0
cuda-pathfinder              1.5.4
cuda-python                  13.2.0
cuda-toolkit                 13.0.2
cycler                       0.12.1
datasets                     4.8.5
dill                         0.4.1
distro                       1.9.0
einops                       0.8.2
filelock                     3.29.0
fonttools                    4.62.1
frozendict                   2.4.6
frozenlist                   1.8.0
fsspec                       2026.2.0
gdown                        6.0.0
gram-newton-schulz           0.1.4
h11                          0.16.0
hf-xet                       1.5.0
httpcore                     1.0.9
httpx                        0.28.1
huggingface_hub              1.14.0
idna                         3.11
jaraco.classes               3.4.0
jaraco.context               6.1.0
jaraco.functools             4.4.0
jeepney                      0.7.1
Jinja2                       3.1.6
joblib                       1.5.3
jsonpatch                    1.33
jsonpointer                  3.1.1
keyring                      25.7.0
kiwisolver                   1.5.0
libmambapy                   2.3.2
markdown-it-py               4.0.0
MarkupSafe                   3.0.3
matplotlib                   3.10.9
mdurl                        0.1.2
menuinst                     2.4.2
ml_dtypes                    0.5.4
modelscope                   1.36.3
more-itertools               11.0.2
mpmath                       1.3.0
msgpack                      1.1.1
multidict                    6.7.1
multiprocess                 0.70.19
networkx                     3.6.1
numpy                        2.4.4
nvidia-cublas                13.1.0.3
nvidia-cuda-cupti            13.0.85
nvidia-cuda-nvrtc            13.0.88
nvidia-cuda-runtime          13.0.96
nvidia-cudnn-cu13            9.19.0.56
nvidia-cufft                 12.0.0.61
nvidia-cufile                1.15.1.6
nvidia-curand                10.4.0.35
nvidia-cusolver              12.0.4.66
nvidia-cusparse              12.6.3.3
nvidia-cusparselt-cu13       0.8.0
nvidia-cutlass-dsl           4.4.2
nvidia-cutlass-dsl-libs-base 4.4.2
nvidia-nccl-cu13             2.28.9
nvidia-nvjitlink             13.0.88
nvidia-nvshmem-cu13          3.4.5
nvidia-nvtx                  13.0.85
packaging                    26.0
pandas                       3.0.2
pillow                       12.2.0
pip                          26.0.1
pkce                         1.0.3
platformdirs                 4.9.4
pluggy                       1.6.0
propcache                    0.4.1
psutil                       7.2.2
pyarrow                      24.0.0
pycosat                      0.6.6
pycparser                    3.0
pydantic                     2.13.2
pydantic_core                2.46.2
pydantic-settings            2.12.0
Pygments                     2.20.0
PyJWT                        2.12.1
pyparsing                    3.3.2
PySocks                      1.7.1
python-dateutil              2.9.0.post0
python-dotenv                1.2.1
PyYAML                       6.0.3
quack-kernels                0.4.1
readchar                     4.2.1
requests                     2.33.1
rich                         14.2.0
ruamel.yaml                  0.18.16
ruamel.yaml.clib             0.2.14
scikit-learn                 1.8.0
scipy                        1.17.1
SecretStorage                3.5.0
semver                       3.0.4
setuptools                   81.0.0
shellingham                  1.5.4
six                          1.17.0
soupsieve                    2.8.3
sympy                        1.14.0
threadpoolctl                3.6.0
tilelang                     0.1.9
tomli                        2.4.0
tomlkit                      0.13.3
torch                        2.11.0
torch_c_dlpack_ext           0.1.5
torchvision                  0.26.0
tqdm                         4.67.3
triton                       3.6.0
truststore                   0.10.1
typer                        0.20.0
typer-slim                   0.20.0
typing_extensions            4.15.0
typing-inspection            0.4.2
urllib3                      2.6.3
wheel                        0.46.3
xxhash                       3.7.0
yarl                         1.23.0
z3-solver                    4.15.4.0
zstandard                    0.25.0

Collecting environment information...
PyTorch version: 2.11.0+cu130
Is debug build: False
CUDA used to build PyTorch: 13.0
ROCM used to build PyTorch: N/A

OS: Arch Linux (x86_64)
GCC version: (GCC) 16.1.1 20260430
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.43

Python version: 3.13.13 | packaged by Anaconda, Inc. | (main, Apr 14 2026, 06:19:41) [GCC 14.3.0] (64-bit runtime)
Python platform: Linux-7.0.3-arch1-2-x86_64-with-glibc2.43
Is CUDA available: True
CUDA runtime version: 13.2.78
CUDA_MODULE_LOADING set to: 
GPU models and configuration: GPU 0: NVIDIA GeForce RTX 4090
Nvidia driver version: 595.71.05
cuDNN version: Could not collect
Is XPU available: False
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
Caching allocator config: N/A

CPU:
Architecture:                            x86_64
CPU op-mode(s):                          32-bit, 64-bit
Address sizes:                           48 bits physical, 48 bits virtual
Byte Order:                              Little Endian
CPU(s):                                  16
On-line CPU(s) list:                     0-15
Vendor ID:                               AuthenticAMD
Model name:                              AMD Ryzen 9 7950X 16-Core Processor
CPU family:                              25
Model:                                   97
Thread(s) per core:                      1
Core(s) per socket:                      1
Socket(s):                               16
Stepping:                                2
Microcode version:                       0xa60120a
BogoMIPS:                                8962.93
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm rep_good nopl xtopology cpuid extd_apicid pni pclmulqdq ssse3 fma cx16 sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm cmp_legacy svm cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw perfctr_core ssbd perfmon_v2 ibrs ibpb stibp ibrs_enhanced vmmcall fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves avx512_bf16 clzero xsaveerptr wbnoinvd arat npt lbrv nrip_save tsc_scale vmcb_clean flushbyasid pausefilter pfthreshold vgif vnmi avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq rdpid overflow_recov succor fsrm flush_l1d
Virtualization:                          AMD-V
L1d cache:                               1 MiB (16 instances)
L1i cache:                               1 MiB (16 instances)
L2 cache:                                8 MiB (16 instances)
L3 cache:                                256 MiB (16 instances)
NUMA node(s):                            1
NUMA node0 CPU(s):                       0-15
Vulnerability Gather data sampling:      Not affected
Vulnerability Ghostwrite:                Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Old microcode:             Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Mitigation; Safe RET
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; Enhanced / Automatic IBRS; IBPB conditional; STIBP disabled; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Vulnerable: No microcode
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Not affected

Versions of relevant libraries:
[pip3] numpy==2.4.4
[pip3] nvidia-cublas==13.1.0.3
[pip3] nvidia-cuda-cupti==13.0.85
[pip3] nvidia-cuda-nvrtc==13.0.88
[pip3] nvidia-cuda-runtime==13.0.96
[pip3] nvidia-cudnn-cu13==9.19.0.56
[pip3] nvidia-cufft==12.0.0.61
[pip3] nvidia-curand==10.4.0.35
[pip3] nvidia-cusolver==12.0.4.66
[pip3] nvidia-cusparse==12.6.3.3
[pip3] nvidia-cusparselt-cu13==0.8.0
[pip3] nvidia-nccl-cu13==2.28.9
[pip3] nvidia-nvjitlink==13.0.88
[pip3] nvidia-nvtx==13.0.85
[pip3] torch==2.11.0
[pip3] torch_c_dlpack_ext==0.1.5
[pip3] torchvision==0.26.0
[pip3] triton==3.6.0
[conda] numpy                           2.4.4            pypi_0              pypi
[conda] nvidia-cublas                   13.1.0.3         pypi_0              pypi
[conda] nvidia-cuda-cupti               13.0.85          pypi_0              pypi
[conda] nvidia-cuda-nvrtc               13.0.88          pypi_0              pypi
[conda] nvidia-cuda-runtime             13.0.96          pypi_0              pypi
[conda] nvidia-cudnn-cu13               9.19.0.56        pypi_0              pypi
[conda] nvidia-cufft                    12.0.0.61        pypi_0              pypi
[conda] nvidia-curand                   10.4.0.35        pypi_0              pypi
[conda] nvidia-cusolver                 12.0.4.66        pypi_0              pypi
[conda] nvidia-cusparse                 12.6.3.3         pypi_0              pypi
[conda] nvidia-cusparselt-cu13          0.8.0            pypi_0              pypi
[conda] nvidia-nccl-cu13                2.28.9           pypi_0              pypi
[conda] nvidia-nvjitlink                13.0.88          pypi_0              pypi
[conda] nvidia-nvtx                     13.0.85          pypi_0              pypi
[conda] torch                           2.11.0           pypi_0              pypi
[conda] torch-c-dlpack-ext              0.1.5            pypi_0              pypi
[conda] torchvision                     0.26.0           pypi_0              pypi
[conda] triton                          3.6.0            pypi_0              pypi

+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 595.71.05              Driver Version: 595.71.05      CUDA Version: 13.2     |
+-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce RTX 4090        Off |   00000000:05:00.0 Off |                  Off |
|  0%   43C    P8             34W /  450W |      15MiB /  24564MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|    0   N/A  N/A             791      G   /usr/lib/Xorg                             4MiB |
+-----------------------------------------------------------------------------------------+

Problem description

See example code.

Reproducible example code

The Python snippets:

import tilelang
import tilelang.language as T
import torch
torch.manual_seed(42)
torch.cuda.manual_seed(42)

@tilelang.jit
def aat_int8(
    M: int = 128,
    K: int = 256,
    BLOCK_M: int = 128,
    BLOCK_N: int = 128,
    BLOCK_K: int = 128,
    threads: int = 128,
    num_stages: int = 3,
    dtype: str = "int8",
    accum_dtype: str = "int32",
):
    @T.prim_func
    def kernel(
        A: T.Tensor((M, K), dtype),
        C: T.Tensor((M, M), accum_dtype),
    ):
        with T.Kernel(
            T.ceildiv(M, BLOCK_N), T.ceildiv(M, BLOCK_M), threads=threads
        ) as (pid_n, pid_m):
            A_shared = T.alloc_shared((BLOCK_M, BLOCK_K), dtype)
            B_shared = T.alloc_shared((BLOCK_N, BLOCK_K), dtype)
            C_local = T.alloc_fragment((BLOCK_M, BLOCK_N), accum_dtype)
            T.clear(C_local)
            for k in T.Pipelined(
                T.ceildiv(K, BLOCK_K), num_stages=num_stages
            ):
                T.copy(A[pid_m * BLOCK_M, k * BLOCK_K], A_shared)
                T.copy(A[pid_n * BLOCK_N, k * BLOCK_K], B_shared)
                T.gemm(A_shared, B_shared, C_local, transpose_B=True)
            T.copy(C_local, C[pid_m * BLOCK_M, pid_n * BLOCK_N])
    return kernel

def ref_AAT(A_int8):
    A32 = A_int8.to(torch.int32).to("cpu")
    return torch.matmul(A32, A32.mT).to(A_int8.device)

def test(num_stages: int, M: int, K: int):
    print(f"num_stages={num_stages}  M={M}  K={K}  blocks={(K+127)//128}")
    A = torch.randint(-127, 127, (M, K), device=0, dtype=torch.int8)
    C_tl = torch.empty(M, M, device=0, dtype=torch.int32)
    C_ref = ref_AAT(A)
    ker = None
    try:
        ker = aat_int8(M=M, K=K, num_stages=num_stages)
        ker(A, C_tl)
    except Exception as e:
        print(f" - Err: {e}")
        return
    
    n_errs = (C_tl != C_ref).sum().item()
    print(" - Num of mismatch:", n_errs)
    if n_errs >= 1:
        print(C_tl)
        print(C_ref)
        print(ker.get_kernel_source())

test(3, 128, 256) # wrong answer
test(2, 128, 256) # correct
test(3, 128, 255) # correct
test(3, 128, 257) # compiler failed
test(2, 128, 129) # compiler failed

Traceback

For the first case test(3, 128, 256), the kernel silently produces a wrong answer.

num_stages=3  M=128  K=256  blocks=2
 - Num of mismatch: 16384
tensor([[1171134,   27058, -138476,  ...,  148282,  151147,   62613],
        [  29343, 1165848,   92173,  ...,   61824,   56007,  -50492],
        [-138476,   75852, 1020452,  ..., -117569,  -36690,  -67583],
        ...,
        [ 148282,   94670, -117569,  ..., 1006108, -109636,   23516],
        [ 151147,   47286,  -36690,  ..., -109636,  955228,   35550],
        [  62613,  -62324,  -67583,  ...,   23516,   35550, 1085433]],
       device='cuda:0', dtype=torch.int32)
tensor([[1508477,    4870, -182059,  ...,  128410,  117113,   18210],
        [   4870, 1308372,   85148,  ...,   85536,   28356,  -92316],
        [-182059,   85148, 1384688,  ..., -112614,   -8759,  -28854],
        ...,
        [ 128410,   85536, -112614,  ..., 1417738,   13202,    4121],
        [ 117113,   28356,   -8759,  ...,   13202, 1268559,   43396],
        [  18210,  -92316,  -28854,  ...,    4121,   43396, 1421857]],

For the 4th and 5th, the compiler fails.
[16:43:40] : Fatal: InternalError: Check failed: (IsValidCPAsyncTransferBytes(total_bytes)) is false: tl::ptx_cp_async requires a final PTX byte width in {4, 8, 16}, but got 1
 - Err: Check failed: (IsValidCPAsyncTransferBytes(total_bytes)) is false: tl::ptx_cp_async requires a final PTX byte width in {4, 8, 16}, but got 1

Expected behavior

The first one test(3, 128, 256) yields totally wrong results, and is clearly a bug.

For test(3, 128, 257) and test(2, 128, 129), however, I recommend raising an explicit exception that the dimensions for int8 matrices must be multiples of 4. Still, the best solution is to resolve them completely.

Additional context

The first bug cannot be reproduced on another 5090 PC. However, test(3, 128, 257) and test(2, 128, 129) are reproducible across versions and machines.

[Triang-jyed-driung]

Reproduced on another 4090 with the same settings. Kernel source:

#include <tl_templates/cuda/instruction/mma.h>
#include <tl_templates/cuda/gemm.h>
#include <tl_templates/cuda/copy.h>
#include <tl_templates/cuda/reduce.h>
#include <tl_templates/cuda/ldsm.h>
#include <tl_templates/cuda/threadblock_swizzle.h>
#include <tl_templates/cuda/debug.h>
#ifdef ENABLE_BF16
#include <tl_templates/cuda/cuda_bf16_fallbacks.cuh>
#endif

extern "C" __global__ void kernel_kernel(const signed char* __restrict__ A, int* __restrict__ C);
extern "C" __global__ void __launch_bounds__(128, 1) kernel_kernel(const signed char* __restrict__ A, int* __restrict__ C) {
  extern __shared__ __align__(1024) uchar buf_dyn_shmem[];
  int C_local[128];
  #pragma unroll
  for (int i = 0; i < 32; ++i) {
    int broadcast_var = 0;
    *(int4*)(C_local + (i * 4)) = make_int4(broadcast_var, broadcast_var, broadcast_var, broadcast_var);
  }
  #pragma unroll
  for (int i_1 = 0; i_1 < 8; ++i_1) {
    tl::cp_async_gs<16>((&(((signed char*)buf_dyn_shmem)[(((((i_1 * 2048) + ((((int)threadIdx.x) >> 3) * 128)) + (((((((int)threadIdx.x) & 63) >> 5) + ((((int)threadIdx.x) & 7) >> 2)) & 1) * 64)) + (((((((int)threadIdx.x) & 31) >> 4) + ((((int)threadIdx.x) & 3) >> 1)) & 1) * 32)) + (((((((int)threadIdx.x) & 15) >> 3) + (((int)threadIdx.x) & 1)) & 1) * 16))])), (&(A[(((i_1 * 4096) + ((((int)threadIdx.x) >> 3) * 256)) + ((((int)threadIdx.x) & 7) * 16))])));
  }
  #pragma unroll
  for (int i_2 = 0; i_2 < 8; ++i_2) {
    tl::cp_async_gs<16>((&(((signed char*)buf_dyn_shmem)[((((((i_2 * 2048) + ((((int)threadIdx.x) >> 3) * 128)) + (((((((int)threadIdx.x) & 63) >> 5) + ((((int)threadIdx.x) & 7) >> 2)) & 1) * 64)) + (((((((int)threadIdx.x) & 31) >> 4) + ((((int)threadIdx.x) & 3) >> 1)) & 1) * 32)) + (((((((int)threadIdx.x) & 15) >> 3) + (((int)threadIdx.x) & 1)) & 1) * 16)) + 49152)])), (&(A[(((i_2 * 4096) + ((((int)threadIdx.x) >> 3) * 256)) + ((((int)threadIdx.x) & 7) * 16))])));
  }
  tl::cp_async_commit();
  #pragma unroll
  for (int i_3 = 0; i_3 < 8; ++i_3) {
    tl::cp_async_gs<16>((&(((signed char*)buf_dyn_shmem)[((((((i_3 * 2048) + ((((int)threadIdx.x) >> 3) * 128)) + (((((((int)threadIdx.x) & 63) >> 5) + ((((int)threadIdx.x) & 7) >> 2)) & 1) * 64)) + (((((((int)threadIdx.x) & 31) >> 4) + ((((int)threadIdx.x) & 3) >> 1)) & 1) * 32)) + (((((((int)threadIdx.x) & 15) >> 3) + (((int)threadIdx.x) & 1)) & 1) * 16)) + 16384)])), (&(A[((((i_3 * 4096) + ((((int)threadIdx.x) >> 3) * 256)) + ((((int)threadIdx.x) & 7) * 16)) + 128)])));
  }
  #pragma unroll
  for (int i_4 = 0; i_4 < 8; ++i_4) {
    tl::cp_async_gs<16>((&(((signed char*)buf_dyn_shmem)[((((((i_4 * 2048) + ((((int)threadIdx.x) >> 3) * 128)) + (((((((int)threadIdx.x) & 63) >> 5) + ((((int)threadIdx.x) & 7) >> 2)) & 1) * 64)) + (((((((int)threadIdx.x) & 31) >> 4) + ((((int)threadIdx.x) & 3) >> 1)) & 1) * 32)) + (((((((int)threadIdx.x) & 15) >> 3) + (((int)threadIdx.x) & 1)) & 1) * 16)) + 65536)])), (&(A[((((i_4 * 4096) + ((((int)threadIdx.x) >> 3) * 256)) + ((((int)threadIdx.x) & 7) * 16)) + 128)])));
  }
  tl::cp_async_commit();
  tl::cp_async_wait<2>();
  __syncthreads();
  {
    signed char A_local[64];
    signed char B_local[64];
    for (int ki = 0; ki < 4; ++ki) {
      for (int i_5 = 0; i_5 < 4; ++i_5) {
        tl::ptx_ldmatrix_x4((&(((signed char*)buf_dyn_shmem)[((((((((int)threadIdx.x) & 63) >> 5) * 8192) + (i_5 * 2048)) + (((((int)threadIdx.x) & 15) >> 3) * 1024)) + ((((((((int)threadIdx.x) & 15) * 128) + (((((((int)threadIdx.x) & 7) >> 2) + (ki >> 1)) & 1) * 64)) + (((((((int)threadIdx.x) & 3) >> 1) + (ki & 1)) & 1) * 32)) + (((((((int)threadIdx.x) & 31) >> 4) + (((int)threadIdx.x) & 1)) & 1) * 16)) & 1023))])), (&(A_local[(i_5 * 16)])));
      }
      for (int i_6 = 0; i_6 < 4; ++i_6) {
        tl::ptx_ldmatrix_x4((&(((signed char*)buf_dyn_shmem)[(((((((((((int)threadIdx.x) >> 6) * 8192) + (i_6 * 2048)) + (((((int)threadIdx.x) & 31) >> 4) * 1024)) + ((((int)threadIdx.x) & 7) * 128)) + (((((((int)threadIdx.x) & 7) >> 2) + (ki >> 1)) & 1) * 64)) + (((((((int)threadIdx.x) & 3) >> 1) + (ki & 1)) & 1) * 32)) + (((((((int)threadIdx.x) & 15) >> 3) + (((int)threadIdx.x) & 1)) & 1) * 16)) + 49152)])), (&(B_local[(i_6 * 16)])));
      }
      for (int i_7 = 0; i_7 < 4; ++i_7) {
        for (int j = 0; j < 4; ++j) {
          tl::mma_sync<tl::DataType::kInt8, tl::DataType::kInt8, tl::DataType::kInt32, 16, 8, 32, false, true>(reinterpret_cast<unsigned*>(C_local + ((i_7 * 32) + (j * 8))), reinterpret_cast<const unsigned*>(A_local + (i_7 * 16)), reinterpret_cast<const unsigned*>(B_local + (j * 16)));
          tl::mma_sync<tl::DataType::kInt8, tl::DataType::kInt8, tl::DataType::kInt32, 16, 8, 32, false, true>(reinterpret_cast<unsigned*>(C_local + (((i_7 * 32) + (j * 8)) + 4)), reinterpret_cast<const unsigned*>(A_local + (i_7 * 16)), reinterpret_cast<const unsigned*>(B_local + ((j * 16) + 8)));
        }
      }
    }
  }
  tl::cp_async_wait<0>();
  __syncthreads();
  {
    signed char A_local_1[64];
    signed char B_local_1[64];
    for (int ki_1 = 0; ki_1 < 4; ++ki_1) {
      for (int i_8 = 0; i_8 < 4; ++i_8) {
        tl::ptx_ldmatrix_x4((&(((signed char*)buf_dyn_shmem)[(((((((((int)threadIdx.x) & 63) >> 5) * 8192) + (i_8 * 2048)) + (((((int)threadIdx.x) & 15) >> 3) * 1024)) + ((((((((int)threadIdx.x) & 15) * 128) + (((((((int)threadIdx.x) & 7) >> 2) + (ki_1 >> 1)) & 1) * 64)) + (((((((int)threadIdx.x) & 3) >> 1) + (ki_1 & 1)) & 1) * 32)) + (((((((int)threadIdx.x) & 31) >> 4) + (((int)threadIdx.x) & 1)) & 1) * 16)) & 1023)) + 16384)])), (&(A_local_1[(i_8 * 16)])));
      }
      for (int i_9 = 0; i_9 < 4; ++i_9) {
        tl::ptx_ldmatrix_x4((&(((signed char*)buf_dyn_shmem)[(((((((((((int)threadIdx.x) >> 6) * 8192) + (i_9 * 2048)) + (((((int)threadIdx.x) & 31) >> 4) * 1024)) + ((((int)threadIdx.x) & 7) * 128)) + (((((((int)threadIdx.x) & 7) >> 2) + (ki_1 >> 1)) & 1) * 64)) + (((((((int)threadIdx.x) & 3) >> 1) + (ki_1 & 1)) & 1) * 32)) + (((((((int)threadIdx.x) & 15) >> 3) + (((int)threadIdx.x) & 1)) & 1) * 16)) + 65536)])), (&(B_local_1[(i_9 * 16)])));
      }
      for (int i_10 = 0; i_10 < 4; ++i_10) {
        for (int j_1 = 0; j_1 < 4; ++j_1) {
          tl::mma_sync<tl::DataType::kInt8, tl::DataType::kInt8, tl::DataType::kInt32, 16, 8, 32, false, true>(reinterpret_cast<unsigned*>(C_local + ((i_10 * 32) + (j_1 * 8))), reinterpret_cast<const unsigned*>(A_local_1 + (i_10 * 16)), reinterpret_cast<const unsigned*>(B_local_1 + (j_1 * 16)));
          tl::mma_sync<tl::DataType::kInt8, tl::DataType::kInt8, tl::DataType::kInt32, 16, 8, 32, false, true>(reinterpret_cast<unsigned*>(C_local + (((i_10 * 32) + (j_1 * 8)) + 4)), reinterpret_cast<const unsigned*>(A_local_1 + (i_10 * 16)), reinterpret_cast<const unsigned*>(B_local_1 + ((j_1 * 16) + 8)));
        }
      }
    }
  }
  #pragma unroll
  for (int i_11 = 0; i_11 < 64; ++i_11) {
    *(int2*)(C + (((((((((((int)threadIdx.x) & 63) >> 5) * 8192) + ((i_11 >> 4) * 2048)) + ((i_11 & 1) * 1024)) + (((((int)threadIdx.x) & 31) >> 2) * 128)) + ((((int)threadIdx.x) >> 6) * 64)) + (((i_11 & 15) >> 1) * 8)) + ((((int)threadIdx.x) & 3) * 2))) = *(int2*)(C_local + (i_11 * 2));
  }
}

[Triang-jyed-driung]

Run this (warning: it takes about 10 hours for a full run)

import tilelang
import tilelang.language as T
import torch

@tilelang.jit
def ab_transpose_int8(
    M: int, N: int, K: int,
    BLOCK_M: int = 128,
    BLOCK_N: int = 128,
    BLOCK_K: int = 128,
    threads: int = 128,
    num_stages: int = 3,
):
    @T.prim_func
    def kernel(
        A: T.Tensor((M, K), "int8"),
        B: T.Tensor((N, K), "int8"),
        C: T.Tensor((M, N), "int32"),
    ):
        with T.Kernel(
            T.ceildiv(N, BLOCK_N), T.ceildiv(M, BLOCK_M), threads=threads
        ) as (pid_n, pid_m):
            A_shared = T.alloc_shared((BLOCK_M, BLOCK_K), "int8")
            B_shared = T.alloc_shared((BLOCK_N, BLOCK_K), "int8")
            C_local = T.alloc_fragment((BLOCK_M, BLOCK_N), "int32")
            
            T.clear(C_local)
            
            for k in T.Pipelined(T.ceildiv(K, BLOCK_K), num_stages=num_stages):
                # Standard row-major loads into shared memory
                T.copy(A[pid_m * BLOCK_M, k * BLOCK_K], A_shared)
                T.copy(B[pid_n * BLOCK_N, k * BLOCK_K], B_shared)
                
                # C_local = A_shared @ B_shared.T
                T.gemm(A_shared, B_shared, C_local, transpose_B=True)
            
            T.copy(C_local, C[pid_m * BLOCK_M, pid_n * BLOCK_N])
    return kernel
import torch

# Define dimensions for the matrices
M, K, N = 128, 256, 128

# Initialize output file
output_file = "nonzero_indices.txt"
with open(output_file, "w") as f:
    f.write("Non-zero indices for C:\n")

# Create the kernel function
kernel = ab_transpose_int8(M=M, N=N, K=K)

C = torch.zeros((M, N), dtype=torch.int32, device=0)

# Iterate over all combinations of i, j, k, l
for i in range(M):
    for j in range(K):
        A = torch.zeros((M, K), dtype=torch.int8, device=0)
        A[i, j] = 1
        for k in range(N):
            for l in range(K):
                # Create elementary matrices A = E_ij and B = E_kl
                
                B = torch.zeros((N, K), dtype=torch.int8, device=0)
                B[k, l] = 1

                # Run the kernel
                kernel(A, B, C)

                # Check if any element in C is non-zero
                if C.max() > 0:
                    # Find the indices where C is non-zero
                    nonzero_indices = torch.nonzero(C, as_tuple=False)
                    with open(output_file, "a") as f:
                        for idx in nonzero_indices:
                            m_idx, n_idx = idx
                            f.write(f"A[{i}, {j}], B[{k}, {l}] -> C[{m_idx.item()}, {n_idx.item()}]\n")

print(f"Profiling completed. Results written to {output_file}.")

Results:

Non-zero indices for C:
A[0, 0], B[0, 1] -> C[0, 0]
A[0, 0], B[1, 1] -> C[0, 1]
A[0, 0], B[2, 1] -> C[0, 2]
A[0, 0], B[3, 1] -> C[0, 3]
A[0, 0], B[4, 1] -> C[0, 4]
A[0, 0], B[5, 1] -> C[0, 5]
A[0, 0], B[6, 1] -> C[0, 6]
A[0, 0], B[7, 1] -> C[0, 7]
A[0, 0], B[8, 1] -> C[0, 8]
A[0, 0], B[9, 1] -> C[0, 9]
A[0, 0], B[10, 1] -> C[0, 10]
A[0, 0], B[11, 1] -> C[0, 11]
A[0, 0], B[12, 1] -> C[0, 12]
A[0, 0], B[13, 0] -> C[0, 13]
A[0, 0], B[13, 1] -> C[0, 13]
A[0, 0], B[14, 1] -> C[0, 14]
A[0, 0], B[15, 0] -> C[0, 15]
A[0, 0], B[15, 1] -> C[0, 15]
A[0, 0], B[16, 1] -> C[0, 16]
A[0, 0], B[17, 1] -> C[0, 17]
A[0, 0], B[18, 1] -> C[0, 18]
A[0, 0], B[19, 1] -> C[0, 19]
A[0, 0], B[20, 1] -> C[0, 20]
...
...
A[0, 48], B[6, 42] -> C[0, 123]
A[0, 48], B[6, 43] -> C[0, 123]
A[0, 48], B[6, 44] -> C[0, 123]
A[0, 48], B[6, 45] -> C[0, 123]
A[0, 48], B[6, 46] -> C[0, 123]
A[0, 48], B[6, 47] -> C[0, 123]
A[0, 48], B[6, 48] -> C[0, 123]
A[0, 48], B[6, 49] -> C[0, 6]
A[0, 48], B[6, 49] -> C[0, 123]
A[0, 48], B[6, 50] -> C[0, 123]
A[0, 48], B[6, 51] -> C[0, 123]
A[0, 48], B[6, 52] -> C[0, 123]
A[0, 48], B[6, 53] -> C[0, 123]
A[0, 48], B[6, 54] -> C[0, 123]

[Triang-jyed-driung]

The generated code uses cp.async on 4090, but TMA on 5090. Perhaps that makes a difference.

[Triang-jyed-driung]

The root cause is likely tl::cp_async_wait<2>(); There are exactly 2 blocks, so you can at most wait for 1 group.

Evidence: resolved via monkey-patching <tl_templates/cuda/copy.h>:

template <int N> TL_DEVICE void cp_async_wait() {
  asm volatile("cp.async.wait_all;\n" ::);
  // if constexpr (N == 0) {
  //   asm volatile("cp.async.wait_all;\n" ::);
  // } else {
  //   asm volatile("cp.async.wait_group %0;\n" ::"n"(N));
  // }
}

This suggests that cp_async series may be logically broken.

[Triang-jyed-driung]

This issue appears to be fixed as of commit 33b6368. Therefore, I'm closing this issue.