You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[BUG] Ascend: the deepgemm heuristic models an FP8×FP4 configuration that both entry points refuse #4379
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?
0.1.15+cuda.git210611d
System information
Ascend 950PR, CANN 9.0.0-beta.2, Linux 6.6.0-13, Python 3.12. Source build of main @ 210611d.
Problem description
// examples/ascend/deepgemm/_heuristics.pyx:161-164
elif desc.a_dtype == "float8_e4m3fn" and desc.b_dtype in ("float8_e4m3fn", "float4_e2m1fn_x2"):
...
dtype.dequant_b = desc.b_dtype == "float4_e2m1fn_x2"
// plus :111 fp4_decode_bytes_per_cycle, :496 dequant_bytes, :602-612 dequant cost
Both callers raise instead:
// api.py:71-73
raise NotImplementedError("FP8 x FP4 (FP8 dequant) kernels are not implemented")
// config.py:96-98 (same rejection)
and kernels/fp8_gemm.py has no dequant branch at all. So select_gemm_config returns a fully populated configuration (L2 controls, direct_store, version counts) for a combination that cannot be built — the caller discovers it one layer later, from a different error.
Activity
cn-sys commentedon Nov 15, 2026
/formal