Skip to content

[CuTeDSL] Drop Ampere/Ada from cvt_i8_bf16_intrinsic.supported_archs - #3355

Open
Hughshine wants to merge 1 commit into
NVIDIA:mainfrom
Hughshine:fix/cutedsl-i8-bf16-arch-gate
Open

[CuTeDSL] Drop Ampere/Ada from cvt_i8_bf16_intrinsic.supported_archs#3355
Hughshine wants to merge 1 commit into
NVIDIA:mainfrom
Hughshine:fix/cutedsl-i8-bf16-arch-gate

Conversation

@Hughshine

@Hughshine Hughshine commented Jun 29, 2026

Copy link
Copy Markdown

cvt.rn.bf16.s8 requires sm_90+, so pre-Hopper targets should fall through to the existing itofp software fallback instead of emitting an instruction ptxas rejects. Matches the sibling cvt_i4_bf16_intrinsic gate.

Fixes #3354

cvt.rn.bf16.s8 requires sm_90+, so pre-Hopper targets should fall through to
the existing itofp software fallback instead of emitting an instruction ptxas
rejects. Matches the sibling cvt_i4_bf16_intrinsic gate.

See NVIDIA#3354
@Hughshine
Hughshine force-pushed the fix/cutedsl-i8-bf16-arch-gate branch from 374ec5c to a340fe6 Compare August 8, 2026 03:24
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

[BUG] cvt int8/uint8 -> bfloat16 emits a sm_90-only instruction on Ampere/Ada instead of the software fallback (arch gate too loose)

2 participants