gfx950
A GPU core codename in AMD's CDNA architecture, commonly found in MI300 series accelerator cards, whose compute units and memory subsystem are optimized for matrix operations.
Timeline (last 90 days)
vLLM v0.33.0rc2 includes a bugfix commit (#64567) from AMD engineer Micah Williamson that skips the AITER-tuned GEMM path on the gfx950 architecture when M*N>=2^31.
gfx950 is the codename of a GPU core in AMD's CDNA architecture, commonly found in the MI300 series accelerator cards.
The patch detects on gfx950 whether M*N reaches the 2^31 threshold and, if so, bypasses the AITER-tuned GEMM implementation and falls back to the generic path.
Tying the 2^31 check to the gfx950 architecture indicates that the problem may be related to specific instructions of that architecture or to the kernel implementation AITER generates for it, rather than being universal to all ROCm devices.
Upgrading to v0.33.0rc2 or an official release containing this fix can avoid potential correctness issues with very large GEMMs on gfx950.
All Facts (5)
vLLM v0.33.0rc2 includes a bugfix commit (#64567) from AMD engineer Micah Williamson that skips the AITER-tuned GEMM path on the gfx950 architecture when M*N>=2^31.
50%Unverifiedgfx950 is the codename of a GPU core in AMD's CDNA architecture, commonly found in the MI300 series accelerator cards.
50%UnverifiedThe patch detects on gfx950 whether M*N reaches the 2^31 threshold and, if so, bypasses the AITER-tuned GEMM implementation and falls back to the generic path.
50%UnverifiedTying the 2^31 check to the gfx950 architecture indicates that the problem may be related to specific instructions of that architecture or to the kernel implementation AITER generates for it, rather than being universal to all ROCm devices.
50%UnverifiedUpgrading to v0.33.0rc2 or an official release containing this fix can avoid potential correctness issues with very large GEMMs on gfx950.
50%