Skip to content

vllm.models.deepseek_v41.amd.mono.common.arch

The GPU architecture the mono kernels are compiled for.

FlyDSL compiles for the architecture in the ARCH environment variable, and otherwise for the GPU it finds. The kernels were written for gfx950 (MI355X). gfx942 (MI300X, MI325X) lacks several instructions they use: the scaled f8f6f4 MFMA, the 16x16x32 bf16 MFMA, v_permlane16_swap / v_permlane32_swap and the OCP FP8 conversions. Its FP8 format is FNUZ (bias 8, no negative zero, 0x80 is NaN) instead of OCP. The helpers that differ test GFX942 and take another code path there.

Functions:

  • target_arch –

    The architecture FlyDSL compiles for, as FlyDSL's ROCm backend picks it.

target_arch()

The architecture FlyDSL compiles for, as FlyDSL's ROCm backend picks it.

Source code in vllm/models/deepseek_v41/amd/mono/common/arch.py 
19
20
21
22
def target_arch() -> str:
    """The architecture FlyDSL compiles for, as FlyDSL's ROCm backend picks it."""
    arch = os.environ.get("ARCH") or get_rocm_arch()
    return arch.lower().split(":")[0]