vllm.models.mistral.amd.ops
¶
Hand-fused eager Triton kernels for Mistral models on ROCm.
Same kernels and VLLM_ML4_FUSIONS selection as the NVIDIA package: a
comma-separated subset of :data:FUSIONS, all (default) or none.
Modules:
-
indexer–DSA indexer Q/K epilogues.
-
layernorm–One-launch LayerNorm: fp32 statistics and affine, output in the input dtype.
-
ssmax–SSMax query scaling
q *= s * log(pos + 1) + b, alone or fused with RoPE.