vllm.models.muse_spark.common.ops
¶
Cross-platform MuseSpark1x operators.
Modules:
-
fused_finalize–Fused MuseSpark post-expert RMSNorm, route weighting and top-k sum.
-
fused_qk_norm_rope_kv–Muse-Spark QK-norm + RoPE + KV-cache insert as one launch.
-
fused_rmsnorm–Triton RMSNorm for MuseSpark1x, optionally fused with the gated residual add.
-
fused_routing–Triton kernel for MuseSpark1x's top-k router, one program per token.