[Kernel][Boogu-Image] Fuse Q/K RMSNorm + interleaved RoPE via fused_qk_norm_rope (#6982)

Signed-off-by: Qihan Kang <rollykanggg@gmail.com>
This commit is contained in:
KANG Qihan authored and GitHub committed 2026-09-13 09:39:27 +08:00
1 parent 7e520f9cc4
commit f7d9deb45a
8 files changed
+703 -53

No files matched your search

+6
View File
@@ -15,6 +15,7 @@ if TYPE_CHECKING:
MASTER_PORT: int | None = None
CUDA_HOME: str | None = None
LOCAL_RANK: int = 0
VLLM_OMNI_FUSED_QK_NORM_ROPE_MIN_TOKENS: str | None = None
environment_variables: dict[str, Callable[[], Any]] = {
# ================== Runtime Env Vars ==================
@@ -28,6 +29,11 @@ environment_variables: dict[str, Callable[[], Any]] = {
# local rank of the process in the distributed setting, used to determine
# the GPU device id
"LOCAL_RANK": lambda: int(os.environ.get("LOCAL_RANK", "0")),
# Minimum rotary-table span (B*S positions) at which consumers of the shared
# fused_qk_norm_rope op take the fused path instead of their eager chain;
# "0" = always fuse; unset = each consumer's own measured default. Raw
# string or None; validated by fused_qk_norm_rope_min_tokens().
"VLLM_OMNI_FUSED_QK_NORM_ROPE_MIN_TOKENS": lambda: os.environ.get("VLLM_OMNI_FUSED_QK_NORM_ROPE_MIN_TOKENS", None),
}