[Perf] Kimi K3 nvfp4 Align in_proj weights by 128 to avoid elementwise copy (#55242)

Signed-off-by: wzhao18 <wzhao18.sz@gmail.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Wei Zhao
2026-09-05 10:46:47 +08:00
committed by GitHub
co-authored by OpenAI Codex
parent 385ba6b5b0
commit 16328c7a77
6 changed files with 53 additions and 21 deletions
@@ -22,8 +22,11 @@ from cutlass.utils import get_smem_capacity_in_bytes
from vllm.cute_utils import _tcgen05, mbarrier, simple_tma_copy, to_cta0_smem
from vllm.distributed import get_tp_group
from vllm.logger import init_logger
from vllm.model_executor.layers.linear import LinearBase, UnquantizedLinearMethod
logger = init_logger(__name__)
@dsl_user_op
def nanosleep(ns: int, *, loc=None, ip=None) -> None:
@@ -779,6 +782,13 @@ class GemmRsAr:
and w.is_contiguous()
)
def warn_incompatible_projection(self) -> None:
logger.warning_once(
"Some projections are incompatible with GEMM-RS/AR; using the "
"unfused path instead.",
scope="global",
)
def should_run(self, x: torch.Tensor) -> bool:
# Use the same threshold for RS and AR for now. Small-M shapes are
# supported but faster on the existing LL path.