ba9f6d8f26
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> Co-authored-by: Xiaoyu Zhang <35585791+BBuf@users.noreply.github.com>
79 lines
2.0 KiB
Python
79 lines
2.0 KiB
Python
from __future__ import annotations
|
|
|
|
import logging
|
|
from typing import TYPE_CHECKING, Optional
|
|
|
|
import torch
|
|
|
|
from sglang.jit_kernel.utils import (
|
|
cache_once,
|
|
is_arch_support_pdl,
|
|
load_jit,
|
|
make_cpp_args,
|
|
)
|
|
|
|
if TYPE_CHECKING:
|
|
from tvm_ffi.module import Module
|
|
|
|
|
|
@cache_once
|
|
def _jit_qknorm_module(head_dim: int, dtype: torch.dtype) -> Module:
|
|
args = make_cpp_args(head_dim, is_arch_support_pdl(), dtype)
|
|
return load_jit(
|
|
"qknorm",
|
|
*args,
|
|
cuda_files=["elementwise/qknorm.cuh"],
|
|
cuda_wrappers=[("qknorm", f"QKNormKernel<{args}>::run")],
|
|
)
|
|
|
|
|
|
@cache_once
|
|
def _jit_rmsnorm_module(hidden_size: int, dtype: torch.dtype) -> Module:
|
|
args = make_cpp_args(hidden_size, is_arch_support_pdl(), dtype)
|
|
return load_jit(
|
|
"rmsnorm",
|
|
*args,
|
|
cuda_files=["elementwise/rmsnorm.cuh"],
|
|
cuda_wrappers=[("rmsnorm", f"RMSNormKernel<{args}>::run")],
|
|
)
|
|
|
|
|
|
@cache_once
|
|
def can_use_fused_inplace_qknorm(head_dim: int, dtype: torch.dtype) -> bool:
|
|
logger = logging.getLogger(__name__)
|
|
if head_dim not in [64, 128, 256, 512, 1024]:
|
|
logger.warning(f"Unsupported head_dim={head_dim} for JIT QK-Norm kernel")
|
|
return False
|
|
try:
|
|
_jit_qknorm_module(head_dim, dtype)
|
|
return True
|
|
except Exception as e:
|
|
logger.warning(f"Failed to load JIT QK-Norm kernel: {e}")
|
|
return False
|
|
|
|
|
|
def fused_inplace_qknorm(
|
|
q: torch.Tensor,
|
|
k: torch.Tensor,
|
|
q_weight: torch.Tensor,
|
|
k_weight: torch.Tensor,
|
|
eps: float = 1e-6,
|
|
*,
|
|
head_dim: int = 0,
|
|
) -> None:
|
|
head_dim = head_dim or q.size(-1)
|
|
module = _jit_qknorm_module(head_dim, q.dtype)
|
|
module.qknorm(q, k, q_weight, k_weight, eps)
|
|
|
|
|
|
def rmsnorm(
|
|
input: torch.Tensor,
|
|
weight: torch.Tensor,
|
|
output: Optional[torch.Tensor] = None,
|
|
eps: float = 1e-6,
|
|
) -> None:
|
|
output = output if output is not None else input
|
|
hidden_size = input.size(-1)
|
|
module = _jit_rmsnorm_module(hidden_size, input.dtype)
|
|
module.rmsnorm(input, weight, output, eps)
|