Files
sglang/python/sglang/srt/layers/attention/utils.py
2025-03-04 21:23:47 -08:00

40 lines
1.1 KiB
Python

import triton
import triton.language as tl
@triton.jit
def create_flashinfer_kv_indices_triton(
req_to_token_ptr, # [max_batch, max_context_len]
req_pool_indices_ptr,
page_kernel_lens_ptr,
kv_indptr,
kv_start_idx,
kv_indices_ptr,
req_to_token_ptr_stride: tl.constexpr,
):
BLOCK_SIZE: tl.constexpr = 512
pid = tl.program_id(axis=0)
req_pool_index = tl.load(req_pool_indices_ptr + pid)
kv_indices_offset = tl.load(kv_indptr + pid)
kv_start = 0
kv_end = 0
if kv_start_idx:
kv_start = tl.load(kv_start_idx + pid).to(tl.int32)
kv_end = kv_start
kv_end += tl.load(page_kernel_lens_ptr + pid).to(tl.int32)
num_loop = tl.cdiv(kv_end - kv_start, BLOCK_SIZE)
for i in range(num_loop):
offset = tl.arange(0, BLOCK_SIZE) + i * BLOCK_SIZE
mask = offset < kv_end - kv_start
data = tl.load(
req_to_token_ptr
+ req_pool_index * req_to_token_ptr_stride
+ kv_start
+ offset,
mask=mask,
)
tl.store(kv_indices_ptr + kv_indices_offset + offset, data, mask=mask)