triton_backend.py

from __future__ import annotations

from typing import TYPE_CHECKING, Optional

import torch

from sglang.srt.layers.attention import AttentionBackend
from sglang.srt.layers.attention.flashinfer_backend import (
    create_flashinfer_kv_indices_triton,
)
from sglang.srt.layers.dp_attention import get_attention_tp_size
from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode

if TYPE_CHECKING:
    from sglang.srt.layers.radix_attention import RadixAttention
    from sglang.srt.model_executor.model_runner import ModelRunner
    from sglang.srt.speculative.spec_info import SpecInfo


class TritonAttnBackend(AttentionBackend):
    def __init__(self, model_runner: ModelRunner):
        # Lazy import to avoid the initialization of cuda context
        from sglang.srt.layers.attention.triton_ops.decode_attention import (
            decode_attention_fwd,
        )
        from sglang.srt.layers.attention.triton_ops.extend_attention import (
            extend_attention_fwd,
        )

        super().__init__()

        self.decode_attention_fwd = decode_attention_fwd
        self.extend_attention_fwd = extend_attention_fwd

        max_bs = model_runner.req_to_token_pool.size
        self.kv_indptr = torch.zeros(
            (max_bs + 1,), dtype=torch.int32, device=model_runner.device
        )
        self.req_to_token = model_runner.req_to_token_pool.req_to_token
        self.qo_indptr = torch.zeros(
            (max_bs + 1,), dtype=torch.int32, device=model_runner.device
        )

        self.num_head = (
            model_runner.model_config.num_attention_heads // get_attention_tp_size()
        )

        self.num_kv_splits = model_runner.server_args.triton_attention_num_kv_splits
        self.v_head_dim = model_runner.token_to_kv_pool.get_value_buffer(0).shape[-1]

        self.forward_metadata = None

        self.cuda_graph_max_seq_len = model_runner.model_config.context_len

        self.device = model_runner.device

    def init_forward_metadata(self, forward_batch: ForwardBatch):
        """Init auxiliary variables for triton attention backend."""

        bs = forward_batch.batch_size
        kv_indptr = self.kv_indptr

        if forward_batch.forward_mode.is_decode():
            attn_logits = torch.empty(
                (
                    forward_batch.batch_size,
                    self.num_head,
                    self.num_kv_splits,
                    self.v_head_dim + 1,
                ),
                dtype=torch.float32,
                device=self.device,
            )

            max_extend_len = None

            kv_indptr[1 : bs + 1] = torch.cumsum(forward_batch.seq_lens, dim=0)
            kv_indptr = kv_indptr[: bs + 1]
            kv_indices = torch.empty(
                forward_batch.seq_lens_sum, dtype=torch.int32, device=self.device
            )
            create_flashinfer_kv_indices_triton[(bs,)](
                self.req_to_token,
                forward_batch.req_pool_indices,
                forward_batch.seq_lens,
                kv_indptr,
                None,
                kv_indices,
                self.req_to_token.stride(0),
            )

            qo_indptr = None
            custom_mask = None
            mask_offsets = None
        else:
            kv_indptr[1 : bs + 1] = torch.cumsum(
                forward_batch.extend_prefix_lens, dim=0
            )
            kv_indptr = kv_indptr[: bs + 1]
            kv_indices = torch.empty(
                forward_batch.extend_prefix_lens.sum().item(),
                dtype=torch.int32,
                device=self.device,
            )
            create_flashinfer_kv_indices_triton[(bs,)](
                self.req_to_token,
                forward_batch.req_pool_indices,
                forward_batch.extend_prefix_lens,
                kv_indptr,
                None,
                kv_indices,
                self.req_to_token.stride(0),
            )

            qo_indptr = self.qo_indptr
            qo_indptr[1 : bs + 1] = torch.cumsum(forward_batch.extend_seq_lens, dim=0)
            qo_indptr = qo_indptr[: bs + 1]
            custom_mask = None
            mask_offsets = None

            attn_logits = None
            max_extend_len = torch.max(forward_batch.extend_seq_lens).item()

        self.forward_metadata = (
            attn_logits,
            max_extend_len,
            kv_indptr,
            kv_indices,
            qo_indptr,
            custom_mask,
            mask_offsets,
        )

    def init_cuda_graph_state(self, max_bs: int):
        self.cuda_graph_max_total_num_tokens = max_bs * self.cuda_graph_max_seq_len

        self.cuda_graph_start_loc = torch.zeros(
            (max_bs,), dtype=torch.int32, device=self.device
        )
        self.cuda_graph_attn_logits = torch.empty(
            (max_bs, self.num_head, self.num_kv_splits, self.v_head_dim + 1),
            dtype=torch.float32,
            device=self.device,
        )
        self.cuda_graph_kv_indices = torch.zeros(
            (max_bs * self.cuda_graph_max_seq_len),
            dtype=torch.int32,
            device=self.device,
        )

    def init_forward_metadata_capture_cuda_graph(
        self,
        bs: int,
        num_tokens: int,
        req_pool_indices: torch.Tensor,
        seq_lens: torch.Tensor,
        encoder_lens: Optional[torch.Tensor],
        forward_mode: ForwardMode,
        spec_info: Optional[SpecInfo],
    ):
        assert encoder_lens is None, "Not supported"
        assert forward_mode.is_decode(), "Not supported"
        assert spec_info is None, "Not supported"

        kv_indptr = self.kv_indptr
        kv_indptr[1 : bs + 1] = torch.cumsum(seq_lens, dim=0)
        kv_indptr = kv_indptr[: bs + 1]
        kv_indices = self.cuda_graph_kv_indices
        create_flashinfer_kv_indices_triton[(bs,)](
            self.req_to_token,
            req_pool_indices,
            seq_lens,
            kv_indptr,
            None,
            kv_indices,
            self.req_to_token.stride(0),
        )

        self.forward_metadata = (
            self.cuda_graph_attn_logits,
            None,
            kv_indptr,
            kv_indices,
            None,
            None,
            None,
        )

    def init_forward_metadata_replay_cuda_graph(
        self,
        bs: int,
        req_pool_indices: torch.Tensor,
        seq_lens: torch.Tensor,
        seq_lens_sum: int,
        encoder_lens: Optional[torch.Tensor],
        forward_mode: ForwardMode,
        spec_info: Optional[SpecInfo],
    ):
        # NOTE: encoder_lens expected to be zeros or None
        self.cuda_graph_start_loc.zero_()
        self.cuda_graph_start_loc[1:bs] = torch.cumsum(seq_lens[: bs - 1], dim=0)

        kv_indptr = self.kv_indptr
        kv_indptr[1 : bs + 1] = torch.cumsum(seq_lens[:bs], dim=0)
        kv_indptr = kv_indptr[: bs + 1]
        kv_indices = self.cuda_graph_kv_indices
        create_flashinfer_kv_indices_triton[(bs,)](
            self.req_to_token,
            req_pool_indices[:bs],
            seq_lens[:bs],
            kv_indptr,
            None,
            kv_indices,
            self.req_to_token.stride(0),
        )

    def get_cuda_graph_seq_len_fill_value(self):
        return 1

    def forward_extend(
        self,
        q: torch.Tensor,
        k: torch.Tensor,
        v: torch.Tensor,
        layer: RadixAttention,
        forward_batch: ForwardBatch,
        save_kv_cache=True,
    ):
        # TODO: reuse the buffer across layers
        if layer.qk_head_dim != layer.v_head_dim:
            o = q.new_empty((q.shape[0], layer.tp_q_head_num * layer.v_head_dim))
        else:
            o = torch.empty_like(q)

        if save_kv_cache:
            forward_batch.token_to_kv_pool.set_kv_buffer(
                layer, forward_batch.out_cache_loc, k, v
            )

        (
            _,
            max_extend_len,
            kv_indptr,
            kv_indices,
            qo_indptr,
            custom_mask,
            mask_offsets,
        ) = self.forward_metadata
        self.extend_attention_fwd(
            q.view(-1, layer.tp_q_head_num, layer.qk_head_dim),
            k.contiguous(),
            v.contiguous(),
            o.view(-1, layer.tp_q_head_num, layer.v_head_dim),
            forward_batch.token_to_kv_pool.get_key_buffer(layer.layer_id),
            forward_batch.token_to_kv_pool.get_value_buffer(layer.layer_id),
            qo_indptr,
            kv_indptr,
            kv_indices,
            custom_mask,
            mask_offsets,
            max_extend_len,
            layer.scaling,
            layer.logit_cap,
        )
        return o

    def forward_decode(
        self,
        q: torch.Tensor,
        k: torch.Tensor,
        v: torch.Tensor,
        layer: RadixAttention,
        forward_batch: ForwardBatch,
        save_kv_cache=True,
    ):
        # During torch.compile, there is a bug in rotary_emb that causes the
        # output value to have a 3D tensor shape. This reshapes the output correctly.
        q = q.reshape(-1, layer.tp_q_head_num * layer.qk_head_dim)

        # TODO: reuse the buffer across layers
        if layer.qk_head_dim != layer.v_head_dim:
            o = q.new_empty((q.shape[0], layer.tp_q_head_num * layer.v_head_dim))
        else:
            o = torch.empty_like(q)

        attn_logits, _, kv_indptr, kv_indices, _, _, _ = self.forward_metadata

        if save_kv_cache:
            forward_batch.token_to_kv_pool.set_kv_buffer(
                layer, forward_batch.out_cache_loc, k, v
            )

        self.decode_attention_fwd(
            q.view(-1, layer.tp_q_head_num, layer.qk_head_dim),
            forward_batch.token_to_kv_pool.get_key_buffer(layer.layer_id),
            forward_batch.token_to_kv_pool.get_value_buffer(layer.layer_id),
            o.view(-1, layer.tp_q_head_num, layer.v_head_dim),
            kv_indptr,
            kv_indices,
            attn_logits,
            self.num_kv_splits,
            layer.scaling,
            layer.logit_cap,
        )
        return o