EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[Fix] Window attention compatible with RadixAttention and chunked prefill (#1112)

Browse Source
This commit is contained in:
Ying Sheng
2024-08-15 10:33:20 -07:00
committed by GitHub
parent 9195d1362a
commit 93d4e354d8
5 changed files with 37 additions and 56 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
python/sglang/srt/layers/radix_attention.py
View File

@@ -15,6 +15,8 @@ limitations under the License.
"""Radix attention.""" """Radix attention."""
from typing import Optional
import torch import torch
from flashinfer.cascade import merge_state from flashinfer.cascade import merge_state
from torch import nn from torch import nn
@@ -34,8 +36,7 @@ class RadixAttention(nn.Module):
scaling: float, scaling: float,
num_kv_heads: int, num_kv_heads: int,
layer_id: int, layer_id: int,
reuse: bool = False, sliding_window_size: Optional[int] = None,
sliding_window_size: int = -1,
logit_cap: int = -1, logit_cap: int = -1,
v_head_dim: int = -1, v_head_dim: int = -1,
): ):
@@ -48,8 +49,7 @@ class RadixAttention(nn.Module):
self.v_head_dim = v_head_dim if v_head_dim != -1 else head_dim self.v_head_dim = v_head_dim if v_head_dim != -1 else head_dim
self.scaling = scaling self.scaling = scaling
self.layer_id = layer_id self.layer_id = layer_id
self.reuse = reuse self.sliding_window_size = sliding_window_size if sliding_window_size else -1
self.sliding_window_size = sliding_window_size
if ( if (
not global_server_args_dict.get("disable_flashinfer", False) not global_server_args_dict.get("disable_flashinfer", False)
@@ -118,16 +118,16 @@ class RadixAttention(nn.Module):
def extend_forward_flashinfer(self, q, k, v, input_metadata: InputMetadata): def extend_forward_flashinfer(self, q, k, v, input_metadata: InputMetadata):
# using two wrappers is unnecessary in the current PR, but are prepared for future PRs # using two wrappers is unnecessary in the current PR, but are prepared for future PRs
prefill_wrapper_ragged = input_metadata.flashinfer_prefill_wrapper_ragged
prefill_wrapper_paged = input_metadata.flashinfer_prefill_wrapper_paged prefill_wrapper_paged = input_metadata.flashinfer_prefill_wrapper_paged
if self.sliding_window_size != -1 or self.reuse: if self.sliding_window_size != -1:
prefill_wrapper_paged = prefill_wrapper_paged[0] prefill_wrapper_paged = prefill_wrapper_paged[0]
else: else:
if isinstance(prefill_wrapper_paged, list): if isinstance(prefill_wrapper_paged, list):
prefill_wrapper_paged = prefill_wrapper_paged[1] prefill_wrapper_paged = prefill_wrapper_paged[1]
if not input_metadata.flashinfer_use_ragged or self.reuse: if not input_metadata.flashinfer_use_ragged:
if not self.reuse: if k is not None:
assert v is not None
self.store_kv_cache(k, v, input_metadata) self.store_kv_cache(k, v, input_metadata)
o = prefill_wrapper_paged.forward( o = prefill_wrapper_paged.forward(
@@ -139,21 +139,20 @@ class RadixAttention(nn.Module):
logits_soft_cap=self.logit_cap, logits_soft_cap=self.logit_cap,
) )
else: else:
o1, s1 = prefill_wrapper_ragged.forward_return_lse( o1, s1 = (
q.contiguous().view(-1, self.tp_q_head_num, self.head_dim), input_metadata.flashinfer_prefill_wrapper_ragged.forward_return_lse(
k.contiguous().view(-1, self.tp_k_head_num, self.head_dim), q.contiguous().view(-1, self.tp_q_head_num, self.head_dim),
v.contiguous().view(-1, self.tp_v_head_num, self.head_dim), k.contiguous().view(-1, self.tp_k_head_num, self.head_dim),
causal=True, v.contiguous().view(-1, self.tp_v_head_num, self.head_dim),
sm_scale=self.scaling, causal=True,
window_left=self.sliding_window_size, sm_scale=self.scaling,
logits_soft_cap=self.logit_cap, logits_soft_cap=self.logit_cap,
)
) )
if input_metadata.extend_no_prefix: if input_metadata.extend_no_prefix:
o = o1 o = o1
else: else:
# TODO window attention + radix attention will come up in next PR
assert self.sliding_window_size == -1
o2, s2 = prefill_wrapper_paged.forward_return_lse( o2, s2 = prefill_wrapper_paged.forward_return_lse(
q.contiguous().view(-1, self.tp_q_head_num, self.head_dim), q.contiguous().view(-1, self.tp_q_head_num, self.head_dim),
input_metadata.token_to_kv_pool.get_kv_buffer(self.layer_id), input_metadata.token_to_kv_pool.get_kv_buffer(self.layer_id),
@@ -179,7 +178,8 @@ class RadixAttention(nn.Module):
if isinstance(decode_wrapper, list): if isinstance(decode_wrapper, list):
decode_wrapper = decode_wrapper[1] decode_wrapper = decode_wrapper[1]
if not self.reuse: if k is not None:
assert v is not None
self.store_kv_cache(k, v, input_metadata) self.store_kv_cache(k, v, input_metadata)
o = decode_wrapper.forward( o = decode_wrapper.forward(

37
python/sglang/srt/model_executor/forward_batch_info.py
View File

@@ -194,6 +194,7 @@ class InputMetadata:
if ( if (
forward_mode != ForwardMode.DECODE forward_mode != ForwardMode.DECODE
and int(torch.sum(ret.seq_lens)) > 4096 and int(torch.sum(ret.seq_lens)) > 4096
and model_runner.sliding_window_size is None
): ):
flashinfer_use_ragged = True flashinfer_use_ragged = True
ret.init_flashinfer_handlers( ret.init_flashinfer_handlers(
@@ -322,22 +323,25 @@ def update_flashinfer_indices(
1, 1,
) )
else: else:
# window attention use paged only
kv_last_page_len = torch.ones((batch_size,), dtype=torch.int32, device="cuda") kv_last_page_len = torch.ones((batch_size,), dtype=torch.int32, device="cuda")
for wrapper_id in range(2): for wrapper_id in range(2):
if flashinfer_use_ragged and wrapper_id == 1: if wrapper_id == 0:
# full attention use ragged+paged if forward_mode == ForwardMode.DECODE:
paged_kernel_lens = prefix_lens paged_kernel_lens = torch.minimum(
seq_lens, torch.tensor(model_runner.sliding_window_size + 1)
)
else:
paged_kernel_lens = torch.minimum(
seq_lens,
torch.tensor(model_runner.sliding_window_size)
+ seq_lens
- prefix_lens,
)
else: else:
# window attention use paged only
paged_kernel_lens = seq_lens paged_kernel_lens = seq_lens
if wrapper_id == 0 and forward_mode == ForwardMode.DECODE: kv_start_idx = seq_lens - paged_kernel_lens
paged_kernel_lens = torch.minimum(
paged_kernel_lens, torch.tensor(model_runner.sliding_window_size)
)
kv_start_idx = seq_lens - paged_kernel_lens
else:
kv_start_idx = torch.zeros(batch_size, dtype=torch.int32, device="cuda")
kv_indptr = torch.zeros((batch_size + 1,), dtype=torch.int32, device="cuda") kv_indptr = torch.zeros((batch_size + 1,), dtype=torch.int32, device="cuda")
kv_indptr[1:] = torch.cumsum(paged_kernel_lens, dim=0) kv_indptr[1:] = torch.cumsum(paged_kernel_lens, dim=0)
@@ -376,17 +380,6 @@ def update_flashinfer_indices(
) )
qo_indptr[1:] = torch.cumsum(seq_lens - prefix_lens, dim=0) qo_indptr[1:] = torch.cumsum(seq_lens - prefix_lens, dim=0)
if flashinfer_use_ragged and wrapper_id == 1:
model_runner.flashinfer_prefill_wrapper_ragged.end_forward()
model_runner.flashinfer_prefill_wrapper_ragged.begin_forward(
qo_indptr,
qo_indptr,
num_qo_heads,
num_kv_heads,
head_dim,
)
# cached part
model_runner.flashinfer_prefill_wrapper_paged[wrapper_id].end_forward() model_runner.flashinfer_prefill_wrapper_paged[wrapper_id].end_forward()
model_runner.flashinfer_prefill_wrapper_paged[wrapper_id].begin_forward( model_runner.flashinfer_prefill_wrapper_paged[wrapper_id].begin_forward(
qo_indptr, qo_indptr,

6
python/sglang/srt/model_executor/model_runner.py
View File

@@ -334,11 +334,7 @@ class ModelRunner:
dtype=torch.uint8, dtype=torch.uint8,
device="cuda", device="cuda",
) )
self.flashinfer_prefill_wrapper_ragged = ( self.flashinfer_prefill_wrapper_ragged = None
BatchPrefillWithRaggedKVCacheWrapper(
self.flashinfer_workspace_buffer, "NHD"
)
)
self.flashinfer_prefill_wrapper_paged = [] self.flashinfer_prefill_wrapper_paged = []
self.flashinfer_decode_wrapper = [] self.flashinfer_decode_wrapper = []
for i in range(2): for i in range(2):

2
python/sglang/srt/models/gemma2.py
View File

@@ -213,7 +213,7 @@ class Gemma2Attention(nn.Module):
self.scaling, self.scaling,
num_kv_heads=self.num_kv_heads, num_kv_heads=self.num_kv_heads,
layer_id=layer_idx, layer_id=layer_idx,
sliding_window_size=get_window_size(config) if use_sliding_window else -1, sliding_window_size=get_window_size(config) if use_sliding_window else None,
logit_cap=self.config.attn_logit_softcapping, logit_cap=self.config.attn_logit_softcapping,
) )

10
python/sglang/srt/server_args.py
View File

@@ -450,16 +450,8 @@ class ServerArgs:
self.dp_size > 1 and self.node_rank is not None self.dp_size > 1 and self.node_rank is not None
), "multi-node data parallel is not supported" ), "multi-node data parallel is not supported"
if "gemma-2" in self.model_path.lower(): if "gemma-2" in self.model_path.lower():
logger.info( logger.info(f"When using sliding window in gemma-2, turn on flashinfer.")
f"When using sliding window in gemma-2, disable radix_cache, regex_jump_forward, and turn on flashinfer."
)
# FIXME: compatibility with radix attention
self.disable_radix_cache = True
# FIXME: compatibility with jump forward
self.disable_regex_jump_forward = True
self.disable_flashinfer = False self.disable_flashinfer = False
# FIXME: compatibility with chunked prefill
self.chunked_prefill_size = None
@dataclasses.dataclass @dataclasses.dataclass