chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

MLA prefill w/o weight absorption (#2349)

Browse Source
This commit is contained in:
Ke Bao
2024-12-05 01:50:28 +08:00
committed by GitHub
parent eb0c1f5373
commit ec52464dde
8 changed files with 166 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/__init__.py
+5 -2
View File
@@ -52,12 +52,13 @@ class AttentionBackend(ABC):
v: torch.Tensor,
layer: RadixAttention,
forward_batch: ForwardBatch,
save_kv_cache: bool = True,
):
"""Run forward on an attention layer."""
if forward_batch.forward_mode.is_decode():
return self.forward_decode(q, k, v, layer, forward_batch)
return self.forward_decode(q, k, v, layer, forward_batch, save_kv_cache)
else:
return self.forward_extend(q, k, v, layer, forward_batch)
return self.forward_extend(q, k, v, layer, forward_batch, save_kv_cache)
def forward_decode(
self,
@@ -66,6 +67,7 @@ class AttentionBackend(ABC):
v: torch.Tensor,
layer: RadixAttention,
forward_batch: ForwardBatch,
save_kv_cache: bool = True,
):
"""Run a forward for decode."""
raise NotImplementedError()
@@ -77,6 +79,7 @@ class AttentionBackend(ABC):
v: torch.Tensor,
layer: RadixAttention,
forward_batch: ForwardBatch,
save_kv_cache: bool = True,
):
"""Run a forward for extend."""
raise NotImplementedError()
python/sglang/srt/layers/attention/double_sparsity_backend.py
+22 -8
View File
@@ -165,7 +165,13 @@ class DoubleSparseAttnBackend(AttentionBackend):
return 1
def forward_extend(
self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
self,
q,
k,
v,
layer: RadixAttention,
forward_batch: ForwardBatch,
save_kv_cache=True,
):
# TODO: reuse the buffer across layers
if layer.qk_head_dim != layer.v_head_dim:
@@ -181,9 +187,10 @@ class DoubleSparseAttnBackend(AttentionBackend):
.expand(k.shape[0], -1, -1),
)
forward_batch.token_to_kv_pool.set_kv_buffer(
layer, forward_batch.out_cache_loc, k, v, k_label
)
if save_kv_cache:
forward_batch.token_to_kv_pool.set_kv_buffer(
layer, forward_batch.out_cache_loc, k, v, k_label
)
(
start_loc,
@@ -212,7 +219,13 @@ class DoubleSparseAttnBackend(AttentionBackend):
return o
def forward_decode(
self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
self,
q,
k,
v,
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.
@@ -242,9 +255,10 @@ class DoubleSparseAttnBackend(AttentionBackend):
.expand(k.shape[0], -1, -1),
)
forward_batch.token_to_kv_pool.set_kv_buffer(
layer, forward_batch.out_cache_loc, k, v, k_label
)
if save_kv_cache:
forward_batch.token_to_kv_pool.set_kv_buffer(
layer, forward_batch.out_cache_loc, k, v, k_label
)
# NOTE(Andy) shouldn't be used when max_len_in_batch < heavy_token_num
# and set a minimum value for sparse_decode
python/sglang/srt/layers/attention/flashinfer_backend.py
+20 -5
View File
@@ -221,7 +221,13 @@ class FlashInferAttnBackend(AttentionBackend):
return 0
def forward_extend(
self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
self,
q,
k,
v,
layer: RadixAttention,
forward_batch: ForwardBatch,
save_kv_cache=True,
):
prefill_wrapper_paged = self.prefill_wrappers_paged[
self._get_wrapper_idx(layer)
@@ -237,7 +243,8 @@ class FlashInferAttnBackend(AttentionBackend):
if not use_ragged:
if k is not None:
assert v is not None
forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
if save_kv_cache:
forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
o = prefill_wrapper_paged.forward(
q.contiguous().view(-1, layer.tp_q_head_num, layer.head_dim),
@@ -270,12 +277,19 @@ class FlashInferAttnBackend(AttentionBackend):
o, _ = merge_state(o1, s1, o2, s2)
forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
if save_kv_cache:
forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
return o.view(-1, layer.tp_q_head_num * layer.head_dim)
def forward_decode(
self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
self,
q,
k,
v,
layer: RadixAttention,
forward_batch: ForwardBatch,
save_kv_cache=True,
):
decode_wrapper = self.forward_metadata[0][self._get_wrapper_idx(layer)]
cache_loc = (
@@ -286,7 +300,8 @@ class FlashInferAttnBackend(AttentionBackend):
if k is not None:
assert v is not None
forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
if save_kv_cache:
forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
o = decode_wrapper.forward(
q.contiguous().view(-1, layer.tp_q_head_num, layer.head_dim),
python/sglang/srt/layers/attention/torch_native_backend.py
+22 -8
View File
@@ -216,16 +216,23 @@ class TorchNativeAttnBackend(AttentionBackend):
return output
def forward_extend(
self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
self,
q,
k,
v,
layer: RadixAttention,
forward_batch: ForwardBatch,
save_kv_cache=True,
):
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)
forward_batch.token_to_kv_pool.set_kv_buffer(
layer, forward_batch.out_cache_loc, k, v
)
if save_kv_cache:
forward_batch.token_to_kv_pool.set_kv_buffer(
layer, forward_batch.out_cache_loc, k, v
)
use_gqa = layer.tp_q_head_num != layer.tp_k_head_num
@@ -249,7 +256,13 @@ class TorchNativeAttnBackend(AttentionBackend):
return o
def forward_decode(
self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
self,
q,
k,
v,
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.
@@ -260,9 +273,10 @@ class TorchNativeAttnBackend(AttentionBackend):
else:
o = torch.empty_like(q)
forward_batch.token_to_kv_pool.set_kv_buffer(
layer, forward_batch.out_cache_loc, k, v
)
if save_kv_cache:
forward_batch.token_to_kv_pool.set_kv_buffer(
layer, forward_batch.out_cache_loc, k, v
)
use_gqa = layer.tp_q_head_num != layer.tp_k_head_num
python/sglang/srt/layers/attention/triton_backend.py
+22 -8
View File
@@ -114,7 +114,13 @@ class TritonAttnBackend(AttentionBackend):
return 1
def forward_extend(
self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
self,
q,
k,
v,
layer: RadixAttention,
forward_batch: ForwardBatch,
save_kv_cache=True,
):
# TODO: reuse the buffer across layers
if layer.qk_head_dim != layer.v_head_dim:
@@ -122,9 +128,10 @@ class TritonAttnBackend(AttentionBackend):
else:
o = torch.empty_like(q)
forward_batch.token_to_kv_pool.set_kv_buffer(
layer, forward_batch.out_cache_loc, k, v
)
if save_kv_cache:
forward_batch.token_to_kv_pool.set_kv_buffer(
layer, forward_batch.out_cache_loc, k, v
)
start_loc, attn_logits, max_seq_len, max_extend_len = self.forward_metadata
self.extend_attention_fwd(
@@ -146,7 +153,13 @@ class TritonAttnBackend(AttentionBackend):
return o
def forward_decode(
self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
self,
q,
k,
v,
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.
@@ -160,9 +173,10 @@ class TritonAttnBackend(AttentionBackend):
start_loc, attn_logits, max_seq_len, max_extend_len = self.forward_metadata
forward_batch.token_to_kv_pool.set_kv_buffer(
layer, forward_batch.out_cache_loc, k, v
)
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),
python/sglang/srt/layers/attention/triton_ops/extend_attention.py
+3
View File
@@ -284,6 +284,9 @@ def extend_attention_fwd(
elif Lq == 288:
BLOCK_DMODEL = 256
BLOCK_DPE = 32
elif Lq == 192:
BLOCK_DMODEL = 128
BLOCK_DPE = 64
else:
BLOCK_DMODEL = triton.next_power_of_2(Lq)
BLOCK_DPE = 0
python/sglang/srt/layers/radix_attention.py
+4 -2
View File
@@ -48,11 +48,13 @@ class RadixAttention(nn.Module):
self.sliding_window_size = sliding_window_size or -1
self.is_cross_attention = is_cross_attention
def forward(self, q, k, v, forward_batch: ForwardBatch):
def forward(self, q, k, v, forward_batch: ForwardBatch, save_kv_cache=True):
if k is not None:
# For cross-layer sharing, kv can be None
assert v is not None
k = k.view(-1, self.tp_k_head_num, self.qk_head_dim)
v = v.view(-1, self.tp_v_head_num, self.v_head_dim)
return forward_batch.attn_backend.forward(q, k, v, self, forward_batch)
return forward_batch.attn_backend.forward(
q, k, v, self, forward_batch, save_kv_cache
)