Llama3.2 vision model support (#1551)

This commit is contained in:
Liangsheng Yin
2024-10-21 15:01:21 -07:00
committed by GitHub
parent 00611286a1
commit 94cde10920
21 changed files with 1562 additions and 122 deletions
@@ -10,6 +10,7 @@ from sglang.srt.managers.schedule_batch import global_server_args_dict
from sglang.srt.model_executor.forward_batch_info import ForwardBatch
if TYPE_CHECKING:
from sglang.srt.layers.radix_attention import RadixAttention
from sglang.srt.model_executor.model_runner import ModelRunner
@@ -134,8 +135,13 @@ class DoubleSparseAttnBackend(AttentionBackend):
)
def init_forward_metadata_capture_cuda_graph(
self, bs: int, req_pool_indices: torch.Tensor, seq_lens: torch.Tensor
self,
bs: int,
req_pool_indices: torch.Tensor,
seq_lens: torch.Tensor,
encoder_lens=None,
):
# NOTE: encoder_lens expected to be zeros or None
self.forward_metadata = (
self.cuda_graph_start_loc,
self.cuda_graph_attn_logits,
@@ -149,14 +155,18 @@ class DoubleSparseAttnBackend(AttentionBackend):
req_pool_indices: torch.Tensor,
seq_lens: torch.Tensor,
seq_lens_sum: int,
encoder_lens=None,
):
# 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)
def get_cuda_graph_seq_len_fill_value(self):
return 1
def forward_extend(self, q, k, v, layer: nn.Module, forward_batch: ForwardBatch):
def forward_extend(
self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
):
# 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))
@@ -172,7 +182,7 @@ class DoubleSparseAttnBackend(AttentionBackend):
)
forward_batch.token_to_kv_pool.set_kv_buffer(
layer.layer_id, forward_batch.out_cache_loc, k, v, k_label
layer, forward_batch.out_cache_loc, k, v, k_label
)
(
@@ -201,7 +211,9 @@ class DoubleSparseAttnBackend(AttentionBackend):
)
return o
def forward_decode(self, q, k, v, layer: nn.Module, forward_batch: ForwardBatch):
def forward_decode(
self, q, k, v, layer: RadixAttention, forward_batch: ForwardBatch
):
# 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)
@@ -231,7 +243,7 @@ class DoubleSparseAttnBackend(AttentionBackend):
)
forward_batch.token_to_kv_pool.set_kv_buffer(
layer.layer_id, forward_batch.out_cache_loc, k, v, k_label
layer, forward_batch.out_cache_loc, k, v, k_label
)
# NOTE(Andy) shouldn't be used when max_len_in_batch < heavy_token_num