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

Fix flashinfer (#700)

Browse Source
This commit is contained in:
Ying Sheng
2024-07-23 01:27:01 -07:00
committed by GitHub
parent 9fdea29d05
commit cf99eab7d5
2 changed files with 32 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

47
python/sglang/srt/layers/radix_attention.py
View File

@@ -85,32 +85,45 @@ class RadixAttention(nn.Module):
return o
def extend_forward_flashinfer(self, q, k, v, input_metadata: InputMetadata):
o1, s1 = input_metadata.flashinfer_prefill_wrapper_ragged.forward_return_lse(
q.contiguous().view(-1, self.tp_q_head_num, self.head_dim),
k.contiguous().view(-1, self.tp_k_head_num, self.head_dim),
v.contiguous().view(-1, self.tp_v_head_num, self.head_dim),
causal=True,
sm_scale=self.scaling,
logits_soft_cap=self.logit_cap,
)
self.store_kv_cache(k, v, input_metadata)
if input_metadata.extend_no_prefix:
o = o1
else:
o2, s2 = input_metadata.flashinfer_prefill_wrapper_paged.forward_return_lse(
if input_metadata.total_num_tokens <= 4096:
o = input_metadata.flashinfer_prefill_wrapper_paged.forward(
q.contiguous().view(-1, self.tp_q_head_num, self.head_dim),
input_metadata.token_to_kv_pool.get_kv_buffer(self.layer_id),
causal=False,
causal=True,
sm_scale=self.scaling,
logits_soft_cap=self.logit_cap,
)
else:
o1, s1 = (
input_metadata.flashinfer_prefill_wrapper_ragged.forward_return_lse(
q.contiguous().view(-1, self.tp_q_head_num, self.head_dim),
k.contiguous().view(-1, self.tp_k_head_num, self.head_dim),
v.contiguous().view(-1, self.tp_v_head_num, self.head_dim),
causal=True,
sm_scale=self.scaling,
logits_soft_cap=self.logit_cap,
)
)
o, _ = merge_state(o1, s1, o2, s2)
if input_metadata.extend_no_prefix:
o = o1
else:
o2, s2 = (
input_metadata.flashinfer_prefill_wrapper_paged.forward_return_lse(
q.contiguous().view(-1, self.tp_q_head_num, self.head_dim),
input_metadata.token_to_kv_pool.get_kv_buffer(self.layer_id),
causal=False,
sm_scale=self.scaling,
logits_soft_cap=self.logit_cap,
)
)
self.store_kv_cache(k, v, input_metadata)
o, _ = merge_state(o1, s1, o2, s2)
if input_metadata.total_num_tokens >= global_config.layer_sync_threshold:
torch.cuda.synchronize()
if input_metadata.total_num_tokens >= global_config.layer_sync_threshold:
torch.cuda.synchronize()
return o.view(-1, self.tp_q_head_num * self.head_dim)

3
python/sglang/srt/managers/controller/infer_batch.py
View File

@@ -829,8 +829,9 @@ def init_flashinfer_args(
num_kv_heads = model_runner.model_config.get_num_kv_heads(model_runner.tp_size)
head_dim = model_runner.model_config.head_dim
batch_size = len(req_pool_indices)
total_num_tokens = int(torch.sum(seq_lens))
if forward_mode == ForwardMode.DECODE:
if forward_mode == ForwardMode.DECODE or total_num_tokens <= 4096:
paged_kernel_lens = seq_lens
else:
paged_kernel_lens = prefix_lens