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

Reduce number of workspaces (#601)

Browse Source
This commit is contained in:
Mingyi
2024-07-07 19:35:22 -07:00
committed by GitHub
parent 0877f1e75b
commit f4e885b7c3
2 changed files with 6 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
python/sglang/srt/layers/radix_attention.py
View File

@@ -4,6 +4,8 @@ import numpy as np
import torch
from torch import nn
from flashinfer.cascade import merge_state
from sglang.global_config import global_config
from sglang.srt.layers.extend_attention import extend_attention_fwd
from sglang.srt.layers.token_attention import token_attention_fwd
@@ -95,8 +97,6 @@ class RadixAttention(nn.Module):
return o
def prefill_forward_flashinfer(self, q, k, v, input_metadata: InputMetadata):
self.store_kv_cache(k, v, input_metadata)
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),
@@ -117,10 +117,10 @@ class RadixAttention(nn.Module):
logits_soft_cap=self.logit_cap,
)
from flashinfer.cascade import merge_state
o, _ = merge_state(o1, s1, o2, s2)
self.store_kv_cache(k, v, input_metadata)
if input_metadata.total_num_tokens >= global_config.layer_sync_threshold:
torch.cuda.synchronize()

4
python/sglang/srt/managers/controller/model_runner.py
View File

@@ -408,7 +408,7 @@ class ModelRunner:
use_tensor_cores = False
workspace_buffers = torch.empty(
3, 96 * 1024 * 1024, dtype=torch.uint8, device="cuda"
2, 96 * 1024 * 1024, dtype=torch.uint8, device="cuda"
)
self.flashinfer_prefill_wrapper_ragged = (
BatchPrefillWithRaggedKVCacheWrapper(workspace_buffers[0], "NHD")
@@ -417,7 +417,7 @@ class ModelRunner:
workspace_buffers[1], "NHD"
)
self.flashinfer_decode_wrapper = BatchDecodeWithPagedKVCacheWrapper(
workspace_buffers[2], "NHD", use_tensor_cores=use_tensor_cores
workspace_buffers[0], "NHD", use_tensor_cores=use_tensor_cores
)
else:
self.flashinfer_prefill_wrapper_ragged = (