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

[PD]decode: add CLIP_MAX_NEW_TOKEN for pop_preallocated (#8866)

Browse Source
This commit is contained in:
Jimmy
2025-08-12 04:08:11 +08:00
committed by GitHub
parent 3d64fda376
commit 0d9e89ec69
2 changed files with 11 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
docs/advanced_features/pd_disaggregation.md
View File

@@ -67,6 +67,7 @@ Please be aware that this setting will cause prefill instances to take a longer
| **`SGLANG_DISAGGREGATION_HEARTBEAT_INTERVAL`** | Interval (seconds) between health checks to prefill bootstrap servers | `5.0` | | **`SGLANG_DISAGGREGATION_HEARTBEAT_INTERVAL`** | Interval (seconds) between health checks to prefill bootstrap servers | `5.0` |
| **`SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE`** | Consecutive heartbeat failures before marking prefill server offline | `2` | | **`SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE`** | Consecutive heartbeat failures before marking prefill server offline | `2` |
| **`SGLANG_DISAGGREGATION_WAITING_TIMEOUT`** | Timeout (seconds) for receiving KV Cache after request initialization | `300` | | **`SGLANG_DISAGGREGATION_WAITING_TIMEOUT`** | Timeout (seconds) for receiving KV Cache after request initialization | `300` |
| **`SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION`** | Clip request param "max_tokens" to pre_allocate | `4096` |
If a greater mean TTFT is acceptable, you can `export SGLANG_DISAGGREGATION_WAITING_TIMEOUT=600` (10 minutes) to relax the timeout condition. If a greater mean TTFT is acceptable, you can `export SGLANG_DISAGGREGATION_WAITING_TIMEOUT=600` (10 minutes) to relax the timeout condition.

13
python/sglang/srt/disaggregation/decode.py
View File

@@ -51,7 +51,7 @@ from sglang.srt.mem_cache.base_prefix_cache import BasePrefixCache
from sglang.srt.mem_cache.memory_pool import KVCache, ReqToTokenPool from sglang.srt.mem_cache.memory_pool import KVCache, ReqToTokenPool
from sglang.srt.model_executor.forward_batch_info import ForwardMode from sglang.srt.model_executor.forward_batch_info import ForwardMode
from sglang.srt.torch_memory_saver_adapter import TorchMemorySaverAdapter from sglang.srt.torch_memory_saver_adapter import TorchMemorySaverAdapter
from sglang.srt.utils import require_mlp_sync from sglang.srt.utils import get_int_env_var, require_mlp_sync
logger = logging.getLogger(__name__) logger = logging.getLogger(__name__)
@@ -59,6 +59,10 @@ if TYPE_CHECKING:
from sglang.srt.managers.schedule_batch import Req from sglang.srt.managers.schedule_batch import Req
from sglang.srt.managers.scheduler import Scheduler from sglang.srt.managers.scheduler import Scheduler
DECODE_CLIP_MAX_NEW_TOKEN = get_int_env_var(
"SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION", 4096
)
class DecodeReqToTokenPool: class DecodeReqToTokenPool:
""" """
@@ -384,7 +388,10 @@ class DecodePreallocQueue:
max( max(
required_tokens_for_request, required_tokens_for_request,
origin_input_len origin_input_len
+ decode_req.req.sampling_params.max_new_tokens + min(
decode_req.req.sampling_params.max_new_tokens,
DECODE_CLIP_MAX_NEW_TOKEN,
)
- retractable_tokens, - retractable_tokens,
) )
> allocatable_tokens > allocatable_tokens
@@ -433,7 +440,7 @@ class DecodePreallocQueue:
need_space_for_single_req = ( need_space_for_single_req = (
max( max(
[ [
x.sampling_params.max_new_tokens min(x.sampling_params.max_new_tokens, DECODE_CLIP_MAX_NEW_TOKEN)
+ len(x.origin_input_ids) + len(x.origin_input_ids)
- retractable_tokens - retractable_tokens
for x in self.scheduler.running_batch.reqs for x in self.scheduler.running_batch.reqs