apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-17 14:11:14 +08:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

[Bug fix] fix block num setting in scheduler v1 for develop (#3303)

Browse Source 
* fix block num setting in scheduler v1

* fix block num setting in scheduler v1

* fix max_block_num and max_num_batched_tokens setting

* fix max_block_num and max_num_batched_tokens setting

* fix max_block_num and max_num_batched_tokens setting

* fix max_block_num and max_num_batched_tokens setting
This commit is contained in:
chenjian
2025-08-12 10:38:51 +08:00
committed by GitHub
parent 183e3863e8
commit b21272d9ff
4 changed files with 30 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
fastdeploy/engine/args_utils.py
View File

@@ -18,6 +18,7 @@ import json
from dataclasses import asdict, dataclass
from dataclasses import fields as dataclass_fields
from typing import Any, Dict, List, Optional
import os
from fastdeploy.config import (
CacheConfig,
@@ -884,7 +885,10 @@ class EngineArgs:
if self.enable_chunked_prefill:
self.max_num_batched_tokens = 2048
else:
self.max_num_batched_tokens = self.max_model_len
if not int(os.getenv('ENABLE_V1_KVCACHE_SCHEDULER', '0')):
self.max_num_batched_tokens = self.max_model_len
else:
self.max_num_batched_tokens = 8192 # if set to max_model_len, it's easy to be OOM
all_dict = asdict(self)
all_dict["model_cfg"] = model_cfg

14
fastdeploy/engine/config.py
View File

@@ -245,7 +245,10 @@ class Config:
if self.cache_config.enable_chunked_prefill:
self.max_num_batched_tokens = 2048
else:
self.max_num_batched_tokens = self.max_model_len
if not int(os.getenv('ENABLE_V1_KVCACHE_SCHEDULER', '0')):
self.max_num_batched_tokens = self.max_model_len
else:
self.max_num_batched_tokens = 8192 # if set to max_model_len, it's easy to be OOM
if self.long_prefill_token_threshold == 0:
self.long_prefill_token_threshold = int(self.max_model_len * 0.04)
@@ -293,10 +296,11 @@ class Config:
)
if not self.cache_config.enable_chunked_prefill:
assert self.max_num_batched_tokens >= self.max_model_len, (
f"max_num_batched_tokens: {self.max_num_batched_tokens} "
f"should be larger than or equal to max_model_len: {self.max_model_len}"
)
if not int(os.getenv('ENABLE_V1_KVCACHE_SCHEDULER', '0')):
assert self.max_num_batched_tokens >= self.max_model_len, (
f"max_num_batched_tokens: {self.max_num_batched_tokens} "
f"should be larger than or equal to max_model_len: {self.max_model_len}"
)
else:
assert self.max_num_batched_tokens >= self.cache_config.block_size, (
f"max_num_batched_tokens: {self.max_num_batched_tokens} "