Skip to content

Support MLA_CACHE & Fix V1_Schedule Bug #4318

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 2 commits into
base: develop
Choose a base branch
from
Open

Support MLA_CACHE & Fix V1_Schedule Bug #4318

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a145665
Support MLA_CACHE & Fix V1_Schedule Bug
chang-wenbin Sep 29, 2025
51f924f
merge develop
chang-wenbin Sep 29, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 1 addition & 2 deletions fastdeploy/engine/sched/resource_manager_v1.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -460,8 +460,7 @@ def schedule(self):
# Prepare decoding task
scheduled_reqs.append(self._prepare_decode_task(request))
num_decoding_req_nums += 1
token_budget -= 1

token_budget -= 1
if (
request.use_extend_tables
and request.request_id not in self.using_extend_tables_req_id
Expand Down
52 changes: 40 additions & 12 deletions fastdeploy/worker/gpu_model_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1192,31 +1192,48 @@ def initialize_kv_cache(self, profile: bool = False) -> None:

logger.info(f"Initializing kv cache for all layers. {cache_ready_signal.value}")
cache_kvs_list = []

# NOTE:(changwenbin) Determine whether it is Multi-Head Latent Attention,
# To rationalize the allocation of kvcache.
from fastdeploy import envs

self.mla_cache = envs.FD_ATTENTION_BACKEND == "MLA_ATTN"
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

这里是用 MLA 的模型自动设置此环境变量，还是需要手动设置？

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

目前是启动脚本手动设置 export FD_ATTENTION_BACKEND="MLA_ATTN"，
后面会根据config.json中的model_type 自动设置backend,这项修改计划和mla默认开启tensor_core一起提交。

for i in range(self.model_config.num_hidden_layers):
key_cache_name = f"key_caches_{i}_rank{local_rank}.device{self.device_id}"
val_cache_name = f"value_caches_{i}_rank{local_rank}.device{self.device_id}"
if not self.mla_cache:
val_cache_name = f"value_caches_{i}_rank{local_rank}.device{self.device_id}"
if create_cache_tensor:
logger.info(f"..creating kv cache for layer {i}: {kv_cache_shape}")
key_cache = paddle.full(shape=kv_cache_shape, fill_value=0, dtype=cache_type)
val_cache = paddle.full(shape=kv_cache_shape, fill_value=0, dtype=cache_type)
set_data_ipc(key_cache, key_cache_name)
set_data_ipc(val_cache, val_cache_name)
cache_kvs_list.extend([key_cache, val_cache])
if not self.mla_cache:
val_cache = paddle.full(shape=kv_cache_shape, fill_value=0, dtype=cache_type)
set_data_ipc(val_cache, val_cache_name)
cache_kvs_list.extend([key_cache, val_cache])
else:
cache_kvs_list.extend([key_cache])
if kv_cache_quant_type == "block_wise_fp8":
key_cache_scales = paddle.full(
shape=kv_cache_scale_shape, fill_value=0, dtype=paddle.get_default_dtype()
)
val_cache_scales = paddle.full(
shape=kv_cache_scale_shape, fill_value=0, dtype=paddle.get_default_dtype()
)
cache_kvs_list.extend([key_cache_scales, val_cache_scales])
if not self.mla_cache:
val_cache_scales = paddle.full(
shape=kv_cache_scale_shape, fill_value=0, dtype=paddle.get_default_dtype()
)
cache_kvs_list.extend([key_cache_scales, val_cache_scales])
else:
cache_kvs_list.extend([key_cache_scales])
else:
logger.info(f"..attaching kv cache for layer {i}: {kv_cache_shape}")
key_cache = paddle.empty(shape=[], dtype=cache_type)
val_cache = paddle.empty(shape=[], dtype=cache_type)
key_cache = share_external_data(key_cache, key_cache_name, kv_cache_shape)
val_cache = share_external_data(val_cache, val_cache_name, kv_cache_shape)
cache_kvs_list.extend([key_cache, val_cache])
if not self.mla_cache:
val_cache = paddle.empty(shape=[], dtype=cache_type)
val_cache = share_external_data(val_cache, val_cache_name, kv_cache_shape)
cache_kvs_list.extend([key_cache, val_cache])
else:
cache_kvs_list.extend([key_cache])

self.share_inputs["caches"] = cache_kvs_list

if not profile and create_cache_tensor:
Expand Down Expand Up @@ -1936,7 +1953,18 @@ def cal_theortical_kvcache(self):
if self.speculative_method in ["mtp"]
else self.model_config.num_hidden_layers
)
required_memory = byte_of_dtype * 2 * (self.cache_config.block_size * hidden_dim) * num_layers # k + v

# NOTE:(changwenbin) Determie whether it is Multi-Head Latent Attention,
# To rationalize the allocation of kvcache.
if self.mla_cache:
required_memory = (
byte_of_dtype
* (self.fd_config.model_config.kv_lora_rank + self.fd_config.model_config.qk_rope_head_dim)
* (self.cache_config.block_size)
* num_layers
) # compress_kv + k_pe
else:
required_memory = byte_of_dtype * 2 * (self.cache_config.block_size * hidden_dim) * num_layers # k + v
return required_memory

def not_need_stop(self) -> bool:
Expand Down
Loading