Skip to content
GitLab
Menu
Projects
Groups
Snippets
Loading...
Help
Help
Support
Community forum
Keyboard shortcuts
?
Submit feedback
Contribute to GitLab
Sign in / Register
Toggle navigation
Menu
Open sidebar
change
sglang
Commits
d22d0447
Unverified
Commit
d22d0447
authored
Nov 04, 2025
by
Baizhou Zhang
Committed by
GitHub
Nov 04, 2025
Browse files
Revert "Enable memory saver for hybrid model" (#12648)
parent
887742a1
Changes
3
Hide whitespace changes
Inline
Side-by-side
Showing
3 changed files
with
8 additions
and
15 deletions
+8
-15
python/sglang/srt/mem_cache/memory_pool.py
python/sglang/srt/mem_cache/memory_pool.py
+8
-12
python/sglang/srt/model_executor/model_runner.py
python/sglang/srt/model_executor/model_runner.py
+0
-1
test/srt/test_mamba_unittest.py
test/srt/test_mamba_unittest.py
+0
-2
No files found.
python/sglang/srt/mem_cache/memory_pool.py
View file @
d22d0447
...
@@ -154,19 +154,12 @@ class MambaPool:
...
@@ -154,19 +154,12 @@ class MambaPool:
size
:
int
,
size
:
int
,
cache_params
:
Union
[
"Mamba2CacheParams"
,
"KimiLinearCacheParams"
],
cache_params
:
Union
[
"Mamba2CacheParams"
,
"KimiLinearCacheParams"
],
device
:
str
,
device
:
str
,
enable_memory_saver
:
bool
,
speculative_num_draft_tokens
:
Optional
[
int
]
=
None
,
speculative_num_draft_tokens
:
Optional
[
int
]
=
None
,
):
):
conv_state_shape
=
cache_params
.
shape
.
conv
conv_state_shape
=
cache_params
.
shape
.
conv
temporal_state_shape
=
cache_params
.
shape
.
temporal
temporal_state_shape
=
cache_params
.
shape
.
temporal
conv_dtype
=
cache_params
.
dtype
.
conv
conv_dtype
=
cache_params
.
dtype
.
conv
ssm_dtype
=
cache_params
.
dtype
.
temporal
ssm_dtype
=
cache_params
.
dtype
.
temporal
self
.
size
=
size
self
.
device
=
device
self
.
free_slots
=
torch
.
arange
(
self
.
size
,
dtype
=
torch
.
int64
,
device
=
self
.
device
)
self
.
memory_saver_adapter
=
TorchMemorySaverAdapter
.
create
(
enable
=
enable_memory_saver
)
num_mamba_layers
=
len
(
cache_params
.
layers
)
num_mamba_layers
=
len
(
cache_params
.
layers
)
# for disagg with nvlink
# for disagg with nvlink
...
@@ -181,8 +174,9 @@ class MambaPool:
...
@@ -181,8 +174,9 @@ class MambaPool:
self
.
custom_mem_pool
=
torch
.
cuda
.
MemPool
(
allocator
.
allocator
())
self
.
custom_mem_pool
=
torch
.
cuda
.
MemPool
(
allocator
.
allocator
())
else
:
else
:
self
.
custom_mem_pool
=
None
self
.
custom_mem_pool
=
None
self
.
is_kda_cache
=
isinstance
(
cache_params
,
KimiLinearCacheParams
)
self
.
is_kda_cache
=
isinstance
(
cache_params
,
KimiLinearCacheParams
)
with
self
.
memory_saver_adapter
.
region
(
GPU_MEMORY_TYPE_KV_CACHE
),
(
with
(
torch
.
cuda
.
use_mem_pool
(
self
.
custom_mem_pool
)
torch
.
cuda
.
use_mem_pool
(
self
.
custom_mem_pool
)
if
self
.
enable_custom_mem_pool
if
self
.
enable_custom_mem_pool
else
nullcontext
()
else
nullcontext
()
...
@@ -276,6 +270,11 @@ class MambaPool:
...
@@ -276,6 +270,11 @@ class MambaPool:
f
"conv_state size:
{
get_tensor_size_bytes
(
conv_state
)
/
GB
:.
2
f
}
GB, "
f
"conv_state size:
{
get_tensor_size_bytes
(
conv_state
)
/
GB
:.
2
f
}
GB, "
f
"ssm_state size:
{
get_tensor_size_bytes
(
temporal_state
)
/
GB
:.
2
f
}
GB "
f
"ssm_state size:
{
get_tensor_size_bytes
(
temporal_state
)
/
GB
:.
2
f
}
GB "
)
)
self
.
size
=
size
self
.
device
=
device
self
.
free_slots
=
torch
.
arange
(
self
.
size
,
dtype
=
torch
.
int64
,
device
=
self
.
device
)
self
.
mem_usage
=
self
.
mamba_cache
.
mem_usage_bytes
()
/
GB
self
.
mem_usage
=
self
.
mamba_cache
.
mem_usage_bytes
()
/
GB
self
.
num_mamba_layers
=
num_mamba_layers
self
.
num_mamba_layers
=
num_mamba_layers
...
@@ -370,7 +369,6 @@ class HybridReqToTokenPool(ReqToTokenPool):
...
@@ -370,7 +369,6 @@ class HybridReqToTokenPool(ReqToTokenPool):
device
=
device
,
device
=
device
,
enable_memory_saver
=
enable_memory_saver
,
enable_memory_saver
=
enable_memory_saver
,
)
)
self
.
enable_memory_saver
=
enable_memory_saver
self
.
_init_mamba_pool
(
self
.
_init_mamba_pool
(
size
=
mamba_size
,
size
=
mamba_size
,
cache_params
=
cache_params
,
cache_params
=
cache_params
,
...
@@ -389,7 +387,6 @@ class HybridReqToTokenPool(ReqToTokenPool):
...
@@ -389,7 +387,6 @@ class HybridReqToTokenPool(ReqToTokenPool):
size
=
size
,
size
=
size
,
cache_params
=
cache_params
,
cache_params
=
cache_params
,
device
=
device
,
device
=
device
,
enable_memory_saver
=
self
.
enable_memory_saver
,
speculative_num_draft_tokens
=
speculative_num_draft_tokens
,
speculative_num_draft_tokens
=
speculative_num_draft_tokens
,
)
)
self
.
mamba_map
=
{
layer_id
:
i
for
i
,
layer_id
in
enumerate
(
cache_params
.
layers
)}
self
.
mamba_map
=
{
layer_id
:
i
for
i
,
layer_id
in
enumerate
(
cache_params
.
layers
)}
...
@@ -870,7 +867,6 @@ class HybridLinearKVPool(KVCache):
...
@@ -870,7 +867,6 @@ class HybridLinearKVPool(KVCache):
full_attention_layer_ids
:
List
[
int
],
full_attention_layer_ids
:
List
[
int
],
enable_kvcache_transpose
:
bool
,
enable_kvcache_transpose
:
bool
,
device
:
str
,
device
:
str
,
enable_memory_saver
:
bool
,
mamba_pool
:
MambaPool
,
mamba_pool
:
MambaPool
,
# TODO: refactor mla related args
# TODO: refactor mla related args
use_mla
:
bool
=
False
,
use_mla
:
bool
=
False
,
...
@@ -903,7 +899,7 @@ class HybridLinearKVPool(KVCache):
...
@@ -903,7 +899,7 @@ class HybridLinearKVPool(KVCache):
head_dim
=
head_dim
,
head_dim
=
head_dim
,
layer_num
=
self
.
full_layer_nums
,
layer_num
=
self
.
full_layer_nums
,
device
=
device
,
device
=
device
,
enable_memory_saver
=
enable_memory_saver
,
enable_memory_saver
=
False
,
)
)
else
:
else
:
TokenToKVPoolClass
=
MLATokenToKVPool
TokenToKVPoolClass
=
MLATokenToKVPool
...
...
python/sglang/srt/model_executor/model_runner.py
View file @
d22d0447
...
@@ -1798,7 +1798,6 @@ class ModelRunner:
...
@@ -1798,7 +1798,6 @@ class ModelRunner:
),
),
enable_kvcache_transpose
=
False
,
enable_kvcache_transpose
=
False
,
device
=
self
.
device
,
device
=
self
.
device
,
enable_memory_saver
=
self
.
server_args
.
enable_memory_saver
,
mamba_pool
=
self
.
req_to_token_pool
.
mamba_pool
,
mamba_pool
=
self
.
req_to_token_pool
.
mamba_pool
,
use_mla
=
self
.
use_mla_backend
,
use_mla
=
self
.
use_mla_backend
,
**
extra_args
,
**
extra_args
,
...
...
test/srt/test_mamba_unittest.py
View file @
d22d0447
...
@@ -42,7 +42,6 @@ class TestMamba(unittest.TestCase):
...
@@ -42,7 +42,6 @@ class TestMamba(unittest.TestCase):
full_attention_layer_ids
=
full_attention_layer_ids
,
full_attention_layer_ids
=
full_attention_layer_ids
,
enable_kvcache_transpose
=
False
,
enable_kvcache_transpose
=
False
,
device
=
device
,
device
=
device
,
enable_memory_saver
=
False
,
mamba_pool
=
None
,
mamba_pool
=
None
,
)
)
assert
pool
.
_transfer_full_attention_id
(
global_interval
-
1
)
==
0
assert
pool
.
_transfer_full_attention_id
(
global_interval
-
1
)
==
0
...
@@ -175,7 +174,6 @@ class TestMamba(unittest.TestCase):
...
@@ -175,7 +174,6 @@ class TestMamba(unittest.TestCase):
full_attention_layer_ids
=
full_attention_layer_ids
,
full_attention_layer_ids
=
full_attention_layer_ids
,
enable_kvcache_transpose
=
False
,
enable_kvcache_transpose
=
False
,
device
=
device
,
device
=
device
,
enable_memory_saver
=
False
,
mamba_pool
=
req_to_token_pool
.
mamba_pool
,
mamba_pool
=
req_to_token_pool
.
mamba_pool
,
)
)
...
...
Write
Preview
Markdown
is supported
0%
Try again
or
attach a new file
.
Attach a file
Cancel
You are about to add
0
people
to the discussion. Proceed with caution.
Finish editing this message first!
Cancel
Please
register
or
sign in
to comment